Energy storage system arc fault diagnosis method and system based on voiceprint detection

Abstract

The invention relates to the technical field of arc fault diagnosis, and provides an energy storage system arc fault diagnosis method and system based on voiceprint detection, and the method comprises the following steps: collecting a voiceprint signal in real time, and carrying out the anti-interference preprocessing, and obtaining a time-frequency domain spectrogram; identifying the time-frequency domain spectrogram based on a multi-branch convolutional neural network and weighting the time-frequency domain spectrogram to obtain a time sequence feature sequence; inputting the time sequence feature sequence into a time sequence convolutional network model, and outputting global context features; performing dynamic weighted pooling on the global context features to obtain key frame feature vectors representing the arc fault; and inputting the key frame feature vector into an arc fault diagnosis classification model to obtain an arc fault type, an arc fault severity grade and sound source positioning information, and generating a corresponding alarm signal and a control instruction based on the arc fault severity grade. According to the invention, the detection sensitivity and the anti-interference capability are improved, and the safety protection level and the fault early warning capability of the energy storage system are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of arc fault diagnosis, and in particular to a method and system for diagnosing arc faults in an energy storage system based on voiceprint detection. Background Art

[0002] Energy storage systems are a key component of modern power systems, playing a vital role in energy transition and ensuring grid stability. However, due to the high voltage and high current characteristics of energy storage systems, arc faults, once occurring, pose serious safety risks, including fire, explosion, and even casualties. Arc faults are a common and dangerous type of fault in energy storage systems, primarily manifesting as incomplete discharges caused by poor contact, cable aging, or insulation damage. Due to the intermittent, low-energy, and diverse nature of arc faults, traditional current and voltage monitoring methods often struggle to effectively detect and identify them.

[0003] In existing technologies, arc fault detection in energy storage systems primarily utilizes methods based on electrical parameter monitoring, spectrum analysis, and simple machine learning. Initial fault identification is achieved by monitoring abnormal current / voltage fluctuations, arc characteristics are captured using fixed-band spectrum analysis, and fault diagnosis is performed in conjunction with shallow classifiers. However, while this method can achieve basic fault detection, it lacks sensitivity to low-energy arcs and is susceptible to interference from load fluctuations. Fixed-band spectrum analysis struggles to analyze the dynamic characteristics of arcs in the time-frequency domain, while shallow machine learning models lack the ability to resist complex noise interference and model time series features. Consequently, existing technologies generally suffer from high false alarm rates, missed detection rates, and a lack of fault classification and location capabilities, making them unable to meet the high-precision safety warning requirements of energy storage systems. Summary of the Invention

[0004] In view of this, the present invention proposes a method and system for diagnosing arc faults in energy storage systems based on voiceprint detection, which solves the problems that the existing technology is insufficiently sensitive to low-energy arcs and is easily disturbed by load fluctuations, fixed-band spectrum analysis is difficult to analyze the dynamic characteristics of arcs in the time and frequency domains, and shallow machine learning models lack the ability to resist interference from complex noises and model time series characteristics. As a result, the existing technology generally has high false alarm rates, missed detection rates, and lacks fault classification and positioning capabilities, and cannot meet the high-precision safety warning needs of energy storage systems.

[0005] The technical solution of the present invention is implemented as follows: In a first aspect, the present invention provides a method for diagnosing arc faults in an energy storage system based on voiceprint detection, comprising the following steps:

[0006] Acquiring the soundprint signal of the energy storage system in real time based on an acoustic sensor array, performing anti-interference preprocessing on the soundprint signal, and obtaining a time-frequency domain spectrogram;

[0007] Constructing a multi-branch convolutional neural network, identifying the time-frequency domain spectrogram based on the multi-branch convolutional neural network, obtaining the time-frequency characteristics of high-frequency spike pulses and low-frequency discharge whistles, and weighting the time-frequency characteristics of high-frequency spike pulses and low-frequency discharge whistles using a channel-by-channel frequency band weighting method to obtain a time series feature sequence;

[0008] Constructing a temporal convolutional network model, inputting the temporal feature sequence into the temporal convolutional network model, and outputting global context features;

[0009] Dynamically weighted pooling the global context features is performed through a multi-head temporal attention network to obtain a key frame feature vector representing the arc fault;

[0010] An arc fault diagnosis and classification model is constructed, and the key frame feature vector is input into the arc fault diagnosis and classification model to obtain the arc fault type, arc fault severity grade and sound source location information, and a corresponding alarm signal and control instruction are generated based on the arc fault severity grade.

[0011] On the basis of the above technical solution, preferably, the real-time acquisition of the voiceprint signal of the energy storage system based on the acoustic sensor array and the anti-interference preprocessing of the voiceprint signal to obtain the time-frequency domain spectrogram include:

[0012] Arrange acoustic sensor arrays with anti-electromagnetic interference capabilities on the energy storage system busbars and key connection points, and determine the number, type, and spatial layout of the sensors;

[0013] Based on the acoustic sensor array, the soundprint signals during the operation of the energy storage system are collected in real time, and the collected soundprint signals are preprocessed with anti-interference to obtain a time-frequency domain spectrogram. The anti-interference preprocessing includes adaptive gain control and dual-passband notch filtering.

[0014] On the basis of the above technical solution, preferably, the multi-branch convolutional neural network is constructed, the time-frequency domain spectrogram is identified based on the multi-branch convolutional neural network, the time-frequency characteristics of high-frequency spike pulses and low-frequency discharge whistles are obtained, and the time-frequency characteristics of high-frequency spike pulses and low-frequency discharge whistles are weighted using a channel-by-channel frequency band weighting method to obtain a time series feature sequence, including:

[0015] Constructing a multi-branch convolutional neural network comprising multiple parallel branches, wherein each branch is provided with a convolution kernel structure of a different size, and extracting the high-frequency spike feature map and the low-frequency discharge whistling feature map in the time-frequency domain spectrogram in parallel;

[0016] The feature maps extracted from each branch are subjected to the channel-by-channel frequency band weighting method, the importance weight of each frequency band is calculated and weighted fusion is performed to obtain the temporal feature sequence.

[0017] On the basis of the above technical solution, preferably, the multi-branch convolutional neural network includes three parallel branches, wherein the first branch is used to capture high-frequency spike pulse features, the second branch is used to capture medium-frequency transition features, and the third branch is used to capture low-frequency discharge whistling features; each parallel branch includes two layers of convolution operations, wherein the first layer of convolution is followed by a ReLU activation function and a batch normalization layer, and the second layer of convolution is followed by a maximum pooling operation;

[0018] The channel-by-channel frequency band weighting method adopts a compression-excitation network structure, compresses the spatial information of each feature channel through global average pooling, learns the dependency between channels through a two-layer fully connected network, outputs the frequency band weight coefficient of each channel, multiplies the frequency band weight coefficient with the original feature map for channel-by-channel weighting, and cascades the weighted feature maps of the three branches to obtain a comprehensive temporal feature sequence.

[0019] On the basis of the above technical solution, preferably, the constructing of a temporal convolutional network model, inputting the temporal feature sequence into the temporal convolutional network model, and outputting global context features include:

[0020] Constructing a temporal convolutional network model based on causal convolution and dilated convolution, wherein the temporal convolutional network model includes multiple layers of residual blocks, each of which has a dilated convolution layer with an exponentially increasing dilation rate.

[0021] The temporal feature sequence is input into the temporal convolutional network model, and the temporal dependency features are extracted through the stacking of multiple layers of residual blocks and the skip connection structure to obtain the global context features covering the entire sequence time span.

[0022] Based on the above technical solution, preferably, the global context features are dynamically weighted pooled by a multi-head temporal attention network to obtain a key frame feature vector representing the arc fault, including:

[0023] Construct a multi-head temporal attention network to adaptively evaluate the importance of global context features at different time steps and build long-range temporal dependencies;

[0024] Based on the multi-head temporal attention network, the attention weight of each time step is calculated, the global context features are dynamically weighted pooled, and the key frame feature vector representing the characteristics of the arc fault is extracted.

[0025] Based on the above technical solution, preferably, the multi-head temporal attention network includes 4 parallel attention heads, each of which consists of a query-key-value triple structure;

[0026] Each attention head generates a query matrix, a key matrix, and a value matrix through linear projection. The attention score is obtained by calculating the similarity between the query matrix and the key matrix. The attention score is normalized using the temperature-scaled Softmax function to obtain the attention weight. The output of each attention head is obtained through the attention weight value matrix.

[0027] The outputs of each attention head are spliced and linearly transformed to obtain the multi-head fusion attention weight. The global context features are dynamically weighted and summed using the multi-head fusion attention weight to obtain a key frame feature vector of fixed dimension.

[0028] Based on the above technical solution, preferably, the arc fault diagnosis classification model is constructed, the key frame feature vector is input into the arc fault diagnosis classification model, the arc fault type, arc fault severity level and sound source location information are obtained, and the corresponding alarm signal and control instruction are generated based on the arc fault severity level, including:

[0029] Constructing an arc fault diagnosis and classification model based on a multi-task learning framework, inputting the key frame feature vector into the arc fault diagnosis and classification model, and simultaneously performing fault type identification, severity classification, and sound source location by sharing underlying feature representation to obtain arc fault type, arc fault severity classification, and sound source location information;

[0030] Based on the fault diagnosis results output by the arc fault diagnosis classification model, the danger level of the arc fault is determined according to the preset multi-level threshold value, and a graded alarm signal and corresponding control instructions are generated. The control instructions are sent to the energy storage system management system through a secure communication channel to trigger the corresponding protection measures.

[0031] In a second aspect, the present invention further provides an energy storage system arc fault diagnosis system based on voiceprint detection, the system comprising:

[0032] A voiceprint signal acquisition module is used to collect the voiceprint signal of the energy storage system in real time based on the acoustic sensor array, perform anti-interference preprocessing on the voiceprint signal, and obtain a time-frequency domain spectrogram;

[0033] a time series feature extraction module, configured to construct a multi-branch convolutional neural network, identify the time-frequency domain spectrogram based on the multi-branch convolutional neural network, obtain the time-frequency features of high-frequency spike pulses and low-frequency discharge whistles, and weight the time-frequency features of high-frequency spike pulses and low-frequency discharge whistles using a channel-by-channel frequency band weighting method to obtain a time series feature sequence;

[0034] A global feature capture module is used to build a temporal convolutional network model, input the temporal feature sequence into the temporal convolutional network model, and output global context features;

[0035] An arc fault characterization module is configured to dynamically weight pool the global context features through a multi-head temporal attention network to obtain a key frame feature vector representing the arc fault;

[0036] The fault diagnosis and alarm module is used to build an arc fault diagnosis and classification model, input the key frame feature vector into the arc fault diagnosis and classification model, obtain the arc fault type, arc fault severity level and sound source location information, and generate corresponding alarm signals and control instructions based on the arc fault severity level.

[0037] In a third aspect, the present invention further provides an electronic device comprising: at least one processor, at least one memory, a communication interface, and a bus;

[0038] Among them, the processor, memory, and communication interface communicate with each other through the bus, the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to implement the steps of a method for diagnosing arc faults in an energy storage system based on voiceprint detection.

[0039] The energy storage system arc fault diagnosis method and system based on voiceprint detection of the present invention have the following beneficial effects compared with the prior art:

[0040] (1) By utilizing the unique acoustic characteristics of arcs in voiceprint signals, combined with a deep learning architecture that uses a multi-branch convolutional neural network to extract high and low frequency features, a temporal convolutional network to capture long-range dependencies, and a multi-head temporal attention mechanism to extract key frame features, the system achieves sensitive identification, precise classification, and accurate positioning of arc faults. Especially in complex noise environments, the system improves detection sensitivity and anti-interference capabilities through anti-interference preprocessing and channel-by-channel frequency band weighting. At the same time, it supports fault severity classification and precise sound source positioning, providing a hierarchical protection response capability for the safe operation of the energy storage system and improving the safety protection level and fault warning capability of the energy storage system.

[0041] (2) Through three parallel branches, high-frequency spike pulse features, medium-frequency transition features, and low-frequency discharge whistling features are accurately identified respectively. The adaptive weighting of frequency band importance is realized by combining the channel-by-channel attention mechanism, which improves the recognition accuracy of arc features in different frequency bands. The compression-excitation network structure is used to model the inter-channel dependency relationship, suppress the interference of irrelevant frequency bands, enhance the expression of key frequency band features, and retain the complete time-frequency information through feature cascade fusion. Thus, the high-frequency spike pulse and low-frequency discharge whistling features of arc faults can be accurately extracted in a complex noise environment.

[0042] (3) Through the organic combination of the temporal convolutional network model and the multi-head temporal attention network, the temporal characteristics and key information of the arc fault are fully captured and accurately extracted. The exponentially growing dilated convolution in the temporal convolutional network significantly expands the receptive field range and can capture the arc temporal dependency of up to hundreds of milliseconds. The stacking structure and jump connection mechanism of the four-layer residual block effectively prevent the gradient from disappearing and enhance feature reuse. The multi-head temporal attention network uses four parallel attention heads to adaptively evaluate different temporal positions to achieve accurate recognition of key arc feature frames. By introducing an adaptive threshold screening mechanism, the arc outbreak moment is highlighted in complex background noise, thereby improving the detection sensitivity and anti-interference ability of intermittent arc faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0044] Figure 1 This is a flow chart of a method for diagnosing arc faults in an energy storage system based on voiceprint detection according to the present invention;

[0045] Figure 2 This is a structural diagram of an energy storage system arc fault diagnosis system based on voiceprint detection according to the present invention. DETAILED DESCRIPTION

[0046] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0047] See also Figure 1 The present invention provides a method for diagnosing arc faults in an energy storage system based on voiceprint detection, comprising the following steps:

[0048] Acquiring the soundprint signal of the energy storage system in real time based on an acoustic sensor array, performing anti-interference preprocessing on the soundprint signal, and obtaining a time-frequency domain spectrogram;

[0049] Constructing a multi-branch convolutional neural network, identifying the time-frequency domain spectrogram based on the multi-branch convolutional neural network, obtaining the time-frequency characteristics of high-frequency spike pulses and low-frequency discharge whistles, and weighting the time-frequency characteristics of high-frequency spike pulses and low-frequency discharge whistles using a channel-by-channel frequency band weighting method to obtain a time series feature sequence;

[0050] Constructing a temporal convolutional network model, inputting the temporal feature sequence into the temporal convolutional network model, and outputting global context features;

[0051] Dynamically weighted pooling the global context features is performed through a multi-head temporal attention network to obtain a key frame feature vector representing the arc fault;

[0052] An arc fault diagnosis and classification model is constructed, and the key frame feature vector is input into the arc fault diagnosis and classification model to obtain the arc fault type, arc fault severity grade and sound source location information, and a corresponding alarm signal and control instruction are generated based on the arc fault severity grade.

[0053] Specifically, this embodiment utilizes the unique acoustic characteristics of arcs in voiceprint signals, combines a deep learning architecture with a multi-branch convolutional neural network to extract high- and low-frequency features, a temporal convolutional network to capture long-range dependencies, and a multi-head temporal attention mechanism to extract key frame features. This achieves sensitive identification, precise classification, and accurate positioning of arc faults. Especially in complex noise environments, anti-interference preprocessing and channel-by-channel frequency band weighting improve detection sensitivity and anti-interference capability. At the same time, it supports fault severity classification and precise sound source positioning, providing a hierarchical protection response capability for the safe operation of the energy storage system and improving the safety protection level and fault warning capability of the energy storage system.

[0054] The method of collecting the voiceprint signal of the energy storage system in real time based on the acoustic sensor array and performing anti-interference preprocessing on the voiceprint signal to obtain a time-frequency domain spectrogram includes:

[0055] Arrange acoustic sensor arrays with anti-electromagnetic interference capabilities on the energy storage system bus and key connection points, and determine the number, type and spatial layout of sensors.

[0056] In a specific embodiment, the acoustic sensor array uses microacoustic sensors with a working bandwidth of 20kHz to 100kHz. By performing a point-distribution layout on the energy storage system bus, battery connection points and key connectors, and setting an electromagnetic shielding layer and grounding measures around the sensors and signal cables, the anti-interference capability and signal acquisition quality are improved.

[0057] Based on the acoustic sensor array, the soundprint signals during the operation of the energy storage system are collected in real time, and the collected soundprint signals are preprocessed with anti-interference to obtain a time-frequency domain spectrogram. The anti-interference preprocessing includes adaptive gain control and dual-passband notch filtering.

[0058] In a specific embodiment, the collected voiceprint signal is digitized at a sampling rate of not less than 256 kHz, the signal strength is dynamically adjusted using adaptive gain control, and background narrowband noise is suppressed by combining bandpass filtering with dual-passband notch filtering. Finally, a short-time Fourier transform is performed on the processed signal to obtain a time-frequency domain spectrogram.

[0059] Specifically, this embodiment achieves high-quality soundprint signal acquisition through acoustic sensor array deployment and signal preprocessing strategies. By deploying a microacoustic sensor array with an operating bandwidth of 20kHz to 100kHz on the energy storage system busbars and key connection points, using a point-distributed layout and equipped with electromagnetic shielding layers and grounding measures, the acquisition accuracy and anti-interference capability of arc fault soundprint signals are improved. At the same time, a high sampling rate of no less than 256kHz ensures the complete capture of high-frequency spikes. Combined with adaptive gain control, dynamic adjustment of soundprint signals of varying intensities is achieved. Dual-band notch filtering is used to suppress the narrowband noise inherent in the energy storage system, such as cooling fan and transformer vibration, ultimately generating a high-quality time-frequency domain spectrogram, thus overcoming the problem of insufficient sensitivity of traditional electrical parameter monitoring methods in complex noisy environments.

[0060] The method comprises: constructing a multi-branch convolutional neural network, identifying the time-frequency domain spectrogram based on the multi-branch convolutional neural network, obtaining the time-frequency features of high-frequency spike pulses and low-frequency discharge whistles, and weighting the time-frequency features of high-frequency spike pulses and low-frequency discharge whistles using a channel-by-channel frequency band weighting method to obtain a time series feature sequence, including:

[0061] A multi-branch convolutional neural network containing multiple parallel branches is constructed, each branch is set with a convolution kernel structure of a different size, and the high-frequency spike pulse feature map and the low-frequency discharge whistling feature map in the time-frequency domain spectrogram are extracted in parallel.

[0062] In a specific embodiment, the multi-branch convolutional neural network includes three parallel branches, with convolution kernels of sizes 1×3, 1×5 and 1×7 respectively set, wherein the 1×3 convolution kernel branch is used to capture high-frequency spike pulse features, the 1×5 convolution kernel branch is used to capture medium-frequency transition features, and the 1×7 convolution kernel branch is used to capture low-frequency discharge whistling features; each parallel branch includes two layers of convolution operations, wherein the first layer of convolution is followed by a ReLU activation function and a batch normalization layer, and the second layer of convolution is followed by a maximum pooling operation.

[0063] The feature maps extracted from each branch are subjected to the channel-by-channel frequency band weighting method, the importance weight of each frequency band is calculated and weighted fusion is performed to obtain the temporal feature sequence.

[0064] In a specific embodiment, the channel-by-channel attention mechanism adopts a compression-excitation network structure. First, the spatial information of each feature channel is compressed through global average pooling, and then the dependency between channels is learned through a two-layer fully connected network to output the weight coefficient of each frequency band. Finally, the weight coefficient is multiplied by the original feature map to complete the channel-by-channel weighting, and the weighted feature maps of the three branches are cascaded and fused to generate a comprehensive temporal feature sequence.

[0065] In a specific embodiment, the convolution operations of the three branches are:

[0066] Y1=ReLU(BN(W1X+b1));

[0067] Y2=ReLU(BN(W2X+b2));

[0068] Y3=ReLU(BN(W3X+b3));

[0069] Where Y1 is the high-frequency spike feature map, Y2 is the medium-frequency transition feature map, Y3 is the low-frequency discharge whistling feature map, X is the time-frequency domain spectrogram, W1 is the 1×3 convolution kernel weight, W2 is the 1×5 convolution kernel weight, W3 is the 1×7 convolution kernel weight, b1 is the high-frequency spike bias term, b2 is the medium-frequency transition bias term, b3 is the low-frequency discharge whistling bias term, BN(·) is the batch normalization operation, and ReLU(·) is the rectified linear unit activation function;

[0070] The calculation formula for the weight coefficient of the channel-by-channel attention mechanism is:

[0071] Z i =AvgPool(Y i ),i∈{1,2,3};

[0072]

[0073] V i =U i ⊙Y i ;

[0074] Among them, Z i is the channel statistic vector obtained after pooling the feature map of the i-th branch, AvgPool(·) is the global average pooling function, U i is the frequency band channel importance weight vector corresponding to the i-th branch feature map, σ(·) is the sigmoid activation function, W1 att and are the weight matrices of the first and second layers of the attention module, and are the first layer bias item of the attention module and the second layer bias item of the attention module, V i is the channel importance weighted feature map corresponding to the i-th branch feature map, and ⊙ is the channel-by-channel multiplication operation;

[0075] The calculation formula of the time series feature sequence is:

[0076] F0=Concat[V1,V2,V3]·W f +b f ;

[0077] Among them, F0 is the time series feature sequence, Concat[·] is the feature cascade function, and W f is the feature fusion layer weight matrix, b f is the feature fusion layer bias term.

[0078] Specifically, this embodiment uses three parallel branches with different convolution kernel sizes (1×3, 1×5, and 1×7) to fully capture the arc fault's performance in different frequency bands, effectively distinguishing between high-frequency spike pulse characteristics (usually appearing at the initial stage of arcing) and low-frequency discharge whistling characteristics (usually appearing during the arcing phase).

[0079] Each parallel branch is optimized for a specific frequency band, ensuring accurate identification of key frequency band characteristics of arc faults in energy storage systems and improving the pertinence and effectiveness of feature extraction.

[0080] By introducing a compression-excitation network structure, a channel-by-channel attention mechanism is implemented, which automatically learns the importance weights of each frequency band, highlights the key frequency band features and suppresses the interference of irrelevant frequency bands, and improves the feature differentiation ability in complex noise environments; by adopting a feature cascade fusion method, the complete feature information extracted by each branch is retained, avoiding the information loss that may be caused by simple addition or averaging.

[0081] The step of constructing a temporal convolutional network model, inputting the temporal feature sequence into the temporal convolutional network model, and outputting global context features includes:

[0082] A temporal convolutional network model based on causal convolution and dilated convolution is constructed. The temporal convolutional network model includes multiple layers of residual blocks. Each residual block is internally provided with a dilated convolution layer with an exponentially increasing dilation rate to ensure that the receptive field covers the critical temporal range.

[0083] In a specific embodiment, the temporal convolutional network model is composed of 4 stacked residual blocks, each residual block contains two layers of one-dimensional causal convolution layers, the convolution kernel size is 3, and the expansion rates of each residual block are 1, 2, 4, and 8, respectively, to achieve exponential growth; each residual block is also provided with layer normalization, Dropout regularization, and input-output identity mapping connection to form a complete residual learning structure; the model also introduces a channel attention module to adaptively adjust the importance weight of each channel feature.

[0084] The temporal feature sequence is input into the temporal convolutional network model, and the temporal dependency features are extracted through the stacking of multiple layers of residual blocks and the skip connection structure to obtain the global context features covering the entire sequence time span.

[0085] In a specific embodiment, the temporal convolutional network model sets jump connections between the residual blocks of each layer, and forms a multi-scale feature fusion path through the feature reuse mechanism; the output features of each residual block are nonlinearly transformed through the gated linear unit activation function to enhance the ability to capture sparse arc features; finally, the features of each layer are adaptively weighted and fused to obtain global context features, and the global context features simultaneously contain the long-range temporal dependency and transient feature information of the arc event.

[0086] In a specific embodiment, the calculation formula of the global context feature is:

[0087] G=σ(W g S4+b g )⊙S4;

[0088] S j =α j ·F j out +(1-α j )·S j-1 ;

[0089]

[0090] Among them, G is the global context feature, W g is the weight of the global feature extraction layer, b g is the global feature extraction layer bias parameter, S4 is the skip connection output of the last residual block, σ(·) is the sigmoid activation function, S j is the skip connection output of the jth residual block, α j is the adaptive learning skip connection weight coefficient of the jth residual block, is the output feature of the jth residual block, S j-1 is the skip connection output of the j-1th residual block, is the output feature of the k-th layer in the j-th residual block after adding the residual connection, Dropout(·) is the random deactivation function, is the gated activation feature of the kth layer in the jth residual block, GLU(·) is the gated linear unit function, ⊙ is the element-wise multiplication operation, and are the first and second half of the feature tensor in the channel dimension, is the convolution output feature of the kth layer in the jth residual block, conv1d(·) is the one-dimensional convolution function, is the convolution kernel weight parameter of the kth layer in the jth residual block, d j is the dilation rate of the jth residual block, is the normalized output feature of the k-th layer in the j-th residual block, is the input feature of the kth layer in the jth residual block, and LayerNorm(·) is the layer normalization function.

[0091] Specifically, this embodiment constructs a temporal convolutional network model based on causal convolution and dilated convolution, specifically modeling the temporal characteristics of arc fault voiceprint signals, including:

[0092] Overall network architecture: It consists of 4 stacked residual blocks to form a deep temporal feature extraction network.

[0093] Dilated convolution design: A dilated convolution layer with an exponentially increasing dilation rate is used in each residual block. The dilation rates are 1, 2, 4, and 8, respectively, which exponentially expands the network receptive field.

[0094] Internal structure of the residual block: Each residual block contains two one-dimensional causal convolution layers with a convolution kernel size of 3; integrates layer normalization and random dropout regularization mechanisms; uses an identity mapping connection between input and output to form a complete residual learning structure; introduces a gated linear unit activation function for nonlinear transformation; and adds a channel attention module to adaptively adjust the importance weight of each channel feature.

[0095] Cross-layer connection mechanism: Skip connections are set between residual blocks to form a multi-scale feature fusion path, and feature reuse is achieved through adaptive learning weight coefficients.

[0096] This embodiment uses an exponentially increasing expansion rate design (1, 2, 4, 8) to exponentially expand the receptive field, effectively capturing arc timing information within a time span of hundreds of milliseconds. It can simultaneously model the complete life cycle characteristics of arc bursts, duration, and extinction. By adopting causal convolution, the model ensures that only information from the current and past time steps is used for prediction, ensuring the feasibility of real-time detection and meeting the timeliness requirements of energy storage system safety monitoring.

[0097] Through the four-layer residual block structure combined with the skip connection mechanism, the gradient vanishing problem of deep networks is solved, the end-to-end propagation of arc features is guaranteed, and the model training stability and convergence speed are improved.

[0098] Through cross-layer skip connections and adaptive weight fusion, effective integration of features at different time scales is achieved, which can not only capture the transient changes of arcs (up to microseconds), but also identify their continuous evolution patterns (up to milliseconds to seconds); the gated linear unit activation function combined with layer normalization enhances the ability to capture sparse arc features, which is particularly suitable for processing non-stationary and sudden signal characteristics in arc faults.

[0099] By introducing a random dropout regularization mechanism, the model is effectively prevented from overfitting the training data, improving its generalization ability in unseen scenarios and enabling stable operation in different types of energy storage systems. Compared with traditional recurrent neural networks / long short-term memory network structures, the temporal convolutional network of this embodiment supports parallel computing, improves the inference speed, meets the low-latency requirements of real-time monitoring of energy storage systems, and maintains a low computing resource usage.

[0100] The global context features are dynamically weighted pooled by a multi-head temporal attention network to obtain a key frame feature vector representing the arc fault, including:

[0101] Construct a multi-head temporal attention network to adaptively evaluate the importance of global context features at different time steps and build long-range temporal dependencies;

[0102] Based on the multi-head temporal attention network, the attention weight of each time step is calculated, the global context features are dynamically weighted pooled, and the key frame feature vector representing the characteristics of the arc fault is extracted.

[0103] The multi-head temporal attention network includes four parallel attention heads, each of which consists of a query-key-value triple structure;

[0104] Each attention head generates a query matrix, a key matrix, and a value matrix through linear projection. The attention score is obtained by calculating the similarity between the query matrix and the key matrix. The attention score is normalized using a temperature-scaled Softmax function to obtain the attention weight. The output of each attention head is obtained through the weighted value matrix of the attention weight. The multi-head temporal attention network introduces position encoding and residual connection structure to enhance the ability to capture temporal relationships.

[0105] The outputs of each attention head are spliced and linearly transformed to obtain the multi-head fusion attention weight. The multi-head fusion attention weight takes into account both the global context and local peak detection, and accurately captures the critical moment by introducing an adaptive threshold screening mechanism. The global context features are dynamically weighted and summed by the multi-head fusion attention weight to obtain a key frame feature vector of fixed dimension. The key frame feature vector comprehensively encodes the key timing characteristics of the arc fault.

[0106] In a specific embodiment, the calculation formula of the key frame feature vector is:

[0107]

[0108]

[0109] MultiAtt=Concat[Att1,Att2,…,Att4]·W O ;

[0110] Att h =softmax(Score h )·V h r ;

[0111]

[0112] Q h =W h Q G;

[0113] K h =W h K G;

[0114]

[0115] Among them, Q h , K h 、 are the query matrix, key matrix, and value matrix of the h-th attention head, respectively. Q h , K h 、 The corresponding linear projection weight matrix, G is the global context feature, D is the feature dimension, T is the time step, Score h is the attention score matrix of the h-th head, d s is the scaling factor, P mask Encodes a mask for the learnable position, Att his the attention output of the h-th attention head, softmax(·) is the normalization function, MultiAtt is the multi-head attention fusion output, Concat[·] is the concatenation function on the feature dimension, W O is the output projection matrix, β t is the dynamic attention weight at time step t, λ is the temperature parameter, 1(·) is the indicator function, exp(·) is the exponential function, γ is the adaptive threshold coefficient, G t is the vector representation of the global context feature at time step t, F key is the key frame feature vector.

[0116] Specifically, this embodiment describes a multi-head temporal attention network for dynamically weighting global context features and filtering key information, including:

[0117] A multi-head temporal attention network with 4 parallel attention heads was constructed. Each attention head consists of a triple structure of query, key, and value. The core idea of "attention is weight" was adopted to realize dynamic importance evaluation of temporal information.

[0118] The global context features are mapped into query matrix, key matrix and value matrix through linear projection. The original attention score is obtained by calculating the dot product similarity between the query matrix and the key matrix. A learnable position encoding mask is introduced to enhance the perception of temporal position information. The attention score is normalized using the temperature-scaled Softmax function to obtain the attention weight. The attention weight is used to perform weighted summation on the value matrix to obtain the output feature of each attention head.

[0119] This embodiment uses an adaptive threshold screening mechanism to accurately locate the critical moments of arc fault occurrence, development, and disappearance, avoiding the blind spots and redundancy of the traditional fixed window method. The adaptive threshold screening mechanism can focus on the arc outbreak moment in complex noisy environments, thereby improving accuracy.

[0120] This embodiment breaks through the limitation of traditional methods that it is difficult to establish long-term series dependencies. It can simultaneously focus on the arc evolution process within hundreds of milliseconds. The four parallel attention heads can capture the characteristics of arc events from different angles and time scales, forming a multi-dimensional and comprehensive time series perception capability.

[0121] Through the query-key-value mechanism, dynamic evaluation of the importance of different time steps is achieved, key arc features are given higher weights, and the timing patterns of arc faults are adaptively learned without the need for manual pre-setting of key frame recognition rules, thereby improving the adaptability to new arc faults.

[0122] This embodiment compresses variable-length global context features into fixed-dimensional keyframe feature vectors, achieving dimensionality reduction while retaining key information. The generated keyframe feature vectors contain the essential characteristics of arc faults. Through matrix parallel computing and an optimized attention mechanism, efficient keyframe extraction is achieved, meeting the low-latency requirements of real-time monitoring of energy storage systems. Compared with traditional recurrent neural network structures, this embodiment reduces inference time and reduces computing resource usage.

[0123] This embodiment uses multi-head design and feature splicing to capture and effectively fuse information in different feature subspaces, thereby enhancing the expressiveness of the model. Four attention heads are used to focus on features in different frequency ranges and time spans, jointly constructing a comprehensive arc fault characterization.

[0124] The arc fault diagnosis classification model is constructed, the key frame feature vector is input into the arc fault diagnosis classification model, the arc fault type, the arc fault severity level and the sound source location information are obtained, and the corresponding alarm signal and control instruction are generated based on the arc fault severity level, including:

[0125] An arc fault diagnosis and classification model based on a multi-task learning framework is constructed. The key frame feature vector is input into the arc fault diagnosis and classification model. By sharing the underlying feature representation, fault type identification, severity classification and sound source localization are simultaneously performed to obtain arc fault type, arc fault severity classification and sound source localization information.

[0126] In a specific embodiment, the arc fault diagnosis and classification model of the multi-task learning framework adopts a hybrid architecture of "hard parameter sharing + soft parameter sharing", which includes a shared feature extraction backbone network and three task-specific branch networks; the backbone network is composed of three fully connected layers, which are used to extract common features from key frame feature vectors; the fault type identification branch adopts a classifier based on the attention mechanism to accurately distinguish between contact arcs, series arcs and parallel arcs; the severity grading branch adopts a multi-label sequence model to divide arc faults into four levels; the sound source localization branch accurately determines the location of the arc through time difference analysis of multi-band acoustic features and triangulation positioning algorithm.

[0127] Based on the fault diagnosis results output by the arc fault diagnosis classification model, the danger level of the arc fault is determined according to the preset multi-level threshold value, and a graded alarm signal and corresponding control instructions are generated. The control instructions are sent to the energy storage system management system through a secure communication channel to trigger the corresponding protection measures.

[0128] In a specific embodiment, the preset multi-level threshold judgment system sets a four-level arc hazard level classification standard, including observation level, early warning level, alarm level and emergency level; when the arc fault diagnosis result is lower than the preset observation level threshold, the system only records the fault information; when the arc fault diagnosis result reaches the early warning level but lower than the alarm level, a prompt message is generated and the local monitoring frequency is increased; when the arc fault diagnosis result reaches the alarm level but lower than the emergency level, an alarm signal is generated and a local load reduction control instruction is triggered; when the arc fault diagnosis result reaches the emergency level, a high priority alarm is generated and a tripping protection control instruction is triggered; the control instruction ensures execution reliability through encrypted transmission and redundant verification.

[0129] Specifically, this embodiment achieves comprehensive diagnosis and flexible response to arc faults through a multi-task learning framework. This embodiment adopts a multi-task learning framework with a hybrid architecture of "hard parameter sharing + soft parameter sharing". By organically combining a shared feature extraction backbone network with three dedicated branch networks, it achieves the coordinated optimization of the three tasks of fault type identification, severity classification, and sound source localization. By specifically designing different network structures in each branch (including attention mechanism classifiers, multi-label sequence models, and time difference analysis algorithms), the recognition accuracy of each task is improved.

[0130] This embodiment combines the four-level arc hazard classification standard (including observation, warning, alarm, and emergency) with corresponding differentiated response strategies to form a complete protection closed loop from fault monitoring and early warning to local load reduction and emergency tripping. At the same time, the reliability of control command execution is ensured through encrypted transmission and redundant verification mechanisms, thereby achieving precise hierarchical management and differentiated response to arc faults in the energy storage system, improving the safety protection level and emergency response capabilities of the energy storage system.

[0131] See also Figure 2 The present invention also provides an energy storage system arc fault diagnosis system based on voiceprint detection, the system comprising:

[0132] A voiceprint signal acquisition module is used to collect the voiceprint signal of the energy storage system in real time based on the acoustic sensor array, perform anti-interference preprocessing on the voiceprint signal, and obtain a time-frequency domain spectrogram;

[0133] a time series feature extraction module, configured to construct a multi-branch convolutional neural network, identify the time-frequency domain spectrogram based on the multi-branch convolutional neural network, obtain the time-frequency features of high-frequency spike pulses and low-frequency discharge whistles, and weight the time-frequency features of high-frequency spike pulses and low-frequency discharge whistles using a channel-by-channel frequency band weighting method to obtain a time series feature sequence;

[0134] A global feature capture module is used to build a temporal convolutional network model, input the temporal feature sequence into the temporal convolutional network model, and output global context features;

[0135] An arc fault characterization module is configured to dynamically weight pool the global context features through a multi-head temporal attention network to obtain a key frame feature vector representing the arc fault;

[0136] The fault diagnosis and alarm module is used to build an arc fault diagnosis and classification model, input the key frame feature vector into the arc fault diagnosis and classification model, obtain the arc fault type, arc fault severity level and sound source location information, and generate corresponding alarm signals and control instructions based on the arc fault severity level.

[0137] Specifically, the present embodiment provides an energy storage system arc fault diagnosis system based on voiceprint detection, which organically integrates five functional modules: voiceprint acquisition, feature extraction, global feature capture, fault characterization, and diagnostic alarm, to construct a complete closed-loop system from signal acquisition to protection execution. The present embodiment adopts a high-precision acoustic sensor array with electromagnetic shielding at the hardware level, and integrates a deep learning algorithm chain from multi-branch convolutional neural networks, temporal convolutional networks to multi-head temporal attention networks at the software level, forming a dedicated detection system for energy storage system arc faults. The system can automatically complete the entire process from voiceprint signal acquisition, anti-interference preprocessing, feature extraction, key frame recognition to final fault diagnosis and graded response, significantly improving the accuracy and reliability of arc fault detection.

[0138] The present invention also discloses an electronic device, comprising: at least one processor, at least one memory communication interface and a bus: wherein the processor, memory and communication interface communicate with each other through the bus; the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to implement an energy storage system arc fault diagnosis method based on voiceprint detection.

[0139] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for diagnosing arc faults in energy storage systems based on voiceprint detection, characterized in that: The following steps are involved: Acquiring the soundprint signal of the energy storage system in real time based on an acoustic sensor array, performing anti-interference preprocessing on the soundprint signal, and obtaining a time-frequency domain spectrogram; Constructing a multi-branch convolutional neural network, identifying the time-frequency domain spectrogram based on the multi-branch convolutional neural network, obtaining the time-frequency characteristics of high-frequency spike pulses and low-frequency discharge whistles, and weighting the time-frequency characteristics of high-frequency spike pulses and low-frequency discharge whistles using a channel-by-channel frequency band weighting method to obtain a time series feature sequence; Constructing a temporal convolutional network model, inputting the temporal feature sequence into the temporal convolutional network model, and outputting global context features; Dynamically weighted pooling the global context features is performed through a multi-head temporal attention network to obtain a key frame feature vector representing the arc fault; An arc fault diagnosis and classification model is constructed, and the key frame feature vector is input into the arc fault diagnosis and classification model to obtain the arc fault type, arc fault severity grade and sound source location information, and a corresponding alarm signal and control instruction are generated based on the arc fault severity grade.

2. The method for diagnosing arc faults in energy storage systems based on voiceprint detection according to claim 1, characterized in that: The method of collecting the voiceprint signal of the energy storage system in real time based on the acoustic sensor array and performing anti-interference preprocessing on the voiceprint signal to obtain a time-frequency domain spectrogram includes: Arrange acoustic sensor arrays with anti-electromagnetic interference capabilities on the energy storage system busbars and key connection points, and determine the number, type, and spatial layout of the sensors; Based on the acoustic sensor array, the soundprint signals during the operation of the energy storage system are collected in real time, and the collected soundprint signals are preprocessed with anti-interference to obtain a time-frequency domain spectrogram. The anti-interference preprocessing includes adaptive gain control and dual-passband notch filtering.

3. The method for diagnosing arc faults in an energy storage system based on voiceprint detection according to claim 1, characterized in that: The method comprises: constructing a multi-branch convolutional neural network, identifying the time-frequency domain spectrogram based on the multi-branch convolutional neural network, obtaining the time-frequency features of high-frequency spike pulses and low-frequency discharge whistles, and weighting the time-frequency features of high-frequency spike pulses and low-frequency discharge whistles using a channel-by-channel frequency band weighting method to obtain a time series feature sequence, including: Constructing a multi-branch convolutional neural network comprising multiple parallel branches, wherein each branch is provided with a convolution kernel structure of a different size, and extracting the high-frequency spike feature map and the low-frequency discharge whistling feature map in the time-frequency domain spectrogram in parallel; The feature maps extracted from each branch are subjected to the channel-by-channel frequency band weighting method, the importance weight of each frequency band is calculated and weighted fusion is performed to obtain the temporal feature sequence.

4. The method for diagnosing arc faults in an energy storage system based on voiceprint detection according to claim 3, characterized in that: The multi-branch convolutional neural network includes three parallel branches, wherein the first branch is used to capture high-frequency spike pulse features, the second branch is used to capture medium-frequency transition features, and the third branch is used to capture low-frequency discharge whistling features; each parallel branch includes two layers of convolution operations, wherein the first layer of convolution is followed by a ReLU activation function and a batch normalization layer, and the second layer of convolution is followed by a maximum pooling operation; The channel-by-channel frequency band weighting method adopts a compression-excitation network structure, compresses the spatial information of each feature channel through global average pooling, learns the dependency between channels through a two-layer fully connected network, outputs the frequency band weight coefficient of each channel, multiplies the frequency band weight coefficient with the original feature map for channel-by-channel weighting, and cascades the weighted feature maps of the three branches to obtain a comprehensive temporal feature sequence.

5. The method for diagnosing arc faults in an energy storage system based on voiceprint detection according to claim 1, characterized in that: The step of constructing a temporal convolutional network model, inputting the temporal feature sequence into the temporal convolutional network model, and outputting global context features includes: Constructing a temporal convolutional network model based on causal convolution and dilated convolution, wherein the temporal convolutional network model includes multiple layers of residual blocks, each of which has a dilated convolution layer with an exponentially increasing dilation rate. The temporal feature sequence is input into the temporal convolutional network model, and the temporal dependency features are extracted through the stacking of multiple layers of residual blocks and the skip connection structure to obtain the global context features covering the entire sequence time span.

6. The method for diagnosing arc faults in energy storage systems based on voiceprint detection according to claim 1, characterized in that: The global context features are dynamically weighted pooled by a multi-head temporal attention network to obtain a key frame feature vector representing the arc fault, including: Construct a multi-head temporal attention network to adaptively evaluate the importance of global context features at different time steps and build long-range temporal dependencies; Based on the multi-head temporal attention network, the attention weight of each time step is calculated, the global context features are dynamically weighted pooled, and the key frame feature vector representing the characteristics of the arc fault is extracted.

7. The method for diagnosing arc faults in an energy storage system based on voiceprint detection according to claim 6, characterized in that: The multi-head temporal attention network includes four parallel attention heads, each of which consists of a query-key-value triple structure; Each attention head generates a query matrix, a key matrix, and a value matrix through linear projection. The attention score is obtained by calculating the similarity between the query matrix and the key matrix. The attention score is normalized using the temperature-scaled Softmax function to obtain the attention weight. The output of each attention head is obtained through the attention weight value matrix. The outputs of each attention head are spliced and linearly transformed to obtain the multi-head fusion attention weight. The global context features are dynamically weighted and summed using the multi-head fusion attention weight to obtain a key frame feature vector of fixed dimension.

8. The method for diagnosing arc faults in an energy storage system based on voiceprint detection according to claim 1, characterized in that: The arc fault diagnosis classification model is constructed, the key frame feature vector is input into the arc fault diagnosis classification model, the arc fault type, the arc fault severity level and the sound source location information are obtained, and the corresponding alarm signal and control instruction are generated based on the arc fault severity level, including: Constructing an arc fault diagnosis and classification model based on a multi-task learning framework, inputting the key frame feature vector into the arc fault diagnosis and classification model, and simultaneously performing fault type identification, severity classification, and sound source location by sharing underlying feature representation to obtain arc fault type, arc fault severity classification, and sound source location information; Based on the fault diagnosis results output by the arc fault diagnosis classification model, the danger level of the arc fault is determined according to the preset multi-level threshold value, and a graded alarm signal and corresponding control instructions are generated. The control instructions are sent to the energy storage system management system through a secure communication channel to trigger the corresponding protection measures.

9. An energy storage system arc fault diagnosis system based on voiceprint detection, used to execute an energy storage system arc fault diagnosis method based on voiceprint detection according to any one of claims 1 to 8, characterized in that: The system comprises: A voiceprint signal acquisition module is used to collect the voiceprint signal of the energy storage system in real time based on the acoustic sensor array, perform anti-interference preprocessing on the voiceprint signal, and obtain a time-frequency domain spectrogram; a time series feature extraction module, configured to construct a multi-branch convolutional neural network, identify the time-frequency domain spectrogram based on the multi-branch convolutional neural network, obtain the time-frequency features of high-frequency spike pulses and low-frequency discharge whistles, and weight the time-frequency features of high-frequency spike pulses and low-frequency discharge whistles using a channel-by-channel frequency band weighting method to obtain a time series feature sequence; A global feature capture module is used to build a temporal convolutional network model, input the temporal feature sequence into the temporal convolutional network model, and output global context features; An arc fault characterization module is configured to dynamically weight pool the global context features through a multi-head temporal attention network to obtain a key frame feature vector representing the arc fault; The fault diagnosis and alarm module is used to build an arc fault diagnosis and classification model, input the key frame feature vector into the arc fault diagnosis and classification model, obtain the arc fault type, arc fault severity level and sound source location information, and generate corresponding alarm signals and control instructions based on the arc fault severity level.

10. An electronic device, characterized in that: include: at least one processor, at least one memory, a communication interface, and a bus; The processor, memory, and communication interface communicate with each other via the bus, the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to implement the method according to any one of claims 1 to 8.

Images