Emedia: ISSN 1529-7306 Emedia/Online Inc.
Emedia: ISSN 1529-7306 Vol.39_1, 1-28
Copyright © 2026 The Authors. Published by Emedia/Online Inc. This work is licensed under a Creative
Commons Attribution 4.0 International License, which permits unrestricted use, distribution, and
reproduction in any medium, provided the original work is properly cited.
Multi-source Traveling Wave Signal Fusion and Fault
Identification Supported by Cloud Computing
Xingyuan Xu *, Quanhui Liu, ZeMing Chen, Lianchen Li, Na You
Guangdong Power Grid Co., Ltd. Guangzhou Zengcheng Power Supply Bureau, Zengcheng
511300, Guangzhou, China
*Corresponding Author:Xingyuan Xu
Email: xingyuanxu11@outlook.com
Received 20 March 2026; Accepted 23 June 2026
Abstract: The accuracy and real-time performance of fault identification of multi-source
traveling wave signals in current power systems are difficult to guarantee due to problems such
as asynchronous acquisition, data heterogeneity, and processing delay. In view of the above
problems, this paper constructs a multi-source traveling wave signal fusion and fault
identification method based on cloud computing platform, relying on the advantages of cloud
computing in resource elastic scheduling and distributed computing to achieve efficient
concurrent processing of large-scale heterogeneous signals and global information
collaborative analysis. The signal time alignment is completed through the global clock
synchronization mechanism, and the unified data format is used for standardized processing.
The transient characteristics of each source signal are extracted based on wavelet transform,
and the multi-source characteristics are fused in the cloud before inputting into the deep neural
network model to complete fault identification and location. The adopted fault identification
model is constructed based on long short-term memory network and spatio-temporal attention
mechanism, which can realize dynamic weighted fusion of multi-source signal characteristics
in time dimension and space dimension, and output fault type discrimination results and
location estimation values in an end-to-end manner. Experimental results show that the
proposed method achieves a fault identification accuracy of 98.6% on the IEEE 39-node power
system simulation platform, and the average system response time is 1.3 seconds, which meets
the real-time identification requirements. The research results verify the effectiveness of this
method in improving the efficiency and reliability of power system fault handling, and provide
technical support for smart grid fault monitoring in the Internet of Things environment.
Keywords: Power System Fault; Traveling Wave Signal; Data Fusion; Cloud Computing; Time
Synchronization
1. Introduction
With the deep integration of smart grids and the Internet of Things (IoT), power systems are
required to achieve higher levels of real-time performance, accuracy, and intelligence in fault
detection and identification [1-2]. Traveling wave signals, as high-frequency transient signals
Xingyuan Xu et al.
2
with fast response characteristics, have become an important basis for fault identification and
location in modern power systems. However, as grid structures become increasingly complex
and the number of measurement points continues to grow, traveling wave signals exhibit
multi-source distribution, data heterogeneity, and massive data volume characteristics [3-4].
These features pose significant challenges to traditional signal processing and fault
identification methods [5-6].
In practical power system monitoring scenarios, traveling wave signals collected from different
measurement points often suffer from time asynchrony, inconsistent data formats, and strong
noise interference [7-8]. These factors reduce the timeliness, consistency, and availability of
the acquired signals, making it difficult to effectively align and fuse multi-source traveling
wave data within a unified time domain. Consequently, the efficiency and reliability of fault
judgment are adversely affected [9-10]. In addition, most existing signal fusion methods rely
on edge devices or local computing architectures. Owing to limited computing resources, fixed
resource allocation, and processing delays, such methods are often unable to meet the real-time
response requirements of large-scale power grid fault detection [11-12]. In IoT-based power
system environments, the multidimensionality, dynamic variation, and complex processing
flow of monitoring data further intensify these challenges [13-14]. Therefore, efficient
collaborative fusion and processing of multi-source traveling wave signals on a unified
platform, as well as fast and accurate fault identification, have become key issues in power
system monitoring and protection [15].
To address the above problems, this paper proposes a cloud-computing-based multi-source
traveling wave signal fusion and fault identification method. Compared with traditional local
centralized architectures, cloud computing platforms provide advantages in elastic resource
scheduling, distributed computing, and collaborative processing of heterogeneous data.
Specifically, containerized deployment and distributed resource scheduling mechanisms are
adopted to support dynamic task allocation and elastic system scaling through the Kubernetes
cluster management framework. When the number of measurement points increases
significantly, the system can automatically expand feature extraction nodes to ensure timely
signal processing. Meanwhile, distributed message middleware is introduced to realize orderly
distribution and unified caching of multi-channel traveling wave data, thereby reducing time
dislocation caused by transmission jitter. In addition, the heterogeneous resource pool
equipped with graphics processing units (GPUs) provides computational support for deep
neural network models, enabling efficient feature fusion and fault identification.
Based on the above design, this paper constructs an integrated system architecture covering
data acquisition, preprocessing, feature extraction, fusion analysis, and fault identification. A
high-precision clock synchronization mechanism is deployed, and the Global Positioning
System (GPS) is used to achieve time alignment among measurement points. Anti-noise
filtering and standardized data conversion are then performed to construct a unified feature
vector set. In the cloud-based fusion stage, the time-frequency characteristics of signals from
different measurement points are integrated through a multi-channel feature fusion network
[16]. Furthermore, a deep learning model based on the Long Short-Term Memory (LSTM)
Multi-source Traveling Wave Signal Fusion and Fault Identification Supported by Cloud Computing
3
network is employed to identify fault types and estimate fault locations. The proposed method
enables efficient processing and rapid inference of large-scale multi-source traveling wave data,
thereby promoting the development of power system fault diagnosis toward cloud-based
collaborative intelligence.
2. Related Works
The research on multi-source traveling wave signal fusion technology focuses on improving
fault identification accuracy, improving ranging methods, and optimizing fusion mechanisms.
LI H et al. [17] proposed a fault ranging method based on a multi-terminal traveling wave
frequency matrix. By establishing a mathematical relationship between the main frequency
component and the transmission distance, accurate location can be achieved under complex
distribution network conditions, reducing the dependence on line parameter frequency changes
and fault location types. Zhang H [18]’s team constructed an identification model based on
U-Net and multimodal fusion mechanism. By extracting features from time-frequency graphs
and wave frequency graphs, the identification rate of four-signal aliasing exceeds 97.3% under
low signal-to-noise ratio conditions, verifying the effectiveness of multimodal fusion. Sun H
[19]’s team applied multi-source information fusion technology to the fault diagnosis of key
devices in railway signal systems and promoted the development of fault analysis towards
data-driven through sensor data fusion. Although existing research has made progress in some
local links, there are still problems such as lack of unified standards for fusion strategies,
insufficient algorithm generalization ability, and low engineering matching. Traditional local
centralized or edge computing modes are limited by resource constraints and are difficult to
support large-scale heterogeneous data collaborative processing. The multimodal fusion
mechanism of Zhang H’s team relied on single device or local network modeling, which
lacked flexibility in task scheduling in high-concurrency multi-point scenarios and failed to
take advantage of cloud-based distributed computing. In addition, the static weight allocation
strategy is difficult to adapt to the dynamic evolution of complex fault modes in power
systems.
Although the cloud computing architecture of the parallel programming model proposed by
Al-Jumaili A H A et al. [20] emphasized resource elastic scheduling and large-scale concurrent
task management capabilities, it was not optimized for the scenario of burst collection of
massive traveling wave signals in power grids. The idea of Maurya M’s [21] team on
integrating artificial intelligence algorithms in the cloud to improve the level of intelligent
device state perception provided a theoretical basis for the end-to-end signal feature extraction
and fault identification process in this paper, but did not involve the real-time synchronization
problem of streaming data caused by the high-frequency transient characteristics of traveling
wave signals. Although the lightweight deployment strategy on the edge side of Lu S’s [22]
team provided a reference direction for model compression and inference acceleration, it did
not solve the dynamic scheduling mechanism of signal feature fusion in heterogeneous
resource environments. The above studies reveal the general advantages of cloud computing
and edge computing in intelligent fault diagnosis, but do not map their technical characteristics
Xingyuan Xu et al.
4
to the specific needs of traveling wave signal processing in power systems. The traditional
local centralized method is limited by the fixed resource allocation and communication
bottlenecks of the measuring points, and it is difficult to support the global collaborative
analysis of large-scale signals; the edge computing fusion strategy lacks a unified scheduling
mechanism, resulting in insufficient ability to model the temporal correlation across measuring
points, and is difficult to adapt to high concurrent dynamic loads. In contrast, the cloud
computing platform relies on distributed resource scheduling to support unified access and
elastic computing of multi-source data, realizes dynamic scaling of tasks through containerized
deployment, and combines message middleware to ensure orderly reception and caching of
data streams. In view of the adaptability defects of the static weight allocation strategy under
complex power grid fault modes, the method in this paper integrates heterogeneous resource
pools and parallel computing frameworks to achieve spatio-temporal joint optimization of
multi-measurement point signals, significantly enhancing feature expression capabilities and
identification robustness.
3. Methods
3.1 Data Acquisition and Clock Synchronization of Multi-source
Traveling Wave Signals
The acquisition system of multi-source traveling wave signals is based on distributed
intelligent terminal devices deployed at key nodes of the power system [23-24]. Each terminal
device is equipped with a high-precision atomic clock or GPS synchronization module as a
global time reference. During the acquisition process, all signals are timestamped with the
UTC time standard to achieve cross-device timing unification. Clock synchronization adopts a
synchronization mechanism based on the Precision Time Protocol (PTP) and dynamically
adjusts the local clock deviation through the master-slave clock architecture to ensure that the
time error of each terminal is controlled at the sub-microsecond level.
In the process of multi-source signal acquisition, time synchronization is an important
prerequisite for ensuring signal consistency. To achieve high-precision time alignment across
measurement points, this method constructs a global clock synchronization mechanism based
on the cloud computing platform and uses the Precision Time Protocol (PTP) with a
master-slave structure to complete the time calibration of each acquisition terminal. This
mechanism dynamically manages the local clock deviations of multiple edge nodes through a
centralized cloud scheduler and uses a containerized microservice module to execute a
recursive filtering algorithm to improve the stability and response speed of clock
synchronization. Compared with the traditional local centralized architecture, the
synchronization mechanism in the cloud computing environment has higher scalability and
fault tolerance and can maintain sub-microsecond time error control when a large number of
nodes are connected.
Multi-source Traveling Wave Signal Fusion and Fault Identification Supported by Cloud Computing
5
The clock deviation is defined as the difference between the local time of the -th
acquisition device and the standard time. Its adjustment process is shown in Formula 1:
(1)
Among them, is the local acquisition time of the -th device; is the reference
master clock deviation; is the communication delay compensation function; function
achieves stable convergence based on the recursive filtering algorithm of the clock deviation.
This method effectively solves the time asynchrony problem caused by clock drift and network
delay in distributed acquisition nodes, ensures the high consistency of multi-source signal
timestamps, and meets the strict requirements of subsequent fusion algorithms for time
alignment.
The data acquisition module designs a unified data transmission protocol, uses edge computing
nodes to pre-process the collected data, and completes data integrity verification and
preliminary format conversion in real-time. The signal sampling rate is uniformly set to 1MHz
to ensure sufficient time resolution to cover the transient characteristics of the traveling wave
signal. Figure 1 illustrates the clock synchronization and data transmission process framework
in the multi-source signal acquisition system and shows the collaborative working structure of
the synchronization module, edge pre-processing, and cloud reception.
Figure 1. Multi-source traveling wave signal acquisition and clock synchronization system
architecture
Xingyuan Xu et al.
6
Figure 1 shows the system architecture of the entire process of multi-source traveling wave
signal acquisition to cloud fault identification. The multi-source acquisition terminal is
responsible for accessing the primary device signal. The high-precision clock module
composed of the atomic clock and GPS provides a unified time base to the time
synchronization module and completes the timestamp consistency through the PTP
master-slave structure. The synchronized and calibrated signal is pre-processed in the edge
computing node and is transmitted to the local database after integrity verification and format
conversion operations. The data is uploaded to the cloud computing platform through the
message queue middleware and is uniformly cached and reordered in the cloud to construct a
standardized input vector. The system finally inputs the fused signal features into the fault
diagnosis model to achieve precise identification and location tasks based on time-frequency
features.
To quantify the acquisition time difference of multiple measurement points, the signal time
alignment error index is defined, and its expression is:
(2)
This index is used as the core parameter for synchronization performance evaluation in the
system design stage. By adjusting the PTP synchronization period and filtering parameters,
can be controlled below 1 microsecond, meeting the technical requirements of strict
timing alignment of traveling wave signals. Table 1 analyzes the synchronization accuracy
differences of each terminal.
Table 1. Statistics of time synchronization errors of multi-source acquisition terminals
Collection terminal
number
Mean time error
(μs)
Time error standard
deviation (μs)
Maximum time error
(μs)
1
0.45
0.07
0.62
2
0.38
0.06
0.55
3
0.52
0.09
0.74
4
0.47
0.08
0.68
5
0.41
0.05
0.59
Table 1 lists the time error statistical indicators of each acquisition terminal in the system under
the unified time synchronization mechanism. Each terminal number corresponds to a set of
synchronization error data; the mean time error reflects the average offset of the terminal
relative to the master clock; the standard deviation indicates the degree of fluctuation of the
time error; the maximum time error is used to measure the maximum deviation amplitude
during the synchronization process. All data are in microseconds, aiming to quantify the
stability and consistency of terminal synchronization accuracy and provide basic support for
subsequent timing alignment and feature fusion.
During the signal acquisition process, there are differences in the output formats of
Multi-source Traveling Wave Signal Fusion and Fault Identification Supported by Cloud Computing
7
heterogeneous devices. The data standardization conversion mechanism is used to unify the
signal format. To ensure the integrity and real-time performance of the data acquisition process,
an asynchronous multi-threaded data acquisition and time synchronization management
module is designed. The acquisition delay and processing delay are defined, and the
sum of the two does not exceed the system real-time response time threshold. The delay model
formula is as follows:
(3)
where is the maximum allowable delay set by the system, which is controlled at the
millisecond level to ensure the timeliness of fault identification. The priority scheduling
strategy is used to allocate transmission and processing thread resources to achieve high-speed
and stable flow of multi-source data.
3.2 Preprocessing and Standardization of Traveling Wave Signals
In the preprocessing and standardization of multi-source traveling wave signals, the key goal is
to remove the noise interference in the original signal, unify the data format of different
acquisition sources, and provide structured input for subsequent feature extraction and fusion.
The original traveling wave signal is affected by factors such as sensor accuracy, sampling
noise, and electromagnetic interference in the field environment. The time domain waveform
of the signal has interference components such as high-frequency spikes and low-frequency
drift [25-26]. To suppress high-frequency random noise, the signal is first subjected to wavelet
threshold denoising. The db4 wavelet function is selected, and the maximum number of
decomposition layers is set to 5. The high-frequency component is denoised according to the
soft threshold function to filter out high-frequency disturbances not related to faults. In this
process, the approximate coefficient and detail coefficient of the signal in the -th layer
wavelet domain are recorded as and , respectively, and the detail coefficient is
denoised using the following soft threshold function:
(4)
where is the denoising threshold; is the noise standard deviation; is
the sample length. The denoised signal is reconstructed for subsequent processing. The
approximate coefficient contains the main features of the signal and is directly retained
without participating in the threshold processing.
Considering the differences in sampling rate, channel structure and data encapsulation methods
of different devices, a unified standard data reconstruction process is used to normalize
multi-source signals into a unified format. All signals are resampled to 1 MHz and
reconstructed using the bilinear interpolation method. The resampled data is stored in a
three-dimensional tensor structure. It is assumed that represents the signal
sample of the -th measurement point; is the number of channels; is the length of the
time series.
To improve the efficiency of signal preprocessing and standardization conversion, a parallel
data processing process is designed on the cloud computing platform, and the format parsing,
Xingyuan Xu et al.
8
field mapping, and resampling operations of the original signal are distributed to multiple
computing nodes for parallel execution. After the signal is structured, it is further normalized
to eliminate the dimension effect. Using the z-score standardization method, the time series of
each channel is transformed to make its mean 0 and standard deviation 1, and the expression
is:
(5)
where and are the mean and standard deviation of the signal at the -th measurement
point, respectively. This processing step improves the comparability of each source data in
subsequent feature fusion and reduces the deviation caused by scale differences.
Figure 2 shows the waveform comparison of the signal before and after filtering. To quantify
the filtering effect, an ideal traveling wave signal model is constructed as a theoretical
reference, which is defined as follows:
(6)
where represents the fault occurrence time, and , , and are the set
response coefficients, which reflect the slope change and amplitude offset in the transient
process and are used to simulate the traveling wave attenuation and reflection effects. The ideal
model refers to a theoretically constructed noise-free traveling wave signal model, which is
used to simulate the ideal voltage or current waveform when a transient disturbance propagates
in a transmission line.
Figure 2. Comparison of waveforms of multi-source traveling wave signals before and after
filtering
Multi-source Traveling Wave Signal Fusion and Fault Identification Supported by Cloud Computing
9
As shown in Figure 2, the waveform comparison of multi-source traveling wave signals before
and after wavelet filtering is presented. The gray curve in the figure is the noisy original signal,
which has a large fluctuation amplitude and presents significant high-frequency disturbances,
interfering with the identification of transient characteristics of the signal. The blue curve is the
signal after wavelet threshold filtering. On the basis of retaining the main mutation structure,
high-frequency noise is effectively suppressed; the edge contour remains clear; the position of
the mutation point is highly consistent with the ideal model. The green dotted line represents
the ideal traveling wave signal, which is used to test the ability of the filtering algorithm to
restore the mutation structure. The filtered signal is highly consistent with the ideal model at
0.003 seconds, indicating that the filtering process has achieved a good balance between
fidelity and denoising performance. The waveform comparison reflects the basic supporting
role of the preprocessing stage in improving feature stability and location accuracy.
In the process of data structure standardization, a unified data frame format field is set, as
shown in Table 2, which defines the timestamp, channel number, voltage, current key fields,
and their byte lengths in the data frame. This structure ensures the consistency of data
transmission and decoding on the cloud platform.
Table 2. Definition table of unified data frame format for multi-source signals
Field Name
Data Type
Byte Length
Description
Timestamp
uint64
8
Unified timestamp
ChannelID
uint8
1
Signal channel identifier
VoltageValue
float32
4
Voltage sampling value
CurrentValue
float32
4
Current sampling value
As shown in Table 2, the listed fields constitute the data frame structure of the unified
encapsulation of multi-source traveling wave signals during the preprocessing process, in
which each field has a clear functional positioning in the data transmission and analysis
process. The Timestamp field is in the form of a 64-bit unsigned integer, which is used to
identify the global unified time of each frame of data to ensure the alignment consistency of
data across measurement points in the time dimension. The ChannelID field is represented by
an 8-bit unsigned integer to identify the number of the data source channel, which helps to
distinguish and index the channel dimension during the fusion process. The VoltageValue and
CurrentValue fields are represented by 32-bit floating point numbers, respectively, carrying the
real-time sampling values of voltage and current, which are the basic inputs for subsequent
signal analysis and feature extraction. The length of each field is designed based on data
accuracy and storage efficiency. The overall structure supports the standardized encapsulation
of high-frequency, high-density, and multi-channel signals, and maintains the consistency and
stability of the format in the cloud data decoding and model input interface.
3.3 Multi-scale Time-frequency Feature Extraction
Xingyuan Xu et al.
10
In the multi-scale time-frequency feature extraction process, the wavelet packet decomposition
(WPD) method is used to perform multi-resolution analysis on each source signal in view of
the significant transient characteristics and complex frequency band distribution of the
traveling wave signal [27-28]. In the multi-scale time-frequency feature extraction process, to
cope with the computing power requirements brought by the concurrent processing of multiple
source signals, this paper relies on the heterogeneous resource pool of the cloud computing
platform to allocate the wavelet packet decomposition task to multiple central processing unit
(CPU) nodes for parallel execution. The resource scheduling strategy dynamically adjusts the
weight parameters according to the task priority and historical consumption to ensure that the
feature extraction tasks on the critical path obtain sufficient computing resources.
This method achieves a finer frequency domain division by recursively filtering and
downsampling the low-frequency and high-frequency parts of the signal at the same time,
thereby effectively capturing the energy mutation and frequency component changes of the
signal at the moment of fault occurrence. In the specific implementation process, the db4
wavelet is selected as the mother wavelet function, and the number of decomposition layers is
set to 5 layers to balance the time resolution and frequency resolution. Each layer of
decomposition generates a set of subband coefficients, forming a time-frequency distribution
diagram of a binary tree structure. Among them, the subband signal corresponding to the -th
node after the -th layer decomposition is represented as , and its mathematical
expression is:
(7)
where is the original input signal; is the impulse response of the -th filter;
represents the number of decomposition layers; is the discrete time index. By calculating
the energy entropy value of each subband coefficient, a feature vector representing the local
singularity of the signal is constructed. The definition of energy entropy is as follows:
(8)
where represents the energy proportion of the -th time period in the -th subband,
and is the total number of time periods. This indicator can reflect the complexity of the
signal in different frequency bands and help distinguish the dynamic differences between the
signal in normal operation and fault state.
To further enhance the feature expression capability, after completing the wavelet packet
decomposition, the Teager Energy Operator (TEO) is applied to perform nonlinear energy
estimation on each subband signal to highlight the transient mutation characteristics of the
signal. The calculation formula of TEO is as follows:
(9)
This operator can quickly detect the transient impact component in the signal without adding
additional computational burden and is suitable for the task of extracting the features of
traveling wave signals caused by sudden faults such as lightning strikes and short circuits in
power systems.
On this basis, to improve the robustness and discriminability of the features, the principal
Multi-source Traveling Wave Signal Fusion and Fault Identification Supported by Cloud Computing
11
component analysis (PCA) dimensionality reduction processing is performed on the extracted
multi-scale time-frequency feature vectors to remove redundant information and retain the
most representative low-dimensional feature space. It is assumed that the original feature
matrix is: , where is the feature dimension and is the number of samples.
PCA constructs the projection matrix by solving the maximum eigenvector of the
covariance matrix and maps the original data to the low-dimensional space , where
is the target dimension.
3.4 Design of Multi-source Feature Fusion Algorithm
In the process of multi-source traveling wave signal feature fusion, a feature layer fusion
strategy [29] based on dynamic weight allocation and spatial attention mechanism [30] is
designed for the multi-channel feature vector [31] after time alignment. This method captures
the topological dependency between measurement points by constructing a feature correlation
matrix and uses a nonlinear mapping function to achieve dimensional alignment and
information complementarity in the feature space. Specifically, assuming that the wavelet
packet decomposition feature vector of the -th measurement point is , where is
the feature dimension of the -th measurement point, the features of each measurement point
are projected to the common feature space through the shared weight matrix
, and the mapping relationship is expressed as:
(10)
In the formula, is the bias term, and is the common space dimension. This
step eliminates the inconsistency of feature scale caused by the difference in measurement
point position, while retaining the discriminative ability of local features.
To further strengthen the cross-measurement point correlation between features, the spatial
attention weight vector is applied in the fusion process, where is the total number
of measurement points involved in the fusion. The attention weight is dynamically generated
by calculating the cosine similarity between the features of each measurement point and the
global feature mean. The calculation process is as follows:
(11)
where
is the global feature mean vector. This weight reflects the
contribution of each measurement point feature to the overall fusion result, so that the system
can still focus on the high-confidence feature area when there is local noise interference. The
fused global feature vector is obtained by weighted summation:
In the feature interaction modeling link, the gated recurrent unit (GRU) is used to construct a
temporal association network to capture the dynamic evolution of multi-source features in the
time dimension. Assuming that the state of the fused feature sequence at time step is
, its update process is coordinated by the reset gate and the update gate
:
Xingyuan Xu et al.
12
(12)
In the formula, is the Sigmoid activation function; represents the Hadamard product;
,, and are trainable parameter matrices. This process realizes the deep coupling of
multi-source features in time series and enhances the representation consistency of fault
features in continuous time windows.
To quantify the changing characteristics of feature dimensions during the fusion process, a
feature contribution evaluation index is designed, which is defined as:
(13)
where is the output layer weight matrix, and is the target output
dimension. This index is used to guide the optimization direction of the fusion network
parameters to ensure that high-contribution features play a dominant role in subsequent
classification tasks.
3.5 Construction of Fault Identification Model Based on Long
Short-Term Memory Network
After the fusion of multi-source traveling wave signal features is completed, LSTM is used to
build a fault identification model, and fault type discrimination and location estimation are
realized through time series feature modeling [32]. The model input is a global feature vector
sequence fused by the spatial attention mechanism, with a time resolution of , and
the feature dimension is reduced by principal component analysis to retain the first
principal components.
The spatial attention weight output by the feature fusion process is applied into the time series
modeling link of the LSTM model [33] as a dynamic adjustment factor of the input feature to
enhance the dominant role of high-confidence measurement point features in time series
evolution. Specifically, the measurement point contribution vector generated in the fusion stage
is multiplied element by element with the input feature of LSTM, so that the model can capture
the trend of time dimension changes while retaining the selective enhancement ability of key
spatial features. By combining the spatial attention mechanism with the time modeling
capability of LSTM, an end-to-end identification framework with spatio-temporal joint
optimization characteristics [34] is constructed, which enhances the consistency of feature
expression and classification separability of multi-source traveling wave signals in complex
fault scenarios.
The LSTM unit has a triple gate structure, including input gate , forget gate , and output
Multi-source Traveling Wave Signal Fusion and Fault Identification Supported by Cloud Computing
13
gate . The gate parameter matrix , , and bias terms , , are used to
control the information flow. Specifically, the input gate calculates the concatenation vector of
the feature vector at the current moment and the hidden state at the previous
moment and generates the update weight through the Sigmoid activation function:
(14)
The forget gate maintains the long-term dependency through the exponential decay
function, and its weight update follows:
(15)
The cell state update process combines the input gate and forget gate outputs and
generates the candidate state
through the hyperbolic tangent function:
(16)
where is the candidate state weight matrix, and is the bias term. After the candidate
state is weighted by the input gate activation value, it constitutes the cell state update together
with the long-term memory controlled by the forget gate:
(17)
where the operator represents the Hadamard product. The output gate calculates the hidden
layer output based on the updated cell state:
(18)
During the training process, the cross entropy loss function is used to measure the
classification error, and the Adam optimizer is combined for parameter iteration. To enhance
the sensitivity of the model to the fault location, a dual-branch structure is applied in the output
layer. One branch outputs the probability distribution of the fault type through the Softmax
activation function, and the other branch uses the linear regression layer to predict the fault
distance. The loss function is defined as:
(19)
The first term is the classification loss term: and represent the true label and the
predicted probability respectively, and is the number of fault type categories; the second
term is the position regression loss term: and correspond to the actual and predicted
fault distances, and is the balance coefficient.
After the feature sequence is processed by the LSTM network, the hidden layer output vector
is sent to the fully connected layer for nonlinear mapping. The weight matrix and bias term
of the fully connected layer are optimized by the back propagation algorithm, and the output
vector is normalized and input into the classification and regression branches. The
classification branch weight matrix and the regression branch weight matrix are trained
independently. The former generates the probability distribution through the Softmax function:
(20)
The latter directly outputs the fault distance prediction value:
(21)
Xingyuan Xu et al.
14
where and are the -th and -th row transposed vectors of the
classification branch weight matrix, respectively, and is the regression branch bias
term.
3.6 Cloud Computing Platform Deployment and System Architecture
Design
In the cloud computing platform deployment and system architecture design phase, a
distributed computing environment is built based on containerization technology, and the
Kubernetes (K8s) cluster management framework is used to realize dynamic resource
scheduling [35] [36]. The system architecture is divided into data access layer, computing
resource layer, and service application layer. Among them, the data access layer realizes
real-time caching and orderly distribution of multi-source traveling wave signals through
Kafka message queues; the computing resource layer is composed of GPU acceleration nodes
and CPU computing nodes to form a heterogeneous resource pool; the service application layer
encapsulates core algorithm modules such as feature extraction, fusion analysis, and fault
identification through microservice architecture.
At the system deployment level, the cloud computing platform decouples and encapsulates
algorithm modules such as feature extraction, fusion analysis, and model inference through
microservice architecture, and uses Kubernetes to realize independent scheduling and version
control of each component. The data access layer uses Kafka message queues to complete
batch caching and streaming push of multi-source traveling wave signals to ensure data loss
and sequential consistency in high-throughput scenarios. The computing resource layer is
composed of GPU acceleration nodes and CPU computing nodes to form a heterogeneous
resource pool. The former undertakes the matrix operation task of the LSTM model, and the
latter is used to perform wavelet packet decomposition and timestamp alignment operations.
The resource allocation model dynamically adjusts the weight parameters according to the task
priority and historical consumption and gives priority to guaranteeing the computing resource
supply of the critical path when multiple source signals arrive concurrently, thereby improving
the overall system’s response efficiency and operation stability.
To solve the contradiction between large-scale data concurrent processing and low delay
response, a priority-based resource allocation model is designed:
(22)
In the formula, represents the resource allocation weight of the -th task; is the task
priority coefficient; corresponds to the demand intensity of the task for the -th resource;
is the resource constraint factor; reflects the historical consumption time of the -th
task on the -th resource. The model dynamically adjusts the weight parameter and the
constraint factor to achieve the optimal allocation of computing resources between feature
extraction intensive tasks and real-time sensitive tasks.
Multi-source Traveling Wave Signal Fusion and Fault Identification Supported by Cloud Computing
15
In the model deployment stage, model sharding and pipeline parallel technology are used to
optimize the inference efficiency of the LSTM network. The LSTM weight matrix
is divided into three sub-modules: input gate , forget gate , and output gate
according to the gate structure, where is the input feature dimension, and is the number
of hidden layer nodes. Each submodule is deployed on different GPU instances through the
gRPC remote call interface, and the CUDA stream technology is used to implement
asynchronous execution of gated computing. To reduce the cross-node communication
overhead, a feature vector compression transmission protocol is designed to compress the
transmission accuracy of the intermediate feature from 32-bit floating point to 16-bit
half-precision format, and the differential encoding technology is applied to eliminate
redundant information of temporal features.
To improve the system’s scalability, a metadata management mechanism based on etcd
distributed key-value storage is designed, and the elastic expansion and contraction of
computing nodes is realized through heartbeat detection and service registration mechanism.
When the cluster load rate exceeds the threshold , the node expansion strategy is
automatically triggered:
(23)
In the formula, is the number of new nodes, and represents the single node load
bearing coefficient. In the system architecture design, this paper combines the elastic
expansion and contraction mechanism unique to cloud computing to realize the on-demand
allocation of computing resources. When the cluster load rate exceeds the set threshold, the
operation of adding computing nodes is automatically triggered to dynamically adapt to the
burst traffic pressure in multi-source traveling wave signal processing.
4. Experiments
4.1 Experimental Environment and Hardware Configuration
The experimental environment of this paper is built on a distributed cloud computing
architecture, and the containerization technology is used to achieve dynamic resource
scheduling. The overall system framework consists of a data access layer, a computing
resource layer, and a service application layer. The data access layer uses the Kafka message
queue to achieve real-time caching and orderly distribution of multi-source traveling wave
signals, and its throughput is designed to achieve a concurrent processing capacity of 100,000
data packets per second. The computing resource layer consists of a heterogeneous resource
pool, including GPU acceleration nodes and CPU computing nodes. Among them, the GPU
node uses the NVIDIA A100 accelerator card, equipped with a PCIe 4.0 interface and the
third-generation Tensor Core technology, and the computing power of a single card can reach
19.5 TFLOPS (FP16). The CPU node uses the Intel Xeon Platinum 8380 processor, with a
main frequency of 2.3GHz, supporting the 512-bit AVX-512 instruction set, 8 memory
channels, and a bandwidth of 204.8 GB/s.
Xingyuan Xu et al.
16
The service application layer encapsulates core algorithm modules such as feature extraction,
fusion analysis, and fault identification through a microservice architecture. The cluster
management uses the Kubernetes (K8s) framework, configures the etcd distributed key-value
storage system to implement metadata management, and ensures real-time synchronization of
node status through heartbeat detection and service registration mechanisms. The load
balancing module deploys the HAProxy component, uses the minimum connection number
algorithm to allocate task flows, and supports SSL offloading and HTTP/2 protocol
transmission.
Table 3 shows the resource configuration of each functional node of the cloud computing
platform. Among them, the feature extraction node is configured with a high-concurrency
multi-threaded processing unit, and the model inference node uses the NVIDIA A100 GPU
accelerator card to meet the computing power requirements of deep neural networks.
Table 3. Resource configuration table of functional nodes of the cloud computing platform
Node Type
CPU Cores
Memory(GB)
GPU Model
Storage(TB)
Feature Extraction
32
128
None
10
Model Inference
16
64
NVIDIA A100
5
Data Storage
8
32
None
100
Cluster Management
4
16
None
2
Load Balancer
8
32
None
1
As shown in Table 3, the resource configuration parameters of the five types of functional
nodes directly correspond to the core functional module requirements of the cloud computing
platform. The feature extraction node is configured with a 32-core CPU and 128GB memory to
support the high concurrent computing requirements of parallel preprocessing and feature
extraction of multi-source traveling wave signals. The design choice of its GPU-free
configuration is due to the technical characteristics of this stage that mainly focuses on
multi-threaded task scheduling and does not require floating-point operation acceleration. The
model inference node uses a combination of a 16-core CPU with 64GB memory and an
NVIDIA A100 GPU. The matrix parallel computing capability of the CUDA core accelerates
the time series feature processing of LSTM. The memory capacity design meets the storage
requirements of deep learning model parameter cache and intermediate feature tensors. The
data storage node is configured with an 8-core CPU and 32GB memory, with a 100TB
mechanical hard disk array. The RAID5 redundant architecture ensures the persistent storage
reliability of the original data and feature library of multi-source traveling wave signals. Its
hardware selection focuses on storage density and cost control rather than random read and
write performance. The cluster management node implements metadata management and task
scheduling with a streamlined configuration of 4-core CPU and 16GB memory. The control
plane resource consumption is reduced through lightweight service design to ensure the
efficient operation of etcd distributed key-value storage and Kubernetes scheduler. The load
balancing node is configured with 8-core CPU and 32GB memory. The software load
balancing solution is used to implement the traffic distribution of Kafka message queue and
microservice call chain tracking. Its storage capacity design only meets the temporary storage
Multi-source Traveling Wave Signal Fusion and Fault Identification Supported by Cloud Computing
17
requirements of log records and state snapshots.
4.2 Dataset Construction and Parameter Setting
This paper focuses on the construction of multi-source traveling wave signal dataset and the
setting of key parameters. This dataset is generated based on the IEEE 39-node power system
simulation platform and covers the characteristics of traveling wave signals under various fault
types and operating conditions. To ensure the wide applicability and statistical validity of the
experimental results, the design of the dataset fully considers the diverse distribution of fault
locations, fault types, and environmental noise. The fault simulation includes typical power
system fault forms such as single-phase grounding fault, two-phase short circuit fault, and
three-phase short circuit fault. At the same time, by adjusting the fault resistance, fault
occurrence time, and line load level, the complexity of the dataset and the representativeness of
actual engineering are further enhanced.
To accurately reflect the time-space characteristics of multi-source traveling wave signals, the
acquisition terminals are deployed at key monitoring points of the power system, covering
different sections of the transmission lines. Each measuring point is equipped with a
high-precision synchronization module, and the signal is marked with the UTC time standard
to ensure the consistency of multi-source signals in the time domain. The signal sampling rate
is uniformly set to 1 MHz to ensure sufficient time resolution for capturing the transient
characteristics of traveling wave signals. Each record in the dataset contains complete voltage
and current waveform information and is accompanied by a precise timestamp, fault type
identifier, and fault distance parameter. These fields constitute the core input for subsequent
feature extraction and model training.
Table 4 lists the main parameter configurations of the multi-source traveling wave signal
dataset, including key indicators such as signal sampling rate, total number of dataset samples,
number of fault types, fault distance range, and fault resistance value range. The basic scale
and complexity of the dataset are defined, reflecting its adaptability and coverage in simulating
real power grid fault scenarios.
Table 4 describes in detail the specific values or value ranges of each parameter in the dataset,
covering the entire process from signal sampling to data format standardization. The signal
sampling rate is 1 MHz to ensure the integrity of high-frequency transient signals. The total
number of samples in the dataset is 50,000, covering a variety of scenarios with different fault
types. There are three types of faults, corresponding to single-phase grounding faults,
two-phase short circuit faults, and three-phase short circuit faults. The fault distance ranges
from 0.5 km to 150 km, covering the typical conditions of short-distance and long-distance
transmission lines. The fault resistance ranges from 0 Ω to 500 Ω, simulating the transition
state of low-impedance and high-impedance faults. There are five monitoring points,
distributed at multiple key nodes in the power system. The data format standardization part
clarifies the data type and byte length of voltage, current, timestamp, and channel number,
providing structured support for subsequent signal processing.
Xingyuan Xu et al.
18
In the process of dataset construction, a multi-level noise interference mechanism is applied to
simulate the electromagnetic interference and sensor measurement errors that may occur in the
field environment. The noise injection adopts a combination of white noise and impulse noise,
which is controlled between 10 dB and 30 dB to ensure that the dataset has a certain
anti-interference ability while retaining the fault characteristics. This design not only improves
the practicality of the dataset but also provides a basis for evaluating the robustness of the
signal preprocessing algorithm.
Table 4. Multi-source traveling wave signal dataset parameter configuration table
Parameter Description
Value OR Range
Signal Sampling Rate
1 MHz
Total Number of Samples
50,000
Number of Fault Types
3 (Single-phase Grounding, Two-phase Short Circuit,
Three-phase Short Circuit)
Fault Distance Range
0.5 km - 150 km
Fault Resistance Range
0 Ω - 500 Ω
Number of Monitoring Points
5
Data Format Standardization
Voltage: float32, Current: float32, Timestamp: uint64,
ChannelID: uint8
4.3 Experimental Process and Performance Index Definition
This paper focuses on the experimental process and performance index definition and
comprehensively verifies the multi-source traveling wave signal fusion and fault identification
method based on the cloud computing platform through a systematic method.
The experimental process includes data acquisition, preprocessing, feature extraction,
multi-source feature fusion, and fault identification model training and testing. In the data
acquisition stage, all traveling wave signals are marked according to a unified time standard,
and the time consistency of the data at each measuring point is ensured by a high-precision
clock synchronization mechanism. Subsequently, the original signal is filtered, denoised, and
standardized to eliminate noise interference and unify the data format. On this basis, the
wavelet packet decomposition method is used to perform multi-scale time-frequency analysis
on the signal to extract features such as transient energy distribution and frequency band
energy ratio that can reflect the fault characteristics. After these features are integrated by the
feature layer fusion strategy, they are sent to LSTM as input data for fault type discrimination
and location estimation tasks.
To comprehensively evaluate the effectiveness of the method, multiple performance indicators
are defined, covering aspects such as fault identification accuracy, fault location error, system
response time, and multi-source data fusion effect. Among them, the fault identification
accuracy is used to measure the model’s ability to classify different fault types; the fault
location error reflects the system’s estimation accuracy of the fault location; the system
response time counts the total time from data collection to the final output of the fault
Multi-source Traveling Wave Signal Fusion and Fault Identification Supported by Cloud Computing
19
identification result, which is an important basis for evaluating the real-time performance of
the system; the multi-source data fusion effect is reflected by comparing the changes in the
discrimination of the feature vectors before and after fusion, which is used to verify the
contribution of the feature fusion algorithm to improving the overall identification
performance.
During the experiment, attention should also be paid to the resource utilization of the cloud
computing platform and the scalability of the system to ensure that the method has good
deployment capabilities and operating efficiency in actual engineering applications. To this end,
parameters such as resource allocation weight, task priority coefficient, and load rate threshold
are set to dynamically adjust the allocation ratio of computing resources between different
tasks. The parameter configuration of the dataset is detailed in Table 4. Its sampling rate, fault
type, and distance range design ensure the wide applicability of the experiment.
The parameter definition covers the entire process from signal acquisition to data format
normalization, reflecting the adaptability and extensiveness of the dataset in simulating real
power grid fault scenarios. By setting a variety of fault types, distances, and resistance
parameters, the experimental results are ensured to be highly representative and practical,
providing solid data support for subsequent model training and performance evaluation.
5. Results
5.1 Fault Identification Accuracy
In the task of fault identification in power systems, this paper constructs a multi-source signal
fusion and fault identification method based on a cloud computing platform to address the
problems of asynchronous multi-source traveling wave signal acquisition, data heterogeneity,
and processing delay. To verify the identification performance of this method under different
fault types, the experiment selects five representative models for comparative analysis,
including traditional methods (wavelet transform combined with support vector machine
(wavelet+SVM)), single LSTM, multi-source feature splicing multi-layer perceptron (MLP),
fusion model without attention mechanism, and the spatio-temporal long short-term memory
network (ST-LSTM) of this paper. These categories represent the development path from early
shallow classifiers to modern deep neural networks, and their structural complexity and feature
extraction capabilities are gradually improved. The experiment uses the identification accuracy
of three typical fault types as evaluation indicators, covering single-phase grounding,
two-phase short circuit, and three-phase short circuit, aiming to fully reflect the model’s
generalization ability and robustness under different fault modes. The standard deviation is
used to measure the model’s stability in multiple experiments; the error bar reflects the
confidence interval of the result; the red dotted line marks the overall accuracy of the method
in this paper, which is 98.6%.
Xingyuan Xu et al.
20
Figure 3. Comparison of the accuracy of multi-source traveling wave signal fusion fault
identification methods based on cloud computing platform
Figure 3 shows the difference in identification accuracy of five types of models on three types
of faults. Traditional methods are limited by manual feature extraction and the design of
shallow classifiers, and the accuracy of various types of faults is generally lower than 93%.
Although a single LSTM has the ability to model time series, its overall performance still lags
behind the fusion method because it only relies on single measurement point information. After
applying multi-source signal splicing, the MLP model improves the accuracy to more than
96%, but fails to achieve dynamic weight allocation, resulting in limited identification of some
fault types. The fusion model without attention mechanism further improves the identification
performance, but it is still insufficient in capturing detailed features. In contrast, the ST-LSTM
method proposed in this paper applies the spatio-temporal attention mechanism to
simultaneously optimize the feature expression in the time dimension and the space dimension,
so that the identification accuracy of the three types of faults exceeds 98%, among which the
three-phase short circuit fault reaches 99.0%. This result is due to the effective collaborative
use of multi-source signals and the support of cloud-based distributed computing resources for
model training efficiency. Experimental data shows that with the gradual optimization of the
model structure, the identification performance shows a progressive improvement trend, which
confirms the significant enhancement of the fusion strategy and deep learning architecture
improvement on the power system fault diagnosis effect.
5.2 Fault Location Error Evaluation
In the power system fault identification task, a multi-source signal fusion and fault
identification method is constructed based on the cloud computing platform to address the
location error problem caused by asynchronous acquisition and data heterogeneity of
Multi-source Traveling Wave Signal Fusion and Fault Identification Supported by Cloud Computing
21
multi-source traveling wave signals. The experiment uses fault location error (unit: meter) as
the core evaluation indicator to reflect the accuracy of the model’s estimation of the fault
location; the standard deviation is used to quantify the degree of discreteness of multiple
experimental results, and the error band (± standard deviation) further reflects the data
fluctuation range. Figure 4 shows the changing trend of the location error of the five models in
the range of 0.5 km to 150 km, and the semi-transparent shaded area indicates the error
fluctuation range.
Figure 4. Comparison of location errors of multi-source traveling wave signal fusion models at
different fault distances
As shown in Figure 4, the traditional method (Wavelet+SVM) has an error of 105 m at 150 km;
the single LSTM has an error of 97 m; the ST-LSTM has a stable error of 30 m, and the error
band width is significantly lower than other models. This result is due to the fact that
ST-LSTM dynamically allocates multi-source signal weights through the spatio-temporal
attention mechanism, suppressing the influence of noise interference on feature expression in
long-distance scenarios, and the distributed computing resources provided by the cloud
computing platform guarantee the real-time processing capability of complex models. In
contrast, the traditional method is limited by the limitations of manual feature extraction; the
single LSTM fails to fully integrate the information of multiple measurement points; the
multi-source MLP and non-attention fusion models lack dynamic weight adjustment, resulting
in local feature redundancy. As the fault distance increases, the errors of each model show an
increasing trend, but the increase of ST-LSTM is the smallest, and its error band is always
maintained within ±6 m, which verifies the stability of the location accuracy and engineering
applicability of this method in long-distance transmission lines. This result is due to the
effective weighted fusion of multi-point features by the spatio-temporal attention mechanism,
and the support of cloud-based distributed computing resources for model training efficiency.
It can be concluded that the coordinated optimization of the fusion strategy and the deep
Xingyuan Xu et al.
22
learning architecture has a significant effect on improving the accuracy of power system fault
location.
5.3 System Response Time Measurement
In the research on power system fault monitoring and identification, the experiment focuses on
the real-time bottleneck problem caused by asynchronous acquisition, data heterogeneity, and
processing delay of multi-source traveling wave signals. The response characteristics of three
typical faults (single-phase grounding, two-phase short circuit, and three-phase short circuit)
are compared and analyzed around different model architectures. The experiment
quantitatively evaluates the performance differences of various methods under the cloud
computing platform through three core indicators: response time (total time from data
acquisition to fault output), standard deviation (reflecting model stability), and real-time
threshold (1.3 seconds). The horizontal bar chart in Figure 5 distinguishes the three types of
faults with different colors; the error bar reflects the data fluctuation range; the red dotted line
marks the real-time threshold.
As shown in Figure 5, the response time of traditional methods generally exceeds 1.9 seconds;
the single LSTM is shortened to 1.71-1.83 seconds; the multi-source MLP is further reduced to
1.51-1.68 seconds; the non-attention fusion model achieves a breakthrough in the range of
1.33-1.45 seconds; the ST-LSTM achieves real-time requirements with a stable performance of
1.28-1.32 seconds. The difference in response time is due to the hierarchical evolution of the
model architecture and feature fusion mechanism: the traditional method relies on manual
feature extraction, resulting in redundant calculations; the single LSTM is limited to the
single-point information input dimension; the multi-source MLP enhances information
integrity through feature splicing but lacks dynamic optimization capabilities; the non-attention
fusion model initially applies a weight distribution strategy; the ST-LSTM realizes the
coordinated optimization of multi-source features in the time and space dimensions through the
spatio-temporal attention mechanism, and significantly reduces the redundant calculation
overhead in combination with the distributed resource scheduling of the cloud computing
platform. The standard deviation analysis shows that as the model complexity increases, the
response stability shows a gradient enhancement trend, and the standard deviation of
ST-LSTM is controlled within 0.05-0.06 seconds, reflecting its robustness in high-concurrency
scenarios. This result verifies that the multi-source signal fusion framework based on the cloud
computing platform improves the engineering applicability of the fault identification system
while ensuring real-time requirements through algorithm architecture innovation and dynamic
resource allocation.
Multi-source Traveling Wave Signal Fusion and Fault Identification Supported by Cloud Computing
23
Figure 5. Comparison of response time of multi-source traveling wave signal fusion fault
identification system based on cloud computing platform
5.4 Verification of Multi-source Data Fusion Effect
The experiment uses feature discrimination as the core evaluation indicator to quantify the
feature separation ability of fault categories under different models. The higher the value, the
clearer the feature expression. In Figure 6, the X-axis shows the names of five types of models,
and the Y-axis represents the average Euclidean distance.
As shown in Figure 6, the feature discrimination of traditional methods under three types of
faults is less than 0.75, which is limited by the limitations of manual feature extraction and the
discrimination ability of shallow classifiers. Single LSTM improves the discrimination to
above 0.71 through time series modeling, but the gain is limited due to the lack of sufficient
fusion of multi-measurement point information. After applying multi-source feature
concatenation, the discrimination of Multi-Source MLP reaches above 0.80, but the fixed
weight allocation limits the improvement of some fault types (such as three-phase short circuit).
The fusion model without attention mechanism further increases the discrimination to above
0.81 by weighted integration of multi-source features, but there is still redundant feature
interference. In contrast, ST-LSTM dynamically optimizes the measurement point weights
through the spatio-temporal attention mechanism, and the discrimination of the three types of
faults reaches 1.02–1.15, which is an average improvement of more than 60% compared with
traditional methods. This result stems from the ability of ST-LSTM to suppress noise
interference and strengthen high-credibility feature areas in the spatial dimension, as well as
the advantage of capturing dynamic evolution laws in the temporal dimension. At the same
Xingyuan Xu et al.
24
time, the distributed resource scheduling of the cloud computing platform ensures efficient
training and inference of complex models. The data shows that the coordinated optimization of
the multi-source signal fusion strategy and the deep learning architecture significantly
improves the feature expression ability, providing key support for the robustness and real-time
performance of power system fault identification.
Figure 6. Comparison of feature discrimination of five models under three types of faults
5.5 Resource Utilization of Cloud Computing Platform
In the study of multi-source traveling wave signal fusion and fault identification in power
systems, resource utilization aims to reveal the dynamic correlation characteristics between
heterogeneous resource scheduling strategies and load changes. The experiment collects
resource utilization data of CPU, memory, and model inference GPU by time period. The
resource utilization index quantifies the computing power requirements of each component in
data preprocessing, feature fusion, and deep learning inference by calculating the core
occupancy rate of the calculation unit, memory bandwidth consumption, and GPU video
memory peak ratio.
Multi-source Traveling Wave Signal Fusion and Fault Identification Supported by Cloud Computing
25
Figure 7. Comparative analysis of resource utilization of multi-source traveling wave
signal processing on cloud computing platform by time period. (Figure 7(a). 01:00–06:00
(Low-Load Period), Figure 7(b). 07:00–12:00 (Working Period), Figure 7(c). 13:00–18:00
(Peak Load Period), Figure 7(d). 19:00–24:00 (Post-Peak Period).)
As shown in Figure 7, during the peak load period, the GPU utilization rate of model inference
climbs to 80-88%, and the low load period increases by more than 30%. This increase is
directly related to the gradient adjustment mechanism of the priority coefficient in the dynamic
resource allocation model: the highly complex spatio-temporal attention calculation triggers
the elastic expansion of GPU resources when the load rate exceeds the threshold, and the CPU
and memory utilization rates of the feature extraction module simultaneously climbed to 52-58%
and 58-63%, reflecting its continuous computing power demand as a preprocessing core. The
GPU and CPU utilization fluctuate in a positive correlation during the working cycle,
indicating that the feedback optimization of the historical consumption time parameters
effectively balances the resource competition between feature extraction and deep learning
inference. During the fallback period, the GPU utilization rate drops to 60-70%, and the CPU
utilization rate drops to 35-41%, revealing the priority guarantee characteristics of the cloud
resource recovery strategy for deep learning tasks. The load rate-driven node expansion
strategy effectively balances the computing load requirements of the deep model through
dynamic resource allocation, avoiding performance bottlenecks in heterogeneous resource
scheduling. The study verifies the effectiveness of dynamic scheduling of containerized
heterogeneous resource pools through Kubernetes cluster management, and combined with the
feature weight allocation of the spatio-temporal attention mechanism, significantly improves
the resource adaptation efficiency and system expansion capabilities of multi-source signal
processing.
Xingyuan Xu et al.
26
6. Conclusions
This paper proposed a cloud-computing-based method for multi-source traveling wave signal
fusion and fault identification in power systems. The proposed framework uses global clock
synchronization for signal time alignment, wavelet packet decomposition for multi-scale
time-frequency feature extraction, and an LSTM model enhanced by a spatio-temporal
attention mechanism for fault type identification and location estimation. At the system level,
cloud-native technologies such as containerized deployment, automatic scaling, heterogeneous
resource scheduling, and microservice-based module decoupling are introduced to improve the
flexibility, robustness, and scalability of multi-source signal processing. These mechanisms
enable dynamic resource allocation and efficient concurrent processing, thereby enhancing
system response capability and maintainability. Experimental results on the IEEE 39-bus
power system simulation platform show that the proposed method achieves a fault
identification accuracy of 98.6% and an average response time of 1.3 s, meeting real-time fault
identification requirements. The dynamic-weight-based feature fusion strategy improves the
discrimination of three typical fault types to 1.02–1.15, which is over 60% higher than that of
traditional methods. However, the peak GPU utilization reaches 88%, indicating a potential
energy consumption bottleneck for computation-intensive deep learning tasks. Future work
will focus on lightweight feature extraction, cloud-edge collaborative computing, federated
learning, and 5G-based low-latency transmission to further improve the efficiency and
scalability of smart grid fault monitoring in IoT environments.
Funding
This work was supported by the Science and Technology Project of China Southern Power
Grid Company Limited under (GDKJXM20231090).
References
[1] Koohang A, Sargent C S, Nord J H, et al. Internet of Things (IoT): From awareness to
continued use. International Journal of Information Management, 2022, (62), pp. 102442.
[2] Yu F, Liu J, Yu H, et al. Multimodal Wearable System With Dual-Frequency Enhancement
Network for Risk Recognition. IEEE Internet of Things Journal, 2025, 12(11),
pp.17634-17648.
[3] Cui H, Yang Q, Sun J, et al. Electromagnetic time reversal fault location method based on
characteristic frequency of traveling waves. IEEE Transactions on Power Delivery, 2023,
38(5), pp. 3033-3044.
[4] Deng F, Zhong Y, Zeng Z, et al. A single-ended fault location method for transmission line
based on full waveform features extractions of traveling waves. IEEE Transactions on
Power Delivery, 2023, 38(4), pp.2585-2595.
[5] Yu F, Rao Z, Chen N, et al. ArmBCIsys: Robot Arm BCI System With Time–Frequency
Network for Multiobject Grasping. IEEE Transactions on Neural Networks and Learning
Systems, 2025, 36 (10), pp.18327-18341..
Multi-source Traveling Wave Signal Fusion and Fault Identification Supported by Cloud Computing
27
[6] Feng Y, Xiao Z, Liu L, et al. BiaCanDet: Bioelectrical impedance analysis for breast cancer
detection with space–time attention neural network. Expert Systems with Applications,
2025, (269), pp. 126223.
[7] Zhao H, Gong Z, Yan J, et al. Unsynchronized underwater localization with isogradient
sound speed profile and anchor location uncertainties. IEEE Transactions on Vehicular
Technology, 2024, 73(6): 8864-8877.
[8] Shu H, Deng Y, Cao P, et al. Transient pilot protection based on directional traveling wave
principle for wind power transmission line. International Journal of Electrical Power &
Energy Systems, 2024, 157, pp. 109835-109846.
[9] Wu J, Li Q, Chen Q, et al. Evaluation, analysis and diagnosis for HVDC transmission
system faults via knowledge graph under new energy systems construction: A critical
review. Energies, 2022, 15(21), pp. 8031-8051.
[10] Liu R, Fu R, Xu K, et al. A review of knowledge graph-based reasoning technology in the
operation of power systems. Applied Sciences, 2023, 13(7), pp. 4357-4381.
[11] Hassani H, Razavi-Far R, Saif M, et al. Intelligent decision support and fusion models for
fault detection and location in power grids. IEEE Transactions on Emerging Topics in
Computational Intelligence, 2021, 6(3), pp. 530-543.
[12] Ni Q, Zhan Z, Li X, et al. A real-time fault diagnosis method for grid-side overcurrent in
train traction system using signal time series feature pattern recognition. IEEE Transactions
on Industrial Electronics, 2023, 71(4), pp. 4210-4218.
[13] Sun R, Ren Y. A multi-source heterogeneous data fusion method for intelligent systems in
the Internet of Things. Intelligent Systems with Applications, 2024, 23, pp.
200424-200433.
[14] Sahu A, Mao Z, Wlazlo P, et al. Multi-source multi-domain data fusion for cyberattack
detection in power systems. IEEE Access, 2021, 9, pp. 119118-119138.
[15] De La Cruz J, Gómez-Luna E, Ali M, et al. Fault location for distribution smart grids:
Literature overview, challenges, solutions, and future trends. Energies, 2023, 16(5), pp.
2280-2317.
[16] F. Yu, Y. Zhang, H. Li, C, et al. Phase Contour Enhancement Network for Clothing
Parsing. IEEE Transactions on Consumer Electronics, 2024, 70(1), pp. 2784-2793.
[17] LI H, ZENG H, YU K, et al. Fault location method for complex distribution networks
based on multi-terminal traveling wave frequency matrix. Journal of Electric Power
Science and Technology, 2024, 39(3), pp. 19-30, 37.
[18] Zhang H, Li L, Pan H, et al. Time-Frequency Aliased Signal Identification Based on
Multimodal Feature Fusion. Sensors, 2024, 24(8), pp. 2558-2581.
[19] Sun H, Qin B, Wei Z, et al. A review of fault diagnosis for train signal system based on
multi-source information fusion. Engineering Research Express, 2025, 7(2), pp.
022503-022528.
[20] Al-Jumaili A H A, Muniyandi R C, Hasan M K, et al. Big data analytics using cloud
computing based frameworks for power management systems: Status, constraints, and
future recommendations. Sensors, 2023, 23(6), pp. 2952-2989.
[21] Maurya M, Panigrahi I, Dash D, et al. Intelligent fault diagnostic system for rotating
machinery based on IoT with cloud computing and artificial intelligence techniques: a
review. Soft Computing, 2024, 28(1), pp. 477-494.
Xingyuan Xu et al.
28
[22] Lu S, Lu J, An K, et al. Edge computing on IoT for machine signal processing and fault
diagnosis: A review. IEEE Internet of Things Journal, 2023, 10(13), pp. 11093-11116.
[23] Li B, Tang L, Gu Z, et al. Research on Distribution Network Fault Location Based on
Electric Field Coupling Voltage Sensing and Multi-Source Information Fusion. Energies,
2025, 18(4), pp. 913-939.
[24] Yuan Q, Pi Y, Kou L, et al. Multi-source data processing and fusion method for power
distribution internet of things based on edge intelligence. Frontiers in Energy Research,
2022, 10, pp. 891867-891876.
[25] Li T, Sun C, Fink O, et al. Filter-informed spectral graph wavelet networks for multiscale
feature extraction and intelligent fault diagnosis. IEEE Transactions on Cybernetics, 2023,
54(1), pp. 506-518.
[26] Ahmed N, Hashmani A A, Khokhar S, et al. Fault detection through discrete wavelet
transform in overhead power transmission lines. Energy Science & Engineering, 2023,
11(11), pp. 4181-4197.
[27] Shao H, Xia M, Wan J, et al. Modified stacked autoencoder using adaptive Morlet
wavelet for intelligent fault diagnosis of rotating machinery. IEEE/ASME Transactions on
Mechatronics, 2021, 27(1), pp. 24-33.
[28] Huang D, Zhang W A, Guo F, et al. Wavelet packet decomposition-based multiscale CNN
for fault diagnosis of wind turbine gearbox. IEEE Transactions on Cybernetics, 2021, 53(1),
pp. 443-453.
[29] Yu F, Cao J, Liu L, et al. SuperLightNet: Lightweight Parameter Aggregation Network for
Multimodal Brain Tumor Segmentation. Proceedings of the Computer Vision and Pattern
Recognition Conference. 2025, pp. 5197-5206.
[30] Liao C, Du C, Liu Y, et al. SterAF: A Scene Text Recognizer with Appearance-Flow
Rectification. International Journal of Pattern Recognition and Artificial Intelligence, 2024,
38(06): 2450011.
[31] Yu F, Yu C, Tian Z, et al. Intelligent wearable system with motion and emotion
recognition based on digital twin technology. IEEE internet of things journal, 2024, 11(15),
pp. 26314-26328.
[32] Al-Selwi S M, Hassan M F, Abdulkadir S J, et al. RNN-LSTM: From applications to
modeling techniques and beyond—Systematic review. Journal of King Saud
University-Computer and Information Sciences, 2024, 36(5), pp. 102068.
[33] Yu F, Chen Z, Jiang M, et al. Smart clothing system with multiple sensors based on
digital twin technology. IEEE Internet of Things Journal, 2022, 10(7), pp. 6377-6387.
[34] Cheng W, Zhou F, Liao C, et al. STHM: A skinning transformer approach for predicting
clothing deformation from human motion. Journal of Computational Methods in Sciences
and Engineering, 2025, pp. 14727978251385168.
[35] Soni D, Kumar N. Machine learning techniques in emerging cloud computing integrated
paradigms: A survey and taxonomy. Journal of Network and Computer Applications, 2022,
205, pp. 103419.
[36] Parast F K, Sindhav C, Nikam S, et al. Cloud computing security: A survey of
service-based models. Computers & Security, 2022, 114: 102580.
