Blog

Apr 05, 2025

Partial discharge defect recognition method of switchgear based on cloud-edge collaborative deep learning | Scientific Reports

Scientific Reports volume 15, Article number: 10956 (2025) Cite this article 602 Accesses 1 Altmetric Metrics details To address the limitations of traditional partial discharge (PD) detection methods

Scientific Reports volume 15, Article number: 10956 (2025) Cite this article

602 Accesses

1 Altmetric

Metrics details

To address the limitations of traditional partial discharge (PD) detection methods for switchgear, which fail to meet the requirements for real-time monitoring, rapid assessment, sample fusion, and joint analysis in practical applications, a joint PD recognition method of switchgear based on edge computing and deep learning is proposed. An edge collaborative defect identification architecture for switchgear is constructed, which includes the terminal device side, terminal collection side, edge-computing side, and cloud-computing side. The PD signal of switchgear is extracted based on UHF sensor and broadband pulse current sensor on the terminal collection side. Multidimensional features are obtained from these signals and a high-dimensional feature space is constructed based on feature extraction and dimensionality reduction on the edge-computing side. On the cloud side, the deep belief network (DBN)-based switchgear PD defect identification method is proposed and the PD samples acquired on the edge side are transmitted in real time to the cloud for training. Upon completion of the training, the resulting model is transmitted back to the edge side for inference, thereby facilitating real-time joint analysis of PD defects across multiple switchgear units. Verification of the proposed method is conducted using PD samples simulated in the laboratory. The results indicate that the DBN proposed in this paper can recognize PDs in switchgear with an accuracy of 88.03%, and under the edge computing architecture, the training time of the switchgear PD defect type classifier can be reduced by 44.28%, overcoming the challenges associated with traditional diagnostic models, which are characterized by long training durations, low identification efficiency, and weak collaborative analysis capabilities.

Switchgear is an important component of the power supply system, widely distributed in converter stations, substations, and distribution stations. It is key equipment for power system control and protection, characterized by large data, multiple types, and wide dispersion. Due to the large difference in switchgear operation and maintenance standards, high difficulty in operation and maintenance, and limited by the operation and maintenance capacity of the distribution network, it leads to poor switchgear reliability and high failure rate1,2. Partial discharge (PD) detection is an important means to detect the health status and insulation condition of switchgear, which is widely used due to the advantages of high sensitivity and timeliness. With the rapid development of power grids, the number of switchgears shows rapid growth. To ensure the accuracy of PD recognition methods, various detection means such as ultrasonic detection method and ultra-high frequency (UHF) detection method are widely used to obtain various heterogeneous data such as map type, picture type, real-time monitoring, etc3.

In practical applications, multi-dimensional, heterogeneous, and massive data from many devices need to be processed to obtain accurate PD recognition results for switch-gear. However, the large number of widely distributed switchgear and massive data pose a great challenge to the calculation method and the design of recognition algorithms. At present, when diagnosing and identifying the insulation status of the switchgear, the original signal of PD inside the switchgear is first obtained by sensors. After that, the signal is transmitted to the data processing center for extraction, dimensionality reduction, recognition, diagnosis and other processing. However, this method requires a large amount of network bandwidth, transmission rate and data center processing capacity. Relying on the existing on-site network configuration and calculation configuration, it is difficult to meet the requirements for timeliness and accuracy. In addition, a large amount of fault recording and map data are transmitted to a node for centralized processing, which will lead to excessive load on the node, poor delay, and poor stability. Cloud-edge collaboration calculation technology, which has been widely used in recent years, can well solve the above problems.

Cloud-edge collaboration calculation refers to a distributed calculation model based on the “cloud-edge-end” architecture. Among them, “cloud” is the center node of traditional cloud calculation and the control end of edge computing; “edge” is the core of cloud calculation, which is divided into infrastructure edge and equipment edge; “end” is the terminal equipment, and in the PD recognition of switchgear, the end is the UHF sensor and broadband pulse current sensor that obtains the PD signal. Compared with the traditional center-based calculation model, cloud-edge collaboration calculation adds the edge computing link between the distributed end-side sensors and the traditional cloud calculation center. It utilizes the excess data processing capabilities of collection devices, communication devices, etc. on the edge side to realize a distributed data processing technology that integrates network, computing, storage and applications. edge computing performs preliminary processing of collected data at the edge of the network to achieve the distribution of massive data, which can effectively reduce the network transmission pressure and the data center processing pressure4,5. Based on the above advantages, cloud-edge collaboration can solve the impact of large-scale multidimensional and heterogeneous data formed by massive equipment on the traditional equipment diagnosis model relying on data centers. It safeguards the requirements of switch-gear PD recognition for real-time, timeliness and accuracy. Currently, cloud-edge collaboration has been applied in various aspects such as distribution network management and operation6, intelligent transportation7, face recognition8, and so on.

In terms of power equipment operation and maintenance management, literature9 implemented the localization of homologous local amplifier signals based on mean drift algorithm on the cloud edge collaboration platform to eliminate interfering signals and retain the validity signals. Literature10 integrated edge computing technology with optimized back propagation neural network (BPNN) model and used it in the prediction and assessment of power grid status. Literature11 proposed a state recognition system for the thermal characteristics of gas-insulated switchgear based on cloud-edge collaboration, and identified the recognition strategy by simulating the shell temperature of gas insulated switchgear. Literature12 developed a PD monitoring unit for cable joints based on edge computing, and constructed a PD defect detection method based on an adaptive neuro-fuzzy system. However, none of the above literature involves the recognition of PD defects in switchgears, and does not deeply introduce the specific implementation process and application effect of cloud-edge collaboration. To summarize, the current cloud-edge-end is rarely used in the field of PD defect type recognition in switchgears.

Relevant scholars have also conducted a lot of research on the recognition of PD fault types. Literature13 utilized recurrent neural network to train the skewness, steepness, and Weibull parameters of PD signals to achieve PD pattern recognition. Literature14 constructed a fault tree based on the shape parameters, discharge volume center and discharge phase center of the PD signal. Literature15 used a clustering algorithm to analyze 34-dimensional image features including color moments, gray gradient co-occurrence matrix, and pseudo-Zernike moments of the PD spectrum. Methods such as support vector machine (SVM)16, fractal theory17, random forest (RF)18, and neural network (NN)19,20 have also been widely used. However, all of the above methods are shallow learning (SL) methods. In the training process, there are problems such as difficulty in determine initial parameters, insufficient training, and poor generalization ability. They are not suitable for identifying PD defects in a large number of switchgears in the context of edge computing.

In order to solve the above problems, this paper proposes a method for diagnosing and recognizing PD in switchgear based on cloud-edge collaboration and deep learning (DL). A cloud-edge collaboration recognition architecture containing the collection layer, the edge layer and the data processing cloud layer is constructed. UHF sensors and broadband pulse current sensors obtain the original PD signal from the switchgear. The signal is transmitted to the edge layer for feature extraction, feature dimensionality reduction and other operations. After that, it is sent to the data processing cloud layer for training based on DL methods. In order to apply to the cloud-edge collaboration calculation, a PD feature dimensionality reduction method based on locally linear embedding (LLE) is pro-posed to eliminate redundant state quantities and improve the recognition speed and accuracy. A method for identifying PD defects in switchgears based on deep belief net-work (DBN) is proposed. It overcomes the difficulties in determining the parameters and insufficient training of the traditional SL, and improves the accuracy and generalization ability of the recognition method. Finally, this paper simulates various PD defects based on actual switchgears, and verifies the accuracy of the recognition based on the DBN method and the usability of the cloud-side collaboration calculation architecture.

The concept of cloud-edge collaboration calculation evolved from content distribution networks. After 2015, cloud-edge collaboration calculation technology has been widely used and rapidly developed, and the related technical system has become more and more complete. edge computing is defined as a new calculation model that deploys smart platforms with calculation, storage, and application capabilities on the network side close to the data source20,21. In cloud-edge collaboration calculation, the edge computing model is a supplement to the centralized cloud calculation model. The complementary collaboration of cloud calculation and edge computing can better meet the needs of various application scenarios. A generalized calculation framework for cloud-edge cooperation calculation is given in related studies23,24, as shown in Fig. 1.

The general architecture of cloud-edge collaboration calculation.

Among them, the terminal layer corresponds to specific devices in application scenarios, such as smartphones, virtual reality devices, industrial equipment, etc. In the context of the rapid development of the internet of things, the terminal layer contains billions of devices, generating massive amounts of data. The edge layer is composed of edge nodes. The edge nodes can be devices with calculation capabilities, such as mobile phones, tablets, etc., or can be deployed in the gateway, router and other equipment in the network connection. The cloud layer, which is the data processing center, is used to process the settlement results of edge computing nodes or tasks that cannot be processed by edge computing nodes.

According to the cloud-edge collaboration calculation architecture shown in Fig. 1, combined with the actual situation of the on-site switchgear, a switchgear recognition architecture based on cloud-edge collaboration calculation is constructed, as shown in Fig. 2.

PD defect recognition architecture for switchgear based on cloud-edge collaboration calculation.

As shown in Fig. 2, the PD defect recognition architecture of the switchgear based on edge computing technology includes three parts: terminal layer, edge layer and cloud layer. The terminal layer also includes terminal equipment layer and terminal collection layer. The terminal equipment layer is the switchgear which is the research object of this paper. The terminal collection layer is various types of sensors used to collect PD signals in the switchgear, such as optical sensors, photomultipliers, ultrasonic sensors, ultra-high frequency sensors, etc. The edge layer is located on the top of the terminal layer for analyzing and processing the collected signals. In practical applications, since the sensing devices of the terminal collection layer do not have calculation capability. Therefore, the edge layer is not only responsible for processing the original signals, but also, at the same time, undertakes the computational tasks issued by the cloud calculation center layer. That is, the edge layer offloads the huge computation volume of the traditional data processing center and reduces the load of the data processing center. In order to ensure data security and consider the impact of electromagnetic interference in the field, the architecture proposed in this paper uses optical fiber for data transmission, instead of signal transmission based on wireless transmission. In the signal transmission method, it is a unidirectional data transmission from the terminal layer to the edge collection layer. while it is a bidirectional data transmission between the terminal collection layer and the edge layer. On the one hand, the edge computing node obtains signals from the collection device for processing. On the other hand, the edge computing node sends the analysis results or decision-making instructions to the terminal collection device. There is also a bi-directional data transmission between the cloud calculation center layer and the edge layer. The edge computing nodes transmit the preprocessing results to the cloud calculation center for centralized computation. And the cloud calculation center sends the results of centralized computation or the tasks that need to be processed to the edge layer for processing.

The PD diagnosis framework in switchgear based on cloud-edge collaboration calculation is a kind of guarantee to reduce the load of cloud calculation center and improve the real-time, timeliness and accuracy. However, the recognition of PD defect types still needs the corresponding recognition methods. Typical defects in switchgear are corona discharge, air gap discharge, discharge along dielectric surface, and suspended discharge. Among them, corona discharge is caused by the presence of metal burrs and tips on the inner surface of conductors and cabinets due to irregular component processing and assembly processes. Air gap discharge is generated by the existence of gaps inside or on the surface of insulators in the switchgear. Discharge along dielectric surface is due to the operation of the equipment insulation parts on the existence of dirt, moisture and condensation and produced. Suspended discharge is caused by improper installation techniques on site. It is caused by problems such as insufficient insulation distance between conductors and failure to tighten busbar bolts due to regulations. After determining the type of PD, it is necessary to choose a suitable PD recognition method according to the characteristics of edge computing technology. For this reason, this paper proposes a PD recognition method based on DBN. The method includes three parts: feature extraction, feature dimensionality reduction, and pattern recognition.

In the field applications, UHF sensors and broadband pulse current sensors are ar-ranged in the edge collection layer to obtain signals such as the discharge amount q, discharge pulse repetition rate n, discharge phase φ and discharge waveform of PDs. The PD RGB spectrum, n-φ histogram, qmax-φ histogram, qave-φ histogram and waveform chart are formed. Various feature quantities extracted from the above spectra are shown in Table 1.

Considering the skewness direction and degree of data distribution in the positive and negative half cycles of the three types of bar charts, namely n-φ, qmax-φ, and qave-φ, skewness is selected as the feature quantity. The skewness of the positive and negative half cycles of the three types of bar charts are Skn+, Skn-, Skm+, Skm-, Ska+, and Ska-. The kurtosis is selected as a feature to assess the concentration of data in both half-cycles and its smoothness relative to a normal distribution, denoted as kun+, kun-, kum+, kum-, kua+, and kua-. The correlation of shape profiles for both half-cycles is evaluated using the cross-correlation coefficient, represented as ccn, ccm, and cca. The asymmetry or nonlinear relationship of the data in both half-cycles is characterized by the asymmetry coefficient, indicated as qqn, qqm, and qqa.

From the PD RGB spectra, the first, second, and third moments of color are employed to describe the characteristics of color distribution, specifically the intensity and distribution features of discharge. The moments for the R, G, and B colors are represented as µ1, µ2, µ3, σ1, σ2, σ3, s1, s2, and s3, respectively.

Additionally, the gray-level gradient co-occurrence matrix method is utilized to obtain the joint statistical distribution of pixel intensity and edge gradients, allowing for the analysis of the resolution patterns of pixel intensities and gradients within the image, as well as the spatial relationships between each pixel and its neighboring pixels. The following state variables are used to characterize these patterns and relationships: average intensity (T1), average gradient (T2), intensity variance (T3), gradient variance (T4), small gradient dominance (T5), large gradient dominance (T6), intensity distribution uniformity (T7), gradient distribution uniformity (T8), energy (T9), correlation (T10), intensity entropy (T11), gradient entropy (T12), mixed entropy (T13), inertia (T14), and inverse difference moment (T15).

Furthermore, ten Hu invariant moments (R1, R2, R3, R4, R5, R6, R7, R8, R9, R10) and fifteen Zernike moments (A11, A21, A22, A31, A32, A33, A41, A42, A43, A44, A51, A52, A53, A54, A55) are utilized as shape features for the PD RGB spectra.

For the waveform spectra of partial discharge, thirteen waveform features are extracted from the time domain, frequency domain, time-frequency joint analysis, and shape characteristics. These features include: the rise time ts(t901−t101)(difference between the moments when the waveform reaches 90% and 10% of its maximum amplitude), and the fall time tf(t102−t902) difference between the moments when the waveform drops back to 10% and 90% of its maximum amplitude); the duration tc (the pulse remains above a defined threshold until returning to the initial threshold); the duration Δt(t502−t501) time from when the pulse first reaches 50% amplitude to when it returns to 50%); the frequency fc (corresponding to the maximum amplitude of the discharge time-domain waveform after Fourier transform, ); the skewness Skw (characterizes the direction of waveform data distribution); the kurtosis Kuw (indicates the concentration of data and its flatness relative to a normal distribution); the peak factor Qc (characterizes the relationship between the peak value and the RMS value); the time centroid N (representing the density of time domain distribution); the frequency centroid N (describes the frequency distribution density and frequency centroid); the equivalent duration T (representing the signal’s duration); the equivalent bandwidth F (describing the spectral range of the signal).

According to Table 1, a total of 80 feature quantities can be obtained for the PD signal of switchgear. However, not all of these feature quantities can improve the accuracy of PD recognition. On the contrary, the existence of redundant feature quantities not only increases the amount of computation and affects recognition efficiency, but also the correlation between the feature quantities affects the recognition process. Therefore, this paper proposes a feature space dimensionality reduction method based on LLE. LLE dimensionality reduction is a nonlinear dimensionality reduction algorithm, which is a typical manifold learning method. It believes that there is a low-dimensional space (local space) in the high-dimensional feature space. The local space is the same embryo as the Euclidean space, that is, the local space has the properties of the Euclidean space, and the local space is called a “manifold”. Many research results show that LLE method can maintain the original manifold structure after dimensionality reduction, which is more in line with the characteristics of the high-dimensional feature space of PD25,26,27. The LLE dimensionality reduction method has relatively small computational complexity and is easy to implement. It is suitable for high-dimensional feature space dimensionality reduction in edge computing situations.

After completing the dimensionality reduction of the feature space, it is necessary to construct the mapping relationship between the feature space and PD types based on the corresponding recognition algorithm. This paper uses an recognition method based on DL to construct a mapping relationship. Compared with traditional SVM, artificial neural network (ANN), RF and other methods, the DL method can dig deeper into the characteristics of the PD feature space and build a more accurate mapping relationship. It can solve problems such as local optimality and diffusion gradient of traditional NN methods. In practical applications, DBN is a widely used method, which deeply integrates probability theory, NN and machine learning (ML), and has achieved better application results in many fields28,29. The DBN is stacked by multiple restricted boltzmann machines (RBM) and a classification network.

There is no connection between neurons in the same layer of RBM. The activation of neurons in the visible layer V and hidden layer H are independent of each other. Assume that the input of the visible layer unit is v={v1,v2,…,vm}, the output of the hidden layer unit is h={h1,h2,…,hn}, and the network parametersθ={ω, a, b}. Define the RBM energy function as:

where ωij represents the connection weight between ai and bj, ai represents the bias of the i-th node in the visible layer, and bj represents the bias of the j-th node in the hidden layer.

According to the definition of the energy function, the joint probability distribution of each node in the visible layer V and hidden layer H is:

The Sigmoid function is used as the activation function. When the samples in the visible layer V (input samples) are known, the probability that hj is 1 is calculated as:

Similarly, when the hidden layer H (output result) is known, the probability that vi is 1 is calculated as:

During the training process, the objective function of RBM is:

In the formula, <⋅>P(h|v, θ) is the mathematical expectation of the joint probability distribution of the visible layer and the hidden layer, and <⋅>P(v|h,θ) is the mathematical expectation of the probability distribution of the hidden layer when the input sample data is known. The training process of RBM is the process of maximizing Eq. (10). Currently, the sampling method based on the contrastive divergence (CD) algorithm is the most widely used training method30. After training on the RBM layer, classification is performed using a classification network based on the softmax function. The structure of the DBN used for PD recognition in switchgear is shown in Fig. 3.

Basic structure of DBN for PD recognition of switchgear.

The feature collection, feature dimensionality reduction, and deep belief network (DBN)-based recognition processes with the switchgear edge computing architecture shown in Fig. 2. The switchgear PD recognition process based on edge computing and DBN that can be applied to actual on-site scenarios is shown in Fig. 4.

PD recognition process of switchgear based on edge computing and DBN.

A collection unit including a UHF detection device and a broadband pulse current detection device is arranged on the on-site switchgear. Each switchgear corresponds to a collection unit. After the collection unit obtains the data, it is connected to the edge computing node. An edge computing node may correspond to multiple collection units. The edge computing node is equipped with a feature extraction module, a feature dimensionality reduction module, and an in-situ recognition module. The feature extraction module receives the original PD feature data obtained from the collection unit and per-forms feature extraction to extract 80-dimensional features as shown in Table 1. After that, the 80-dimensional features are transmitted to the feature dimensionality reduction module for dimensionality reduction according to rules. The post-dimensional feature quantities are transmitted to the data processing center, where a PD defect recognition classifier based on DBN is trained. After completing the training of the PD defect recog-nition classifier, the classifier is transmitted to an in-situ recognition module in the edge computing node. The in-situ recognition module takes the dimensionally reduced feature quantities as the input for recognition. The recognition results are fed back to the data processing center to expand the sample base for training the DBN. At the same time, a corresponding warning signal is given based on the recognition results. The warning signal can be transmitted to the collection unit or switchgear for display, or can also be displayed in the data processing center. In Fig. 5, arrows with different colors are used to characterize different data streams. The blue arrows represent the data flow from the bottom up, and the red arrows represent the data flow after processing by the data processing center and grounding recognition module.

Comparison of accuracy of different PD recognition methods.

In order to verify the effectiveness and accuracy of the PD method for switchgear based on edge computing and DBN, a verification platform was built based on a 35 kV switchgear (model: XGN-40.5) in a laboratory environment. Four types of defects, air gap discharge, creeping discharge, corona discharge, and suspension discharge, are simulated in the switchgear respectively. The defect simulation method is consistent with the reference31. A PD detection system based on UHF and a PD detection system based on broadband pulse current are arranged in the collection unit. Huawei’s Atlas200 DK ap-plication development board is used as the calculation node in the edge computing layer. The development board integrates the Ascend 310 AI processor with a computing power of up to 16 TOPS. The feature extraction module, feature dimensionality reduction mod-ule and in-situ recognition module on the board are developed and deployed based on Python3.5. An Inspur SA5212 M4 server is used as the main computing device in the data processing center. The server uses Intel Xeon E5-2620V3 processor with 64GB of calculation memory. The DBN algorithm for the data processing center is also developed and deployed based on Python3.5.

The experimental simulation includes two stages: classifier pre-training and re-al-time recognition. In the classifier pre-training stage, sample data are obtained by simulating various defects, and an initial classifier for recognition is trained based on the sample data. This classifier can be continuously improved and supplemented by the subsequent recognition process. In the pre-training stage of the classifier, 350 sets of data are collected as samples for each type of defect respectively. In addition, 350 sets of data are also used as defect-free samples when no defects are present. In the feature extraction module, feature extraction is performed on the above sample data and a matrix of 175,080 is obtained. Further, the feature dimensionality reduction module in the edge computing node reduces the dimensionality of the matrix of 175,080 to obtain a feature matrix of 175,037. Subsequently, the dimensionally reduced feature matrix is uploaded to the data processing center. The DBN is invoked for training to obtain an initial classifier, which is synchronized to the in-situ recognition module in edge computing. In the real-time recognition stage, 117 sets of data are simulated in the switchgear for each defect as well as for the no-defect case, for a total of 585 sets of data. During the recognition process, for every set of data generated, the feature extraction module in the edge computing is in-voked for feature extraction. Afterwards, the feature quantity after dimensionality re-duction is obtained using the dimensionality reduction extraction rules recorded in the feature dimensionality reduction module. Further, the recognition is performed directly using the pre-trained classifiers stored in the in-situ recognition module to obtain the recognition results of the PD type. The recognition results of the above process are shown in Table 2. According to the results shown in Table 2, the accuracy rate of PD defect recognition in switchgears based on edge computing and DBN is 532/585 = 88.03%. Among them, the recognition rate of corona discharge is the highest, which can reach to 109/117 = 93.16%.

In order to compare the accuracy of different recognition methods, the classification methods based on BPNN, SVM and RF are arranged to recognize 585 sets of data. The recognition accuracy of different methods for each defect is shown in Table 3, and an intuitive comparison chart is drawn as shown in Fig. 5. According to the results in Table 2; Fig. 5, it can be seen that the switchgear PD recognition method based on DBN pro-posed in this paper has a higher recognition accuracy compared to the traditional identification method based on BPNN, SVM and RF. The accuracy be improved by 7.69%. It also has a better recognition accuracy when each defect type is recognized individually.

Furthermore, in order to verify the application effectiveness of edge computing in the recognition of PD defects in switchgear, a data center-based calculation method is deployed based on the same computing server. The feature extraction, feature dimensionality reduction, classifier training and classifier-based recognition processes are deployed on calculation servers in the data center. The computation time counted is from the time after the PD signal is output from the collection unit until separator training is completed or the single case recognition is completed. The computation time of different recognition methods is also calculated, as shown in Table 4. According to results in Table 4, it can be seen that under the edge computing architecture, the DBN-based recognition method can reduce (987 − 550)/987 = 44.28% relative to the traditional data center-based processing method when training the classifier. The time required by other recognition methods is also significantly reduced. In terms of single case recognition, since the computation is significantly lower than that of classifier training, there is little difference between edge computing and data center-based calculation, and the difference between different methods is also small. In summary, the training time of the switchgear PD defect recognition model is significantly reduced after the introduction of edge computing architecture. Therefore, the PD defect recognition method of switchgears based on edge computing can be well applied to the scene of rapid and synchronous recognition of multiple switchgears on site.

This paper proposes a switchgear PD defect recognition method based on edge computing and DL, and constructs an on-site switchgear PD defect recognition architecture including the terminal layer, collection layer, edge layer and data processing center. The edge layer undertakes the original signals processing as well as the calculation tasks issued by the data processing center, offloading the huge computational volume of the traditional data processing center and reducing the load on the data processing center. It is suitable for the scenario of synchronized and rapid PD detection, analysis and recognition of a large number of switchgear equipment.

A method of dimensionality reduction of high-dimensional PD feature space based on LLE and a method of PD defect recognition based on DBN are proposed. The method eliminates redundant feature quantities and significantly improves the PD defect recognition method. The actual switchgear defect simulation data shows that the accuracy of the PD recognition method based on DBN can reach 88.03%, which is 7.69% higher than that of the traditional method.

The time required for training the switchgear PD defect recognition model is significantly reduced under the edge computing architecture. The reduction can be up to 44.28% when using DBN to train the recognition model, which is significantly better than the traditional computation based on data processing center.

Data Statements：The datasets generated and/or analysed during the current study are not publicly available due (The Status Data of the Equipment Involves Numerous Private Information such as Power Companies, Equipment Manufacturers, and End Users) but are available from the corresponding author on reasonable request.

Yang, F. et al. Evaluation for switchgear health states based on multi-dimensional feature with optimal rank. High Voltage Engineering 46(11), 3934–3942 (2020).

MATH Google Scholar

Li, J. et al. Insulation characteristics of compound insulation in environmental gas switch cabinet. Science Technology and Engineering 46(11), 3934–3942 (2020).

MATH Google Scholar

Tang, Z. et al. Review on partial discharge pattern recognition of electrical equipment. High Voltage Engineering 43(7), 2263–2277 (2017).

MATH Google Scholar

Wang, L. et al. A fine-grained multi-acess edge computing architecture for cloud-network integration. Journal of Computer Research and Development 58(6), 1275–1290 (2021).

MATH Google Scholar

Ni, J. & Lv, M. Distributed multi-task edge computation offloading algorithm assisted by unmanned aerial vehicle and vehicle. Science Technology and Engineering 21(3), 1045–1051 (2021).

MATH Google Scholar

Guan, G. et al. Research on distribution network management and operation and maintenance system based on edge computing. Advances of Power System & Hydroelectric Engineering 36(10), 90–96 (2020).

MATH Google Scholar

Song, H., Jiao, J. & Liu, Y. Research on mobile edge computing in expressway. Computer Science 48(S1), 383–386 (2021).

MATH Google Scholar

Qian, T. & Zhang, F. Emotion recognition system bases on distributed edge computing. Computer Science 48(S1), 638–643 (2021).

MATH Google Scholar

Tian, J. et al. Edge computing method of partial discharge based on mean-shift. Power System Technology 45(6), 2449–2456 (2021).

MATH Google Scholar

Xu, Y. et al. Edge computing application in power network state prediction. Microcontrollers & Embedded Systems 20(11), 47–50 (2020).

MATH Google Scholar

Cheng, Z. et al. Research on state identification of thermal characteristics for GIS busbar based on edge computing. Power Systems and Big Data 23(11), 31–37 (2020).

MATH Google Scholar

Cao, P. et al. Intelligent sensing and monitoring of cable joints’ state based on edge computing. High Voltage Apparatus 56(9), 26–32 (2020).

MATH Google Scholar

Tang, J. et al. Application of the weibull distribution for the recognition of transformer PD phenomena. Journal of Chongqing University 4, 87–90 (2002).

MATH Google Scholar

Contin, A. & Rabach, G. PD analysis of rotating AC machine. IEEE Transactions on Electrical Insulation 28(6), 1033–1104 (1993).

MATH Google Scholar

Tang, J. et al. Statistical parameter method for PD pattern recognition. High Voltage Engineering 28(8), 4–6 (2002).

MATH Google Scholar

Jin, Z. Optimization of Features of Grey-Scale Maps Obtained from GIS Partial Discharge and Its Pattern Recognition (North China Electric Power University, 2013).

MATH Google Scholar

Krivda, A. et al. Can fractal features be used for recognizing 3-d partial discharge patterns. IEEE Transactions on Dielectrics & Electrical Insulation. 2(3), 352–359 (2002).

Google Scholar

Cheng, Y. & Zhang, Z. Multi-source partial discharge diagnosis of transformer based on random forest. Proceedings of the CSEE 38(17), 280–290 (2018).

MATH Google Scholar

Evagorou, D. et al. Feature extraction of partial discharge signals using the wavelet packet trans-form and classification with a probabilistic neural network. IET Science Measurement & Technology 4(3), 177–192 (2010).

MATH Google Scholar

Zhu, N. et al. Partial discharge failure evaluation based on sf6 decomposition characteristics. Proceedings of the CSEE 39(3), 307–316 (2019).

MATH Google Scholar

Shi, W. S. et al. Edge computing: Vision and challenges. IEEE Internet of Things Journal 3(5), 637–646 (2016).

MATH Google Scholar

Portilla, J. et al. The extreme edge at the bottom of the Internet of things: A review. IEEE Sensors Journal 1, 1–20 (2019).

MATH Google Scholar

Li, L. et al. Architecture, challenges and applications of edge computing. Big Data Research 5(2), 3–16 (2019).

CAS MATH Google Scholar

Edge computing industry alliance, Alliance of industrial internet. Edge Calculation Reference Architecture 3.0[R/OL] (2018-11). http://www.ecconsortium.org/Lists/show/id/334.html.

Fang, T. & Fu, D. A Multi-model MLLE-PCA method for unstable industrial process monitoring. in 2018 IEEE International Conference on Mechatronics, Robotics and Automation, Hefei, China 75–79 (2018).

Liu, S., Chen, Z. & Jiao, F. Detection of maize seed germination rate based on improved locally linear embedding. Computers and Electronics in Agriculture 204, 107514 (2023).

MATH Google Scholar

Duan, Y. F. et al. Study on non-destructive detection method for egg freshness based on LLE-SVR and visible/near-infrared spectrum. Guang Pu Xue Yu Guang Pu Fen Xi 36(4), 981–985 (2016).

CAS PubMed MATH Google Scholar

Zhang, A. & Deng, F. Strength prediction of phosphogypsum filling material based on improved deep belief net-work. Science Technology and Engineering 20(18), 7220–7225 (2020).

MATH Google Scholar

Li, A. & Bi, Z. Multi-objective optimization based on deep belief network model. Science Technology and Engineering 19(16), 8–14 (2019).

MATH Google Scholar

Yang, Y. & Zhang, K. Research on book recommendation algorithm integrating FP-growth and RBM. Journal of Shanxi University of Science & Technology 39(4), 156–160 (2021).

MATH Google Scholar

Wang K. Development of On-Line Partial Discharge Monitoring System for Ring Main Unit (Harbin University of Science and Technology, 2020).

Download references

This research was funded by the Research Project of State Grid Sichuan Electric Power Company (52199719000X),

State Grid Sichuan Electric Power Research Institute, Chengdu, 610000, China

Zhijie Jia & Songhai Fan

State Grid Sichuan Electric Power Company, Chengdu, 610000, China

Zhichuan Wang

State Grid Sichuan Maintenance Company, Chengdu, 610000, China

Shuai Shao & Dameng He

You can also search for this author inPubMed Google Scholar

You can also search for this author inPubMed Google Scholar

You can also search for this author inPubMed Google Scholar

You can also search for this author inPubMed Google Scholar

You can also search for this author inPubMed Google Scholar

Zhijie Jia: Conceptualization, Methodology, Validation, Formal analysis, ResourcesSonghai Fan: Software, Validation, Formal analysis, Investigation, VisualizationZhichuan Wang: Validation, Investigation, VisualizationShuai Shao: Resources, Data Curation, SupervisionDameng He：Resources, Data Curation, Supervision.

Correspondence to Zhijie Jia.

The authors declare no competing interests.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Below is the link to the electronic supplementary material.

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.

Reprints and permissions

Jia, Z., Fan, S., Wang, Z. et al. Partial discharge defect recognition method of switchgear based on cloud-edge collaborative deep learning. Sci Rep 15, 10956 (2025). https://doi.org/10.1038/s41598-024-81478-9

Download citation

Received: 17 March 2024

Accepted: 26 November 2024

Published: 31 March 2025

DOI: https://doi.org/10.1038/s41598-024-81478-9

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative