Power cable fault sound recognition method and system based on multi-network fusion

Abstract

The invention relates to the technical field of power cable fault discharge sound recognition, and provides a power cable fault sound recognition method and system based on multi-network fusion, and the method comprises the steps: carrying out the multi-level feature extraction and data envelope analysis of a sound signal sequence, and generating a multi-scale feature and trend score graph; obtaining a similarity score graph and a value vector by using a self-attention mechanism; constructing gating characteristics and generating a mixed weight graph by splicing the trend and the similar graph; fusing the value vectors based on the mixed weight map to complete feature fusion; and finally, carrying out time sequence modeling and classification, and outputting an identification result. According to the invention, through the multi-network fusion deep learning model, automatic extraction and fusion of multi-level and multi-type features of the power cable discharge sound signals are realized, and the intelligent level of fault identification and the generalization ability of the model are improved. The influence caused by insufficient generalization ability of a classical sound signal description method and a recognition method is overcome, and intelligentization of power cable fault discharge sound recognition is realized.

Description

Technical Field

[0001] This invention relates to the technical field of power cable fault discharge sound recognition, specifically to a power cable fault sound recognition method and system based on multi-network fusion. Background Art

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] As a crucial infrastructure of urban power distribution networks, power cables play a key role in ensuring the safe and stable operation of power systems. Currently, urban power cables are typically laid underground using conduits or direct burial. While this method improves the safety of the lines and the coordination with urban planning, it also makes the cables more susceptible to external environmental influences, such as moisture, corrosion, and mechanical stress, leading to various cable faults such as insulation aging, joint failures, and mechanical damage.

[0004] To promptly detect and locate these faults, the power industry has developed various cable fault detection technologies, mainly including two key aspects: fault location and fault pinpointing. In the fault pinpointing stage, the mainstream methods currently used in engineering projects include acoustic detection and acoustic-magnetic synchronization: acoustic detection uses ground probes to collect sound signals generated by cable discharge, which are then processed by a host computer and played back through headphones for testing personnel to determine the fault location; acoustic-magnetic synchronization simultaneously collects sound and electromagnetic signals generated by discharge, using waveform time-difference analysis to assist in determining the fault location.

[0005] The methods described above all rely on the accurate acquisition and identification of discharge sound signals. However, in practical applications, the sound signals at fault points are susceptible to various interferences, such as complex environmental noise (traffic noise, pedestrian noise, etc.) and the influence of different laying media (such as cement, sand, etc.) on the sound wave propagation characteristics. These factors lead to a decrease in signal quality, making subsequent analysis difficult. To improve the accuracy of signal identification, some current studies use traditional signal processing methods to describe and classify discharge sounds, such as short-time Fourier transform, Mel-frequency cepstral coefficients (MFCC), and wavelet transform. These methods typically rely on fixed feature extraction mechanisms and are difficult to adapt to complex scenarios where multiple media and multi-source noise coexist. Meanwhile, existing deep learning models, such as convolutional neural networks, lack effective modeling capabilities for multi-level structural features in sound signals, such as local mutations and global trends; and they have significant shortcomings in handling long-term dependencies in sound signals, failing to comprehensively characterize the adaptability and generalization capabilities of discharge sound signals in various scenarios. Summary of the Invention

[0006] To address the aforementioned problems, this invention proposes a method and system for power cable fault sound recognition based on multi-network fusion. Through a deep learning network model that integrates multiple networks, it enables the interactive fusion of multiple features, allowing for the automatic extraction of multi-level and multi-type features from cable discharge sound signals. This method highlights local detail features and global trend features while mitigating redundant features and models the long-term dependencies of sound signals. It overcomes the limitations of classical sound signal description and recognition methods in terms of generalization ability, thus achieving intelligent recognition of power cable fault discharge sounds.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: One or more embodiments provide a method for power cable fault sound recognition based on multi-network fusion, including the following steps: The acquired audio signal sequence to be identified Multi-level feature extraction is performed to obtain the extracted multi-scale features. ; The acquired audio signal sequence to be identified Data envelopment feature extraction is performed, and a trend score map is generated based on the obtained envelope. ; Extracted multi-scale features Perform self-attention mechanism operations to generate a similarity score graph between the query and the key. and value vectors; Trend Score Chart Similarity score map The data is concatenated, convolved, and activated to generate gated features; the fusion ratio of the similarity score map and the trend score map is dynamically controlled based on the gated features to generate a hybrid weight map. ; Based on hybrid weight graph For using multi-scale features The generated value vectors are fused to obtain the fused features. ; Targeting fusion features After performing time-series modeling and classification, the classification results of the sound signals are obtained.

[0008] One or more embodiments provide a power cable fault sound recognition system based on multi-network fusion, including: A multi-level feature extraction network is configured to extract the acquired sequence of sound signals to be identified. Multi-level feature extraction is performed to obtain the extracted multi-scale features. ; The envelope feature extraction module is configured to extract the acquired sequence of sound signals to be recognized. Data envelopment feature extraction is performed, and a trend score map is generated based on the obtained envelope. ; The fusion attention module is configured to extract multi-scale features. Perform self-attention mechanism operations to generate a similarity score graph between the query and the key. and value vectors; The gating fusion module is configured to convert the trend score map Similarity score map The data is concatenated, convolved, and activated to generate gated features; the fusion ratio of the similarity score map and the trend score map is dynamically controlled based on the gated features to generate a hybrid weight map. Based on hybrid weight graph For using multi-scale features The generated value vectors are fused to obtain the fused features. ; The timing capture and signal classification modules are configured to target fused features. After performing time-series modeling and classification, the classification results of the sound signals are obtained.

[0009] One or more embodiments provide a power cable fault sound recognition system based on multi-network fusion, including: a sound acquisition device and a processor; The processor is configured to perform the steps of the above-described method for identifying power cable fault sounds based on multi-network fusion.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: The proposed power cable fault discharge sound recognition method can automatically extract multi-level and multi-type features from discharge sound signals, highlighting local detail features and global trend features, effectively weakening redundant information, and improving the pertinence and discriminativeness of feature representation. By fusing and modeling multi-scale structural information and trend features, it overcomes the limitations of traditional methods that rely on single features. The introduced gating fusion mechanism dynamically adjusts the fusion ratio based on the complementarity between features, enhancing the system's adaptability and robustness to different sound patterns. Furthermore, by utilizing a self-attention mechanism to model the long-term dependencies of sound signals, it achieves accurate characterization of complex dynamic evolution processes, significantly improving fault identification and location accuracy under multi-source noise interference and media differences.

[0011] The advantages of the present invention, as well as its additional advantages, will be described in detail in the following specific embodiments. Attached Figure Description

[0012] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute a limitation thereof.

[0013] Figure 1 This is a flowchart of the power cable fault sound recognition method according to Embodiment 1 of the present invention; Figure 2 This is a model architecture diagram of the deep learning network model with multi-network fusion in Embodiment 1 of the present invention; Figure 3 This is an example waveform diagram of the sound signal of a fault discharge of a power cable laid at a manhole cover during the construction of the training set in Embodiment 1 of the present invention; Figure 4 This is an example waveform diagram of the fault discharge sound signal of power cables laid on the ground and underground during the construction of the training set in Embodiment 1 of the present invention; Figure 5 This is an example waveform diagram of the non-discharge sound signal during the construction of the training set in Embodiment 1 of the present invention; Figure 6 This is a network structure diagram of the multi-level feature extraction network in Embodiment 1 of the present invention; Figure 7 This is a schematic diagram of the fusion attention module according to Embodiment 1 of the present invention; Figure 8 This is a schematic diagram of the gated fusion module according to Embodiment 1 of the present invention; Figure 9 This is a schematic diagram of the long short-term memory network of Embodiment 1 of the present invention; Figure 10 This is a schematic diagram of the timing capture and signal classification module of Embodiment 1 of the present invention. Detailed Implementation

[0014] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0015] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0016] It should be noted that the terminology used herein is for describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof. It should be noted that, without conflict, the various embodiments and features within those embodiments can be combined with each other. The embodiments will now be described in detail with reference to the accompanying drawings.

[0017] Example 1 In one or more of the technical solutions disclosed in the embodiments, such as Figures 1 to 10 As shown, a method for power cable fault sound recognition based on multi-network fusion includes the following steps: Step 1: Process the acquired audio signal sequence to be recognized. Multi-level feature extraction is performed to obtain the extracted multi-scale features. ; Step 2: Process the acquired audio signal sequence to be recognized. Data envelopment feature extraction is performed, and a trend score map is generated based on the obtained envelope. ; Step 3: Extract the multi-scale features Perform self-attention mechanism operations to generate a similarity score graph between the query and the key. and value vectors; Step 4: Develop the trend score chart. Similarity score map The data is concatenated, convolved, and activated to generate gated features; the fusion ratio of the similarity score map and the trend score map is dynamically controlled based on the gated features to generate a hybrid weight map. ; Step 5: Based on the hybrid weight graph For using multi-scale features The generated value vectors are fused to obtain the fused features. ; Step 6: Targeting fusion features After performing time-series modeling and classification, the classification results of the sound signals are obtained; based on the identified faulty discharge sound signals, further localization can be performed to obtain the fault localization results.

[0018] This implementation addresses the problem of insufficient modeling capabilities caused by the large amount of environmental noise mixed in the collected raw sound data. It proposes a multi-level feature extraction method, which significantly improves the extraction capability of sound signal features and enhances feature representation. To address the difficulty of convolutional neural networks in depicting the overall trend of signal changes, a data envelopment analysis (DEA) method is proposed to model the temporal contour of the sound signal and generate a trend score map, thereby effectively modeling the overall trend of the sound signal. To address the model's poor ability to capture long-term dependencies of the sound signal, a fusion attention module is proposed. The extracted multi-scale features are used to calculate the relationship between queries, keys, and values, generating a similarity score map and its corresponding value vector. By dynamically controlling the fusion ratio of the trend score map and the similarity score map through gating fusion, the ability of the model's extracted features to represent global trends and local abrupt changes is significantly enhanced. Finally, the above method achieves effective modeling and classification of fault sounds, and the classification results can then be used for fault signal localization.

[0019] The power cable fault discharge sound recognition method proposed in this embodiment can automatically extract multi-level and multi-type features from the discharge sound signal, highlighting local detail features and global trend features, effectively weakening redundant information, and improving the pertinence and discriminativeness of feature expression. By fusing and modeling multi-scale structural information and trend features, it overcomes the limitations of traditional methods that rely on single features. The introduced gating fusion mechanism dynamically adjusts the fusion ratio based on the complementarity between features, enhancing the system's adaptability and robustness to different sound patterns. Furthermore, by utilizing a self-attention mechanism to model the long-term dependency of the sound signal, it achieves accurate characterization of complex dynamic evolution processes, significantly improving the fault identification and location accuracy under multi-source noise interference and media differences.

[0020] Step 1 can be achieved by constructing a multi-level feature extraction network, such as... Figure 6 As shown, the structure of the multi-level feature extraction network includes a first CR module, a splitting module, a lightweight processing branch and a deep extraction branch, a splicing module, and a second CR module. The two CR modules include convolutional layers and activation layers, which perform convolution and activation operations on the input features. In this embodiment, the activation layer uses the ReLU activation function. The multi-level feature extraction network can significantly improve the ability to extract sound signal features and enhance the feature expression ability.

[0021] Based on the constructed multi-level feature extraction network, the acquired sound signal sequence to be identified is processed. Multi-level feature extraction is performed to obtain the extracted multi-scale features. Multiscale features The extraction method includes the following steps: Step 11: Process the acquired audio signal sequence to be recognized. Convolution and activation processing are performed, and the features are split along the channel dimension and then input into the lightweight processing branch and the deep extraction branch; Specifically, the input sound signal sequence is The input is fed into a multi-level feature extraction network with two branches to capture multi-level local detail information in the sound signal. A CR module combining convolution and ReLU activation functions processes the input data, and the features are split along the channel dimension before being fed into a lightweight processing branch and a deep extraction branch; the feature processing formula is as follows: (1); (2); in, This represents the input audio signal sequence data. This represents the convolution operation. Represents the ReLU activation function. This represents a combination of convolution operation and ReLU activation function. Indicates a splitting operation; This indicates the characteristics of the input lightweight processing branch after processing by the CR module. This indicates the features of the deep extraction branch after processing by the CR module.

[0022] Step 12: Extract branches at a deeper level for the input features. A deep, progressive, multi-layer encoding based on convolutional neural networks is used to obtain multi-scale information, and skip connection blocks are used to connect the multi-scale information to obtain multi-level features. The formula for deep branch extraction is: (3); (4); (5); in, This represents the features of the input deep extraction branches. This represents the multi-level features obtained by progressively extracting them through skip connect blocks. This represents the convolution operation. Represents the ReLU activation function. This represents a combination of convolution operation and ReLU activation function. This indicates a jump connection block.

[0023] Figure 6 In this context, n represents the dimension of the input signal; c represents the feature dimension of the signal; and m represents the number of skip connections in deep feature extraction. n × c / 2 represents the feature dimension after the splitting operation; n × ( m +2) c / 2 represents the feature dimension after splicing the deep extraction branch and the lightweight processing branch.

[0024] Step 13: Concatenate the multi-level features obtained from the deep extraction branch and the low-level information features obtained from the lightweight processing branch along the channel dimension, and then reduce the dimensionality of the concatenated features to obtain the fused multi-scale features. The fusion process, the formula is: (6); (7); in, This represents the low-level characteristics of lightweight processing branches. This represents the features of the input deep extraction branches. This represents the multi-level features obtained by progressively extracting them through skip connect blocks. This represents the convolution operation. Represents the ReLU activation function. This represents a combination of convolution operation and ReLU activation function. This indicates a splicing operation along the channel dimension.

[0025] In the feature extraction scheme described in this embodiment, the concatenated features are dimensionality reduced, balancing network performance and efficiency by decreasing the number of network parameters. The dual-branch, multi-level feature network achieves efficient and comprehensive feature extraction through the synergistic effect of a deep extraction branch and a lightweight processing branch. The deep extraction branch extracts multi-scale information and deep features through multi-layer convolutional operations and skip connection blocks, while the lightweight processing branch quickly extracts basic feature information. The fusion of these two approaches enables the model to more comprehensively understand the input data, improving its performance and robustness, and making it suitable for various complex tasks.

[0026] In step 2, the acquired audio signal sequence to be recognized is processed. Data envelopment feature extraction is performed, and a trend score map is generated based on the obtained envelope. Trend score chart The generation method is as follows: Step 21: Extract the envelope features from the acquired sound signal sequence x to be recognized, i.e., perform Hilbert transform to obtain the envelope line, as shown in the following formula: (8); in, Indicates the input signal Hilbert transform, .

[0027] Step 22: Generate a trend score chart using the envelope line. The formula is: (9); in, .

[0028] In step 3, the obtained multi-scale features are processed. The process of generating associated features by performing self-attention mechanisms includes the following steps: Step 31: Combine multi-scale features Mapped to query vector key vector Sum value vector ; Step 32: Apply the mapped matrix to the learnable weight matrix. , , Perform a linear transformation; the formula is as follows: (10); (11); (12); Step 33: Calculate the similarity between each query vector and all key vectors using the dot product, and obtain the similarity score graph between queries and keys. Calculated using the following formula: (13); (14); in, It is the dimension of the key vector. express The transpose of .

[0029] Step 4 is implemented through the gating fusion module, as shown in the structure diagram below. Figure 8 As shown, the trend score map obtained in step 2 and the similarity score map obtained in step 3 are concatenated, and a gate feature is generated using convolution and the Sigmoid activation function. The gate feature is used to dynamically control the ratio of the similarity score map to the trend score map, and a mixed weight map is generated using convolution. The formula is as follows: (15); (16); in, , The weight matrix represents the convolution. and This represents the bias of the convolution. Represents the dot product operation. , gate represents the gated feature; In step 5, based on the hybrid weight graph Using multi-scale features The generated value vectors are fused to obtain the final self-attention output features. for: (17); Step 6, targeting fusion features Methods for performing time series modeling and classification include the following steps: Step 61: fusion features after dynamically assigning weights based on the self-attention mechanism Long Short-Term Memory (LSTM) networks are used to model time-series relationships, capture temporal features, and obtain output features. To better understand the dynamic changes in sound signals during the fault occurrence process.

[0030] LSTM is used to model the time-series relationships of the fused features after dynamic weight allocation via a self-attention mechanism, resulting in the final output features. These are feature representations after time series modeling. These features can better reflect the dynamic changes of sound signals during the fault occurrence process, including long-term dependencies and short-term dynamic changes, thereby improving the accuracy of fault detection and the robustness of the model.

[0031] like Figure 9 As shown, the fusion features are based on a long short-term memory network. The process of performing time series relationship modeling is as follows: Three gates—the forget gate, the input gate, and the output gate—are used to construct the Long Short-Term Memory (LSTM) network to control the flow of information and capture long-term dependencies. The specific formula is as follows: (18); in, It is the output of the forget gate. It is the sigmoid activation function. It is the weight matrix of the forget gate. It is the bias vector of the forget gate. It is the hidden state of the previous time step. These are the input features at the current time step; (19); (20); in, It is the output of the input gate. It is a candidate cell state. It is the hyperbolic tangent activation function. and These are the weight matrices for the input gate and the candidate cell state, respectively. and These are the bias vectors for the input gate and the candidate cell state, respectively; (twenty one); in, It represents the cell state at the current time step. It represents the cell state at the previous time step.

[0032] (twenty two); (twenty three); in, It is the output of the output gate. It is the hidden state at the current time step. It is the weight matrix of the output gate. It is the bias vector of the output gate.

[0033] Through the gating mechanism described above, LSTM can process input features step by step over time. and update the hidden state. and cell state Ultimately, the output features of LSTM It contains information about the dynamic changes in the sequence.

[0034] Step 62: Output features of the Long Short-Term Memory (LSTM) network After activation processing, the output is transformed into a probability distribution through a fully connected layer and a softmax function to obtain the result of the sound signal type determination, which may include faulty or faultless. Optionally, the activation operation processing can be a ReLU activation operation; like Figure 10 The diagram shows the fully connected classification structure. The Long Short-Term Memory network outputs features. After processing by the ReLU activation function, the features are mapped to the classification space through a fully connected layer, and the output is transformed into a probability distribution using the softmax function, thereby determining the type of sound signal.

[0035] Further technical solutions also include constructing a deep learning network that integrates multiple networks to implement the processes of steps 1 to 6, including a feature extraction module, a fusion attention module, and a temporal capture and signal classification module; The feature extraction module, configured to perform steps 1 and 2, includes a multi-level feature extraction network and an envelope feature extraction module. A multi-level feature extraction network is configured to extract the acquired sequence of sound signals to be identified. Multi-level feature extraction is performed to obtain the extracted multi-scale features. ; The envelope feature extraction module is configured to extract the acquired sequence of sound signals to be recognized. Data envelopment feature extraction is performed, and a trend score map is generated based on the obtained envelope. ; The fusion attention module is configured to extract multi-scale features. Perform self-attention mechanism operations to generate a similarity score graph between the query and the key. and value vectors; The gating fusion module is configured to convert the trend score map Similarity score map The data is concatenated, convolved, and activated to generate gated features; the fusion ratio of the similarity score map and the trend score map is dynamically controlled based on the gated features to generate a hybrid weight map. Based on hybrid weight graph Using multi-scale features The generated value vectors are fused to obtain the fused features. ; The timing capture and signal classification modules are configured to target fused features. After performing temporal modeling and classification, the classification results of the sound signals are obtained; this includes a long short-term memory network, a fully connected layer, and an activation module; the activation module can be activated using the Softmax function, which transforms the output into a probability distribution; Furthermore, it also includes methods for training deep learning networks that fuse multiple networks, comprising the following steps: Step S1: Obtain cable fault discharge sound signals and non-discharge sound signals, label them, and construct a training dataset; Optionally, a cable fault locator can be used to collect sound signals, including discharge sound signals and non-discharge sound signals of cable faults at different locations. In this embodiment, signals from two locations were selected: one at the manhole cover and the other underground. The manhole cover is where the cable is installed or accessed for maintenance, and the sound signal is transmitted through the cavity under the manhole cover. The sound signal from the underground location is the sound signal that propagates through the soil. The dataset constructed using signals from different locations is used to verify the generalization performance of the model.

[0036] Specifically, we can collect both the discharge and non-discharge sound signals of cable faults at manhole covers to construct dataset A; and collect both the discharge and non-discharge sound signals of underground and surface cable faults to construct dataset B. For datasets A and B respectively, we can label the cable fault discharge sound signal as 1 and the non-discharge sound signal as 0.

[0037] Figure 3 The waveform of the sound signal of cable fault discharge at the manhole cover is shown. Figure 4 The waveform of the sound signal of fault discharge in ground and underground cables is shown. Figure 5 The waveform of the non-discharge sound signal is shown.

[0038] Step S2: Divide the dataset into training dataset and test dataset according to the set ratio; Specifically, dataset A and dataset B are divided into training dataset and test dataset in a 7:3 ratio.

[0039] For datasets A and B, each dataset has 1260 training samples, including 630 sets of discharge sounds and 630 sets of non-discharge sounds; the test samples are 540 sets each, including 270 sets of discharge sounds and 270 sets of non-discharge sounds. The training datasets from datasets A and B are used for training... Figure 2The parameters of the multi-network fusion deep learning network are shown. The test dataset is used to verify the recognition performance of the trained network model.

[0040] Step S3: Input the discharge sound and non-discharge sound signals from the training dataset into a deep network model that integrates multiple features. Through multiple iterations of learning and model parameter optimization, obtain the optimal network model, including the following steps: Step S31: Perform multi-level feature extraction on the discharge sound and non-discharge sound signals in the training dataset to obtain the extracted multi-scale features; Step S32: Extract data envelopment features from the discharge sound and non-discharge sound signals in the training dataset, and generate a trend score map based on the obtained envelope. Step S33: Perform self-attention mechanism operation on the extracted multi-scale features to generate a similarity score map and value vector between the query and the key; Step S34: Concatenate the trend score map and the similarity score map, perform convolution and activation to generate gated features; dynamically control the fusion ratio of the similarity score map and the trend score map according to the gated features to generate a hybrid weight map; Step S35: Based on the hybrid weight map, fuse the value vectors generated using multi-scale features to obtain fused features; Step S36: After performing temporal modeling and classification on the fusion features, the classification results of the sound signals are obtained; Step S37: Calculate the loss value of the overall loss function based on the fault identification results, adjust the parameters of each model of the multi-network fusion deep learning network, and obtain the trained multi-network fusion deep learning network model through multiple iterations of training and backpropagation until the convergence condition is met. Furthermore, the constructed loss function includes classification loss and regularization loss: (twenty four); The classification loss is: (25); This loss term is used to optimize the binary classification performance of the model. It's a real label. This is the positive class probability predicted by the model. By minimizing the difference between the predicted probability and the true label, this loss term can improve the model's classification accuracy in fault detection tasks.

[0041] The regularization loss is: (26); This loss term uses L2 regularization to prevent overfitting. These are the parameters of the model. This is the regularization coefficient. By adding a regularization term to the loss function, the complexity of the model can be effectively reduced, and the model's generalization ability can be improved.

[0042] Furthermore, the deep learning network model integrating multiple networks was tested using a test sound dataset, and the model performance evaluation index was calculated using the following formula: (27); Wherein, TP represents the number of samples correctly identified as fault discharge sound signals, TN represents the number of samples correctly identified as non-discharge sound signals, FP represents the number of samples incorrectly identified as fault discharge sound signals, and FN represents the number of samples incorrectly identified as non-discharge sound signals.

[0043] Tables 1 and 2 show the performance of the optimal training model in recognizing cable fault discharge sounds under a single laying medium. Table 1 shows the model built using training samples from dataset A, and the recognition performance was verified using test samples from dataset A. Table 2 shows the model built using training samples from dataset B, and the recognition performance was verified using test samples from dataset B.

[0044] Table 1 shows the recognition performance verified using test samples from dataset A.

[0045] Table 1 shows the confusion matrix of a single experiment when using a multi-feature interactive fusion deep network model to identify both discharge and non-discharge sound signals from cable faults at manhole covers. Ultimately, after ten cross-experiments, the average accuracy rate for sound recognition was 97.03%.

[0046] Table 2 uses test samples from dataset B to verify the recognition effect;

[0047] Table 2 shows the confusion matrix of a single experiment when using a multi-feature interactive fusion deep network model to identify both discharge and non-discharge sound signals from cable faults at manhole covers. Ultimately, after ten cross-experiments, the average accuracy rate for sound recognition was 98.11%.

[0048] Table 3 shows the effect of power cable fault discharge sound recognition across laying media. That is, the model is built using training samples in dataset A and the recognition effect is verified using test samples in dataset B, which verifies the generalization ability of the multi-network fusion deep learning network architecture in sound recognition in this embodiment.

[0049] Table 3 shows the recognition performance of the model built using training samples from dataset A on test samples from dataset B.

[0050] Table 3 shows the multi-feature interactive fusion deep network model trained and optimized using both the discharge sound and non-discharge sound of cable faults at manhole covers. Then, the confusion matrix of one experiment was used to generalize the trained model to the recognition of discharge sounds and non-discharge sounds of underground and surface cable faults. Finally, through ten cross-validation experiments, the average accuracy rate for sound recognition was 87.07%.

[0051] Ablation experiments were conducted to compare the individual effects of the proposed "multi-level feature extraction network" and "signal envelope trend score" on datasets A and B, respectively, as well as the effects of using "fusion attention" to integrate multi-scale features. and trend score chart The accuracy after fusion is shown in Table 4. Table 4. Results of the ablation test;

[0052] The checkmarks in Table 4 indicate the inclusion of the corresponding modules or the use of the corresponding methods. It can be seen that when only the basic models (i.e., the convolutional neural networks CNN and LSTM in this embodiment) are used, the accuracy of the models on dataset A and dataset B is 90.56% and 89.72% respectively, which is relatively low performance.

[0053] After introducing a multi-level feature extraction network, the accuracy was significantly improved to 95.43% (dataset A) and 96.25% (dataset B), indicating that the module can effectively mine deeper feature information.

[0054] When the envelope feature extraction module is used alone to extract the signal envelope trend score, the accuracy reaches 94.32% and 93.35%, which is also higher than the basic model, indicating that this feature can provide important discriminative information.

[0055] Based on this, a fusion attention mechanism was introduced to comprehensively utilize multi-scale features and trend scores. The model ultimately achieved the best performance with accuracies of 97.03% and 98.11% on datasets A and B, respectively.

[0056] In summary, the ablation experiments fully demonstrate that the multi-level feature extraction network and the signal envelope trend score both play important roles in improving model performance, while the fusion attention mechanism further enhances the model's ability to comprehensively utilize different features, significantly improving the accuracy and robustness of fault detection.

[0057] The intelligent identification method for power cable fault discharge sound proposed in this embodiment can automatically extract multi-level and multi-type features of cable discharge sound signals, highlight local detail features and global trend features, weaken redundant features, and model the long-term dependency relationship of sound signals. It overcomes the influence of insufficient generalization ability of classical sound signal description and identification methods, realizes intelligent identification of power cable fault discharge sound, avoids the subjective influence caused by relying on the experience of testers, and provides a strong guarantee for the safe and stable operation of the power system.

[0058] Example 2 Based on Embodiment 1, this embodiment provides a power cable fault sound recognition system based on multi-network fusion, including: A multi-level feature extraction network is configured to extract the acquired sequence of sound signals to be identified. Multi-level feature extraction is performed to obtain the extracted multi-scale features. ; The envelope feature extraction module is configured to extract the acquired sequence of sound signals to be recognized. Data envelopment feature extraction is performed, and a trend score map is generated based on the obtained envelope. ; The fusion attention module is configured to extract multi-scale features. Perform self-attention mechanism operations to generate a similarity score graph between the query and the key. and value vectors; The gating fusion module is configured to convert the trend score map Similarity score map The data is concatenated, convolved, and activated to generate gated features; the fusion ratio of the similarity score map and the trend score map is dynamically controlled based on the gated features to generate a hybrid weight map. Based on hybrid weight graph For using multi-scale features The generated value vectors are fused to obtain the fused features. ; The timing capture and signal classification modules are configured to target fused features. After performing time-series modeling and classification, the classification results of the sound signals are obtained.

[0059] It should be noted that each module in this embodiment corresponds one-to-one with each step in embodiment 1, and their specific implementation process is the same, so it will not be repeated here.

[0060] Example 3 Based on Embodiment 1, this embodiment provides a power cable fault sound recognition system based on multi-network fusion, including: a sound acquisition device and a processor; The sound acquisition device can be a cable fault locator, used to acquire fault discharge sound signals and non-discharge sound signals; The processor is configured to perform the steps of the power cable fault sound recognition method based on multi-network fusion described in Embodiment 1.

[0061] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0062] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for identifying power cable fault sounds based on multi-network fusion, characterized in that, Includes the following steps: The acquired audio signal sequence to be identified Multi-level feature extraction is performed to obtain the extracted multi-scale features. ; The acquired audio signal sequence to be identified Data envelopment feature extraction is performed, and a trend score map is generated based on the obtained envelope. ; Extracted multi-scale features Perform self-attention mechanism operations to generate a similarity score graph between the query and the key. and value vectors; Trend Score Chart Similarity score map The data is concatenated, convolved, and activated to generate gated features; the fusion ratio of the similarity score map and the trend score map is dynamically controlled based on the gated features to generate a hybrid weight map. ; Based on hybrid weight graph For using multi-scale features The generated value vectors are fused to obtain fused features. ; Targeting fusion features After performing time-series modeling and classification, the classification results of the sound signals are obtained.

2. The power cable fault sound recognition method based on multi-network fusion as described in claim 1, characterized in that: Extracting multi-scale features This is achieved by constructing a multi-level feature extraction network, which includes a first CR module, a splitting module, a lightweight processing branch and a deep extraction branch, a splicing module and a second CR module connected in sequence; the two CR modules include convolutional layers and activation layers, which perform convolution and activation operations on the input features.

3. The power cable fault sound recognition method based on multi-network fusion as described in claim 2, characterized in that: Multiscale features The extraction method includes the following steps: The acquired audio signal sequence to be identified Convolution and activation processing are performed, and the features are split along the channel dimension and then input into the lightweight processing branch and the deep extraction branch; In the deep extraction branch, the input features are encoded using a deep progressive multi-layer encoding based on a convolutional neural network to obtain multi-scale information, and the multi-scale information is connected using skip connection blocks to obtain multi-level features. The multi-level features obtained from the deep extraction branch and the low-level information features obtained from the lightweight processing branch are concatenated along the channel dimension, and the concatenated features are then dimensionality-reduced to obtain the fused multi-scale features. .

4. The power cable fault sound recognition method based on multi-network fusion as described in claim 1, characterized in that, Trend Score Chart The generation method is as follows: The acquired audio signal sequence to be identified Perform Hilbert transform, then extract envelope features to obtain the envelope line; Generate a trend score chart using the envelope.

5. The power cable fault sound recognition method based on multi-network fusion as described in claim 1, characterized in that, The obtained multi-scale features Performing self-attention mechanism operations includes the following steps: Multiscale features Mapped to a matrix, including the query matrix Key matrix Sum matrix ; The mapped matrix is ​​then linearly transformed using a learnable weight matrix. The similarity between each query vector and all key vectors is calculated using the dot product, resulting in a similarity score graph between queries and keys. .

6. The power cable fault sound recognition method based on multi-network fusion as described in claim 1, characterized in that: The ratio of similarity score maps to trend score maps is dynamically controlled using gating features, and a mixed weight map is generated using convolution. , the formula is as follows: ； ； in, , The weight matrix represents the convolution. and This represents the bias of the convolution. Represents the dot product operation. , gate represents the gated feature.

7. The power cable fault sound recognition method based on multi-network fusion as described in claim 1, characterized in that: Targeting fusion features Methods for performing time series modeling and classification include the following steps: Fusion features derived from dynamically assigning weights to the self-attention mechanism using a Long Short-Term Memory network Perform time series relationship modeling, capture time series features, and obtain output features. ; Output features of Long Short-Term Memory Network After activation processing, the output is transformed into a probability distribution through a fully connected layer and a softmax function to obtain the result of determining the type of sound signal.

8. A power cable fault sound recognition system based on multi-network fusion, characterized in that, include: A multi-level feature extraction network is configured to extract the acquired sequence of sound signals to be identified. Multi-level feature extraction is performed to obtain the extracted multi-scale features. ; The envelope feature extraction module is configured to extract the acquired sequence of sound signals to be recognized. Data envelopment feature extraction is performed, and a trend score map is generated based on the obtained envelope. ; The fusion attention module is configured to extract multi-scale features. Perform self-attention mechanism operations to generate a similarity score graph between the query and the key. and value vectors; The gating fusion module is configured to convert the trend score map Similarity score map The data is concatenated, convolved, and activated to generate gated features; the fusion ratio of the similarity score map and the trend score map is dynamically controlled based on the gated features to generate a hybrid weight map. Based on hybrid weight graph For using multi-scale features The generated value vectors are fused to obtain fused features. ; The timing capture and signal classification modules are configured to target fused features. After performing time-series modeling and classification, the classification results of the sound signals are obtained.

9. The power cable fault sound recognition system based on multi-network fusion as described in claim 8, characterized in that: The ratio of similarity score maps to trend score maps is dynamically controlled using gating features, and a mixed weight map is generated using convolution. , the formula is as follows: ； ； in, , The weight matrix represents the convolution. and This represents the bias of the convolution. Represents the dot product operation. , gate represents the gated feature.

10. A power cable fault sound recognition system based on multi-network fusion, characterized in that, include: Sound acquisition device and processor; The processor is configured to perform the steps of the power cable fault sound recognition method based on multi-network fusion as described in any one of claims 1-7.

Images