A substation electrical equipment partial discharge detection and diagnosis method based on deep learning

Abstract

This invention discloses a deep learning-based method for detecting and diagnosing partial discharge in substation electrical equipment. It includes acquiring partial discharge signals from the electrical equipment and denoising the signals using a threshold optimization method based on wavelet packet decomposition and sparrow search algorithm, effectively suppressing periodic narrowband interference and Gaussian white noise in the partial discharge signals. Features are extracted from the denoised partial discharge signals through two-dimensional variational mode decomposition and Hilbert transform to generate grayscale texture feature vectors. These feature vectors are then input into a partial discharge signal pattern recognition model based on VMD-Hilbert marginal spectrum and sparse autoencoder-deep neural network to achieve intelligent identification of partial discharge types. This invention significantly improves the anti-interference capability and recognition accuracy of partial discharge detection in high-noise environments through the synergistic optimization of signal denoising, time-frequency feature extraction, and deep adaptive learning, providing a reliable technical means for insulation status assessment and fault early warning of substation electrical equipment.

Description

Technical Field

[0001] This invention relates to a partial discharge detection technology for substations, and more particularly to a method for detecting and diagnosing partial discharge in substation electrical equipment based on deep learning. Background Technology

[0002] Partial discharge is an early sign of insulation degradation and potential fault development in electrical equipment. Its condition monitoring and accurate identification are crucial for the safe operation of power systems. Partial discharge detection technologies mainly include electrical detection methods, ultrasonic detection methods, and optical detection methods. Electrical detection methods measure the pulse current or electromagnetic wave signals generated by partial discharge, but they are susceptible to electromagnetic interference and have difficulty accurately locating the discharge point. Ultrasonic detection methods capture the sound wave signals generated by partial discharge, offering high positioning accuracy, but are easily affected by mechanical vibration and noise under complex operating conditions. Optical detection methods collect the optical signals generated during partial discharge using a light guide rod and convert them into electrical signals for analysis. They offer significant advantages such as strong resistance to electromagnetic interference, high sensitivity, and non-contact measurement, and have gradually become a research hotspot in the field of partial discharge detection in recent years.

[0003] Currently, due to the diverse sources of noise within equipment, including corona discharge, mechanical vibration, and ambient light, these noises may overlap with partial discharge signals in both the time and frequency domains, making it difficult for traditional filtering methods to effectively separate partial discharge signals from noise. Furthermore, power systems have high real-time requirements for partial discharge detection, while existing algorithms have high computational complexity, making them unsuitable for online monitoring. Traditional signal processing methods (such as Fourier transform and wavelet transform) have limitations in feature extraction, failing to fully exploit the multi-scale features of partial discharge signals, resulting in insufficient detection accuracy. They also suffer from limited ability to suppress complex noise, feature extraction often relies on manual design and has a single dimension, and the recognition model is sensitive to noise and lacks adaptability and generalization capabilities. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to overcome the shortcomings of existing calibration technology and provide a deep learning-based method for partial discharge detection and diagnosis of substation electrical equipment that is reasonably designed, has a significant noise reduction effect, and high discharge identification accuracy.

[0005] The technical solution of this invention is:

[0006] A deep learning-based method for detecting and diagnosing partial discharge in substation electrical equipment includes the following steps:

[0007] S1. Deploy multiple types of sensors in key parts of electrical equipment and collect signals synchronously, while outputting the collected multi-channel raw signals.

[0008] S2. A denoising method based on wavelet packet decomposition and sparrow search algorithm to optimize threshold is used to preprocess the multi-channel original signal, adaptively suppressing periodic narrowband interference and Gaussian white noise, and retaining useful discharge components to the greatest extent. The denoised signal is then subjected to baseline correction and normalization to make the multi-source signals comparable, and the denoised preprocessed signal is output.

[0009] S3. Construct time-frequency features based on variational mode decomposition and Hilbert marginal spectrum, and PRPD grayscale image texture features based on two-dimensional variational mode decomposition and quaternion Hilbert transform in parallel for the preprocessed signal, and fuse the two to form a highly discriminative feature vector.

[0010] S4. Construct a sparse autoencoder deep neural network (SAE-DNN) model as the core classifier. This model first uses the sparse autoencoder (SAE) to perform unsupervised pre-training on high-dimensional fusion features, automatically learning the inherent sparse representation and robust features of the data. Then, a deep neural network (DNN) is initialized using the weights obtained from the SAE training, and a supervised fine-tuning is performed using an adaptive step size learning rate strategy to accelerate model convergence and optimize performance.

[0011] S5. Use a dataset of historically known partial discharge samples to train and validate the SAE-DNN model. Take the fusion feature vector generated in step S3 as input and the discharge type label as supervision signal. Iteratively optimize the model parameters by minimizing the cross-entropy loss function until the model performs stably on the validation set and obtain the trained intelligent diagnostic model.

[0012] S6. The real-time signals obtained from online monitoring of electrical equipment are processed in steps S2 and S3 and then input into the trained SAE-DNN model. The model automatically extracts deep features and outputs the identification results of partial discharge type.

[0013] Further: Step S2 specifically includes: first, inputting the original partial discharge signal of the noisy electrical equipment, initializing the sparrow search algorithm, setting core parameters such as population size and maximum number of iterations, and defining the range of wavelet packet parameters to be optimized: decomposition level and threshold coefficient; the position of each sparrow represents a specific set of wavelet packet parameter combinations. For each sparrow in the current population, the original signal is decomposed using the wavelet packet parameter combination it represents to obtain multiple sub-band coefficients. The signal-to-noise ratio of the denoised signal is used as the fitness function to evaluate the merits of each set of parameters. SSA achieves threshold optimization by updating the sparrow position.

[0014] Discoverer location update:

[0015]

[0016] Where t is the number of iterations, T is the maximum number of iterations, α∈(0,1] is the step size factor, and randn is a standard normally distributed random number. At the same time, the position is updated by random perturbation of standard normal distribution, which increases the randomness and diversity of the search and prevents getting stuck in local regions;

[0017] Follower location update:

[0018]

[0019] in, This is the current worst threshold. The current optimal threshold, To accommodate the perturbation factor that follows a normal distribution, a distance-weighted random perturbation is introduced, increasing the flexibility of the search.

[0020] Guardian location update:

[0021]

[0022] in, To be the globally optimal threshold, Let be the fitness value of the i-th sparrow. To avoid the minimum value where the denominator is 0.

[0023] Further, step S3 specifically includes: For the partial discharge signals of each channel preprocessed in step S2, variational mode decomposition (VMD) is used to adaptively decompose them into multiple finite-bandwidth intrinsic mode components with different center frequencies; Hilbert transform is performed on each decomposed mode component to construct its time-frequency distribution, and integration along the time axis generates a Hilbert marginal spectrum characterizing the signal energy distribution with frequency; statistical features are extracted from the Hilbert marginal spectrum to form an eigenvector F1 describing the time-frequency characteristics of the signal; in parallel, a phase-resolved partial discharge signal is generated using the pulse current sensor signal. The grayscale image of the discharge PRPD is obtained. A two-dimensional variational mode decomposition algorithm is used to decompose the PRPD grayscale image to obtain a series of two-dimensional mode components with different center frequencies. Quaternion Hilbert transform is performed on each two-dimensional mode component to extract multi-dimensional image texture features such as instantaneous amplitude, phase and frequency. The texture features of all mode components are integrated to form a feature vector F2 describing the discharge image mode. The time-frequency feature vector F1 and the image texture feature vector F2 are concatenated and subjected to principal component analysis for dimensionality reduction to form a unified and highly discriminative fusion feature vector, which is used as the input for subsequent pattern recognition.

[0024] Further, step S4 specifically includes: constructing a sparse autoencoder network structure, where the number of neurons in the input and output layers is equal to the dimension of the fused feature vector obtained in step S3, and the number of neurons in the hidden layers is less than that in the input layer; introducing a sparse penalty term and a weight normalization term based on KL divergence into the network cost function; performing unsupervised pre-training on the SAE using the unlabeled fused feature vector; minimizing the reconstruction error and sparsity constraints through forward and backward propagation algorithms; iteratively optimizing the network weight matrix and bias vector until the network can learn the high-dimensional sparse feature representation of the input data; and using the weight matrix and bias vector of the trained SAE hidden layers as the depth... The initial parameters of the hidden layers of the neural network (DNN) are initialized. A DNN classifier with multiple hidden layers is constructed, with the number of neurons in its output layer equal to the number of known partial discharge types. The Softmax function is used to output the class probabilities. Supervised fine-tuning of the initialized DNN is performed using training samples with labeled classes. The classification error is calculated using forward propagation, and an adaptive step-size learning rate strategy is used to dynamically adjust the parameter update step size during backpropagation, accelerating network convergence and optimizing model performance. The training process is repeated iteratively until the model's error converges on the validation set, resulting in the trained SAE-DNN classifier model.

[0025] Further, step S5 specifically includes: collecting known partial discharge signal samples of known types that have been manually labeled during historical operation, and extracting the fused feature vector of each sample using the method described in step S3, constructing a standard supervised training sample set in which the feature vector and type label correspond one-to-one; dividing the sample set into training set, validation set and test set according to the sample source and category distribution, and ensuring that there is no overlap between samples in each set for model training, tuning and performance evaluation; setting the hyperparameters required for model training; initializing the constructed SAE-DNN model, using the weight matrix and bias vector obtained from the unsupervised pre-trained sparse autoencoder SAE as the initial parameters of the corresponding hidden layer of the deep neural network DNN; reading fused feature vector samples from the training set in batches and inputting them into the initialized DNN for forward feeding. The propagation computation involves multiple hidden layer nonlinear transformations, followed by the Softmax function at the output layer to obtain the predicted class probability distribution. The cross-entropy loss between the predicted probability distribution and the actual class label is then used, and the backpropagation algorithm is executed based on this loss value to calculate the gradients of all weights and bias parameters in the network. According to the set optimizer and adaptive learning rate strategy, the calculated gradients are used to iteratively update the model parameters while monitoring the loss changes on the training set. After each training round, the validation set is used to evaluate the current model performance, recording the validation loss and classification accuracy, and saving a snapshot of the model parameters at the point of optimal validation performance. When the validation set performance no longer significantly improves in multiple consecutive iterations, or when the preset maximum number of iterations is reached, training is terminated, the final model structure and parameters are saved, and the trained intelligent diagnostic model is obtained.

[0026] Further, step S6 specifically includes: inputting the real-time fused feature vector into the final SAE-DNN model that has been trained and saved, performing forward propagation and nonlinear transformation through multiple hidden layers in sequence, and calculating the probability distribution corresponding to each known discharge type by the output layer Softmax function; selecting the category corresponding to the maximum value in the probability distribution as the model's preliminary recognition result, and using the maximum probability value as the confidence reference for this recognition; outputting the structured recognition result to complete the intelligent type identification of the real-time partial discharge signal.

[0027] The beneficial effects of this invention are:

[0028] 1. This invention effectively improves the denoising effect and pattern recognition accuracy of partial discharge signals in strong noise environments by integrating multi-scale decomposition, time-frequency feature extraction and deep adaptive learning, providing a reliable means for intelligent assessment and fault early warning of the insulation status of electrical equipment in substations.

[0029] 2. This invention achieves adaptive separation and suppression of periodic narrowband interference and Gaussian white noise at the partial discharge signal preprocessing level. While preserving useful discharge components to the maximum extent, it significantly improves the signal-to-noise ratio and waveform fidelity, providing a high-quality, high-fidelity signal foundation for subsequent analysis.

[0030] 3. In the feature extraction stage, this invention innovatively integrates dual feature extraction paths in the time-frequency domain and the image domain. Through variational mode decomposition and Hilbert marginal spectrum construction, the time-frequency energy distribution of the discharge signal is accurately characterized; through two-dimensional variational mode decomposition and quaternion Hilbert transform, the deep texture patterns of the PRPD grayscale image are deeply explored. This multi-dimensional and complementary fused feature vector greatly enhances the ability to distinguish and represent different discharge modes.

[0031] 4. In the intelligent diagnosis of partial discharge, this invention constructs a sparse autoencoder deep neural network model. It utilizes unsupervised pre-training to automatically learn the inherent sparse structure and robust expression of high-dimensional features, avoiding the subjectivity and limitations of traditional manual feature design. Combined with deep network fine-tuning using an adaptive learning rate strategy, the model possesses powerful complex pattern learning capabilities and generalization performance, thereby achieving high-precision and high-reliability identification of various known partial discharge types.

[0032] 5. This invention integrates advanced signal adaptive noise reduction, multi-dimensional feature fusion, and deep feature self-learning mechanisms to construct a complete technical solution from high-quality signal preprocessing and robust feature extraction to intelligent pattern recognition. This significantly improves the accuracy and reliability of partial discharge type identification in complex noise environments and provides an effective means for online assessment and intelligent operation and maintenance of the insulation status of substation electrical equipment.

[0033] 6. This invention systematically optimizes the entire process of partial discharge signals from substation electrical equipment, from preprocessing and feature construction to intelligent identification, significantly improving the automation, intelligence, and accuracy of the diagnostic process. It effectively overcomes the technical bottlenecks of traditional methods, such as signal extraction distortion under strong noise, limited ability of manual feature representation, and insufficient model generalization. It has high practical value and application prospects, is easy to promote and implement, and has good economic benefits. Attached Figure Description

[0034] Figure 1 This is an overall flowchart of a deep learning-based method for detecting and diagnosing partial discharge in substation electrical equipment.

[0035] Figure 2 This is a schematic diagram of the signal denoising principle based on wavelet packet decomposition and sparrow search algorithm (SSA).

[0036] Figure 3This is a flowchart of the recognition process based on Hilbert marginal spectrum and SAE-DNN. Detailed Implementation

[0037] Example: See Figure 1 -- Figure 3 The figure shows a method for detecting and diagnosing partial discharge in substation electrical equipment based on deep learning, which includes the following steps:

[0038] S1. Deploy multiple sensors in key parts of the electrical equipment body and collect partial discharge signals simultaneously;

[0039] S2. A denoising method based on wavelet packet decomposition and sparrow search algorithm to optimize threshold is used to preprocess the original multi-channel signals, adaptively suppressing periodic narrowband interference and Gaussian white noise, and retaining useful discharge components to the greatest extent. The denoised signals are then subjected to baseline correction and normalization to make the multi-source signals comparable.

[0040] S3. Construct time-frequency features based on variational mode decomposition and Hilbert marginal spectrum, and PRPD grayscale image texture features based on two-dimensional variational mode decomposition and quaternion Hilbert transform in parallel for the denoised signal, and fuse the two to form a highly discriminative feature vector.

[0041] S4. Construct a Sparse Autoencoder Deep Neural Network (SAE-DNN) model as the core classifier. This model first uses a sparse autoencoder (SAE) to perform unsupervised pre-training on high-dimensional fusion features, automatically learning the inherent sparse representation and robust features of the data. Subsequently, a deep neural network (DNN) is initialized using the weights obtained from the SAE training, and a supervised fine-tuning is performed using an adaptive step size learning rate strategy to accelerate model convergence and optimize performance.

[0042] S5. Train and validate the SAE-DNN model using a dataset of historically known partial discharge samples. Using the fused feature vector generated in step S3 as input and the discharge type label as the supervision signal, iteratively optimize the model parameters by minimizing the cross-entropy loss function until the model performs stably on the validation set, thus obtaining a well-trained intelligent diagnostic model.

[0043] S6. The real-time signals obtained from online monitoring of electrical equipment are processed in steps S2 and S3 and then input into the trained SAE-DNN model. The model automatically extracts deep features and outputs the identification results of partial discharge type.

[0044] Step S1 specifically includes: deploying sensors in key parts of the electrical equipment body, and using a multi-channel synchronous data acquisition device to synchronously acquire and record the output signals of all the above sensors, thereby continuously acquiring the original signals of each channel under the normal operation of the substation electrical equipment to form a multi-source synchronous partial discharge raw dataset.

[0045] like Figure 2 As shown, step S2 specifically includes: First, inputting the original partial discharge signal of the noisy electrical equipment. Initializing the sparrow search algorithm, setting core parameters such as population size and maximum number of iterations. Simultaneously, defining the range of wavelet packet parameters to be optimized: decomposition level and threshold coefficients. The position of each sparrow represents a specific set of wavelet packet parameter combinations. For each sparrow in the current population, the original signal is decomposed using its representative wavelet packet parameter combination to obtain multiple sub-band coefficients. The signal-to-noise ratio of the denoised signal is used as the fitness function to evaluate the quality of each set of parameters. SSA achieves threshold optimization by updating the "sparrow" position.

[0046] Discoverer location update (responsible for exploring the optimal threshold region):

[0047]

[0048] Where t is the number of iterations, T is the maximum number of iterations, α∈(0,1] is the step size factor, and randn is a standard normally distributed random number. At the same time, the position is updated by random perturbation of standard normal distribution, which increases the randomness and diversity of the search and prevents getting stuck in local regions;

[0049] Follower location update (follow the discoverer search):

[0050]

[0051] in, This is the current worst threshold. The current optimal threshold, To accommodate the perturbation factor that follows a normal distribution, a distance-weighted random perturbation is introduced, increasing the flexibility of the search.

[0052] Guardian position update (avoiding local optima):

[0053]

[0054] in, To be the globally optimal threshold, Let be the fitness value of the i-th sparrow. To avoid the minimum value where the denominator is 0.

[0055] The sparrow search algorithm iteratively optimizes the position of the sparrow population, i.e., the wavelet packet parameter combination, based on fitness values ​​through a collaborative update mechanism of producers, followers, and scouts. This process is repeated until the maximum number of iterations is reached, ultimately outputting the optimal wavelet packet parameters corresponding to the sparrow with the highest fitness. The original signal is decomposed using the optimal wavelet packet parameters, and thresholding is applied to the coefficients of each frequency band. The correlation coefficient between each thresholded frequency band and the original signal is calculated. If the correlation coefficient of a frequency band is higher than a preset threshold, it is marked as a "high-quality frequency band" and directly retained; for "low-quality frequency bands" with sub-correlation coefficients below the threshold, a second wavelet packet decomposition is performed, and effective sub-components are selected based on energy proportion. Finally, all high-quality frequency band coefficients and the effective sub-components extracted from the low-quality frequency bands are integrated, and a pure partial discharge signal is reconstructed through inverse wavelet packet transform.

[0056] Step S3 specifically includes: For the partial discharge signals of each channel preprocessed in step S2, variational mode decomposition (VMD) is used to adaptively decompose them into multiple finite-bandwidth intrinsic mode components with different center frequencies; Hilbert transform is performed on each decomposed mode component to construct its time-frequency distribution, and integration along the time axis generates a Hilbert marginal spectrum characterizing the signal energy distribution with frequency; statistical features are extracted from the Hilbert marginal spectrum to form an eigenvector F1 describing the time-frequency characteristics of the signal; and in parallel, phase-resolved partial discharge (PRPD) is generated using the pulse current sensor signal. The grayscale image is decomposed using a two-dimensional variational mode decomposition (2D-VMD) algorithm to obtain a series of two-dimensional mode components with different center frequencies. Quaternion Hilbert transform is performed on each two-dimensional mode component to extract multi-dimensional image texture features such as instantaneous amplitude, phase, and frequency. The texture features of all mode components are integrated to form a feature vector F2 describing the discharge image pattern. The time-frequency feature vector F1 and the image texture feature vector F2 are concatenated and subjected to principal component analysis (PCA) for dimensionality reduction to form a unified, highly discriminative fusion feature vector, which serves as the input for subsequent pattern recognition.

[0057] Step S4 specifically includes: constructing a sparse autoencoder (SAE) network structure, where the number of neurons in the input and output layers is equal to the dimension of the fused feature vector obtained in step S3, and the number of neurons in the hidden layers is less than that in the input layer; introducing a sparse penalty term and a weight normalization term based on KL divergence into the network cost function; performing unsupervised pre-training on the SAE using the unlabeled fused feature vector; minimizing the reconstruction error and sparsity constraints through forward and backward propagation algorithms; iteratively optimizing the network weight matrix and bias vector until the network can learn the high-dimensional sparse feature representation of the input data; and using the trained SAE hidden layer weight matrix and bias vector as the deep neural network. The DNN network parameters are initialized by setting the initial parameters of the corresponding hidden layers. A DNN classifier with multiple hidden layers is constructed, with the number of neurons in its output layer equal to the number of known partial discharge types, and the class probability is output using the Softmax function. The initialized DNN is then fine-tuned using training samples with labeled classes. The classification error is calculated using forward propagation, and the parameter update step size during backpropagation is dynamically adjusted using an adaptive step size learning rate strategy to accelerate network convergence and optimize model performance. The training process is repeated iteratively until the model's error converges on the validation set, resulting in the trained SAE-DNN classifier model.

[0058] like Figure 3As shown, step S5 specifically includes: collecting known partial discharge signal samples of known types that have been manually labeled during historical operations, and extracting the fusion feature vector of each sample using the method described in step S3, constructing a standard supervised training sample set in which the feature vector and type label correspond one-to-one; dividing the sample set into training set, validation set and test set according to the sample source and category distribution, and ensuring that the samples in each set do not overlap, for model training, tuning and performance evaluation; setting the hyperparameters required for model training, including optimizer type, cross-entropy loss function, batch size, maximum number of iterations, and specifically configuring an adaptive step size learning rate strategy to dynamically adjust the learning rate to accelerate convergence; initializing the constructed SAE-DNN model, using the weight matrix and bias vector obtained from the unsupervised pre-trained sparse autoencoder (SAE) as the initial parameters of the corresponding hidden layer of the deep neural network (DNN); from the training set The system reads fused feature vector samples in batches and inputs them into an initialized DNN for forward propagation computation. After nonlinear transformations through multiple hidden layers, the predicted class probability distribution is obtained at the output layer using the Softmax function. The system calculates the cross-entropy loss between the predicted probability distribution and the actual class label, and then performs a backpropagation algorithm based on this loss value to calculate the gradients of all weights and bias parameters in the network. According to the set optimizer and adaptive learning rate strategy, the calculated gradients are used to iteratively update the model parameters while monitoring the loss changes on the training set. After each training round, the model performance is evaluated using a validation set, and the validation loss and classification accuracy are recorded. A snapshot of the model parameters when the validation performance is optimal is saved. When the validation set performance no longer improves significantly in multiple consecutive iterations, or when the preset maximum number of iterations is reached, training is terminated, the final model structure and parameters are saved, and the trained intelligent diagnostic model is obtained.

[0059] Step S6 specifically includes: inputting the real-time fused feature vector into the final SAE-DNN model that has been trained and saved, performing forward propagation and nonlinear transformation through multiple hidden layers in sequence, and calculating the probability distribution corresponding to each known discharge type by the output layer Softmax function; selecting the category corresponding to the maximum value in the probability distribution as the model's preliminary recognition result, and using the maximum probability value as the confidence reference for this recognition; outputting the structured recognition result to complete the intelligent type identification of the real-time partial discharge signal.

[0060] This invention integrates advanced signal adaptive noise reduction, multi-dimensional feature fusion, and deep feature self-learning mechanisms to construct a complete technical solution from high-quality signal preprocessing and robust feature extraction to intelligent pattern recognition. It significantly improves the accuracy and reliability of partial discharge type identification in complex noise environments and provides an effective means for online assessment and intelligent operation and maintenance of the insulation status of substation electrical equipment.

[0061] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications made based on the technical essence of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A method for detecting and diagnosing partial discharge in substation electrical equipment based on deep learning, comprising the following steps: S1. Deploy multiple types of sensors in key parts of electrical equipment and collect signals synchronously, while outputting the collected multi-channel raw signals. S2. A denoising method based on wavelet packet decomposition and sparrow search algorithm to optimize threshold is used to preprocess the multi-channel original signal, adaptively suppressing periodic narrowband interference and Gaussian white noise, and retaining useful discharge components to the greatest extent. The denoised signal is then subjected to baseline correction and normalization to make the multi-source signals comparable, and the denoised preprocessed signal is output. S3. Construct time-frequency features based on variational mode decomposition and Hilbert marginal spectrum, and PRPD grayscale image texture features based on two-dimensional variational mode decomposition and quaternion Hilbert transform in parallel for the preprocessed signal, and fuse the two to form a highly discriminative feature vector. S4. Construct a sparse autoencoder deep neural network (SAE-DNN) model as the core classifier. This model first uses the sparse autoencoder (SAE) to perform unsupervised pre-training on high-dimensional fusion features, automatically learning the inherent sparse representation and robust features of the data. Then, a deep neural network (DNN) is initialized using the weights obtained from the SAE training, and a supervised fine-tuning is performed using an adaptive step size learning rate strategy to accelerate model convergence and optimize performance. S5. Use a dataset of historically known partial discharge samples to train and validate the SAE-DNN model. Take the fusion feature vector generated in step S3 as input and the discharge type label as supervision signal. Iteratively optimize the model parameters by minimizing the cross-entropy loss function until the model performs stably on the validation set and obtain the trained intelligent diagnostic model. S6. The real-time signals obtained from online monitoring of electrical equipment are processed in steps S2 and S3 and then input into the trained SAE-DNN model. The model automatically extracts deep features and outputs the identification results of partial discharge type.

2. The substation electrical equipment partial discharge detection and diagnosis method based on deep learning according to claim 1, characterized in that: Step S2 specifically includes: First, inputting the original partial discharge signal of the noisy electrical equipment, initializing the sparrow search algorithm, setting core parameters such as population size and maximum number of iterations, and defining the range of wavelet packet parameters to be optimized: decomposition level and threshold coefficient; the position of each sparrow represents a specific set of wavelet packet parameter combinations. For each sparrow in the current population, the original signal is decomposed using the wavelet packet parameter combination it represents to obtain multiple sub-band coefficients. The signal-to-noise ratio of the denoised signal is used as the fitness function to evaluate the merits of each set of parameters. SSA achieves threshold optimization by updating the sparrow position. Discoverer location update: ； where t is the iteration number, T is the maximum iteration number, a e (0, 1] is the step factor, and randn is a standard normal distribution random number. When the position is updated by a standard normal distribution random disturbance, the randomness and diversity of the search are increased, and falling into a local area is prevented. Follower location update: ； wherein, is the current worst threshold, is the current best threshold, is a perturbation factor obeying normal distribution, by introducing distance-weighted random perturbation, the flexibility of search is increased; Guardian location update: ； wherein, is a global optimal threshold value, is a fitness value of the i-th sparrow, is a minimum value to avoid a denominator of 0.

3. The substation electrical equipment partial discharge detection and diagnosis method based on deep learning according to claim 1, characterized in that: Step S3 specifically includes: For the partial discharge signals of each channel preprocessed in step S2, variational mode decomposition (VMD) is used to adaptively decompose them into multiple finite-bandwidth intrinsic mode components with different center frequencies; Hilbert transform is performed on each decomposed mode component to construct its time-frequency distribution, and integration along the time axis generates a Hilbert marginal spectrum characterizing the signal energy distribution with frequency; statistical features are extracted from the Hilbert marginal spectrum to form an eigenvector F1 describing the time-frequency characteristics of the signal; in parallel, a phase-resolved partial discharge signal is generated using the pulse current sensor signal. The PRPD grayscale image is obtained by using a two-dimensional variational mode decomposition algorithm to decompose the PRPD grayscale image and obtain a series of two-dimensional mode components with different center frequencies. The quaternion Hilbert transform is performed on each two-dimensional mode component to extract multi-dimensional image texture features such as instantaneous amplitude, phase and frequency. The texture features of all mode components are integrated to form a feature vector F2 describing the discharge image mode. The time-frequency feature vector F1 and the image texture feature vector F2 are concatenated and subjected to principal component analysis for dimensionality reduction to form a unified and highly discriminative fusion feature vector, which is used as the input for subsequent pattern recognition.

4. The substation electrical equipment partial discharge detection and diagnosis method based on deep learning according to claim 1, characterized in that: Step S4 specifically includes: constructing a sparse autoencoder network structure, where the number of neurons in the input and output layers is equal to the dimension of the fused feature vector obtained in step S3, and the number of neurons in the hidden layers is less than that in the input layer; introducing a sparse penalty term and a weight normalization term based on KL divergence into the network cost function; performing unsupervised pre-training on the SAE using the unlabeled fused feature vector; minimizing the reconstruction error and sparsity constraints through forward and backward propagation algorithms; iteratively optimizing the network weight matrix and bias vector until the network can learn the high-dimensional sparse feature representation of the input data; and using the weight matrix and bias vector of the trained SAE hidden layers as a deep neural network. The initial parameters of the hidden layers of the DNN network are initialized to complete the parameter initialization of the DNN network. A DNN classifier with multiple hidden layers is constructed, the number of neurons in its output layer is equal to the number of known partial discharge types, and the class probability is output using the Softmax function. The initialized DNN is then fine-tuned using training samples with labeled classes. The classification error is calculated by forward propagation, and the parameter update step size during backpropagation is dynamically adjusted using an adaptive step size learning rate strategy to accelerate network convergence and optimize model performance. The training process is repeated iteratively until the model's error converges on the validation set, resulting in the trained SAE-DNN classifier model.

5. The substation electrical equipment partial discharge detection and diagnosis method based on deep learning according to claim 1, characterized in that: Step S5 specifically includes: collecting known partial discharge signal samples of known types that have been manually labeled during historical operation, and extracting the fused feature vectors of each sample using the method described in step S3, constructing a standard supervised training sample set in which the feature vectors correspond one-to-one with the type labels; dividing the sample set into training set, validation set and test set according to the sample source and category distribution, ensuring that there is no overlap between samples in each set, for model training, tuning and performance evaluation; setting the hyperparameters required for model training; initializing the constructed SAE-DNN model, using the weight matrix and bias vector obtained from the unsupervised pre-trained sparse autoencoder SAE as the initial parameters of the corresponding hidden layer of the deep neural network DNN; reading fused feature vector samples from the training set in batches and inputting them into the initialized DNN for forward propagation. The calculation process involves calculating the predicted class probability distribution through a Softmax function at the output layer after multiple nonlinear transformations in the hidden layers. The cross-entropy loss between the predicted probability distribution and the actual class label is then used, and a backpropagation algorithm is executed based on this loss value to calculate the gradients of all weights and bias parameters in the network. According to the set optimizer and adaptive learning rate strategy, the calculated gradients are used to iteratively update the model parameters while monitoring the loss changes on the training set. After each training round, the validation set is used to evaluate the current model performance, recording the validation loss and classification accuracy, and saving a snapshot of the model parameters at the point of optimal validation performance. When the validation set performance no longer significantly improves in multiple consecutive iterations, or when the preset maximum number of iterations is reached, training is terminated, the final model structure and parameters are saved, and the trained intelligent diagnostic model is obtained.

6. The substation electrical equipment partial discharge detection and diagnosis method based on deep learning according to claim 1, characterized in that: Step S6 specifically includes: inputting the real-time fused feature vector into the final SAE-DNN model that has been trained and saved, performing forward propagation and nonlinear transformation through multiple hidden layers in sequence, and calculating the probability distribution corresponding to each known discharge type by the output layer Softmax function; selecting the category corresponding to the maximum value in the probability distribution as the model's preliminary recognition result, and using the maximum probability value as the confidence reference for this recognition; outputting the structured recognition result to complete the intelligent type identification of the real-time partial discharge signal.

Images