Cable partial discharge mode identification method and device based on multi-mode fusion, electronic equipment and storage medium

Abstract

The invention discloses a cable partial discharge mode identification method and device based on multi-mode fusion, electronic equipment and a storage medium, and belongs to the technical field of cable partial discharge online monitoring. And converting the original signal into a multi-modal spectrum through phase resolution analysis, wavelet packet energy analysis, skewness and kurtosis statistics and recurrence plot analysis. The group of maps are input into a multi-modal fusion recognition model, the model analyzes the spatial distribution characteristics and the time sequence characteristics of the maps at the same time through parallel vision and time sequence coding modules, and two analysis results are subjected to weighted fusion, so that final classification recognition of the discharge mode is realized. Through the implementation of the method and the device, the problems that key information is lost due to dependence on single-dimensional feature representation and the identification accuracy is limited due to the fact that an existing identification model cannot analyze the spatial distribution characteristic and the time sequence evolution characteristic of the signal at the same time in the prior art can be solved.

Description

Technical Field

[0001] This invention relates to the technical field of cable partial discharge online monitoring, specifically to a cable partial discharge pattern recognition method, device, electronic device, and storage medium based on multimodal fusion. Background Technology

[0002] As a crucial component of modern power systems, the health of the insulation of power cables directly affects the safety and stable operation of the entire power grid. Partial discharge (PD) is a key physical phenomenon in the early deterioration of cable insulation and is widely recognized as one of the most effective indicators for predicting insulation faults. Therefore, acquiring partial discharge signals through online monitoring and accurately identifying signal patterns to determine the specific type of defect is of vital practical significance for achieving predictive maintenance of power equipment, avoiding sudden power outages, and reducing economic losses.

[0003] However, existing methods for identifying partial discharge patterns in cables still have certain limitations. On the one hand, many methods rely solely on a single-dimensional analysis and presentation of the original signal during the feature extraction stage. This approach essentially involves a "dimensionality reduction" observation of the complex, high-dimensional original signal, inevitably losing key feature information that may exist in other dimensions. On the other hand, partial discharge signals themselves possess both complex spatial distribution characteristics (such as the aggregation pattern of discharge pulses in the phase-amplitude two-dimensional plane) and temporal evolution characteristics (such as the evolution of discharge pulses over time). Existing identification model architectures often excel at handling only one of these characteristics, resulting in an insufficient and incomplete understanding of the signal's intrinsic physical properties, fundamentally limiting further improvements in identification accuracy. Summary of the Invention

[0004] This invention provides a cable partial discharge pattern recognition method, device, electronic device, and storage medium based on multimodal fusion, which can solve the problems of key information loss caused by relying on single-dimensional feature representation in the prior art, and the limited recognition accuracy caused by the inability of existing recognition models to simultaneously and effectively analyze the spatial distribution characteristics and temporal evolution characteristics of signals due to their architectural limitations.

[0005] One embodiment of the present invention provides a cable partial discharge pattern recognition method based on multimodal fusion, comprising:

[0006] Acquire the raw electrical signal of partial discharge from the cable to be identified;

[0007] Phase-resolved partial discharge analysis is performed on the raw partial discharge electrical signal to generate a phase-resolved partial discharge map.

[0008] Wavelet packet energy analysis is performed on the original partial discharge electrical signal to generate a wavelet packet energy map.

[0009] The skewness and kurtosis statistical characteristics of the original partial discharge electrical signal are extracted, and a two-dimensional skewness-kurtosis distribution map is generated based on the skewness and kurtosis statistical characteristics.

[0010] Recurrence graph analysis was performed on the original electrical signal of partial discharge to obtain a time series recurrence graph;

[0011] The phase-resolved partial discharge map, the wavelet packet energy map, the skewness-kurtosis two-dimensional distribution map, and the time series recursive map are combined into a multimodal map.

[0012] The multimodal map is input into a preset multimodal fusion recognition model, so that the multimodal fusion recognition model can perform spatial feature analysis on the multimodal map through the built-in visual feature encoding module and generate visual recognition results.

[0013] The built-in time-series feature encoding module performs time-series feature analysis on the multimodal graph to generate time-series recognition results.

[0014] The built-in weighted fusion module performs weighted fusion on the visual recognition results and the temporal recognition results to generate a comprehensive fusion result.

[0015] The built-in classification and prediction module predicts the discharge type based on the comprehensive fusion results, generating cable partial discharge pattern recognition results.

[0016] Furthermore, after acquiring the original partial discharge electrical signal of the cable to be identified, the process further includes:

[0017] Add white noise of a preset intensity to the original partial discharge electrical signal to generate a set of disturbance signals;

[0018] The set of disturbance signals is subjected to CEEMDAN decomposition to generate a set of final intrinsic mode function components;

[0019] Calculate the sample entropy of each component in the final set of intrinsic mode function components, and generate the sample entropy value corresponding to each component;

[0020] The final set of intrinsic mode function components are sorted according to the sample entropy value, and a preset number of intrinsic mode function components are selected as a set of filtered intrinsic mode function components.

[0021] The selected intrinsic mode function components are subjected to signal reconstruction processing to generate a denoised PD signal;

[0022] The denoised PD signal is used to replace the original partial discharge electrical signal.

[0023] Furthermore, the training of the multimodal fusion recognition model includes representation pre-training and fine-tuning training;

[0024] The representation pre-training includes:

[0025] A first training dataset containing several historical partial discharge signals of cables is obtained, and a set of unlabeled multimodal maps are generated based on the historical partial discharge signals of cables in the first training dataset.

[0026] Multiple enhanced views are generated by performing at least two different random augmentation processes on each of the set of unlabeled multimodal maps.

[0027] From the multiple augmented views corresponding to each map, any two augmented views originating from the same map are defined as positive sample pairs, and any two augmented views originating from different maps are defined as negative sample pairs.

[0028] According to the preset batch size, the generated positive sample pairs and negative sample pairs are divided into several batches of first training samples;

[0029] The first training samples of each batch are sequentially input into the visual feature encoding module, so that when the visual feature encoding module receives the first training samples of each batch, it extracts features from the positive and negative sample pairs in the current batch to generate the corresponding first feature vector; based on the preset contrast loss function, the first loss value is calculated according to the first feature vector; the learnable parameters in the visual feature encoding module are updated according to the first loss value using a preset optimizer until the preset number of training rounds is reached.

[0030] The first training samples of each batch are sequentially input into the temporal feature encoding module, so that when the temporal feature encoding module receives the first training samples of each batch, it extracts features from the positive and negative sample pairs in the current batch to generate the corresponding second feature vector; based on the preset contrast loss function, the second loss value is calculated according to the second feature vector; the learnable parameters in the temporal feature encoding module are updated according to the second loss value using a preset optimizer until the preset number of training rounds is reached.

[0031] Furthermore, the fine-tuning training is performed based on the completion of the representation pre-training;

[0032] The fine-tuning training includes:

[0033] A second training dataset containing several historical partial discharge signals of cables and their corresponding real discharge type labels is obtained, and a set of labeled multimodal maps is generated based on the historical partial discharge signals of cables in the second training dataset.

[0034] According to the preset batch size, the labeled multimodal map is randomly divided into several batches of second training samples;

[0035] The second training samples of each batch are sequentially input into the multimodal fusion recognition model, and the multimodal fusion recognition model is subjected to end-to-end iterative training until the preset number of training rounds is reached.

[0036] The multimodal fusion recognition model, upon receiving a batch of the second training samples, performs the following:

[0037] The visual feature encoding module performs spatial feature analysis on the map samples in the second training sample of the current batch to generate visual recognition results during the training process.

[0038] The temporal feature encoding module performs time series feature analysis on the spectral samples in the second training sample of the current batch to generate temporal recognition results during the training process.

[0039] The weighted fusion module performs weighted fusion of the visual recognition results and the temporal recognition results during the training process to generate a comprehensive fusion result during the training process.

[0040] The classification prediction module predicts the discharge type based on the comprehensive fusion results during the training process, and generates the cable partial discharge pattern recognition result during the training process corresponding to the current second training sample.

[0041] The third loss function value is calculated and generated based on the cable partial discharge pattern recognition results and the corresponding real discharge type labels during the training process using a preset loss function.

[0042] Using a preset optimizer, gradient backpropagation and updates are performed on the learnable parameters in the visual feature encoding module, the temporal feature encoding module, the weighted fusion module, and the classification prediction module based on the value of the third loss function, in order to complete fine-tuning training.

[0043] Furthermore, the contrastive loss function is specifically as follows:

[0044]

[0045] In the formula, l i，j z is the loss value calculated for a positive sample pair (i, j); i and j are the indices of the two augmented views in the positive sample pair; i With z j is the feature vector corresponding to the two enhanced views in the positive sample pair; sim(·,·) is the cosine similarity function; τ is the temperature hyperparameter; N is the batch size; 2N is the total number of views in the first training sample of the current batch; k is the index of other views in the current batch.

[0046] Furthermore, the loss function is specifically as follows:

[0047]

[0048] In the formula, L is the loss value of the current second sample; M is the total number of discharge type categories; c is the category index; y c p is a sign function that has a value of 1 when class c is the true label of the sample, and 0 otherwise; c The predicted probability of the multimodal fusion recognition model for category c is given.

[0049] Furthermore, the visual feature encoding module is a Vision Transformer-based encoder; the temporal feature encoding module is a temporal Transformer-based encoder; and the classification prediction module is a multilayer perceptron classifier.

[0050] Based on the above method embodiments, the present invention provides corresponding apparatus embodiments.

[0051] One embodiment of the present invention provides a cable partial discharge pattern recognition device based on multimodal fusion, comprising: a partial discharge raw electrical signal acquisition module, a multimodal spectrum generation module, and a cable partial discharge pattern recognition module;

[0052] The partial discharge raw electrical signal acquisition module is used to acquire the partial discharge raw electrical signal of the cable to be identified;

[0053] The multimodal spectrum generation module is used to perform phase-resolved partial discharge analysis on the original partial discharge electrical signal to generate a phase-resolved partial discharge spectrum; perform wavelet packet energy analysis on the original partial discharge electrical signal to generate a wavelet packet energy map; extract the skewness and kurtosis statistical features of the original partial discharge electrical signal and generate a skewness-kurtosis two-dimensional distribution map based on the skewness and kurtosis statistical features; perform recursive graph analysis on the original partial discharge electrical signal to obtain a time series recursive graph; and combine the phase-resolved partial discharge spectrum, the wavelet packet energy map, the skewness-kurtosis two-dimensional distribution map, and the time series recursive graph into a multimodal spectrum.

[0054] The cable partial discharge pattern recognition module is used to input the multimodal spectrum into a preset multimodal fusion recognition model, so that the multimodal fusion recognition model can perform spatial feature analysis on the multimodal spectrum through a built-in visual feature encoding module to generate a visual recognition result; perform time series feature analysis on the multimodal spectrum through a built-in temporal feature encoding module to generate a temporal recognition result; perform weighted fusion of the visual recognition result and the temporal recognition result through a built-in weighted fusion module to generate a comprehensive fusion result; and perform discharge type prediction based on the comprehensive fusion result through a built-in classification prediction module to generate a cable partial discharge pattern recognition result.

[0055] Based on the above method embodiments, the present invention provides corresponding electronic device embodiments.

[0056] An embodiment of the present invention provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the cable partial discharge pattern recognition method based on multimodal fusion as described in any of the above-described method embodiments.

[0057] Based on the above method embodiments, the present invention provides corresponding storage medium embodiments.

[0058] One embodiment of the present invention provides a storage medium storing a computer program thereon, wherein, when the computer program is running, it controls the device where the storage medium is located to execute any of the cable partial discharge pattern recognition methods based on multimodal fusion described in the above-described method embodiments.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] This invention provides a method, apparatus, electronic device, and storage medium for cable partial discharge pattern recognition based on multimodal fusion. The method...

[0061] This invention creatively transforms a single partial discharge signal into multiple graphical representations, such as phase-resolved spectra and wavelet packet energy maps, and combines them into an information-rich multimodal spectrum. This fundamentally solves the problem of incomplete information caused by existing technologies relying on single feature representations. Furthermore, this invention utilizes parallel visual and temporal feature encoding modules to perform in-depth analysis and fusion of the spatial distribution characteristics and temporal evolution patterns of the multimodal spectrum, effectively overcoming the technical deficiency of existing models that cannot simultaneously consider both characteristics, resulting in limited recognition capabilities. Attached Figure Description

[0062] Figure 1This is a flowchart illustrating a cable partial discharge pattern recognition method based on multimodal fusion, provided by an embodiment of the present invention.

[0063] Figure 2 This is a schematic diagram of a cable partial discharge pattern recognition device based on multimodal fusion provided in an embodiment of the present invention. Detailed Implementation

[0064] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] like Figure 1 As shown, to address the problems of key information loss due to reliance on single-dimensional feature representation in existing technologies, and the limited recognition accuracy caused by the inability of existing recognition models to simultaneously and effectively analyze the spatial distribution characteristics and temporal evolution characteristics of signals due to their architectural limitations, an embodiment of the present invention provides a cable partial discharge pattern recognition method based on multimodal fusion, which includes at least the following steps:

[0066] Step S1: Obtain the original electrical signal of partial discharge of the cable to be identified;

[0067] In a preferred embodiment, after acquiring the original partial discharge electrical signal of the cable to be identified, the method further includes:

[0068] Add white noise of a preset intensity to the original partial discharge electrical signal to generate a set of disturbance signals;

[0069] The set of disturbance signals is subjected to CEEMDAN decomposition to generate a set of final intrinsic mode function components;

[0070] Calculate the sample entropy of each component in the final set of intrinsic mode function components, and generate the sample entropy value corresponding to each component;

[0071] The final set of intrinsic mode function components are sorted according to the sample entropy value, and a preset number of intrinsic mode function components are selected as a set of filtered intrinsic mode function components.

[0072] The selected intrinsic mode function components are subjected to signal reconstruction processing to generate a denoised PD signal;

[0073] The denoised PD signal is used to replace the original partial discharge electrical signal.

[0074] Specifically, the raw partial discharge electrical signal is a time-series signal collected by an online monitoring system deployed on power cables. This raw signal typically contains real partial discharge pulse signals as well as white noise or narrowband interference signals from various sources, resulting in a low signal-to-noise ratio. Preprocessing is required to extract effective feature information.

[0075] In a preferred embodiment, to improve signal quality, after acquiring the original partial discharge electrical signal of the cable to be identified, a series of signal denoising preprocessing steps are included. This preprocessing aims to effectively separate the real partial discharge signal from background noise. Specifically, white noise with a root mean square (RMS) value of 0.3 is actively injected into the acquired original partial discharge electrical signal. The introduction of white noise can assist the subsequent decomposition algorithm in overcoming the mode mixing problem, thereby generating a set of perturbation signals for analysis.

[0076] Next, the perturbation signal is processed using CEEMDAN (Completely Adaptive Empirical Mode Decomposition of Noise Ensemble) with a decomposition level of 8. CEEMDAN is an adaptive decomposition method for non-stationary signals that decomposes the complex original signal x(t) into a finite number of intrinsic mode function (IMF) components c arranged from high to low frequencies. i The original signal contains a component r(t) and a residual component r(t). Each IMF component represents a vibrational mode with specific physical meaning in the original signal. The decomposition process can be expressed by the following equation:

[0077]

[0078] In the formula, x(t) is the original electrical signal of partial discharge; c i r(t) represents the i-th intrinsic mode function (IMF) component after decomposition; r(t) represents the residual component; and q represents the total number of decomposed intrinsic mode function components.

[0079] After obtaining a final set of intrinsic mode function (IMF) components, these components need to be filtered to distinguish which mainly contain real signals and which mainly contain noise. This embodiment uses sample entropy to measure the complexity of each IMF component. Generally, IMF components dominated by real discharge signals have higher complexity, while IMF components dominated by noise exhibit stronger randomness and lower complexity. The formula for calculating sample entropy reflects the probability of generating new patterns in the time series, and its specific calculation method is as follows:

[0080]

[0081] In the formula, S Edenoted as sample entropy; m is the embedding dimension; ∈ is the similarity tolerance; N is the time series length; A is the number of vector pairs with a distance less than ∈ in the m+1 dimension; B is the number of vector pairs with a distance less than ∈ in the m dimension.

[0082] After calculating the sample entropy value for each component in the final set of IMF components, all IMF components are sorted in descending order based on the calculated sample entropy values. Since the IMF components containing more useful information have higher entropy values, the top 5 IMF components are selected as a set of filtered intrinsic mode function components. Finally, this set of filtered IMF components containing the main signal information is linearly superimposed for signal reconstruction, thereby generating a denoised PD signal with a significantly improved signal-to-noise ratio. This reconstructed signal will replace the original partial discharge signal and be used in all subsequent spectrum generation and feature analysis steps.

[0083] Through the above series of preprocessing steps, the real partial discharge signal can be effectively separated and reconstructed from the noisy original signal, which significantly improves the purity and quality of the signal and lays a solid foundation for subsequent high-precision pattern recognition.

[0084] Step S2: Perform phase-resolved partial discharge analysis on the original partial discharge electrical signal to generate a phase-resolved partial discharge spectrum; perform wavelet packet energy analysis on the original partial discharge electrical signal to generate a wavelet packet energy map; extract the skewness and kurtosis statistical features of the original partial discharge electrical signal, and generate a skewness-kurtosis two-dimensional distribution map based on the skewness and kurtosis statistical features; perform recursive graph analysis on the original partial discharge electrical signal to obtain a time series recursive graph; combine the phase-resolved partial discharge spectrum, the wavelet packet energy map, the skewness-kurtosis two-dimensional distribution map, and the time series recursive graph into a multimodal spectrum;

[0085] Specifically, after acquiring and denoising the partial discharge electrical signal, this invention continues to perform multi-dimensional feature extraction and graphical representation of the signal. Specifically, this embodiment uses four different analysis methods to transform the one-dimensional time-series signal into a set of two-dimensional maps that can more comprehensively reflect its intrinsic physical characteristics.

[0086] First, phase-resolved partial discharge (PRPD) analysis is performed on the partial discharge signal to generate a phase-resolved partial discharge spectrum. PRPD analysis is a statistical method that correlates the amplitude, phase, and frequency of discharge pulses. In its implementation, a power frequency cycle (typically 360°) is divided into multiple phase panes, and the amplitude range of the discharge signal is quantized into multiple amplitude panes, thus forming a two-dimensional matrix. Then, each detected discharge pulse is accumulated into the corresponding cell of this two-dimensional matrix based on its phase angle and corresponding amplitude. After statistical analysis of a large number of pulse signals, the final spectrum can intuitively show the distribution pattern and concentration of discharge pulses at different phases, which is an important basis for determining the discharge type.

[0087] Simultaneously, wavelet packet energy analysis is performed on the partial discharge electrical signal to generate a wavelet packet energy map. Wavelet packet analysis is a signal processing technique capable of finely dividing a signal across its entire frequency band. In its implementation, the signal undergoes multi-level wavelet packet decomposition, decomposing it into multiple preset frequency bands to obtain the wavelet packet decomposition coefficients corresponding to each frequency band. Then, the energy of the wavelet packet decomposition coefficients within each frequency band is calculated, thereby obtaining an energy vector that reflects the distribution of signal energy across different frequency bands. The wavelet packet energy E of the j-th layer and k-th frequency band is... j，k The calculation formula is:

[0088]

[0089] In the formula, E j,k d represents the wavelet packet energy of the k-th frequency band in the j-th layer; j,k (p) represents the wavelet packet decomposition coefficient of the signal in the k-th frequency band of the j-th layer; P represents the total number of coefficients in this frequency band.

[0090] By combining the energy values ​​of all frequency bands, a wavelet packet energy map reflecting the energy-frequency distribution characteristics of the signal can be generated.

[0091] Furthermore, the skewness and kurtosis statistical characteristics of the partial discharge electrical signal were extracted, and a skewness-kurtosis two-dimensional distribution map was generated based on these characteristics. Skewness and kurtosis are high-order statistical measures describing the distribution shape of data. Skewness measures the asymmetry of the signal amplitude distribution, while kurtosis measures the steepness of the signal amplitude distribution. Skewness S k With kurtosis K u The calculation formula is:

[0092]

[0093] In the formula, S k K represents the skewness value. u kurtosis value; x a This represents the value of the a-th sampling point in the signal sequence; is the average value of the signal sequence; n is the total number of sampling points in the signal sequence.

[0094] By calculating the skewness and kurtosis values ​​of the signal within multiple time windows and plotting these (skewness, kurtosis) values ​​as points in a two-dimensional coordinate system, a skewness-kurtosis two-dimensional distribution map can be generated. This map can reflect the changing patterns of the signal's statistical characteristics.

[0095] Finally, recursive graph analysis was performed on the partial discharge signal to obtain a time series recursive graph. A recursive graph is a method for visualizing the dynamic behavior of a time series in a high-dimensional phase space. In its implementation, the one-dimensional time series is first reconstructed in phase space to obtain a set of high-dimensional vectors. Then, a recursive matrix is ​​constructed by calculating whether the distance between any two vectors is less than a preset threshold. This recursive matrix R... f,g The calculation formula is:

[0096] R f，g =Θ(δ-||v f -v g ||)

[0097] In the formula, R f,g v is the value of the element in the f-th row and g-th column of the recursive matrix; f v is the f-th state vector in phase space; g Let g be the g-th state vector in phase space; δ be the preset distance threshold; ||·|| be the vector norm or distance calculation; Θ(·) be the Heaviside step function, which is 1 when the independent variable is greater than or equal to 0, and 0 otherwise.

[0098] By representing the 1s and 0s in the recurrence matrix with different colors, a time series recurrence graph can be generated. The texture and structure of this graph can reveal the nonlinear dynamic characteristics such as periodicity and chaos in the signal.

[0099] After generating the four types of maps mentioned above, the phase-resolved partial discharge map, the wavelet packet energy map, the skewness-kurtosis two-dimensional distribution map, and the time series recursive map are combined to form an information-rich multimodal map. This combination process is not a physical splicing, but rather treats these four maps, which describe the same signal from different dimensions and perspectives, as a whole data set, providing input for the subsequent fusion recognition model.

[0100] Through the above steps, the original one-dimensional electrical signal is creatively transformed into a set of multi-dimensional, multi-modal graphical features that can comprehensively reflect its statistical characteristics, time-frequency characteristics, phase characteristics, and nonlinear dynamic characteristics. This greatly enriches the amount of information that can be used for pattern recognition and provides a solid data foundation for achieving higher-precision recognition.

[0101] Step S3: Input the multimodal graph into a preset multimodal fusion recognition model, so that the multimodal fusion recognition model can perform spatial feature analysis on the multimodal graph through the built-in visual feature encoding module to generate visual recognition results; perform time series feature analysis on the multimodal graph through the built-in temporal feature encoding module to generate temporal recognition results; perform weighted fusion on the visual recognition results and temporal recognition results through the built-in weighted fusion module to generate comprehensive fusion results; and perform discharge type prediction based on the comprehensive fusion results through the built-in classification prediction module to generate cable partial discharge pattern recognition results.

[0102] In a preferred embodiment, training the multimodal fusion recognition model includes representation pre-training and fine-tuning training;

[0103] The representation pre-training includes:

[0104] A first training dataset containing several historical partial discharge signals of cables is obtained, and a set of unlabeled multimodal maps are generated based on the historical partial discharge signals of cables in the first training dataset.

[0105] Multiple enhanced views are generated by performing at least two different random augmentation processes on each of the set of unlabeled multimodal maps.

[0106] From the multiple augmented views corresponding to each map, any two augmented views originating from the same map are defined as positive sample pairs, and any two augmented views originating from different maps are defined as negative sample pairs.

[0107] According to the preset batch size, the generated positive sample pairs and negative sample pairs are divided into several batches of first training samples;

[0108] The first training samples of each batch are sequentially input into the visual feature encoding module, so that when the visual feature encoding module receives the first training samples of each batch, it extracts features from the positive and negative sample pairs in the current batch to generate the corresponding first feature vector; based on the preset contrast loss function, the first loss value is calculated according to the first feature vector; the learnable parameters in the visual feature encoding module are updated according to the first loss value using a preset optimizer until the preset number of training rounds is reached.

[0109] The first training samples of each batch are sequentially input into the temporal feature encoding module, so that when the temporal feature encoding module receives the first training samples of each batch, it extracts features from the positive and negative sample pairs in the current batch to generate the corresponding second feature vector; based on the preset contrast loss function, the second loss value is calculated according to the second feature vector; the learnable parameters in the temporal feature encoding module are updated according to the second loss value using a preset optimizer until the preset number of training rounds is reached.

[0110] In a preferred embodiment, the fine-tuning training is performed after the representation pre-training is completed;

[0111] The fine-tuning training includes:

[0112] A second training dataset containing several historical partial discharge signals of cables and their corresponding real discharge type labels is obtained, and a set of labeled multimodal maps is generated based on the historical partial discharge signals of cables in the second training dataset.

[0113] According to the preset batch size, the labeled multimodal map is randomly divided into several batches of second training samples;

[0114] The second training samples of each batch are sequentially input into the multimodal fusion recognition model, and the multimodal fusion recognition model is subjected to end-to-end iterative training until the preset number of training rounds is reached.

[0115] The multimodal fusion recognition model, upon receiving a batch of the second training samples, performs the following:

[0116] The visual feature encoding module performs spatial feature analysis on the map samples in the second training sample of the current batch to generate visual recognition results during the training process.

[0117] The temporal feature encoding module performs time series feature analysis on the spectral samples in the second training sample of the current batch to generate temporal recognition results during the training process.

[0118] The weighted fusion module performs weighted fusion of the visual recognition results and the temporal recognition results during the training process to generate a comprehensive fusion result during the training process.

[0119] The classification prediction module predicts the discharge type based on the comprehensive fusion results during the training process, and generates the cable partial discharge pattern recognition result during the training process corresponding to the current second training sample.

[0120] The third loss function value is calculated and generated based on the cable partial discharge pattern recognition results and the corresponding real discharge type labels during the training process using a preset loss function.

[0121] Using a preset optimizer, gradient backpropagation and updates are performed on the learnable parameters in the visual feature encoding module, the temporal feature encoding module, the weighted fusion module, and the classification prediction module based on the value of the third loss function, in order to complete fine-tuning training.

[0122] In a preferred embodiment, the contrastive loss function is specifically:

[0123]

[0124] In the formula, l i，j z is the loss value calculated for a positive sample pair (i, j); i and j are the indices of the two augmented views in the positive sample pair; i With z j is the feature vector corresponding to the two enhanced views in the positive sample pair; sim(·,·) is the cosine similarity function; τ is the temperature hyperparameter; N is the batch size; 2N is the total number of views in the first training sample of the current batch; k is the index of other views in the current batch.

[0125] In a preferred embodiment, the loss function is specifically:

[0126]

[0127] In the formula, L is the loss value of the current second sample; M is the total number of discharge type categories; c is the category index; y c p is a sign function that has a value of 1 when class c is the true label of the sample, and 0 otherwise; c The predicted probability of the multimodal fusion recognition model for category c is given.

[0128] In a preferred embodiment, the visual feature encoding module is a Vision Transformer-based encoder; the temporal feature encoding module is a temporal Transformer-based encoder; and the classification prediction module is a multilayer perceptron classifier.

[0129] Specifically, after the multimodal atlas is generated, the core recognition process of this invention is initiated. In this step, the multimodal atlas generated as a whole dataset is input into a pre-trained multimodal fusion recognition model. This model employs a parallel dual-channel architecture, designed to simultaneously perform deep feature extraction and analysis on the input atlas information from both spatial and temporal dimensions.

[0130] Specifically, the model incorporates a visual feature encoding module and a temporal feature encoding module. In a preferred embodiment, the visual feature encoding module is a Vision Transformer-based encoder, which processes the input multimodal maps as two-dimensional images, focusing on analyzing spatial features such as the aggregation morphology and distribution patterns of discharge points in the maps to generate visual recognition results. Simultaneously, the temporal feature encoding module is a temporal Transformer-based encoder, which treats the same set of multimodal maps as a time series, focusing on mining the temporal dependencies and evolution patterns of data points between and within the maps to generate temporal recognition results.

[0131] After the two encoding modules output their respective analysis results in parallel, the model intelligently weights and fuses the visual recognition results carrying spatial information and the temporal recognition results carrying temporal information through a built-in weighted fusion module to generate a comprehensive fusion result that combines features from both dimensions. Finally, this comprehensive fusion result is fed into the built-in classification prediction module. In a preferred embodiment, the classification prediction module is a multilayer perceptron (MLP) classifier, which performs the final discharge type prediction based on the comprehensive fusion result and outputs a deterministic cable partial discharge pattern recognition result.

[0132] In a preferred embodiment, the acquisition of the above-mentioned multimodal fusion recognition model includes two core stages: representation pre-training and fine-tuning training.

[0133] First, representation pre-training is performed. The purpose of this stage is to utilize unlabeled data and, through self-supervised learning, enable the visual feature encoding module and the temporal feature encoding module to fully learn the inherent data structure and deep features of cable partial discharge patterns. This process includes: acquiring a first training dataset containing several historical partial discharge signals from cables, and generating a set of unlabeled multimodal patterns based on the historical signals in this dataset through the aforementioned four graphical transformation processes. Next, at least two different random augmentation processes are performed on each pattern in this set of unlabeled multimodal patterns to generate multiple augmented views. Then, from the multiple augmented views corresponding to each pattern, any two augmented views originating from the same pattern are defined as positive sample pairs, and any two augmented views originating from different patterns are defined as negative sample pairs. After dividing the generated positive and negative sample pairs into several batches of first training samples, the two encoding modules are independently iteratively trained.

[0134] Training the visual feature encoding module involves sequentially inputting the first training samples from each batch, allowing it to extract features from positive and negative sample pairs to generate corresponding first feature vectors. A first loss value is then calculated based on a preset first contrastive loss function, and a preset optimizer updates the learnable parameters in the visual feature encoding module according to this loss value. Similarly, training the temporal feature encoding module involves inputting the first training samples from the same batch to generate corresponding second feature vectors, calculating a second loss value based on a preset second contrastive loss function, and then updating the parameters of the temporal feature encoding module using an optimizer. Training for both modules continues until a preset number of training epochs is reached.

[0135] In a preferred embodiment, the contrastive loss function is a function designed to bring the feature vectors of positive sample pairs closer together and push the feature vectors of negative sample pairs further apart. The loss function is calculated as follows: for a positive sample pair, the similarity of its feature vectors is compared and normalized with the similarity of the feature vectors of all corresponding negative sample pairs within the batch. The loss value is calculated by maximizing the similarity ratio of positive sample pairs, and this loss value is used to guide the optimization of model parameters.

[0136] Based on the completed pre-training of the representation, fine-tuning training continues. The purpose of this stage is to use labeled data to perform end-to-end supervised training on the entire multimodal fusion recognition model, enabling it to learn the specific classification task. This process includes: acquiring a second training dataset containing several historical partial discharge signals of cables and their corresponding real discharge type labels, and generating a set of labeled multimodal maps. After dividing this set of labeled maps into several batches of second training samples, they are sequentially input into the entire multimodal fusion recognition model for end-to-end iterative training. In each iteration, all modules of the model work together to perform a complete forward propagation of the input map samples until the classification prediction module outputs a predicted discharge type.

[0137] In a preferred embodiment, the loss function used for fine-tuning training is a function used to evaluate the difference between the model's prediction results and the true labels. This loss function is calculated by comparing the model's predicted probabilities for each discharge type with the one-hot encoded labels corresponding to the samples, representing the true discharge types, to calculate the cross-entropy between the two, and using this cross-entropy as the loss value.

[0138] Finally, based on the loss value calculated by the loss function, the learnable parameters of all modules in the model (including the visual feature encoding module, the temporal feature encoding module, the weighted fusion module, and the classification prediction module) are uniformly backpropagated and updated using a preset optimizer until the preset number of training rounds is reached, so as to complete the fine-tuning training of the entire model.

[0139] Through the construction, pre-training, and fine-tuning of the multimodal fusion recognition model described above, this invention can perform comprehensive and in-depth analysis of the input partial discharge signal from multiple dimensions, fully explore the spatial and temporal features contained in the signal, and thus significantly improve the recognition accuracy and reliability of complex cable partial discharge modes.

[0140] Based on the above method embodiments, the present invention provides corresponding apparatus embodiments.

[0141] like Figure 2 As shown, an embodiment of the present invention provides a cable partial discharge pattern recognition device based on multimodal fusion, including: a partial discharge raw electrical signal acquisition module, a multimodal map generation module, and a cable partial discharge pattern recognition module;

[0142] The partial discharge raw electrical signal acquisition module is used to acquire the partial discharge raw electrical signal of the cable to be identified;

[0143] The multimodal spectrum generation module is used to perform phase-resolved partial discharge analysis on the original partial discharge electrical signal to generate a phase-resolved partial discharge spectrum; perform wavelet packet energy analysis on the original partial discharge electrical signal to generate a wavelet packet energy map; extract the skewness and kurtosis statistical features of the original partial discharge electrical signal and generate a skewness-kurtosis two-dimensional distribution map based on the skewness and kurtosis statistical features; perform recursive graph analysis on the original partial discharge electrical signal to obtain a time series recursive graph; and combine the phase-resolved partial discharge spectrum, the wavelet packet energy map, the skewness-kurtosis two-dimensional distribution map, and the time series recursive graph into a multimodal spectrum.

[0144] The cable partial discharge pattern recognition module is used to input the multimodal spectrum into a preset multimodal fusion recognition model, so that the multimodal fusion recognition model can perform spatial feature analysis on the multimodal spectrum through a built-in visual feature encoding module to generate a visual recognition result; perform time series feature analysis on the multimodal spectrum through a built-in temporal feature encoding module to generate a temporal recognition result; perform weighted fusion of the visual recognition result and the temporal recognition result through a built-in weighted fusion module to generate a comprehensive fusion result; and perform discharge type prediction based on the comprehensive fusion result through a built-in classification prediction module to generate a cable partial discharge pattern recognition result.

[0145] It should be noted that the embodiments of the device described above correspond to the embodiments of the present invention described above, and can realize the cable partial discharge pattern recognition method based on multimodal fusion described in any one of the present invention. Furthermore, the embodiments of the device described above are merely illustrative. The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the accompanying drawings of the device embodiments provided by the present invention, the connection relationship between modules indicates that they have a communication connection, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without creative effort.

[0146] Based on the above-described method embodiments of the present invention, a corresponding embodiment of an electronic device is provided.

[0147] An embodiment of the present invention provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the cable partial discharge pattern recognition method based on multimodal fusion as described in any one of the present invention, or the processor executes the computer program to implement the functions of each module in the above-described device embodiments.

[0148] For example, the computer program may be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in the terminal device.

[0149] The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0150] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting all parts of the terminal device via various interfaces and lines.

[0151] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the terminal device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function, etc.; the data storage area may store data created based on the use of the mobile phone, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0152] Based on the above method embodiments, the present invention provides corresponding storage medium embodiments;

[0153] Another embodiment of the present invention provides a storage medium including a stored computer program, wherein, when the computer program is running, it controls the device where the storage medium is located to execute any of the above-described cable partial discharge pattern recognition methods based on multimodal fusion.

[0154] The aforementioned storage medium is a computer-readable storage medium, and the computer program includes computer program code, which may be in the form of source code, object code, executable file, or certain intermediate forms. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0155] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0156] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A cable partial discharge pattern recognition method based on multimodal fusion, characterized in that, include: Acquire the raw electrical signal of partial discharge from the cable to be identified; Phase-resolved partial discharge analysis is performed on the raw partial discharge electrical signal to generate a phase-resolved partial discharge map. Wavelet packet energy analysis is performed on the original partial discharge electrical signal to generate a wavelet packet energy map. The skewness and kurtosis statistical characteristics of the original partial discharge electrical signal are extracted, and a two-dimensional skewness-kurtosis distribution map is generated based on the skewness and kurtosis statistical characteristics. Recurrence graph analysis was performed on the original electrical signal of partial discharge to obtain a time series recurrence graph; The phase-resolved partial discharge map, the wavelet packet energy map, the skewness-kurtosis two-dimensional distribution map, and the time series recursive map are combined into a multimodal map. The multimodal map is input into a preset multimodal fusion recognition model, so that the multimodal fusion recognition model can perform spatial feature analysis on the multimodal map through the built-in visual feature encoding module and generate visual recognition results. The built-in time-series feature encoding module performs time-series feature analysis on the multimodal graph to generate time-series recognition results. The built-in weighted fusion module performs weighted fusion on the visual recognition results and the temporal recognition results to generate a comprehensive fusion result. The built-in classification and prediction module predicts the discharge type based on the comprehensive fusion results, generating cable partial discharge pattern recognition results.

2. The cable partial discharge pattern recognition method based on multimodal fusion as described in claim 1, characterized in that, After acquiring the original partial discharge electrical signal of the cable to be identified, the method further includes: Add white noise of a preset intensity to the original partial discharge electrical signal to generate a set of disturbance signals; The set of disturbance signals is subjected to CEEMDAN decomposition to generate a set of final intrinsic mode function components; Calculate the sample entropy of each component in the final set of intrinsic mode function components, and generate the sample entropy value corresponding to each component; The final set of intrinsic mode function components are sorted according to the sample entropy value, and a preset number of intrinsic mode function components are selected as a set of filtered intrinsic mode function components. The selected intrinsic mode function components are subjected to signal reconstruction processing to generate a denoised PD signal; The denoised PD signal is used to replace the original partial discharge electrical signal.

3. The cable partial discharge pattern recognition method based on multimodal fusion as described in claim 2, characterized in that, Training a multimodal fusion recognition model includes representation pre-training and fine-tuning training. The representation pre-training includes: A first training dataset containing several historical partial discharge signals of cables is obtained, and a set of unlabeled multimodal maps are generated based on the historical partial discharge signals of cables in the first training dataset. Multiple enhanced views are generated by performing at least two different random augmentation processes on each of the set of unlabeled multimodal maps. From the multiple augmented views corresponding to each map, any two augmented views originating from the same map are defined as positive sample pairs, and any two augmented views originating from different maps are defined as negative sample pairs. According to the preset batch size, the generated positive sample pairs and negative sample pairs are divided into several batches of first training samples; The first training samples of each batch are sequentially input into the visual feature encoding module, so that when the visual feature encoding module receives the first training samples of each batch, it extracts features from the positive and negative sample pairs in the current batch to generate the corresponding first feature vector; based on the preset contrast loss function, the first loss value is calculated according to the first feature vector; the learnable parameters in the visual feature encoding module are updated according to the first loss value using a preset optimizer until the preset number of training rounds is reached. The first training samples of each batch are sequentially input into the temporal feature encoding module, so that when the temporal feature encoding module receives the first training samples of each batch, it extracts features from the positive and negative sample pairs in the current batch to generate the corresponding second feature vector; based on the preset contrast loss function, the second loss value is calculated according to the second feature vector; the learnable parameters in the temporal feature encoding module are updated according to the second loss value using a preset optimizer until the preset number of training rounds is reached.

4. The cable partial discharge pattern recognition method based on multimodal fusion as described in claim 3, characterized in that, The fine-tuning training is performed after the representation pre-training is completed; The fine-tuning training includes: A second training dataset containing several historical partial discharge signals of cables and their corresponding real discharge type labels is obtained, and a set of labeled multimodal maps is generated based on the historical partial discharge signals of cables in the second training dataset. According to the preset batch size, the labeled multimodal map is randomly divided into several batches of second training samples; The second training samples of each batch are sequentially input into the multimodal fusion recognition model, and the multimodal fusion recognition model is subjected to end-to-end iterative training until the preset number of training rounds is reached. The multimodal fusion recognition model, upon receiving a batch of the second training samples, performs the following: The visual feature encoding module performs spatial feature analysis on the map samples in the second training sample of the current batch to generate visual recognition results during the training process. The temporal feature encoding module performs time series feature analysis on the spectral samples in the second training sample of the current batch to generate the temporal recognition results during the training process. The weighted fusion module performs weighted fusion of the visual recognition results and the temporal recognition results during the training process to generate a comprehensive fusion result during the training process. The classification prediction module predicts the discharge type based on the comprehensive fusion results during the training process, and generates the cable partial discharge pattern recognition result during the training process corresponding to the current second training sample. The third loss function value is calculated and generated based on the cable partial discharge pattern recognition results and the corresponding real discharge type labels during the training process using a preset loss function. Using a preset optimizer, gradient backpropagation and updates are performed on the learnable parameters in the visual feature encoding module, the temporal feature encoding module, the weighted fusion module, and the classification prediction module based on the value of the third loss function, in order to complete fine-tuning training.

5. The cable partial discharge pattern recognition method based on multimodal fusion as described in claim 4, characterized in that, The contrastive loss function is specifically as follows: In the formula, l i，j This is the loss value calculated for a positive sample pair (i, j); i and j are the indices of the two enhanced views in the positive sample pair; z i With z j is the feature vector corresponding to the two enhanced views in the positive sample pair; sim(·,·) is the cosine similarity function; τ is the temperature hyperparameter; N is the batch size; 2N is the total number of views in the first training sample of the current batch; k is the index of other views in the current batch.

6. The cable partial discharge pattern recognition method based on multimodal fusion as described in claim 5, characterized in that, The loss function is specifically as follows: In the formula, L is the loss value of the current second sample; M is the total number of discharge type categories; c is the category index; y c p is a sign function that has a value of 1 when class c is the true label of the sample, and 0 otherwise; c The predicted probability of the multimodal fusion recognition model for category c is given.

7. The cable partial discharge pattern recognition method based on multimodal fusion as described in claim 6, characterized in that, The visual feature encoding module is a Vision Transformer-based encoder; the temporal feature encoding module is a temporal Transformer-based encoder; and the classification prediction module is a multilayer perceptron classifier.

8. A cable partial discharge pattern recognition device based on multimodal fusion, characterized in that, include: The module includes a partial discharge raw electrical signal acquisition module, a multimodal spectrum generation module, and a cable partial discharge mode recognition module. The partial discharge raw electrical signal acquisition module is used to acquire the partial discharge raw electrical signal of the cable to be identified; The multimodal spectrum generation module is used to perform phase-resolved partial discharge analysis on the original partial discharge electrical signal to generate a phase-resolved partial discharge spectrum; perform wavelet packet energy analysis on the original partial discharge electrical signal to generate a wavelet packet energy map; extract the skewness and kurtosis statistical features of the original partial discharge electrical signal and generate a skewness-kurtosis two-dimensional distribution map based on the skewness and kurtosis statistical features; perform recursive graph analysis on the original partial discharge electrical signal to obtain a time series recursive graph; and combine the phase-resolved partial discharge spectrum, the wavelet packet energy map, the skewness-kurtosis two-dimensional distribution map, and the time series recursive graph into a multimodal spectrum. The cable partial discharge pattern recognition module is used to input the multimodal map into a preset multimodal fusion recognition model, so that the multimodal fusion recognition model can perform spatial feature analysis on the multimodal map through the built-in visual feature encoding module and generate visual recognition results. The built-in temporal feature encoding module performs time-series feature analysis on the multimodal map to generate temporal recognition results; the built-in weighted fusion module performs weighted fusion on the visual recognition results and the temporal recognition results to generate a comprehensive fusion result. The built-in classification and prediction module predicts the discharge type based on the comprehensive fusion results, generating cable partial discharge pattern recognition results.

9. An electronic device, characterized in that, The method includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the cable partial discharge pattern recognition method based on multimodal fusion as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device where the storage medium is located to perform the cable partial discharge pattern recognition method based on multimodal fusion as described in any one of claims 1 to 7.

Images