Online Partial Discharge Recognition Method Based on Knowledge Distillation 1D-CNN

Abstract

This invention belongs to the field of power equipment fault diagnosis technology, specifically disclosing an online partial discharge identification method based on knowledge distillation and 1D-CNN. The method includes the following steps: performing Fourier transform on the original signal and then implementing four data augmentation techniques: amplitude modulation, phase modulation, bandwidth enhancement, and frequency domain noise injection; constructing a multi-branch one-dimensional convolutional neural network to extract time-frequency features, and training a teacher model using residual connections and SE attention mechanisms; compressing the teacher model into a lightweight student model using a knowledge distillation strategy with preset temperature parameters and weight coefficients; converting the student model into a mobile-compatible format and building a mobile APP using a QT host computer; and achieving online real-time identification. This invention effectively alleviates the sample imbalance problem. Furthermore, knowledge distillation maintains high accuracy while achieving model size compression, improved inference speed, low computational cost, strong real-time performance, and high precision.

Description

Technical Field

[0001] This invention belongs to the field of power equipment fault diagnosis technology, specifically relating to an online partial discharge identification method based on knowledge distillation 1D-CNN, which is applicable to online partial discharge identification of high-voltage equipment such as gas-insulated switchgear (GIS). Background Technology

[0002] The insulation status of high-voltage equipment such as gas-insulated switchgear (GIS) is directly related to the safe and stable operation of the power grid. Partial discharge (PD), as a key indicator of insulation degradation, can effectively reflect the development trend of insulation defects. Therefore, online partial discharge identification of GIS is of great engineering significance.

[0003] Existing partial discharge (PD) identification methods mainly rely on manual feature extraction and shallow classifiers. While these methods show some effectiveness under laboratory conditions, they are susceptible to noise interference in the complex electromagnetic environment of actual substations and lack robustness to equipment differences. Furthermore, because PD failures are low-probability events, normal samples far outnumber fault samples, leading to extreme imbalances between different fault types. This results in low model accuracy and misclassification issues.

[0004] With the development of deep learning technology, end-to-end methods such as one-dimensional convolutional neural networks (1D-CNN) can automatically extract discriminative features of time-series signals. However, high-precision models have a large number of parameters, making them difficult to deploy in real time on mobile terminals. Meanwhile, partial discharge pulse signals have multi-scale characteristics, and a single convolutional kernel cannot simultaneously capture long-term trends and short-term high-frequency details.

[0005] Knowledge distillation technology, through a teacher-student model, can significantly reduce model complexity while maintaining accuracy, providing a feasible solution for edge device deployment. Furthermore, traditional data augmentation is mostly performed in the time domain, while the diagnostic information of PD signals is highly dependent on the spectral structure. Therefore, frequency domain augmentation is more in line with the characteristic mechanisms of PD and helps improve the robustness of the model to different scenarios. Currently, there is no technology utilizing frequency domain data for augmentation in online partial discharge identification technology for high-voltage equipment such as GIS.

[0006] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art. Summary of the Invention

[0007] The purpose of this invention is to provide an online partial discharge identification method based on knowledge distillation and 1D-CNN. This method effectively alleviates the sample imbalance problem through frequency domain data augmentation technology, and at the same time helps to improve the ability of one-dimensional convolutional neural networks to extract time-frequency features of partial discharge pulse signals. In addition, this invention also adopts a multi-branch architecture to extract features at different scales in parallel, which helps to enhance the model's ability to represent multi-scale patterns of partial discharge pulse signals. At the same time, it combines knowledge distillation to achieve model lightweighting, which helps to solve the technical problems of low identification accuracy and difficult model deployment of existing methods under complex working conditions.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] The online partial discharge identification method based on knowledge distillation 1D-CNN includes the following steps:

[0010] Step 1. Acquire the raw signal, i.e., the partial discharge pulse signal, and perform labeling and classification preprocessing;

[0011] Step 2. Extract the amplitude spectrum and phase spectrum of the original signal through Fourier transform, and perform data augmentation using four methods: amplitude modulation, phase modulation, frequency band modulation, and noise injection. Obtain the augmented dataset through inverse Fourier transform.

[0012] Step 3. Construct an online partial discharge identification model with a multi-branch structure as the teacher model;

[0013] The teacher model includes three feature extraction branches, four residual blocks, a fusion module, and a Softmax layer;

[0014] All three feature extraction branches are one-dimensional convolutional extraction branches, defined as the trunk branch, detail branch, and high-frequency branch, respectively, and are used to extract feature maps at different scales. The outputs of the three feature extraction branches are concatenated using Concat.

[0015] The stitched feature map is processed sequentially through each residual block. After each residual block, an SE channel attention mechanism is combined, and the feature processed by the last SE channel attention mechanism is input into the fusion module.

[0016] The fusion module uses one-dimensional global max pooling and one-dimensional global average pooling for parallel processing, and then merges the processing results of global max pooling and global average pooling. Finally, it outputs the predicted probabilities of each category through the Softmax layer.

[0017] The teacher model is trained using the expanded dataset to obtain a well-trained teacher model.

[0018] Step 4. Construct a lightweight student model and generate soft labels from the trained teacher model; preset temperature parameters and weight coefficients, perform knowledge distillation by combining hard and soft label losses, and obtain the optimal compressed model through an early stopping mechanism;

[0019] Based on the trained student model, online partial discharge identification is performed to obtain the specific discharge type.

[0020] Furthermore, based on the aforementioned online partial discharge identification method using 1D-CNN based on knowledge distillation, this invention also proposes a corresponding online partial discharge identification system based on 1D-CNN based on knowledge distillation, the technical solution of which is as follows:

[0021] An online partial discharge recognition system based on knowledge distillation and 1D-CNN includes the following modules:

[0022] The preprocessing module is used to acquire partial discharge pulse signals and perform labeling and classification preprocessing.

[0023] The data augmentation module is used to extract the amplitude spectrum and phase spectrum of the partial discharge pulse signal through Fourier transform. It uses four methods for data augmentation: amplitude modulation, phase modulation, frequency band modulation, and noise injection. The augmented dataset is obtained through inverse transform.

[0024] The teacher model building module is used to construct an online partial discharge identification model with a multi-branch structure, which serves as the teacher model.

[0025] The teacher model includes three feature extraction branches, four residual blocks, a fusion module, and a Softmax layer;

[0026] All three feature extraction branches are one-dimensional convolutional extraction branches, defined as the trunk branch, detail branch, and high-frequency branch, respectively, and are used to extract feature maps at different scales. The outputs of the three feature extraction branches are concatenated using Concat.

[0027] The stitched feature map is processed sequentially through each residual block. After each residual block, an SE channel attention mechanism is combined, and the feature processed by the last SE channel attention mechanism is input into the fusion module.

[0028] The fusion module uses one-dimensional global max pooling and one-dimensional global average pooling for parallel processing, and then merges the processing results of global max pooling and global average pooling. Finally, it outputs the predicted probabilities of each category through the Softmax layer.

[0029] The teacher model is trained using the expanded dataset to obtain a well-trained teacher model.

[0030] The prediction module is used to build a lightweight student model and generate soft labels from the trained teacher model. It presets temperature parameters and weight coefficients, combines hard and soft label losses to perform knowledge distillation, and obtains the optimal compressed model through an early stopping mechanism.

[0031] Based on the trained student model, online partial discharge identification is performed to obtain the specific discharge type.

[0032] Furthermore, based on the above-mentioned online partial discharge identification method based on knowledge distillation 1D-CNN, this invention also proposes a computer device, including a memory and one or more processors, wherein executable code is stored in the memory, and when the processor executes the executable code, it is used to implement the steps of the above-mentioned online partial discharge identification method based on knowledge distillation 1D-CNN.

[0033] Furthermore, based on the above-mentioned online partial discharge identification method based on knowledge distillation 1D-CNN, this invention also proposes a computer-readable storage medium storing a program thereon; when executed by a processor, the program is used to implement the steps of the above-mentioned online partial discharge identification method based on knowledge distillation 1D-CNN.

[0034] The present invention has the following advantages:

[0035] As described above, to address the issues of sample imbalance, noise interference, and mobile deployment constraints in partial discharge identification of gas-insulated switchgear, this invention proposes an online partial discharge identification method based on frequency domain data augmentation, one-dimensional convolutional neural networks (1D-CNN), and knowledge distillation. This method focuses on the frequency domain features of partial discharge pulse signals, effectively mitigating sample imbalance through frequency domain data augmentation. Simultaneously, it proposes an online partial discharge identification model with a multi-branch structure as the teacher model. This teacher model, combining a multi-branch one-dimensional convolutional neural network structure, residual connections, and SE attention mechanism, can automatically extract the time-frequency features of partial discharge pulse signals, significantly improving identification accuracy compared to traditional manual feature methods and providing a high-quality knowledge source for knowledge distillation. To meet the requirements of lightweight deployment, this invention also proposes a student model. Using the trained teacher model, the student model undergoes distillation learning. Through soft-label distillation with a temperature parameter T=2.5 and a weight balancing strategy with α=0.6, the model is significantly compressed while maintaining identification accuracy, resulting in a substantial reduction in the number of parameters and laying the foundation for mobile deployment. The lightweight student model can be deployed on mobile terminals to achieve end-to-end real-time inference. Compared with traditional two-dimensional convolutional networks, the one-dimensional structure adopted in this invention requires less computation and has a faster inference speed, meeting the immediacy requirements of on-site detection and providing a portable, high-precision intelligent solution for power equipment condition monitoring. Attached Figure Description

[0036] Figure 1 This is a flowchart of the online partial discharge identification method based on knowledge distillation 1D-CNN in an embodiment of the present invention;

[0037] Figure 2 This is a schematic diagram of the teacher model framework constructed in an embodiment of the present invention;

[0038] Figure 3 This is a schematic diagram of the student model framework built in an embodiment of the present invention;

[0039] Figure 4 This is a PRPS diagram of tip discharge data in an embodiment of the present invention;

[0040] in Figure 4 In the diagram, (a) is the PRPS diagram for tip discharge type A, (b) is the PRPS diagram for tip discharge type B, (c) is the PRPS diagram for tip discharge type C, and (d) is the PRPS diagram for tip discharge type D.

[0041] Figure 5 This is a PRPS diagram of particle discharge data in an embodiment of the present invention;

[0042] in Figure 5 In the diagram, (a) is the PRPS diagram for particle discharge type A, (b) is the PRPS diagram for particle discharge type B, (c) is the PRPS diagram for particle discharge type C, and (d) is the PRPS diagram for particle discharge type D.

[0043] Figure 6 This is a PRPS diagram of the air gap discharge data in an embodiment of the present invention;

[0044] in Figure 6 In the diagram, (a) is the PRPS diagram for air gap discharge type A, (b) is the PRPS diagram for air gap discharge type B, (c) is the PRPS diagram for air gap discharge type C, and (d) is the PRPS diagram for air gap discharge type D.

[0045] Figure 7 This is a PRPS diagram of the levitation discharge data in an embodiment of the present invention;

[0046] in Figure 7 In the diagram, (a) is the PRPS diagram for suspended discharge type D, (b) is the PRPS diagram for suspended discharge type B, (c) is the PRPS diagram for suspended discharge type C, and (d) is the PRPS diagram for suspended discharge type A.

[0047] Figure 8 This is a PRPS diagram of surface discharge data in an embodiment of the present invention;

[0048] in Figure 8In the diagram, (a) is the PRPS diagram for surface discharge type A, (b) is the PRPS diagram for surface discharge type B, (c) is the PRPS diagram for surface discharge type C, and (d) is the PRPS diagram for surface discharge type D. Detailed Implementation

[0049] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0050] Example 1

[0051] This embodiment 1 describes an online partial discharge identification method based on knowledge distillation using 1D-CNN. First, the method focuses on the frequency domain features of partial discharge pulse signals, extracting the amplitude and phase spectra through FFT transformation. Data augmentation techniques such as amplitude modulation, phase modulation, bandwidth enhancement, and frequency domain noise injection are employed to effectively alleviate sample imbalance and improve the model's robustness to complex electromagnetic environments. Second, a multi-branch one-dimensional convolutional neural network is constructed as the teacher model. Combined with residual connections and attention mechanisms, it automatically extracts the time-frequency features of partial discharge pulse signals, avoiding the subjectivity and limitations of manual feature engineering and significantly improving recognition accuracy. Finally, knowledge distillation technology is used. Through soft-label distillation with a temperature parameter T=2.5 and a weight balancing strategy with α=0.6, the teacher model is compressed into a lightweight student model. This significantly reduces model complexity while maintaining recognition accuracy, enabling end-to-end deployment from a QT host computer to a mobile device. This meets the real-time detection requirements on-site and provides a portable, high-precision, intelligent solution for power equipment condition monitoring.

[0052] like Figure 1 As shown, the online partial discharge identification method based on knowledge distillation 1D-CNN includes the following steps:

[0053] Step 1. Acquire the raw signal, i.e., the partial discharge pulse signal, and perform labeling and classification preprocessing.

[0054] Step 1.1. First, collect the original signals of the gas-insulated switchgear, namely the partial discharge pulse signals. Each type of partial discharge pulse signal has the same length and is arranged in a 5000×1 manner to obtain one-dimensional partial discharge data.

[0055] The data for each type of partial discharge pulse signal includes 5,000 sampling points.

[0056] Step 1.2. Extract discharge data for different fault types from the partial discharge pulse signal data, label them as tip discharge, particle discharge, air gap discharge, suspension discharge and surface discharge, and save the data with the same data length.

[0057] Step 2. Extract the amplitude and phase spectra of the original signal through Fourier transform, and perform data augmentation using four methods: amplitude modulation, phase modulation, band modulation, and noise injection. Obtain the augmented dataset through inverse Fourier transform.

[0058] Traditional data augmentation is mostly performed in the time domain, while the diagnostic information of PD signals is highly dependent on the spectral structure. Therefore, frequency domain augmentation is more in line with the characteristic mechanism of PD and helps to improve the robustness of the model to different scenarios.

[0059] Step 2.1. First, process the original signal. Perform a Fourier transform to extract the amplitude spectrum and phase spectrum separately, as shown in the following formula:

[0060] .

[0061] Using the above formula, time-domain data can be converted into frequency-domain data through Fourier transform.

[0062] in For the frequency domain representation of the signal, This represents a data point in the original signal, i.e., a sampling point. For ease of subsequent enhancement processing, [the following is used:] Represented as a combination of replication and phase, the formula is as follows:

[0063] ;

[0064] in and These represent the amplitude and phase of the spectrum, respectively.

[0065] Step 2.2. Based on the frequency domain representation of the signal from Step 2.1, four data augmentation methods are designed:

[0066] I. First, the transformed signal is amplitude modulated. A preset scaling factor range is used. For the amplitude spectrum, different scaling factors are selected from the scaling factor range to perform amplitude scaling, generating multiple variants.

[0067] Scaling factor in this embodiment For example, a range of 0.85-1.15 generates 6 variants, expressed by the following formula:

[0068] ;

[0069] in It is the enhanced spectrum. It is a uniform random function on the interval [-1, 1].

[0070] Amplitude modulation (AM) generates new samples by modulating the amplitude spectrum of a signal while keeping the phase information unchanged. This AM method is mainly used to simulate changes in the intensity of a discharge signal.

[0071] II. Next, phase modulation is performed:

[0072] A preset radian range is defined, and phase shifts of different radians within that range are added to the phase spectrum. For example, adding a phase shift within the range of [-0.2, 0.2] radians to the phase spectrum generates five variants, as expressed by the following formula:

[0073] ;

[0074] in This is the spectrum after phase modulation. is the phase modulation coefficient, with a value range of [-0.2, 0.2]. Let be a uniform random phase function on the interval [-π, π].

[0075] Phase modulation methods generate new samples by modulating the phase spectrum of the signal, which are used to simulate the randomness of the discharge occurrence time.

[0076] III. Then perform bandwidth enhancement, expressed by the following formula:

[0077] The original signal is transformed by Fourier transform to obtain multiple frequency domain data points, which are then evenly divided into multiple intervals according to the sampling time. For each interval of frequency domain data points, a preset multiple of energy is applied to enhance them individually.

[0078] In this embodiment, the preset multiple is adopted. This indicates the enhancement coefficient. For example, the value can be 1.2.

[0079] ;

[0080] in The spectrum after band enhancement. For window functions.

[0081] The definition is as follows:

[0082] .

[0083] in Indicates the lower bound of the frequency band. This indicates the upper limit of the frequency band. Frequency band enhancement methods achieve data enhancement by selectively enhancing specific frequency bands, which is significant for the characteristic frequency bands of different types of discharges.

[0084] IV. Finally, perform frequency domain noise injection:

[0085] Gaussian noise of different intensities was added to the amplitude spectrum, and random phase perturbations of different intensities were added to the phase spectrum.

[0086] Taking the amplitude spectrum as an example, the amplitude spectrum with added noise. The formula is expressed as follows:

[0087] ;

[0088] in This represents the noise intensity coefficient, with a value range of [0.01, 0.05]. It is a random function that follows a Gaussian distribution.

[0089] The frequency domain noise injection method enhances data by adding controllable noise in the frequency domain, thereby improving the model's noise resistance.

[0090] Step 2.3. After all four data augmentation methods are completed, the frequency domain data is reconstructed by inverse Fourier transform. The augmented data is then subjected to the following inverse Fourier transform to obtain the expanded discharge dataset.

[0091] ;

[0092] in Represents the spectrum after any type of data augmentation. This is the corresponding enhanced time-domain signal.

[0093] Partial discharge pulse signals are essentially high-frequency pulse superimposed noise. Their amplitude-frequency characteristics directly map the discharge energy and trigger timing. This invention uses Fourier transform to decompose the signal into amplitude and phase for frequency domain data enhancement. This effectively simulates amplitude fluctuations caused by equipment voltage fluctuations, phase drift caused by trigger delays, and the different frequency bands resulting from the specific bandwidths of different defects, as well as the different noise caused by on-site interference. Compared to traditional time-domain data enhancement, direct time shifting and scaling easily disrupt the phase relationship between pulses, even causing pulse overlap and information loss. Random amplitude scaling affects all frequencies simultaneously, failing to simulate the physical mechanisms of energy changes in certain defects only within specific bandwidths. Time-domain noise is generally broadband white noise, making it difficult to reproduce narrowband interference, harmonics, or instrument passband characteristics on-site. Compared to traditional two-dimensional image-based data enhancement, time-frequency mapping itself compresses temporal details; after geometric transformation, it is difficult to revert to the original physical meaning of the discharge. For example, image rotation is equivalent to redefining the phase or time axis, but the corresponding actual physical process does not exist, potentially leading to misleading results. In summary, frequency domain enhancement takes into account both physical realism and correspondence with the actual discharge mechanism, and is more suitable for partial discharge recognition scenarios compared with time domain or two-dimensional image enhancement.

[0094] Step 3. Design of a parallel multi-scale one-dimensional convolutional neural network framework. Figure 2 It is a partial discharge recognition model framework based on a parallel multi-scale one-dimensional convolutional neural network, which is used as a teacher model.

[0095] The teacher model includes three feature extraction branches, four residual blocks, a fusion module, and a Softmax layer.

[0096] Multi-scale feature extraction refers to extracting features from an input signal by using multiple parallel branches and convolutional kernels of different sizes, thereby capturing multiple scale features of the signal, such as high-frequency, low-frequency, and medium-frequency features, at the same level.

[0097] The three feature extraction branches are defined as the trunk branch, detail branch, and high-frequency branch, respectively, and are used to extract feature maps at different scales. The outputs of the three feature extraction branches are concatenated using Concat.

[0098] The stitched feature map is processed sequentially through each residual block. After each residual block, an SE channel attention mechanism is combined, and the feature processed by the last SE channel attention mechanism is input into the fusion module.

[0099] The fusion module uses one-dimensional global max pooling and one-dimensional global average pooling for parallel processing, and then merges the results of global max pooling and global average pooling. Finally, it outputs the predicted probabilities of each category through a Softmax layer.

[0100] The three branches use convolutional kernels of different scales, such as 16-point, 8-point, and 4-point, to extract feature representations of different frequency components, and integrate multi-scale information through a feature fusion mechanism, thereby improving the model's ability to represent complex partial discharge patterns.

[0101] Based on the observed partial discharge data, such as Figure 4 PRPS diagram of mid-tip discharge Figure 5 PRPS diagram of medium particle discharge. Figure 6 PRPS diagram of middle air gap discharge. Figure 7 PRPS diagram of mid-surface discharge and Figure 8 Partial discharge PRPS plot. PRPS is a three-dimensional analysis spectrum used in partial discharge monitoring. It uses the phase angle of the alternating current cycle (0° to 360°) as the horizontal axis (X-axis), the discharge cycle as the vertical axis (Y-axis), and the discharge intensity as the depth axis (Z-axis).

[0102] in Figure 4 In the diagram, (a) is the PRPS diagram for tip discharge type A, (b) is the PRPS diagram for tip discharge type B, (c) is the PRPS diagram for tip discharge type C, and (d) is the PRPS diagram for tip discharge type D.

[0103] in Figure 5 In the diagram, (a) is the PRPS diagram for particle discharge type A, (b) is the PRPS diagram for particle discharge type B, (c) is the PRPS diagram for particle discharge type C, and (d) is the PRPS diagram for particle discharge type D.

[0104] in Figure 6 In the diagram, (a) is the PRPS diagram for air gap discharge type A, (b) is the PRPS diagram for air gap discharge type B, (c) is the PRPS diagram for air gap discharge type C, and (d) is the PRPS diagram for air gap discharge type D.

[0105] in Figure 7 In the diagram, (a) is the PRPS diagram for surface discharge type A, (b) is the PRPS diagram for surface discharge type B, (c) is the PRPS diagram for surface discharge type C, and (d) is the PRPS diagram for surface discharge type D.

[0106] in Figure 8 In the diagram, (a) is the PRPS diagram for suspended discharge type D, (b) is the PRPS diagram for suspended discharge type B, (c) is the PRPS diagram for suspended discharge type C, and (d) is the PRPS diagram for suspended discharge type A.

[0107] Among the above discharge types A, B, C, and D, they represent several different phenomena of the same discharge condition.

[0108] like Figure 4 In the case of tip discharge type B, the main discharge is concentrated between phases of 60° to 120° and 240° to 300°, which shows obvious periodicity. This type is best extracted using mid-frequency scale convolution kernels, which can completely extract features. However, the discharge phenomenon does not always occur in both intervals. For example, types A, C, and D may be concentrated in a certain main interval, while sporadic discharges occasionally occur in another interval. In this case, high-frequency convolution kernels are needed to distinguish it well from other discharge types.

[0109] like Figure 5 Medium-sized particle discharge types A and B both exhibit obvious periods, but the periods are not very long, requiring mid-frequency convolutional kernels to extract features. Type C exhibits no obvious period but has a high discharge frequency, so low-frequency convolutional kernels should be used to extract features. Low frequencies can increase the receptive field and extract more comprehensive global features. Type D discharges sparsely and requires high-frequency convolutional kernels to capture burst pulses.

[0110] like Figure 6In the middle air gap discharge, the air gap discharge pulse is short and has a limited duration. The period of type D is relatively obvious, while the period of other types is not obvious. For this type of rapid discharge, high-frequency convolution kernels are needed to extract features better. However, although type C is short, it has many discharges. This type of discharge is suitable for low-frequency feature extraction.

[0111] like Figure 7 There is a type of levitation discharge that has obvious periodicity and more dense discharge pulses. This is suitable for high-frequency convolutional kernels for feature extraction. However, its period interval is short, so mid-frequency convolutional kernels can also capture the features between periods very well.

[0112] like Figure 8 Surface discharge is similar to tip discharge, but there are significant differences in phase.

[0113] This invention designs a multi-scale parallel branch suitable for all types of discharge data analysis, corresponding to different convolution kernels: high frequency corresponds to a 4-channel convolution kernel, mid frequency corresponds to an 8-channel convolution kernel, and low frequency corresponds to a 16-channel convolution kernel. Each branch includes one-dimensional convolution, batch normalization, and activation function.

[0114] The three outputs are concatenated along the channel dimension to form a multi-scale fused feature. The formula is as follows:

[0115] ;

[0116] ;

[0117] ;

[0118] .

[0119] in They are 4, 8, and 16 respectively. Indicates the kernel size as One-dimensional convolution operation, For batch normalization, For activation function, For channel splicing.

[0120] To further enhance the feature representation capability and training stability of deep networks, this invention introduces a multi-layer residual connection structure in the backbone feature extraction module and integrates a channel attention mechanism (SE) within each residual block.

[0121] The mechanism first processes the original feature map of the input SE. Perform global average pooling to obtain channel descriptors.

[0122] Subsequently, two fully connected networks are used. The first fully connected network first reduces the dimensionality by 1 / 8, and then ReLU is applied. The second fully connected network restores the original number of channels and then Sigmoid activation is applied to generate channel weights.

[0123] Finally, the weights are multiplied channel-by-channel by the original feature map to achieve adaptive feature recalibration. The formula is:

[0124] ;

[0125] .

[0126] in For global average pooling, and For the weights of the fully connected layer, These are the activations of ReLU and Sigmoid, respectively. For channel-by-channel multiplication, Let F be the output feature map of SE, and Reshape(s) means reshaping the weight vector s into a shape that matches the feature map F so that channel-wise multiplication can be performed.

[0127] Through the aforementioned SE mechanism, the model can adaptively enhance its focus on discriminative feature channels, effectively improving its ability to identify complex partial discharge pulse signals. This structure is as follows: Figure 2 As shown in the middle section.

[0128] Following the core feature extraction module, the model employs a global feature fusion strategy to further enhance its feature representation capabilities.

[0129] Specifically, the main output undergoes global average pooling and global max pooling to extract the overall statistical features and extreme value features of the sequence. The results of the two pooling operations are concatenated along the feature dimension to form a globally fused feature, such as... Figure 2 On the right side.

[0130] like Figure 2 As shown, the overall processing flow of the teacher model in this embodiment is as follows:

[0131] Step 3.1. Perform multi-branch feature extraction on the augmented data.

[0132] The main branch, detail branch, and high-frequency branch correspond to different convolution kernels. The high-frequency branch corresponds to 4 kernels, the detail branch corresponds to 8 kernels, and the main branch corresponds to 16 kernels. Each branch contains one-dimensional convolution, batch normalization, and activation functions.

[0133] The main branch outputs a 64-channel feature map, the detail branch outputs a 32-channel feature map, and the high-frequency branch outputs a 32-channel feature map. Then, these three branches are fused to form a 128-channel feature fusion.

[0134] Step 3.2. The channels after multi-branch feature fusion enter the residual connection and attention mechanism module, and feature extraction is performed through incremental residual blocks. The number of convolutional channels in each residual block are 64, 128, 256 and 512, respectively.

[0135] Within each residual block, an SE channel attention mechanism is incorporated to recalibrate the channels. Based on global semantics, important channels are adaptively amplified while secondary channels are suppressed, making features more focused and reducing redundancy. This typically improves expressiveness and robustness.

[0136] Specifically, the SE channel attention mechanism is embedded inside the residual block, located after the convolutional layer and before the residual connection. In the main path of each residual block, the feature map is first weighted by SE, and then the weighted feature map is added to the original feature map to complete the residual connection.

[0137] Step 3.3. Calculate channel weights using global average pooling and global max pooling, and finally proceed to global feature fusion:

[0138] ;

[0139] ;

[0140] in The features are those processed by the attention mechanism in the last SE channel. For global average pooling, For global max pooling, The output features of the backbone network This is a feature splicing operation.

[0141] Step 3.4. The fused global features are processed by activation functions and regularization, and finally the predicted probabilities of each category are output through the Softmax layer. The category with the highest predicted probability is taken as the final prediction output of the teacher model.

[0142] The multi-branch feature extraction proposed in this invention covers key time scales of partial discharge pulse signals. Four-point convolutions focus on dense and rapid discharge signals, eight-point convolutions correspond to moderately sustained pulses, and sixteen-point convolutions provide a wider receptive field for perceiving longer periodic structures and pulse periodic patterns. Simultaneously, the multi-branch approach ensures that the network can extract features at different scales in parallel early on, providing rich underlying representations for the subsequent residual + SE module. After frequency domain augmentation, different frequency bands of the samples are amplified, and the multi-branch approach can match these augmentation patterns in the time domain, improving the model's adaptability to different operating conditions and noise backgrounds. Furthermore, the multi-branch approach of the teacher model provides rich "dark knowledge" for the student model to distill, thus maintaining inter-class distinctions even after compression. Moreover, each data point after frequency domain augmentation enters all three branches simultaneously, rather than entering a specific branch, because multi-scale components can appear under the same defect. If routing to a specific branch only by category, the model will miss the diversity within the same category.

[0143] Step 4. First, generate soft labels for the model trained in Step 2. Use the teacher model to train the student model. The student model structure is as follows: Figure 3 As shown, its structural composition and signal processing flow are as follows:

[0144] The student model consists of a feature extraction block, three residual blocks, a fusion module, and a softmax layer. The feature extraction block is a one-dimensional convolution with a kernel of 16, corresponding to a feature map with 64 output channels.

[0145] The residual blocks in the student model are extracted by the teacher model through progressive residual blocks of 64, 128, 256, and 512, combined with SE attention to extract deep features, compressing the residual channels into 64, 128, and 256 residual blocks, and eliminating the SE attention mechanism.

[0146] Specifically, the processing flow of the student model in this embodiment is as follows:

[0147] The augmented data is first processed using feature extraction blocks; the extracted feature maps are then processed through each residual block; the output feature map of the last residual block is then fed into the fusion module.

[0148] The fusion module uses one-dimensional global max pooling and one-dimensional global average pooling for parallel processing, and then merges the results of global max pooling and global average pooling. Finally, it outputs the predicted probabilities of each category through a Softmax layer.

[0149] The formula for the teacher's temperature soft label is: .

[0150] in The teacher model uses normalized scores, and T is the temperature coefficient.

[0151] The present invention uses T=2.5 to flatten the distribution of the teacher model, resulting in a soft target. The goal is to expose the similarities between each class, so that the student model can directly remember these similarities without needing to identify individual labels.

[0152] The output values ​​of the student model at different temperatures are also consistent with the formula of the output values ​​of the teacher model at different temperatures, with the aim of comparing the distributions of the student model and the teacher model on the same scale.

[0153] The temperature output formula for the student model is: .

[0154] Secondly, to make the student model more similar to the teacher model, and to minimize the difference between the temperature output values ​​of the teacher model and the student model, the soft-label loss function is defined as follows:

[0155] .

[0156] in Represents the soft-label loss function. This indicates a measure of the difference between two distributions. It is gradient scale calibration.

[0157] This will adjust the temperature to make the teacher and student models more similar.

[0158] After labeling with soft labels, hard labels are used to ensure that the student model has sufficient confidence in the correct category and does not deviate from the true label by simply imitating the teacher. This simplifies the training process and reduces the complexity of training the model.

[0159] Finally, the soft labels of the teacher model and the hard labels of the simplified model are used together to apply to the student model, and then the trained student model is used to identify the discharge type of the gas-insulated switchgear.

[0160] Specifically, the process of knowledge distillation is as follows:

[0161] Step 4.1. Use the online partial discharge recognition model trained in Step 3 as the teacher model, and then train the student model; first load the pre-trained teacher model and generate soft labels on the training set and validation set.

[0162] The training set and validation set are obtained by partitioning the augmented dataset obtained in step 2.

[0163] Step 4.2. Construct a lightweight student model. The residual structure used is changed from the original four residual blocks with 64, 128, 256 and 512 convolution channels to three residual blocks with 64, 128 and 256 convolution channels.

[0164] Step 4.3. Set the distillation loss function:

[0165] By combining the label cross-entropy loss and the soft label KL divergence loss, and setting the temperature parameter T=2.5 and the weight coefficient a=0.6, the optimal student model is selected through an early stopping mechanism to balance knowledge transfer and task learning.

[0166] The trained student model is converted into a format that can be deployed on mobile devices. The host computer program is then edited in QT software, including data saving and data plotting functions, which can better display the current discharge effect and type.

[0167] The plotting function includes PRPS and PRPD plots, and it can also be used to create an APK file in QT for download to mobile devices.

[0168] Perform parameter statistics and inference speed verification on the student model; realize data transmission through QT host computer and build a mobile APP; process the signal in real time and use lightweight model inference to output recognition results and confidence scores.

[0169] Furthermore, the process of conducting upper-level computer verification includes the following steps:

[0170] I. Perform parameter statistics and calculate the compression ratio of the student model, calculate whether the model's inference speed meets the sampling rate, and convert the trained student model into a mobile-compatible format;

[0171] II. Data is transmitted via a host computer (QT) and combined with the trained student model to form a mobile app;

[0172] III. Online identification and performance evaluation: Collect partial discharge signals in real time on mobile devices, execute the preprocessing process that runs continuously during the training phase, use a lightweight student model for fast inference, and output the identification results and confidence scores.

[0173] The online partial discharge identification method proposed in this invention provides physically meaningful sample augmentation for one-dimensional time-domain signals through frequency domain data augmentation, laying a data foundation for teacher model training. The one-dimensional direct processing architecture of the teacher model fully utilizes the time-domain characteristics of the augmented samples, avoiding information loss caused by two-dimensional transformation. Knowledge distillation compresses and transfers the discriminative knowledge of the teacher model on the one-dimensional time-domain signal to the student model, achieving a seamless transition from high precision to lightweight design. The core advantage of the entire technical solution lies in its consistent focus on processing one-dimensional time-domain signals, forming a complete and efficient technical closed loop from data augmentation to model training and lightweight deployment, providing an end-to-end solution for online partial discharge monitoring.

[0174] In addition, to verify the effectiveness of the method proposed in this invention, the following specific experiments are also provided:

[0175] Partial discharge data of gas-insulated switchgear at different times were used as generalized test data, including 417 data points of tip discharge, 284 data points of particulate discharge, 276 data points of air gap discharge, 276 data points of suspension discharge, and 250 data points of particulate discharge.

[0176] The accuracy of the teacher and student models was tested using these generalized data, as shown in Table 1.

[0177] Table 1. Accuracy Comparison of Teacher Model and Student Model

[0178]

[0179] As shown in Table 1, the method of this invention achieves excellent recognition performance in terms of accuracy. Both the teacher model and the student model exhibit high accuracy in the recognition of the five types of discharges. In particular, the student model even surpasses the teacher model in the recognition of tip discharges (98.1%) and particle discharges (96.0%) (95.4% and 92.6%, respectively). This fully verifies that knowledge distillation technology, while transferring knowledge from the teacher model, can further enhance the student model's ability to recognize specific categories.

[0180] Meanwhile, model inference speed and model size are also important in online deployment of partial discharge. Table 2 below compares the performance indicators such as inference speed of the teacher model and the student model.

[0181] Table 2 Comparison of performance metrics between the teacher model and the student model

[0182]

[0183] As shown in Table 2, with the accuracy rate remaining almost unchanged (96.87% for teachers and 96.60% for students, a decrease of only 0.27 percentage points), the student model achieved significant weight reduction and acceleration: the model size decreased from 9.6MB to 2.3MB (approximately 4.2 times smaller), the number of parameters decreased from 2,526,589 to 606,789 (approximately 4.2 times less), the single inference time decreased from 1.082s to 0.580s (approximately 1.87 times faster), and the overall efficiency score increased from 1 to 7.742.

[0184] The partial discharge identification method of this invention effectively alleviates the sample imbalance problem through frequency domain enhancement. Knowledge distillation achieves a 76% reduction in model volume and a 46% increase in inference speed while maintaining a high accuracy of 96.60%. It has low computational load, strong real-time performance, and high accuracy, providing a new solution for real-time mobile identification of partial discharge in gas-insulated switchgear.

[0185] Overall, the student model obtained through knowledge distillation in this embodiment can achieve significant improvements in size and model inference speed at a minimal cost to accuracy, making it more suitable for online and real-time deployment on the edge.

[0186] The online partial discharge identification method proposed in this invention, which combines frequency domain data augmentation, multi-branch 1D-CNN, and knowledge distillation, can achieve stable, accurate, and low-latency identification under complex working conditions and meet the deployment requirements of mobile terminals.

[0187] Example 2

[0188] This embodiment 2 describes an online partial discharge identification system based on knowledge distillation 1D-CNN, which is based on the same inventive concept as the online partial discharge identification method based on knowledge distillation 1D-CNN in embodiment 1 above.

[0189] The online partial discharge identification system based on knowledge distillation 1D-CNN in this embodiment includes the following modules:

[0190] The preprocessing module is used to acquire partial discharge pulse signals and perform labeling and classification preprocessing.

[0191] The data augmentation module is used to extract the amplitude spectrum and phase spectrum of the partial discharge pulse signal through Fourier transform. It uses four methods for data augmentation: amplitude modulation, phase modulation, frequency band modulation, and noise injection. The augmented dataset is obtained through inverse transform.

[0192] The teacher model building module is used to construct an online partial discharge identification model with a multi-branch structure, which serves as the teacher model.

[0193] The teacher model includes three feature extraction branches, four residual blocks, a fusion module, and a Softmax layer;

[0194] All three feature extraction branches are one-dimensional convolutional extraction branches, defined as the trunk branch, detail branch, and high-frequency branch, respectively, and are used to extract feature maps at different scales. The outputs of the three feature extraction branches are concatenated using Concat.

[0195] The stitched feature map is processed sequentially through each residual block. After each residual block, an SE channel attention mechanism is combined, and the feature processed by the last SE channel attention mechanism is input into the fusion module.

[0196] The fusion module uses one-dimensional global max pooling and one-dimensional global average pooling for parallel processing, and then merges the processing results of global max pooling and global average pooling. Finally, it outputs the predicted probabilities of each category through the Softmax layer.

[0197] The teacher model is trained using the expanded dataset to obtain a well-trained teacher model.

[0198] The prediction module is used to build a lightweight student model and generate soft labels from the trained teacher model. It presets temperature parameters and weight coefficients, combines hard and soft label losses to perform knowledge distillation, and obtains the optimal compressed model through an early stopping mechanism.

[0199] Based on the trained student model, online partial discharge identification is performed to obtain the specific discharge type.

[0200] It should be noted that any content not mentioned in the above-described functional modules of the system described in Embodiment 2 can be referred to the step description of the corresponding method in Embodiment 1 above, and will not be repeated in detail here.

[0201] Example 3

[0202] This embodiment 3 describes a computer device including a memory and one or more processors. Executable code is stored in the memory. When the processor executes the executable code, it implements the steps of the online partial discharge identification method based on knowledge distillation 1D-CNN described in embodiment 1 above.

[0203] The computer device in this embodiment 3, such as a mobile terminal, is designed to achieve lightweight deployment, requiring less computation and faster inference speed. It can well meet the real-time requirements of on-site detection and provide a portable, high-precision intelligent solution for power equipment condition monitoring.

[0204] In this embodiment, the computer device can be any device or apparatus with data processing capabilities, and will not be described in detail here.

[0205] Example 4

[0206] This embodiment 4 describes a computer-readable storage medium storing a program that, when executed by a processor, implements the steps of the online partial discharge identification method based on knowledge distillation 1D-CNN in embodiment 1 above.

[0207] The computer-readable storage medium can be an internal storage unit of any device or apparatus with data processing capabilities, such as a hard disk or memory, or an external storage device of any device with data processing capabilities, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc.

[0208] Of course, the above description is only a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. It should be noted that any equivalent substitutions or obvious modifications made by those skilled in the art under the guidance of this specification fall within the scope of this specification and should be protected by the present invention.

Claims

1. An online partial discharge recognition method based on knowledge distillation 1D-CNN, characterized in that, Comprising the following steps: Step 1. Collect the original signal, i.e. partial discharge pulse signal, and perform labeling and classification preprocessing; Step 2. Extract the amplitude spectrum and phase spectrum of the original signal by Fourier transform, perform data enhancement in four ways of amplitude modulation, phase modulation, frequency band modulation and noise injection, and obtain the expanded data set by inverse Fourier transform; Step 3. Construct an online partial discharge identification model with a multi-branch structure as a teacher model; wherein the teacher model comprises three feature extraction branches, four residual blocks, a fusion module and a Softmax layer; The three feature extraction branches are one-dimensional convolution extraction branches, respectively defined as a main branch, a detail branch and a high-frequency branch, and are used to extract feature maps of different scales; the outputs of the three feature extraction branches are spliced by Concat; The spliced feature maps are processed in turn by each residual block, and an SE channel attention mechanism is combined after each residual block; The features processed by the last SE channel attention mechanism are input into the fusion module; The fusion module performs parallel processing by one-dimensional global maximum pooling and one-dimensional global average pooling, fuses the processing results of the global maximum pooling and the global average pooling, and finally outputs the prediction probability of each category by the Softmax layer; Train the teacher model based on the expanded data set to obtain the trained teacher model; Step 4. Construct a lightweight student model to generate soft labels from the trained teacher model; Pre-set temperature parameters and weight coefficients, combine hard and soft label losses for knowledge distillation, and obtain the trained student model through early stopping mechanism; Based on the trained student model, perform online partial discharge identification to obtain specific discharge types.

2. The online partial discharge recognition method based on knowledge distillation 1D-CNN according to claim 1, characterized in that, The step 1 is specifically: Step 1.

1. First, collect the original signal of the gas insulated switchgear, i.e. the partial discharge pulse signal; each type of partial discharge pulse signal has the same length, and is arranged in the form of 5000x1 to obtain one-dimensional partial discharge data; Each type of partial discharge pulse signal data includes 5000 sampling points; Step 1.

2. Label and classify the partial discharge pulse signal data by type, which includes five types of discharge, namely, sharp discharge, particle discharge, air gap discharge, suspension discharge and surface discharge. 3.The online partial discharge recognition method based on knowledge distillation 1D-CNN according to claim 1, wherein, The step 2 is specifically: Step 2.

1. First, perform Fourier transform on the original signal to extract the amplitude spectrum and the phase spectrum, as follows: ； wherein is a frequency domain representation of the signal, denotes a data point, i.e. a sample, in the original signal; The represents a combination of copy and phase, as follows: ； wherein and represent the amplitude and phase of the spectrum, respectively; Step 2.

2. Based on the frequency domain representation of the signal in step 2.1, design four data enhancement methods: I. Amplitude modulation is performed on the transformed signal; Pre-set the scaling factor range, and select different scaling factors from the scaling factor range for amplitude scaling of the amplitude spectrum; II. Perform phase modulation; Pre-set the radian range, and add phase offsets of different radian in the radian range to the phase spectrum; III. Perform frequency band enhancement: The multiple frequency domain data points obtained by Fourier transform of the original signal are evenly divided into multiple intervals according to the sampling time, and the frequency domain data points in each interval are individually enhanced by a preset multiple energy; IV. Perform frequency domain noise injection: In the amplitude spectrum, different intensity Gaussian noise is added, and in the phase spectrum, different intensity random phase disturbance is added; Step 2.

3. The enhanced data is subjected to the following inverse Fourier transform to obtain the expanded discharge data set; ； wherein represents any one of the data enhanced frequency spectrum, is the corresponding enhanced time domain signal.

4. The online partial discharge recognition method based on knowledge distillation 1D-CNN according to claim 1, characterized in that, In step 3, the processing flow of the teacher model is as follows: Step 3.

1. Multi-branch feature extraction is performed on the data enhanced data; The main branch, the detail branch and the high frequency branch correspond to different convolution kernels respectively, the high frequency corresponds to 4 channel convolution kernel, the detail corresponds to 8 channel convolution kernel, and the main branch corresponds to 16 channel convolution kernel; Each branch of the above three branches includes one-dimensional convolution, batch normalization and activation function; The main branch outputs a 64 channel feature map, the detail branch outputs a 32 channel feature map, and the high frequency branch outputs a 32 channel feature map, and then the three branches are fused to form a 128 channel feature fusion; Step 3.

2. The channels after multi-branch feature fusion enter the residual connection and attention mechanism module, and the feature extraction is performed through the residual blocks with increasing number of convolution channels, and the number of convolution channels of each residual block is 64, 128, 256 and 512 respectively; Each residual block combines SE channel attention mechanism to reweight the channels; SE channel attention mechanism is embedded in the residual block, located after the convolution layer and before the residual connection; in the main path of each residual block, the channel weighting is performed on the feature map by SE, and then the weighted feature map is added to the original feature map to complete the residual connection; Step 3.

3. The channel weight is calculated through global average pooling and global maximum pooling, and finally the global feature fusion is entered; ； ； wherein is the feature processed by the last SE channel attention mechanism, is the global average pooling, is the global max pooling, is the output feature of the backbone network, is the feature concatenation operation; Step 3.

4. The fused global feature is subjected to activation function and regularization processing, and finally the prediction probability of each class is output through the Softmax layer, and the class with the maximum prediction probability is taken as the final prediction output of the teacher model. 5.The online partial discharge recognition method based on knowledge distillation 1D-CNN according to claim 1, wherein, In step 4, the student model includes a feature extraction block, three residual blocks, a fusion module and a Softmax layer; The feature extraction block is one-dimensional convolution with convolution kernel of 16, corresponding to output channel number of 64 feature map; The convolution channel numbers of the three residual blocks are 64, 128 and 256 respectively; the processing flow of the student model is as follows: The data enhanced data is first subjected to feature extraction by the feature extraction block; the feature map after feature extraction is processed by the residual blocks in turn, and no SE channel attention mechanism is set after each residual block; The output feature map of the last residual block of the above three residual blocks enters the fusion module; The fusion module adopts one-dimensional global maximum pooling and one-dimensional global average pooling for parallel processing, and fuses the processing results of the global maximum pooling and the global average pooling, and finally outputs the prediction probability of each class through the Softmax layer.

6. The online partial discharge recognition method based on knowledge distillation 1D-CNN according to claim 1, characterized in that, In step 4, the process of knowledge distillation is as follows: Step 4.

1. The online partial discharge identification model trained in step 3 is taken as the teacher model, and then the student model is trained; first, load the pre-trained teacher model to generate soft labels on the training set and the validation set; Step 4.

2. Construct a lightweight student model, and the residual structure used is changed from the original four residual blocks with convolution channel numbers of 64, 128, 256, and 512 to three residual blocks with convolution channel numbers of 64, 128, and 256; Step 4.

3. Set the distillation loss function: Combine the label cross-entropy loss and the soft label KL divergence loss; set the temperature parameter T = 2.5 and the weight coefficient a = 0.6 to balance knowledge transfer and task learning, and select the optimal student model through the early stopping mechanism.

7. The online partial discharge recognition method based on knowledge distillation 1D-CNN according to claim 1, characterized in that, In step 3, the processing flow of the channel attention mechanism SE is as follows: First, the original feature map input to SE is globally averaged to obtain a channel descriptor; Then, two fully connected networks are used, where the first fully connected network is first reduced in dimension by a factor of 1 / 8, followed by ReLU, and the second fully connected network is increased back to the original channel number followed by Sigmoid activation to generate channel weights; Finally, the weights are multiplied with the original feature map channel by channel to achieve adaptive re-calibration of the features, with the formula being: ； ； where is the global average pooling, and is the fully connected layer weight, are the activations of ReLU and Sigmoid, respectively, is the channel-wise multiplication, F denotes the feature map, and Reshape(s) denotes reshaping the weight vector s to match the shape of the feature map F.

8. An online partial discharge recognition system based on knowledge distillation 1D-CNN, characterized in that, It includes the following modules: A preprocessing module for collecting partial discharge pulse signals and performing labeling and classification preprocessing; A data enhancement module for extracting the amplitude spectrum and phase spectrum of the partial discharge pulse signal through Fourier transform, using amplitude modulation, phase modulation, frequency band modulation, and noise injection for data enhancement, and obtaining the expanded data set through inverse transform; A teacher model construction module for constructing an online partial discharge recognition model with a multi-branch structure as a teacher model; The teacher model includes three feature extraction branches, four residual blocks, a fusion module, and a Softmax layer; The three feature extraction branches are one-dimensional convolution extraction branches, defined as the main branch, the detail branch, and the high-frequency branch, respectively, and are used to extract feature maps of different scales. The outputs of the three feature extraction branches are concatenated; The concatenated feature map is processed by each residual block in turn, and after each residual block, an SE channel attention mechanism is combined. The feature processed by the last SE channel attention mechanism is input to the fusion module; The fusion module uses one-dimensional global maximum pooling and one-dimensional global average pooling for parallel processing, and fuses the processing results of the global maximum pooling and the global average pooling. Finally, the Softmax layer outputs the prediction probability of each class; Train the teacher model based on the expanded data set to obtain the trained teacher model; And a prediction module for constructing a lightweight student model to generate soft labels from the trained teacher model; Pre-set temperature parameters and weight coefficients, combine hard and soft label losses for knowledge distillation, and obtain the optimal compression model through the early stopping mechanism; Based on the trained student model, perform online partial discharge recognition to obtain the specific discharge type.

9. A computer device comprising a memory and one or more processors; characterized in that, Executable code is stored in the memory, and when the processor executes the executable code, it is used to implement the steps of the online partial discharge recognition method based on knowledge distillation 1D-CNN of any one of claims 1-7.

10. A computer readable storage medium having stored thereon a program; characterized in that, The program, when executed by a processor, is configured to implement the steps of the knowledge distillation-based 1D-CNN online partial discharge identification method of any one of claims 1-7.

Images