Fault recognition method based on adaptive fusion of multi-scale time-frequency features

Abstract

The application discloses a fault recognition method based on multi-scale time-frequency feature adaptive fusion, and relates to the technical field of power fault diagnosis. The method comprises the following steps: extracting signal state features for a to-be-recognized partial discharge signal, and performing multi-scale time-frequency transformation on the partial discharge signal based on a candidate scale library to generate multi-scale time-frequency features; then, the multi-scale time-frequency features are subjected to fault recognition through a trained multi-scale feature fusion diagnosis model, wherein a scale scoring module determines adaptive fusion weights corresponding to each scale; a multi-branch parallel convolution feature extraction module independently extracts features of each scale to maintain the original resolution; a cross-scale self-attention splicing module performs cross-scale global interaction to realize collaborative perception of local details and global information; and a feature pooling module outputs a fault recognition result from a classifier after adaptive fusion weight weighting fusion of classification hidden vectors of each scale, thereby effectively improving the distinguishing precision of the model for similar fault categories.

Description

Technical Field

[0001] This application relates to the field of power fault diagnosis technology, and in particular to a fault identification method based on adaptive fusion of multi-scale time-frequency features. Background Technology

[0002] In the field of power cable fault diagnosis, accurate identification of partial discharge signals is crucial for ensuring the safe operation of the power grid. Existing technologies typically employ data-driven deep learning methods to process partial discharge signals, but they still have significant limitations when dealing with partially discharge signals that exhibit highly non-stationary and transient characteristics. On the one hand, conventional methods often employ a single short-time Fourier transform with fixed parameters or a fixed multi-scale combination pre-set by humans. Due to the uncertainty principle of time-frequency analysis, a single or fixed analysis scale cannot simultaneously take into account high time resolution and high frequency resolution, making it difficult to dynamically adapt to different partial discharge types and noise environments, and easily leading to the loss of weak partial discharge details or scale redundancy.

[0003] On the other hand, in the multi-scale feature fusion stage, existing methods often force the time-frequency spectra of different scales to be scaled to a uniform size to meet the network input requirements by interpolation alignment. This processing method will change the spatial distribution of the original time-frequency spectra, causing local pulse detail distortion and high-frequency energy diffusion, which limits the full utilization of multi-scale complementary information, and thus affects the ability to distinguish similar fault categories and the accuracy of recognition.

[0004] In view of the above problems, this application proposes a new intelligent diagnostic method for partial discharge of cables, which aims to break through the bottleneck of single-scale representation, dynamically select the scale by signal state, and take into account the ability to fuse global feature sequences and local spatial details, so as to meet the needs of high-precision monitoring of distribution network faults. Summary of the Invention

[0005] The purpose of this application is to provide a fault identification method based on multi-scale time-frequency feature adaptive fusion, which aims to solve the problems of low identification accuracy caused by fixed time-frequency resolution, insufficient fusion of local and global features, and inability to adaptively select scale in the prior art.

[0006] To achieve the above objectives, this application provides the following solution: A fault identification method based on adaptive fusion of multi-scale time-frequency features includes the following steps: Acquire the partial discharge signal of the power cable to be identified.

[0007] The signal state features of the partial discharge signal are extracted, and multi-scale time-frequency transformation of the partial discharge signal is performed based on the candidate scale library to generate multi-scale time-frequency features.

[0008] A trained multi-scale feature fusion diagnostic model is used to identify faults in multi-scale time-frequency features, yielding the fault identification results for the power cable to be identified. The multi-scale feature fusion diagnostic model includes a scale scoring module, a multi-branch parallel convolutional feature extraction module, a cross-scale self-attention concatenation module, a feature pooling module, and a classifier. Within the multi-scale feature fusion diagnostic model: The adaptive fusion weights for each scale are determined based on the signal state characteristics through the scale scoring module.

[0009] The multi-branch parallel convolution feature extraction module performs independent feature extraction on the multi-scale time-frequency features at each scale to obtain the local features corresponding to each scale.

[0010] The local features at each scale are sequentially concatenated by a cross-scale self-attention concatenation module, and cross-scale global interaction is performed through a self-attention mechanism to obtain fused features.

[0011] The feature pooling module performs weighted fusion of the latent vectors of each scale in the fused features according to adaptive fusion weights.

[0012] The classifier outputs the fault identification result of the power cable to be identified based on the features obtained after weighted fusion.

[0013] Optionally, the multi-scale feature fusion diagnostic model is obtained through the following training steps: Obtain partial discharge signals of power cables with known historical fault types.

[0014] Signal state features are extracted from partial discharge signals of known historical fault types, and multi-scale time-frequency transformation is performed based on a candidate scale library to construct a multi-scale continuous partial discharge time-frequency feature dataset.

[0015] The multi-scale continuous partial discharge time-frequency feature dataset is divided into training dataset, validation dataset and test dataset.

[0016] The initial multi-scale feature fusion diagnostic model is trained using the training dataset. The model parameters are optimized based on the cross-entropy loss between the fault identification results output by the model and the known fault types until a well-trained multi-scale feature fusion diagnostic model is obtained. During the training process, a data augmentation strategy is introduced to mix the samples in the training dataset to generate augmented samples. The validation dataset is used to select the optimal model, and the test dataset is used to test and evaluate the performance of the model.

[0017] Optionally, the initial multi-scale feature fusion diagnostic model is trained using the training dataset. The model parameters are optimized based on the cross-entropy loss between the fault identification results output by the model and the known fault types until a well-trained multi-scale feature fusion diagnostic model is obtained. This specifically includes the following steps: Input the signal state features of any sample in the training dataset into the scale scoring module to obtain the adaptive fusion weights corresponding to each scale.

[0018] The multi-scale time-frequency features of the sample are split according to the scale dimension and input into the corresponding branches of the multi-branch parallel convolutional feature extraction module to extract local features at each scale.

[0019] Local features at each scale are input into a cross-scale self-attention stitching module for global interaction to obtain fused features.

[0020] The fused features are input into the feature pooling module, and the classification latent vectors of each scale in the fused features are weighted and fused according to the adaptive fusion weights corresponding to each scale to obtain the global feature vector.

[0021] The global feature vector is input into the classifier, and the cross-entropy loss is calculated based on the probability distribution of partial discharge fault categories output by the classifier and the known fault types of the samples. The model parameters are then updated using the backpropagation algorithm.

[0022] Optionally, the formula for calculating cross-entropy loss is: in, For cross-entropy loss, Indicates the number of partial discharge categories. Indicates the true label, This represents the predicted probability of the multi-scale feature fusion diagnostic model; the data augmentation strategy is the Mixup data augmentation strategy.

[0023] Optionally, the signal state features of the partial discharge signal are extracted, and multi-scale time-frequency transformation of the partial discharge signal is performed based on a candidate scale library to generate multi-scale time-frequency features, specifically including the following steps: A sliding window segmentation strategy is used to continuously divide the partial discharge signal into multiple short-time segments.

[0024] Statistical features reflecting the transient characteristics and spectral distribution characteristics of the signal are extracted from each short-time data segment and used as signal state features.

[0025] By utilizing multiple sets of differentiated window parameters from the candidate scale library, short-time Fourier transforms are performed in parallel on each short-time segment data to extract amplitude spectra reflecting different time and frequency resolution dimensions.

[0026] Logarithmically compress the amplitude spectra obtained from each group of differentiated window parameters, and then cascade and combine multiple adjacent logarithmized amplitude spectra on the same time series to form multi-scale time-frequency features.

[0027] Optionally, the candidate scale library contains at least two sets of differentiated window lengths, window overlap lengths, and fast Fourier transform points.

[0028] Signal state characteristics include normalized pulse width, peak factor, kurtosis, normalized spectral centroid, and local energy change rate.

[0029] The formula for standardizing signal state characteristics is: in, For the standardized first 3D eigenvalues For the first 3D eigenvalues For the training dataset, the first The mean of the dimensional features, For the training dataset, the first Standard deviation of dimensional features To prevent overflow of the minimum value when dividing by zero.

[0030] The formula for logarithmically compressing the amplitude spectra obtained from each group of differentiated window parameters is as follows: in, This is the logarithmically compressed amplitude spectrum. This is the result of the short-time Fourier transform. To prevent overflow of the minimum value when dividing by zero.

[0031] Alternatively, the formula for the short-time Fourier transform is: in, This is the result of the short-time Fourier transform. It is a discrete partial discharge signal. For window functions, For window length, For frame shift, For time frame indexing, For frequency index, For discrete-time indexing, The imaginary unit, Pi is the mathematical constant of a circle.

[0032] Optionally, after performing logarithmic compression on the amplitude spectra obtained from each group of differentiated window parameters, the following steps are also included: The continuous time-frequency spectrum sequences after logarithmic compression at each candidate scale are then standardized using the following formula: in, For the first The continuous time-frequency spectrum sequence after scale standardization. For the first A sequence of original continuous time-frequency spectra at each scale. For the first The mean amplitude of a local region in a short time segment of data at each scale. For the first The standard deviation of amplitude within a local region of short-time segment data at various scales. To prevent overflow of the minimum value when dividing by zero.

[0033] Optionally, multiple consecutive logarithmic amplitude spectra on the same time series are concatenated and combined to form multi-scale time-frequency features, including the following steps: For any short-time segment of data, a predetermined number of logarithmic amplitude spectra that are adjacent to each other on the time axis of the short-time segment are stacked in chronological order to form a continuous time-frequency spectrum sequence corresponding to the short-time segment data.

[0034] The continuous time-frequency spectrum sequence corresponding to each short time segment data is used as a multi-scale time-frequency feature; the continuous time-frequency spectrum sequence formed under different candidate scales has different time resolution and frequency resolution.

[0035] Optionally, local features at each scale are sequentially concatenated using a cross-scale self-attention concatenation module, and cross-scale global interaction is performed through a self-attention mechanism to obtain fused features, including the following steps: Local features at each scale are flattened in the spatial dimension and then spliced ​​together to obtain a long sequence across scales.

[0036] By using a self-attention mechanism to perform global interaction on long sequences across scales, fusion features are obtained; the fusion features contain fused features from multiple scales corresponding to each scale.

[0037] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides a fault identification method based on adaptive fusion of multi-scale time-frequency features. In this method, the partial discharge signal of the power cable to be identified is first acquired, and signal state features are extracted. Simultaneously, based on a candidate scale library containing multiple sets of differentiated window length parameters, the partial discharge signal undergoes multi-scale time-frequency transformation to generate multi-scale time-frequency features. The complementary multi-resolution representation is formed through multiple sets of differentiated window length parameters, thus overcoming the bottleneck of single-scale time-frequency analysis in achieving a balance between time and frequency resolution. The extracted signal state features and multi-scale time-frequency features are then input into a trained multi-scale feature fusion diagnostic model. This model includes a scale scoring module, a multi-branch parallel convolutional feature extraction module, a cross-scale self-attention concatenation module, a feature pooling module, and a classifier. The scale scoring module determines the adaptive fusion weights corresponding to each scale based on the signal state features, enabling the model to dynamically adjust the contribution ratio of different time-frequency resolutions according to the transient characteristics and spectral distribution characteristics of the partial discharge signal itself, avoiding the use of fixed scale groups. The multi-branch parallel convolutional feature extraction module extracts multi-scale time-frequency features independently by scale to maintain the original spatial resolution of each scale, obtaining local features corresponding to each scale. This avoids feature edge blurring and pulse detail distortion caused by forcibly interpolating and aligning different scale maps to a uniform size. The cross-scale self-attention stitching module stitches the local features of each scale sequentially and performs cross-scale global interaction through a self-attention mechanism to obtain fused features. This enables the model to achieve complementary fusion of cross-scale information and global perception while retaining local details of each scale. The feature pooling module performs weighted fusion of the classification latent vectors of each scale in the fused features according to adaptive fusion weights. Finally, the classifier outputs the fault identification result of the power cable to be identified based on the features obtained after weighted fusion. This achieves adaptive weighted fusion at the feature level, maximizing the retention of rich information within the features of each scale and effectively improving the model's ability to distinguish similar fault categories and overall identification accuracy. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a flowchart illustrating a fault identification method based on multi-scale time-frequency feature adaptive fusion, as provided in an embodiment of this application.

[0040] Figure 2This is a flowchart of step A2 in a fault identification method based on multi-scale time-frequency feature adaptive fusion, provided as an embodiment of this application.

[0041] Figure 3 This is a schematic diagram of the structure of a multi-scale feature fusion diagnostic model in a fault identification method based on multi-scale time-frequency feature adaptive fusion, provided in an embodiment of this application.

[0042] Figure 4 This is a flowchart illustrating the training of a multi-scale feature fusion diagnostic model in a fault identification method based on multi-scale time-frequency feature adaptive fusion, as provided in an embodiment of this application. Detailed Implementation

[0043] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0044] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0045] This application provides a fault identification method based on multi-scale time-frequency feature adaptive fusion. In one exemplary embodiment, such as... Figure 1 As shown, the method includes the following steps: A1. Obtain the partial discharge signal of the power cable to be identified.

[0046] Specifically, a partial discharge detection device can be used to couple and collect partial discharge signals from a running power cable to obtain the original long-time discrete signal to be identified. This signal contains waveform information of partial discharge activities that may occur in the cable body, intermediate joints, or cable terminals, such as internal discharge, surface discharge, or tip discharge.

[0047] A2. Extract the signal state features of the partial discharge signal, and perform multi-scale time-frequency transformation on the partial discharge signal based on the candidate scale library to generate multi-scale time-frequency features.

[0048] This step aims to generate multi-scale time-frequency representations and extract signal state features. In one embodiment, such as... Figure 2 As shown, step A2 specifically includes the following sub-steps: A21. A sliding window segmentation strategy is used to continuously divide the partial discharge signal to obtain multiple short-time segment data.

[0049] In this sub-step, the partial discharge signal is traversed and truncated using a preset window length and a preset sliding step size, maintaining a preset overlap length between adjacent windows to preserve the complete shape of the partial discharge pulse at the window edges. The preset window length is greater than the preset sliding step size, and the preset overlap length is equal to the preset window length minus the preset sliding step size. For example, using a window length of 4096 sampling points and a sliding step size of 1024 sampling points, the overlap length between adjacent windows is 3072 sampling points. This method converts the original long sequence signal into a large number of fixed-length short-time segments, providing sample support for subsequent multi-scale time-frequency transformation and deep network input.

[0050] A22. Extract statistical features reflecting the transient characteristics and spectral distribution characteristics of the signal from each short-time segment of data, and use them as signal state features.

[0051] In one embodiment, the signal state feature vector comprises five dimensions: normalized pulse width, peak factor, kurtosis, normalized spectral centroid, and local energy change rate. After extraction, the signal state features need to be standardized to eliminate the influence of dimensions. The standardization process can be performed using the following formula: in, For the standardized first 3D eigenvalues For the first 3D eigenvalues For the training dataset, the first The mean of the dimensional features, For the training dataset, the first Standard deviation of dimensional features To prevent the minimum value from overflowing when dividing by zero, in this embodiment... .

[0052] A23. Using multiple sets of differentiated window parameters in the candidate scale library, perform short-time Fourier transforms on each short-time segment data in parallel to extract the amplitude spectrum reflecting different time and frequency resolution dimensions.

[0053] First, a candidate scale library is established, which contains at least two sets of differentiated parameters. Each set of parameters includes the window length, window overlap length (or frame shift), and the number of Fast Fourier Transform points. In one embodiment, the candidate scale library contains at least three sets of differentiated window parameters.

[0054] For any short-time segment of data, its discrete signal is: Its short-time Fourier transform can be expressed as: in, This is the result of the short-time Fourier transform (i.e., the complex spectrum). It is a discrete partial discharge signal. For window functions (such as the Hanning window). For window length, This is the frame shift (i.e., the window overlap length). For time frame indexing, For frequency index, For discrete-time indexing, The imaginary unit, Pi is the mathematical constant of a circle.

[0055] By using multiple sets of differentiated window parameters from the candidate scale library, short-time Fourier transforms are performed in parallel on each short-time segment data to obtain the complex spectrum of each short-time segment data at each candidate scale. Then calculate its amplitude spectrum. .

[0056] A24. The amplitude spectra obtained from the differential window parameters of each group are logarithmically compressed, and multiple adjacent logarithmized amplitude spectra on the same time series are cascaded and combined to form multi-scale time-frequency features.

[0057] To suppress the dynamic range of amplitude and highlight subtle features, logarithmic compression is applied to each amplitude spectrum. The formula is as follows: in, This is the logarithmically compressed amplitude spectrum (instantaneous spectrum). This is the result of the short-time Fourier transform.

[0058] For any of the aforementioned short-time segment data, a predetermined number of logarithmically transformed amplitude spectra (e.g., 4 frames) adjacent to each other on the time axis of the short-time segment data are stacked in chronological order to form a continuous time-frequency spectrum sequence corresponding to the short-time segment data; the continuous time-frequency spectrum sequence corresponding to each of the aforementioned short-time segment data is used as the multi-scale time-frequency feature. This sequence can be regarded as a three-dimensional tensor (time frame dimension, frequency dimension, amplitude dimension), similar to a small video data containing multiple frames of images.

[0059] The continuous time-frequency spectrum sequences formed at different candidate scales have different time and frequency resolutions. Specifically, short window parameters correspond to high time resolution and low frequency resolution, while long window parameters correspond to low time resolution and high frequency resolution. Since different window length parameters correspond to different time-frequency resolutions, complementary multi-scale representations can be formed.

[0060] Before performing subsequent processing on the continuous time-frequency spectrum sequences after logarithmic compression at each candidate scale, the standardization process is also included. The formula for the standardization process is as follows: in, For the first The continuous time-frequency spectrum sequence after scale standardization. For the first A sequence of original continuous time-frequency spectra at each scale. For the first The mean amplitude of a local region in a short time segment of data at each scale. For the first The standard deviation of amplitude within a local region of short-time segment data at various scales. To prevent the minimum value from overflowing during division by zero, the above processing is used to finally construct a multi-scale continuous partial discharge time-frequency spectral feature suitable for deep network input.

[0061] A3. The trained multi-scale feature fusion diagnostic model is used to identify faults in multi-scale time-frequency features, yielding the fault identification results for the power cable to be identified. The multi-scale feature fusion diagnostic model includes a scale scoring module, a multi-branch parallel convolutional feature extraction module, a cross-scale self-attention concatenation module, a feature pooling module, and a classifier. The connection relationships between the modules are as follows: Figure 3 As shown. In the multi-scale feature fusion diagnostic model, fault identification is performed through the following steps: B1. The adaptive fusion weights for each scale are determined based on the signal state characteristics through the scale scoring module.

[0062] The signal state feature vector extracted and standardized in step A2 is input into the scale scoring module. The scale scoring module is a fully connected neural network, comprising an input layer, hidden layers, and an output layer. The number of neurons in the input layer is the same as the dimension of the signal state feature vector (e.g., 5 dimensions). The activation function of the hidden layer is the ReLU function (e.g., the hidden layer can be set to 32 dimensions). The activation function of the output layer is the Softmax function. The number of neurons in the output layer is the same as the number of candidate scales. The output value of the output layer is the adaptive fusion weight corresponding to each scale. These weights sum to 1, representing the model's preference for different time-frequency resolutions based on the current signal state.

[0063] B2. The multi-branch parallel convolution feature extraction module is used to extract independent features of multi-scale time-frequency features at different scales, so as to obtain the local features corresponding to each scale.

[0064] The multi-branch parallel convolutional feature extraction module includes a multi-branch parallel convolutional network, with the number of branches equal to the number of candidate scales. Each branch independently processes a continuous time-frequency map sequence at one scale. For each branch, the continuous time-frequency map sequence at the corresponding candidate scale is input into that branch, and local features with spatiotemporal representation capabilities are extracted through 3D convolution or pseudo-3D convolution (i.e., spatiotemporal separation convolution). The multi-branch parallel convolutional network independently performs convolutional feature extraction on the multi-scale time-frequency features at each scale to maintain the original spatial resolution of each scale, thus obtaining the local features corresponding to each scale. Because each branch processes independently, it avoids forcibly scaling maps of different resolutions to the same size, thereby preserving the original time-frequency details.

[0065] B3. The local features at each scale are sequentially spliced ​​by the cross-scale self-attention splicing module, and cross-scale global interaction is performed through the self-attention mechanism to obtain the fused features.

[0066] First, the local features output from each branch are flattened in spatial dimensions (including height and width) and concatenated to obtain a long, cross-scale sequence. Then, this long, cross-scale sequence is input into a multi-head self-attention module. The self-attention mechanism establishes dependencies between any two positions in the sequence, effectively capturing global interaction information between features at different scales, achieving cross-scale information complementarity, and outputting the fused features. The fused features contain fused features from multiple scales corresponding to each scale; that is, each scale remains an independent feature tensor after cross-scale interaction.

[0067] B4. The latent vectors of classification at each scale in the fused features are weighted and fused according to the adaptive fusion weights through the feature pooling module.

[0068] In this sub-step, the feature maps at each scale of the fused features are first subjected to global average pooling, compressing the spatiotemporal features at each scale into a fixed-dimensional latent vector for classification. Specifically, assuming that after processing by the cross-scale self-attention concatenation module, the first... The feature map corresponding to each scale is Global average pooling is performed on the feature map in terms of time, frequency, and spatial dimensions to obtain the corresponding classification latent vector. The dimension of the classification latent vector is consistent with the number of fused feature channels.

[0069] Then, the classification latent vectors at each scale are weighted and summed according to the adaptive fusion weights corresponding to each scale to obtain the global feature vector. The feature pooling module is specifically implemented as follows: Let the number of candidate scales be... , No. The classification latent vectors at each scale are The corresponding adaptive fusion weights are The weighted fusion of global feature vectors Represented as: Due to each The sum is 1, therefore It is a weighted average or weighted sum of the latent vectors for classification at each scale. Through this weighted fusion, the model can adaptively enhance the classification confidence contribution of appropriate scales and suppress the contribution of inappropriate scales based on the current signal state characteristics.

[0070] B5. Based on the features obtained after weighted fusion, the classifier outputs the fault identification result of the power cable to be identified.

[0071] The global feature vector obtained after weighted fusion The classifier (e.g., a Softmax layer) is input and outputs a probability distribution. The category corresponding to the maximum probability value in the probability distribution is the fault identification result of the power cable.

[0072] In another exemplary embodiment of this application, the multi-scale feature fusion diagnostic model is obtained through the following training steps, such as... Figure 4 As shown, the specific steps include: C1. Obtain partial discharge signals of power cables with known historical fault types.

[0073] Specifically, a partial discharge detection device is used to acquire a large number of original long-time-series partial discharge signals of historically known fault types, and the partial discharge fault type corresponding to each segment of the original signal is labeled to obtain its true fault category label.

[0074] C2. Extract signal state features from partial discharge signals of known historical fault types, and perform multi-scale time-frequency transformation based on the candidate scale library to construct a multi-scale continuous partial discharge time-frequency feature dataset.

[0075] The method employed in step A2 is used to process all the original signals: a sliding window segmentation strategy is used to continuously divide the historical partial discharge signal into multiple short-time segments; signal state features are extracted from each short-time segment; multiple sets of differentiated window parameters from the candidate scale library are used to perform parallel short-time Fourier transforms and logarithmic compression on each short-time segment; and multiple consecutive logarithmically compressed amplitude spectra on the same time series are concatenated and combined to form multi-scale time-frequency features. Finally, the multi-scale time-frequency features, signal state feature vectors, and fault category labels corresponding to each short-time segment constitute a training sample, and all samples form a multi-scale continuous partial discharge time-frequency feature dataset.

[0076] C3. Divide the multi-scale continuous partial discharge time-frequency feature dataset into training, validation, and test datasets. Divide the dataset into training, validation, and test datasets according to a preset ratio (e.g., 3:1:1). The training dataset is used to optimize model parameters, the validation dataset is used to adjust model hyperparameters and select the optimal model, and the test dataset is used for testing and evaluation.

[0077] C4. Train the initial multi-scale feature fusion diagnostic model using the training dataset. Optimize the model parameters based on the cross-entropy loss between the fault identification results output by the model and the known fault types until a well-trained multi-scale feature fusion diagnostic model is obtained. During the training process, a data augmentation strategy is introduced to mix the samples in the training dataset to generate augmented samples.

[0078] In one iteration of step C4, for a single sample in the training dataset, the following operations are performed: C41. Input the signal state features of any sample in the training dataset into the scale scoring module to obtain the adaptive fusion weights corresponding to each scale.

[0079] C42. Split the multi-scale time-frequency features of the sample according to the scale dimension, and input them into the corresponding branches of the multi-branch parallel convolutional feature extraction module to extract local features at each scale.

[0080] C43. Input the local features at each scale into the cross-scale self-attention stitching module for global interaction to obtain fused features.

[0081] C44. Input the fused features into the feature pooling module, and perform weighted fusion of the classification latent vectors at each scale in the fused features according to the adaptive fusion weights to obtain the global feature vector.

[0082] C45. Input the global feature vector into the classifier, and calculate the cross-entropy loss based on the probability distribution of partial discharge fault categories output by the classifier and the known fault types of the samples. Update the model parameters using the backpropagation algorithm. The formula for calculating the cross-entropy loss is: in, For cross-entropy loss, Indicates the number of partial discharge categories. This represents the true label (represented by one-hot encoding, where only the position corresponding to the target category is 1, and the other positions are 0). This represents the predicted probability of the multi-scale feature fusion diagnostic model, i.e., the output value of the Softmax function. With the natural number e as the base, The logarithm of .

[0083] After calculating the loss, the model parameters are updated using the backpropagation algorithm. This process is repeated until the model's performance on the validation set no longer improves, i.e., the convergence condition is met, resulting in a well-trained multi-scale feature fusion diagnostic model.

[0084] During training, a data augmentation strategy is introduced to mix training samples to generate augmented samples. Specifically, the data augmentation strategy is a Mixup data augmentation strategy, which mixes training samples with a Beta distribution at a preset probability (e.g., 0.3), where the parameter α of the Beta distribution is set to 0.6. The generated augmented samples participate in the model parameter update of the current round.

[0085] In addition, the AdamW optimizer was used to optimize the model parameters during training, with an initial learning rate of 0.001, 100 training epochs, and a batch size of 128.

[0086] To better illustrate the application effect of the method provided in this application, the partial discharge signal of the power cable obtained by the PDMaster cable partial discharge detection device validity verification system is used for experimental verification.

[0087] In this embodiment, the PDMaster cable partial discharge detection device validity verification system can simulate and output pulse sequence signals with statistical characteristics of real partial discharge spectra, and simultaneously provide a partial discharge monitoring channel and a synchronization reference channel. The partial discharge monitoring signal channel uses channel CH1, and the synchronization reference signal channel uses channel CH3. Using this system, partial discharge activities such as internal discharge, surface discharge, and tip discharge occurring at locations such as the cable body, cable joint, or cable terminal can be equivalently simulated without applying high voltage, thereby obtaining high-fidelity original partial discharge waveform data. Based on this device, this embodiment constructs 12 categories of partial discharge samples, and each complete detection experimental partial discharge waveform sequence contains 5×10 6 There are 10 sampling points, with a sampling interval of 4×10. -8 Seconds serve as the data foundation for subsequent model training and validation.

[0088] After obtaining the raw partial discharge signal, a sliding window segmentation strategy is used to segment the long time-series waveform. Specifically, continuous partial discharge observation data is traversed and truncated with a window length of 4096 and a sliding step size of 1024, ensuring a certain overlap between adjacent windows. This preserves the complete shape of the partial discharge pulse at the window edges as much as possible and avoids the loss of local features due to truncation. In this way, the raw long sequence signal can be converted into a large number of fixed-length short time segments, providing sample support for subsequent multi-scale time-frequency transformation and deep network input.

[0089] Subsequently, for each short time segment, a signal state feature vector containing five dimensions—normalized pulse width, peak factor, kurtosis, normalized spectral centroid, and local energy change rate—is extracted and standardized using the following formula: in, For the standardized first 3D eigenvalues For the first 3D eigenvalues For the training dataset, the first The mean of the dimensional features, For the training dataset, the first Standard deviation of dimensional features This is a minimum value to prevent overflow when divided by zero.

[0090] For each short-time segment, perform a time-frequency transform under the short-time Fourier transform parameters from the candidate scale library to generate a time-frequency spectrum. Let the original discrete partial discharge signal be... Its short-time Fourier transform can be expressed as: in, This is the result of the short-time Fourier transform. The partial discharge signal is discrete. For window functions, For window length, For frame shift, For time frame indexing, This is a frequency index. In this embodiment, the candidate scale library contains five sets of configuration parameters with window lengths of 16, 24, 32, 40, and 64, corresponding to window overlap lengths of 8, 12, 16, 20, and 32, respectively. The number of FFT points is the same as for each window length. By performing short-time Fourier transform on the same short-time segment using different window length parameters, partial discharge feature representations at different time and frequency resolutions can be obtained simultaneously. Short windows are more conducive to capturing the transient time changes of partial discharge pulses, while long windows are more conducive to extracting relatively stable frequency information. The amplitude spectrum of each set of short-time Fourier transform results is further extracted, and logarithmic compression is used to suppress the difference in amplitude dynamic range, making the weak features in the partial discharge signal easier for subsequent models to learn, as detailed below: in, This is the logarithmically compressed amplitude spectrum. This is the result of a short-time Fourier transform. This processing compresses the amplitude differences in the original time-frequency spectrum to a range more suitable for deep network learning, making it easier to preserve and identify subtle details in the partial discharge signal.

[0091] After obtaining the logarithmically transformed time-frequency spectra at each scale, the time-frequency spectra corresponding to multiple consecutive sub-windows within the same short time segment are combined in chronological order to form the continuous time-frequency spectrum sequence used in this embodiment. In this embodiment, the continuous spectrum sequence consists of 4 frames with an inter-frame step size of 1. Finally, the continuous time-frequency spectrum sequence is standardized using the following formula: in, For the first The continuous time-frequency spectrum sequence after scale standardization. For the first The original continuous time-frequency spectrum sequence at each scale, For the first The mean amplitude of the short-time segment data within a local region at each scale. For the first The amplitude standard deviation of the short-time segment data within a local region at each scale. Through the above processing, a multi-scale continuous partial discharge time-spectrum dataset suitable for deep network input is finally constructed.

[0092] After constructing the multi-scale continuous partial discharge time spectrum dataset, the dataset is divided into training dataset, validation dataset and test dataset according to a preset ratio of 3:1:1. The training dataset is used to optimize model parameters, the validation dataset is used to adjust model hyperparameters and evaluate model performance, and the test dataset is used for testing and evaluation.

[0093] During model training, the scale scoring module first maps the signal state feature vector to adaptive fusion weights α for five candidate scales. This scale scoring module is a fully connected neural network consisting of an input layer (5-dimensional), a hidden layer (32-dimensional, ReLU activation), and a softmax output layer (5-dimensional). The output value of the output layer is the adaptive fusion weight corresponding to each scale. Then, the samples in the training dataset are split according to the scale dimension and input separately into a multi-branch parallel convolutional feature extraction module. This module performs independent feature learning for time-frequency spectra at different scales, ensuring that localized details formed under different window lengths are preserved and extracted in shallow layers. Subsequently, the local features extracted at each scale are flattened in space and concatenated to obtain a long cross-scale sequence, which is then input into a cross-scale self-attention concatenation module to capture global interaction information between features at different scales, outputting a fusion result containing features fused from multiple scales. Then, the feature maps at each scale in the fused features are subjected to global average pooling to obtain the classification latent vectors corresponding to each scale. Finally, the classification latent vectors at each scale are weighted and fused according to the adaptive fusion weights corresponding to each scale to obtain a global feature vector. Finally, the global feature vector is input into the classifier, which outputs the corresponding partial discharge fault category prediction result. A schematic diagram of each module is shown below. Figure 3 As shown.

[0094] During training and optimization, the cross-entropy loss function shown below is used as the model's classification objective function: in, For cross-entropy loss, Indicates the number of partial discharge categories. This represents the true label (represented by one-hot encoding, where only the position corresponding to the target category is 1, and the other positions are 0). This represents the predicted probability of the multi-scale feature fusion diagnostic model, i.e., the output value of the Softmax function. With the natural number e as the base, The logarithm of .

[0095] In this embodiment, a batch size of 128 is used during training. The AdamW optimizer is used to optimize the gradient parameters of the network during training, with an initial learning rate of 0.001 and 100 training epochs. A Mixup data augmentation strategy is introduced during training, mixing the training samples with a Beta distribution (α=0.6) with a probability of 0.3 to generate augmented samples that participate in the model parameter updates of the current epoch.

[0096] After model training, samples from the test dataset are input into the trained model for performance evaluation, and metrics such as accuracy, precision, recall, and F1 score are recorded. Accuracy represents the proportion of correctly predicted samples out of the total number of samples, directly reflecting the model's overall classification ability. Precision represents the proportion of samples predicted as belonging to a certain class that actually belong to that class, reflecting the model's ability to suppress false positives. Recall represents the proportion of samples that actually belong to a certain class that are correctly identified by the model, reflecting the model's ability to reduce false negatives. The F1 score is the harmonic mean of precision and recall, comprehensively reflecting the model's balanced performance in the classification task.

[0097] To verify the effectiveness of the proposed model structure, this application further compares the improved UniFormer model with Random Forest, LSTM, Transformer, CNN, ViT, and the basic UniFormer model. Different models were trained and debugged on training and validation datasets, and then their training performance was evaluated on the test dataset. The training results of the different models are shown in Table 1. Table 1 Training results of different models

[0098] As shown in Table 1 above, the training results demonstrate that the basic UniFormer model outperforms deep learning models such as CNN, ViT, and LSTM, indicating that combining the local inductive ability of convolution with the global modeling ability of the attention mechanism can better adapt to the task of identifying partial discharge time-frequency spectra. This application, based on the basic UniFormer model, further introduces multi-scale time-frequency input, shallow multi-branch parallel convolution, and deep cross-scale feature concatenation mechanisms, enabling the model to simultaneously learn local pulse features and frequency evolution features under different window lengths. Therefore, it achieves further improvements in accuracy, precision, recall, and F1 score. This demonstrates that the proposed scheme can more fully exploit the multi-scale complementary information in partial discharge signals, dynamically adjust the time-frequency resolution preference according to the discharge characteristics of different samples, and improve the discrimination ability between similar fault categories.

[0099] To further verify the rationality of the multi-scale feature fusion method, this application also conducted comparative experiments on single-scale UniFormer, interpolation-aligned multi-scale UniFormer, and the improved UniFormer model proposed in this application. The experimental results are shown in Table 2: Table 2 Comparative experimental results of different multi-scale input methods

[0100] The comparative experimental results shown in Table 2 demonstrate that expanding the single-scale input to a multi-scale input improves the model performance to some extent. This indicates that short-time Fourier transform time-frequency maps at different scales can provide complementary information from both temporal and frequency resolution perspectives, effectively enhancing the partial discharge recognition effect. However, forcibly scaling different scale maps to a uniform size using interpolation alignment, while meeting network input requirements, alters the spatial distribution of the original time-frequency maps to some extent, leading to blurred feature edges, distorted pulse details, and high-frequency energy diffusion, thus limiting the full utilization of multi-scale information.

[0101] This application employs a multi-branch parallel convolutional feature extraction structure to process spectra at different scales independently, avoiding the loss of original features caused by interpolation and scaling. Simultaneously, a cross-scale self-attention splicing module performs global interactive fusion of features at different scales at a deep level, enabling the model to achieve cross-scale information complementarity while preserving local details at each scale. Building upon this, the adaptive multi-scale scheme further generates scale weights through a signal state perceptron, effectively learning and extracting signal features related to scale selection. Experimental results show that this scheme not only improves the overall recognition accuracy of the model but also exhibits better stability in terms of loss curve convergence speed and confusion matrix classification performance, indicating that the multi-scale fusion method proposed in this application is more suitable for modeling spectral features during partial discharge.

[0102] The above embodiments of this application provide a fault identification method based on multi-scale time-frequency feature adaptive fusion. By performing short-time Fourier transforms on multiple sets of differentiated window length parameters in a candidate scale library, multi-resolution time-frequency maps are generated on the same sample. Short windows preserve the transient time localization capability of partial discharge pulses, while long windows extract relatively stable frequency distribution features. The time-frequency maps at different scales complement each other, thereby reducing feature loss and time-frequency aliasing caused by a single fixed window length. Subsequently, the obtained multi-scale maps are input into the multi-scale feature fusion diagnostic model. An adaptive fusion weight for each scale is established based on signal state characteristics through a scale scoring module, enabling the model to dynamically adjust the contribution ratio of different time-frequency resolutions according to the transient and spectral characteristics of the partial discharge segment, avoiding scale redundancy or weakening of key features caused by fixed scale combinations. A multi-branch parallel convolutional network is used to independently extract local detail features from different scale maps, avoiding feature distortion caused by interpolation alignment. After concatenating the feature sequences at each scale, a multi-head self-attention mechanism is used for cross-scale global interaction, achieving collaborative perception of transient local details and long-term evolution patterns of partial discharge, thereby improving the model's ability to distinguish similar partial discharge fault categories.

[0103] Based on the same inventive concept, this application also provides a system for implementing the fault identification method based on multi-scale time-frequency feature adaptive fusion as described above. The solution provided by this system is similar to the implementation scheme described in the above method, and will not be repeated here.

[0104] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0105] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0106] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0107] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with relevant regulations and be authorized by the owner of the corresponding device.

[0108] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0109] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0110] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0111] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A fault identification method based on adaptive fusion of multi-scale time-frequency features, characterized in that, include: Acquire the partial discharge signal of the power cable to be identified; The signal state features of the partial discharge signal are extracted, and the partial discharge signal is subjected to multi-scale time-frequency transformation based on the candidate scale library to generate multi-scale time-frequency features; The multi-scale feature fusion diagnostic model, trained beforehand, is used to identify faults in the multi-scale time-frequency features, resulting in the fault identification of the power cable to be identified. The multi-scale feature fusion diagnostic model includes a scale scoring module, a multi-branch parallel convolutional feature extraction module, a cross-scale self-attention concatenation module, a feature pooling module, and a classifier. Within the multi-scale feature fusion diagnostic model: The adaptive fusion weights for each scale are determined based on the signal state features using the scale scoring module. The multi-branch parallel convolutional feature extraction module performs independent feature extraction on the multi-scale time-frequency features at each scale to obtain the local features corresponding to each scale. The cross-scale self-attention stitching module stitches together the local features at each scale and performs cross-scale global interaction through a self-attention mechanism to obtain fused features. The feature pooling module performs weighted fusion of the latent vectors of each scale in the fused features according to the adaptive fusion weights. The classifier outputs the fault identification result of the power cable to be identified based on the features obtained after weighted fusion. The signal state features of the partial discharge signal are extracted, and multi-scale time-frequency transformation is performed on the partial discharge signal based on a candidate scale library to generate multi-scale time-frequency features, specifically including: The partial discharge signal is continuously segmented using a sliding window segmentation strategy to obtain multiple short-time data segments. Statistical features reflecting the transient characteristics and spectral distribution characteristics of the signal are extracted from each of the short-time data segments and used as the signal state features; Using multiple sets of differentiated window parameters in the candidate scale library, short-time Fourier transforms are performed in parallel on each short-time segment data to extract amplitude spectra reflecting different time and frequency resolution dimensions. The amplitude spectra obtained from the differential window parameters of each group are subjected to logarithmic compression, and multiple adjacent logarithmic amplitude spectra on the same time series are cascaded and combined to form the multi-scale time-frequency features. The multi-scale time-frequency features are formed by concatenating and combining multiple adjacent logarithmic amplitude spectra on the same time series, including: For any of the aforementioned short-time segment data, a predetermined number of logarithmic amplitude spectra that are adjacent to each other on the time axis of the short-time segment data are stacked in chronological order to form a continuous time-frequency spectrum sequence corresponding to the short-time segment data. The continuous time-frequency spectrum sequence corresponding to each short time segment data is used as the multi-scale time-frequency feature; the continuous time-frequency spectrum sequence formed under different candidate scales has different time resolution and frequency resolution; The cross-scale self-attention stitching module concatenates the local features from each scale into sequences and performs cross-scale global interaction through a self-attention mechanism to obtain fused features, including: The local features at each scale are flattened in the spatial dimension and then spliced ​​together to obtain a long sequence across scales; The fusion features are obtained by globally interacting with the long sequence across scales through a self-attention mechanism; the fusion features include fused features of multiple scales corresponding to each scale.

2. The fault identification method based on multi-scale time-frequency feature adaptive fusion according to claim 1, characterized in that, The multi-scale feature fusion diagnostic model is obtained through the following training steps: Obtain partial discharge signals of power cables with known historical fault types; Signal state features are extracted from the partial discharge signals of the known historical fault types, and multi-scale time-frequency transformation is performed based on the candidate scale library to construct a multi-scale continuous partial discharge time-frequency feature dataset. The multi-scale continuous partial discharge time-frequency feature dataset is divided into a training dataset, a validation dataset, and a test dataset. The initial multi-scale feature fusion diagnostic model is trained using the training dataset. The model parameters are optimized based on the cross-entropy loss between the fault identification results output by the model and the known fault types until the trained multi-scale feature fusion diagnostic model is obtained. During the training process, a data augmentation strategy is introduced to mix the samples in the training dataset to generate augmented samples. The validation dataset is used to select the optimal model, and the test dataset is used to test and evaluate the performance of the model.

3. The fault identification method based on multi-scale time-frequency feature adaptive fusion according to claim 2, characterized in that, The initial multi-scale feature fusion diagnostic model is trained using the training dataset. The model parameters are optimized based on the cross-entropy loss between the fault identification results output by the model and the known fault types, until the trained multi-scale feature fusion diagnostic model is obtained. Specifically, this includes: Input the signal state features of any sample in the training dataset into the scale scoring module to obtain the adaptive fusion weights corresponding to each scale; The multi-scale time-frequency features of the sample are split according to the scale dimension and input into the corresponding branches of the multi-branch parallel convolutional feature extraction module to extract local features at each scale. The local features at each scale are input into the cross-scale self-attention stitching module for global interaction to obtain fused features; The fused features are input into the feature pooling module, and the classification latent vectors of each scale in the fused features are weighted and fused according to the adaptive fusion weights corresponding to each scale to obtain the global feature vector; The global feature vector is input into the classifier, and the cross-entropy loss is calculated based on the probability distribution of partial discharge fault categories output by the classifier and the known fault types of the samples. The model parameters are then updated using the backpropagation algorithm.

4. The fault identification method based on multi-scale time-frequency feature adaptive fusion according to claim 3, characterized in that, The formula for calculating the cross-entropy loss is: in, For cross-entropy loss, Indicates the number of partial discharge categories. Indicates the true label, This represents the predicted probability of the multi-scale feature fusion diagnostic model; the data augmentation strategy is the Mixup data augmentation strategy.

5. The fault identification method based on multi-scale time-frequency feature adaptive fusion according to claim 3, characterized in that, The candidate scale library contains at least two sets of differentiated window lengths, window overlap lengths, and fast Fourier transform points; The signal state characteristics include normalized pulse width, peak factor, kurtosis, normalized spectral centroid, and local energy change rate; The formula for standardizing the signal state features is as follows: in, For the standardized first 3D eigenvalues For the first 3D eigenvalues For the training dataset, the first The mean of the dimensional features, For the training dataset, the first Standard deviation of dimensional features To prevent overflow of the minimum value when dividing by zero; The formula for logarithmic compression of the amplitude spectrum obtained from the differential window parameters of each group is as follows: in, This is the logarithmically compressed amplitude spectrum. This is the result of the short-time Fourier transform. To prevent overflow of the minimum value when dividing by zero.

6. The fault identification method based on multi-scale time-frequency feature adaptive fusion according to claim 5, characterized in that, The formula for the short-time Fourier transform is: in, This is the result of the short-time Fourier transform. The partial discharge signal is discrete. For window functions, For window length, For frame shift, For time frame indexing, For frequency index, For discrete-time indexing, The imaginary unit, Pi is the mathematical constant of a circle.

7. The fault identification method based on multi-scale time-frequency feature adaptive fusion according to claim 3, characterized in that, After performing logarithmic compression on the amplitude spectra obtained from the differential window parameters of each group, the method further includes: The continuous time-frequency spectrum sequences after logarithmic compression at each candidate scale are standardized using the following formula: in, For the first The continuous time-frequency spectrum sequence after scale standardization. For the first A sequence of original continuous time-frequency spectra at each scale. For the first The mean amplitude of the short-time segment data within a local region at each scale. For the first The amplitude standard deviation of the short-time segment data within a local region at each scale. To prevent overflow of the minimum value when dividing by zero.

Images