Power transformation device anomaly detection method and apparatus based on voiceprint

Abstract

A power transformation device anomaly detection method and apparatus based on a voiceprint, which method and apparatus relate to the field of power transformation device monitoring, and can solve the technical problem of the accuracy of existing power transformation device anomaly detection based on a voiceprint being low. The method comprises: acquiring historical voiceprint feature data of a plurality of power transformation devices, and on the basis of a two-dimensional convolutional network, extracting local time-frequency features of the historical voiceprint feature data, so as to obtain first voiceprint features (S101); on the basis of a bidirectional recurrent neural network, performing feature extraction on the first voiceprint features, so as to obtain second voiceprint features (S102); on the basis of the second voiceprint features, performing iterative training on a preset initial voiceprint detection model multiple times, and during iteration, performing joint optimization on model parameters of the initial voiceprint detection model on the basis of contrastive learning and perceptron classification, and completing the training when the model parameters meet preset values, so as to obtain a voiceprint detection model (S103); and acquiring a real-time voiceprint feature of a power transformation device to be subjected to detection, and inputting the real-time voiceprint feature into the voiceprint detection model, so as to obtain an anomaly detection result of the power transformation device to be subjected to detection (S104).

Description

A method and apparatus for detecting anomalies in power equipment based on voiceprint Technical Field

[0001] This application relates to the field of power equipment monitoring, and in particular to a method and apparatus for detecting power equipment anomalies based on voiceprints. Background Technology

[0002] As the primary equipment in a substation, the safe and stable operation of power transmission equipment plays a crucial role in the stability of the substation and the power system. Therefore, it is necessary to accurately detect anomalies in power transmission equipment and make timely response decisions to prevent equipment failure.

[0003] In existing technologies, there are generally two methods for anomaly detection in power equipment. The first method involves analyzing the periodic reporting of operational data by the power equipment to determine its operating status. However, due to the periodic nature of this reporting, some emergency situations cannot be detected in a timely manner, resulting in a certain lag. The second method involves collecting operational data from the power equipment, primarily voiceprint data, using sensors, especially sound sensors, to determine its operating status. This method can overcome the lag problem. However, current voiceprint data analysis mainly relies on deep learning prediction methods using a single model, which suffers from difficulties in feature extraction and inadequate model training, leading to low accuracy in voiceprint-based anomaly detection. Therefore, how to achieve more accurate anomaly detection in power equipment based on voiceprint remains a pressing technical problem that needs to be solved. Summary of the Invention

[0004] This application provides a method and apparatus for detecting abnormalities in power equipment based on voiceprint, in order to solve the technical problem that the accuracy of existing voiceprint-based abnormality detection methods for power equipment is not high.

[0005] According to a first aspect of the embodiments of this application, a method for detecting abnormalities in power equipment based on voiceprints is provided, comprising:

[0006] Historical voiceprint feature data of multiple power equipment are acquired, and local time-frequency features of the historical voiceprint feature data are extracted based on a two-dimensional convolutional network to obtain the first voiceprint feature.

[0007] Based on a bidirectional recurrent neural network, feature extraction is performed on the first voiceprint feature to obtain the second voiceprint feature;

[0008] Based on the second voiceprint feature, the preset initial voiceprint detection model is trained iteratively multiple times. During the iteration, the model parameters of the initial voiceprint detection model are jointly optimized based on contrastive learning and perceptron classification until the model parameters meet the preset values, and the training is completed to obtain the voiceprint detection model.

[0009] The real-time voiceprint features of the substation to be detected are obtained, and the real-time voiceprint features are input into the voiceprint detection model to obtain the abnormal detection results of the substation to be detected.

[0010] This application first acquires historical voiceprint feature data of multiple substations, and extracts local time-frequency features based on a two-dimensional convolutional network to obtain the first voiceprint feature. Then, it extracts features based on a bidirectional recurrent neural network to obtain the second voiceprint feature, which can more accurately extract the time-frequency feature part of the voiceprint feature. Based on the second voiceprint feature, the preset initial voiceprint detection model is trained iteratively multiple times based on contrastive learning and perceptron classification to obtain the voiceprint detection model. By using the more accurate second voiceprint feature as the training set, and combining contrastive learning and perceptron classification to optimize the model parameters, the three work together to improve the accuracy of the final output voiceprint detection model. This improves the accuracy of anomaly detection when performing anomaly detection on the substation to be detected based on real-time voiceprint features.

[0011] In some embodiments of this application, obtaining historical voiceprint feature data of multiple power equipment specifically includes:

[0012] Acquire historical audio data from multiple power equipment and divide the historical audio data into multiple sample audio sequences according to a preset sequence length;

[0013] Adaptive time-frequency analysis is performed on the multiple sample audio sequences to obtain historical voiceprint feature data corresponding to the multiple power equipment.

[0014] This application first acquires historical audio data from multiple power equipment, and then segments the historical audio data according to a preset sequence length to obtain multiple sample audio sequences. By segmenting the audio data to obtain short sequences, it is beneficial to improve the accuracy of subsequent time-frequency analysis. In turn, when performing adaptive time-frequency analysis on multiple sample audio data to obtain historical voiceprint feature data, it improves the matching degree between the historical voiceprint feature data and the actual operation of the corresponding power equipment.

[0015] In some embodiments of this application, the step of extracting local time-frequency features of the historical voiceprint feature data based on a two-dimensional convolutional network to obtain the first voiceprint feature specifically includes:

[0016] The historical voiceprint feature data is augmented based on a mask to obtain enhanced voiceprint feature data.

[0017] Based on a two-dimensional convolutional network, the enhanced voiceprint feature data is downsampled to extract local time-frequency features, thus obtaining the first voiceprint feature.

[0018] This application first performs data augmentation on historical voiceprint feature data based on a mask to obtain enhanced voiceprint feature data, which can increase the amount of sample data and prevent insufficient samples from causing a decrease in feature accuracy. Then, when the enhanced voiceprint features are downsampled based on two-dimensional convolution to extract local time-frequency features to obtain the first voiceprint feature, the accuracy of the first voiceprint feature is maintained.

[0019] In some embodiments of this application, the step of iteratively training a preset initial voiceprint detection model based on the second voiceprint feature, wherein the model parameters of the initial voiceprint detection model are jointly optimized based on contrastive learning and perceptron classification during iteration, until the model parameters meet preset values ​​to complete the training and obtain the voiceprint detection model, specifically includes:

[0020] The second voiceprint feature is averaged and pooled to obtain the third voiceprint feature;

[0021] Based on the third voiceprint feature, the preset initial voiceprint detection model is trained iteratively multiple times. During the iteration, the joint total loss is calculated based on contrastive learning and perceptron classification, and the model parameters are optimized according to the joint total loss until the model parameters meet the preset value, thus completing the training and obtaining the voiceprint detection model. The joint total loss includes contrastive loss based on contrastive learning and cross-entropy loss based on perceptron classification.

[0022] The joint total loss is as follows:

[0023] L=α*L contrast +(1-α)*L ce ;

[0024] Among them, L,L contrast L ce These are the joint total loss, the contrastive loss, and the cross-entropy loss, respectively, with α being the loss balance parameter.

[0025] The comparison loss is specifically as follows:

[0026] Among them, H c H is the third voiceprint feature based on the mapping output of a multilayer perceptron. + H - These are the positive and negative samples of the voiceprint features, respectively. sim(a, b) is the similarity between input a and input b, and τ is the temperature parameter.

[0027] This application first performs average pooling on the second voiceprint feature to obtain the third voiceprint feature. Then, based on the third voiceprint feature, and the joint total loss calculated based on the contrastive loss of contrastive learning and the cross-entropy loss of perceptron classification, the preset initial voiceprint detection model is trained multiple times to obtain the voiceprint detection model. The three together can improve the accuracy of the voiceprint detection model, thereby improving the accuracy of abnormal detection of power equipment.

[0028] In some embodiments of this application, the step of inputting the real-time voiceprint features into the voiceprint detection model to obtain the anomaly detection result of the substation to be detected specifically includes:

[0029] The real-time voiceprint features are input into the voiceprint detection model to obtain the anomaly detection score of the substation to be detected.

[0030] When the anomaly detection score is greater than the preset anomaly threshold, the anomaly occurrence time of the substation to be detected is located based on the anomaly detection score and the sliding window method.

[0031] Based on the anomaly detection score and the time period in which the anomaly occurred, the anomaly detection result of the substation to be tested is obtained.

[0032] This application first obtains the anomaly detection score of the substation to be detected based on real-time voiceprint features and voiceprint detection model. When the anomaly detection score is greater than the preset anomaly threshold, the anomaly occurrence time period is located based on the sliding window method, and then the anomaly detection result is obtained. The preset anomaly threshold can eliminate false judgments of anomaly detection, and the sliding window method can more accurately locate the anomaly occurrence time period, so the obtained anomaly detection result is more consistent with the actual anomaly of the substation to be detected.

[0033] According to a second aspect of the embodiments of this application, a substation equipment anomaly detection device based on voiceprint is provided, including a time-frequency feature acquisition module, a voiceprint feature extraction module, a detection model training module, and an equipment anomaly detection module;

[0034] The time-frequency feature acquisition module is used to acquire historical voiceprint feature data of multiple power equipment, and extract local time-frequency features of the historical voiceprint feature data based on a two-dimensional convolutional network to obtain the first voiceprint feature.

[0035] The voiceprint feature extraction module is used to extract features from the first voiceprint feature based on a bidirectional recurrent neural network to obtain the second voiceprint feature;

[0036] The detection model training module is used to perform multiple iterations of training on the preset initial voiceprint detection model based on the second voiceprint feature. During the iteration, the model parameters of the initial voiceprint detection model are jointly optimized based on contrastive learning and perceptron classification until the model parameters meet the preset values, and the training is completed to obtain the voiceprint detection model.

[0037] The equipment anomaly detection module is used to acquire the real-time voiceprint features of the substation equipment to be detected, and input the real-time voiceprint features into the voiceprint detection model to obtain the anomaly detection result of the substation equipment to be detected.

[0038] In some embodiments of this application, the time-frequency feature acquisition module includes a voiceprint feature acquisition submodule; the voiceprint feature acquisition submodule includes an audio acquisition and segmentation unit and a sequence time-frequency analysis unit;

[0039] The audio acquisition and segmentation unit is used to acquire historical audio data from multiple power equipment and segment the historical audio data into multiple sample audio sequences according to a preset sequence length.

[0040] The sequence time-frequency analysis unit is used to perform adaptive time-frequency analysis on the multiple sample audio sequences to obtain historical voiceprint feature data corresponding to the multiple power equipment.

[0041] In some embodiments of this application, the time-frequency feature acquisition module includes a time-frequency feature extraction submodule; the time-frequency feature extraction submodule includes a voiceprint feature enhancement unit and a time-frequency feature extraction unit;

[0042] The voiceprint feature enhancement unit is used to perform data enhancement on the historical voiceprint feature data based on a mask to obtain enhanced voiceprint feature data.

[0043] The time-frequency feature extraction unit is used to downsample the enhanced voiceprint feature data based on a two-dimensional convolutional network to extract local time-frequency features and obtain the first voiceprint feature.

[0044] In some embodiments of this application, the detection model training module includes a voiceprint feature pooling submodule and a detection model training submodule;

[0045] The voiceprint feature pooling submodule is used to perform average pooling on the second voiceprint feature to obtain the third voiceprint feature.

[0046] The detection model training submodule is used to perform multiple iterations of training on a preset initial voiceprint detection model based on the third voiceprint feature. During iteration, the joint total loss is calculated based on contrastive learning and perceptron classification, and the model parameters are optimized according to the joint total loss until the model parameters meet the preset values, thus completing the training and obtaining the voiceprint detection model. The joint total loss includes contrastive loss based on contrastive learning and cross-entropy loss based on perceptron classification.

[0047] The joint total loss is as follows:

[0048] L=α*L contrast +(1-α)* L ce ;

[0049] Among them, L,L contrast L ce These are the joint total loss, the contrastive loss, and the cross-entropy loss, respectively, with α being the loss balance parameter.

[0050] The comparison loss is specifically as follows:

[0051] Among them, H c H is the third voiceprint feature based on the mapping output of a multilayer perceptron. + H - These are the positive and negative samples of the voiceprint features, respectively. sim(a,b) is the similarity between input a and input b, and τ is the temperature parameter.

[0052] In some embodiments of this application, the device anomaly detection module includes an anomaly score acquisition submodule, an anomaly time period location submodule, and an anomaly result acquisition submodule;

[0053] The anomaly score acquisition submodule is used to input the real-time voiceprint features into the voiceprint detection model to obtain the anomaly detection score of the substation to be detected.

[0054] The abnormal time period location submodule is used to locate the abnormal time period of the substation to be detected based on the abnormal detection score and the sliding window method when the abnormal detection score is greater than the preset abnormal threshold.

[0055] The abnormal result acquisition submodule is used to obtain the abnormal detection result of the substation to be detected based on the abnormal detection score and the abnormal occurrence time period.

[0056] This application first acquires historical voiceprint feature data of multiple substations, and extracts local time-frequency features based on a two-dimensional convolutional network to obtain the first voiceprint feature. Then, it extracts features based on a bidirectional recurrent neural network to obtain the second voiceprint feature, which can more accurately extract the time-frequency feature part of the voiceprint feature. Based on the second voiceprint feature, the preset initial voiceprint detection model is trained iteratively multiple times based on contrastive learning and perceptron classification to obtain the voiceprint detection model. By using the more accurate second voiceprint feature as the training set, and combining contrastive learning and perceptron classification to optimize the model parameters, the three work together to improve the accuracy of the final output voiceprint detection model. This improves the accuracy of anomaly detection when performing anomaly detection on the substation to be detected based on real-time voiceprint features. Attached Figure Description

[0057] Figure 1: A flowchart illustrating a method for detecting abnormalities in power equipment based on voiceprints, as shown in some embodiments of this application.

[0058] Figure 2: A module structure diagram of a voiceprint-based abnormality detection device for power equipment, as shown in some embodiments of this application. Detailed Implementation

[0059] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below in conjunction with the accompanying drawings are exemplary and are only used to explain some embodiments of this application, and should not be construed as limiting the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments shown in this application without inventive effort are within the protection scope of this application.

[0060] In the description of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, unless otherwise explicitly specified, "a plurality of" or "several" means two or more.

[0061] In existing technologies, the detection of anomalies in substation equipment based on voiceprint data mainly relies on deep learning prediction methods using a single model. This approach suffers from difficulties in feature extraction and inadequate model training, resulting in low accuracy in voiceprint-based anomaly detection. Therefore, how to achieve more accurate anomaly detection in substation equipment based on voiceprint remains a pressing technical problem that needs to be solved.

[0062] Based on the above technical background, please refer to Figure 1. This application provides a method for detecting abnormalities in power equipment based on voiceprint, including steps S101 to S104, each step of which is as follows:

[0063] Step S101: Obtain historical voiceprint feature data of multiple power equipment, and extract local time-frequency features of the historical voiceprint feature data based on a two-dimensional convolutional network to obtain the first voiceprint feature.

[0064] In some embodiments of this application, obtaining historical voiceprint feature data of multiple power equipment specifically includes:

[0065] Acquire historical audio data from multiple power equipment and divide the historical audio data into multiple sample audio sequences according to a preset sequence length;

[0066] Adaptive time-frequency analysis is performed on the multiple sample audio sequences to obtain historical voiceprint feature data corresponding to the multiple power equipment.

[0067] Specifically, the historical audio data is represented as x(t), with a sampling rate of f. s The preset sequence length is T, and the number of obtained sample audio sequences is k < b, where b is the amount of data processed in a single run. If the length of the currently processed sample audio sequence does not meet the preset sequence length T, the number of 0s is added to the end of the currently processed sample audio sequence up to the preset sequence length T.

[0068] In some embodiments of this application, the step of performing adaptive time-frequency analysis on the plurality of sample audio sequences to obtain historical voiceprint feature data corresponding to the plurality of substation equipment specifically includes:

[0069] The multiple sample audio sequences are combined to obtain the sample audio tensor X[b,1,T];

[0070] Adaptive time-frequency analysis is performed on the sample audio tensor X[b,1,T] to obtain the historical voiceprint feature data S[b,F,T′] corresponding to the multiple substations.

[0071] Where F represents the frequency dimension and T′ represents the time dimension (number of time frames).

[0072] Specifically, in adaptive time-frequency analysis, the preferred value for the frequency dimension F is 2. 8 =256 or 2 9 =512, the time dimension T′ is related to the preset time resolution of the adaptive time-frequency analysis.

[0073] In some embodiments of this application, the preferred embodiment of the adaptive time-frequency analysis is AWPT.

[0074] This application first acquires historical audio data from multiple power equipment, and then segments the historical audio data according to a preset sequence length to obtain multiple sample audio sequences. By segmenting the audio data to obtain short sequences, it is beneficial to improve the accuracy of subsequent time-frequency analysis. In turn, when performing adaptive time-frequency analysis on multiple sample audio data to obtain historical voiceprint feature data, it improves the matching degree between the historical voiceprint feature data and the actual operation of the corresponding power equipment.

[0075] In some embodiments of this application, the step of extracting local time-frequency features of the historical voiceprint feature data based on a two-dimensional convolutional network to obtain the first voiceprint feature specifically includes:

[0076] The historical voiceprint feature data is augmented based on a mask to obtain enhanced voiceprint feature data.

[0077] Based on a two-dimensional convolutional network, the enhanced voiceprint feature data is downsampled to extract local time-frequency features, thus obtaining the first voiceprint feature.

[0078] In some embodiments of this application, the step of performing data augmentation on the historical voiceprint feature data based on a mask to obtain enhanced voiceprint feature data specifically involves:

[0079] Based on the mask, the historical voiceprint feature data S[b,F,T′] is augmented in frequency and time to obtain the augmented voiceprint feature data S. aug [b,F,T′].

[0080] It should be understood that the mask-based data augmentation methods in this application include, but are not limited to, GridMask-based data augmentation methods. The specific implementation methods are well known to those skilled in the art and will not be described in detail here.

[0081] In some embodiments of this application, the step of downsampling the enhanced voiceprint feature data based on a two-dimensional convolutional network to extract local time-frequency features and obtain the first voiceprint feature specifically involves:

[0082] Based on a two-dimensional convolutional network, the enhanced voiceprint feature data S aug [b,F,T′] is downsampled to extract local time-frequency features, resulting in the first voiceprint feature C[b,ch,F′,T″].

[0083] Where ch is the number of channels, and F′ and T″ are the frequency and time dimensions after downsampling, respectively.

[0084] In some embodiments of this application, one possible implementation of the two-dimensional convolutional network is: Conv2d→BatchNorm→ReLU→Pooling, which is two-dimensional convolution, data normalization, data activation, and data pooling.

[0085] This application first performs data augmentation on historical voiceprint feature data based on a mask to obtain enhanced voiceprint feature data, which can increase the amount of sample data and prevent insufficient samples from causing a decrease in feature accuracy. Then, when the enhanced voiceprint features are downsampled based on two-dimensional convolution to extract local time-frequency features to obtain the first voiceprint feature, the accuracy of the first voiceprint feature is maintained.

[0086] Step S102: Based on a bidirectional recurrent neural network, feature extraction is performed on the first voiceprint feature to obtain the second voiceprint feature.

[0087] In some embodiments of this application, the step of extracting features from the first voiceprint feature based on a bidirectional recurrent neural network to obtain the second voiceprint feature specifically involves:

[0088] The first voiceprint feature C[b,ch,F′,T″] is rearranged to obtain the intermediate voiceprint feature C. seq [b,T",ch*F′];

[0089] Based on a bidirectional recurrent neural network, the intermediate voiceprint feature C seq [b,T",ch*F′] is used for feature extraction to obtain the second voiceprint feature H[b,T",2*D].

[0090] Where D is the hidden state dimension of the bidirectional recurrent neural network.

[0091] In some embodiments of this application, the bidirectional recurrent neural network includes, but is not limited to, bidirectional RNN and bidirectional LSTM, with bidirectional LSTM being the preferred embodiment.

[0092] The reason for considering using bidirectional recurrent neural networks for feature extraction is that they can better process sequence information and more accurately represent contextual dependencies by employing multi-head self-attention mechanisms, thereby obtaining more accurate second voiceprint features.

[0093] Step S103: Based on the second voiceprint feature, perform multiple iterations of training on the preset initial voiceprint detection model. During iteration, the model parameters of the initial voiceprint detection model are jointly optimized based on contrastive learning and perceptron classification until the model parameters meet the preset values, thus completing the training and obtaining the voiceprint detection model.

[0094] In some embodiments of this application, the step of iteratively training a preset initial voiceprint detection model based on the second voiceprint feature, wherein the model parameters of the initial voiceprint detection model are jointly optimized based on contrastive learning and perceptron classification during iteration, until the model parameters meet preset values ​​to complete the training and obtain the voiceprint detection model, specifically includes:

[0095] The second voiceprint feature is averaged and pooled to obtain the third voiceprint feature;

[0096] Based on the third voiceprint feature, the preset initial voiceprint detection model is trained iteratively multiple times. During the iteration, the joint total loss is calculated based on contrastive learning and perceptron classification, and the model parameters are optimized according to the joint total loss until the model parameters meet the preset value, thus completing the training and obtaining the voiceprint detection model. The joint total loss includes contrastive loss based on contrastive learning and cross-entropy loss based on perceptron classification.

[0097] The joint total loss is as follows:

[0098] L=α*L contrast +(1-α)*L ce ;

[0099] Among them, L,L contrast L ce These are the joint total loss, the contrastive loss, and the cross-entropy loss, respectively, with α being the loss balance parameter.

[0100] The comparison loss is specifically as follows:

[0101] Among them, H c H is the third voiceprint feature based on the mapping output of a multilayer perceptron. + H - These are the positive and negative samples of the voiceprint features, respectively. sim(a,b) is the similarity between input a and input b, and τ is the temperature parameter.

[0102] In some embodiments of this application, the step of performing average pooling on the second voiceprint feature to obtain the third voiceprint feature specifically involves:

[0103] The second voiceprint feature H[b,T″,2*D] is average pooled to obtain the third voiceprint feature H. avg [b,2*D].

[0104] Specifically, when calculating the contrastive loss based on contrastive learning, the specific steps are as follows:

[0105] The third voiceprint feature H avg [b, 2*D] are input to a multilayer perceptron, and the mapped output H is obtained. c [b,D′];

[0106] The positive and negative samples corresponding to the mapping output are input into a multilayer perceptron to obtain the positive mapping sample H. + and negative mapping sample H - ;

[0107] According to the mapping, output H c The positive mapping sample H of [b,D′] + and the negative mapping sample H - Calculate the contrast loss L contrast .

[0108] Where D′ represents the spatial dimension of contrastive learning.

[0109] In some embodiments of this application, the similarity sim(a,b) includes, but is not limited to, Pearson correlation coefficient, cosine similarity, log-likelihood similarity, and Spearman rank correlation coefficient, with cosine similarity being a preferred embodiment.

[0110] In some embodiments of this application, the preferred sampling object for the positive sample is the normal operation voiceprint feature of the same substation, and the preferred sampling object for the negative sample is the voiceprint feature of other substations or the synthetic abnormal voiceprint feature of the same substation.

[0111] It should be understood that the temperature parameter τ mentioned in this application is a hyperparameter used to control the distribution of positive and negative samples in contrastive learning. The temperature parameter τ is a commonly used hyperparameter for training in the field of neural network models, so it will not be elaborated further here.

[0112] Specifically, when calculating the cross-entropy loss based on perceptron classification, the specific steps are as follows:

[0113] The third voiceprint feature H avg [b,2*D] is input into a multilayer perceptron to obtain the mapped anomaly score [b,1].

[0114] If the third voiceprint feature H avg The current sample [b,2*D] is a labeled sample. The cross-entropy loss L is calculated based on the mapping anomaly score score[b,1]. ce ;

[0115] If the third voiceprint feature H avg The current sample [b,2*D] is an unlabeled sample. Based on the reconstruction error under self-supervision, the cross-entropy loss L is obtained according to the mapping anomaly score [b,1]. ce .

[0116] In some embodiments of this application, the cross-entropy loss L ce The preferred implementation scheme is binary cross-entropy loss.

[0117] This application first performs average pooling on the second voiceprint feature to obtain the third voiceprint feature. Then, based on the third voiceprint feature, and the joint total loss calculated based on the contrastive loss of contrastive learning and the cross-entropy loss of perceptron classification, the preset initial voiceprint detection model is trained multiple times to obtain the voiceprint detection model. The three together can improve the accuracy of the voiceprint detection model, thereby improving the accuracy of abnormal detection of power equipment.

[0118] Step S104: Obtain the real-time voiceprint features of the substation to be detected, and input the real-time voiceprint features into the voiceprint detection model to obtain the abnormal detection results of the substation to be detected.

[0119] In some embodiments of this application, the step of inputting the real-time voiceprint features into the voiceprint detection model to obtain the anomaly detection result of the substation to be detected specifically includes:

[0120] The real-time voiceprint features are input into the voiceprint detection model to obtain the anomaly detection score of the substation to be detected.

[0121] When the anomaly detection score is greater than the preset anomaly threshold, the anomaly occurrence time of the substation to be detected is located based on the anomaly detection score and the sliding window method.

[0122] Based on the anomaly detection score and the time period in which the anomaly occurred, the anomaly detection result of the substation to be tested is obtained.

[0123] This application first obtains the anomaly detection score of the substation to be detected based on real-time voiceprint features and voiceprint detection model. When the anomaly detection score is greater than the preset anomaly threshold, the anomaly occurrence time period is located based on the sliding window method, and then the anomaly detection result is obtained. The preset anomaly threshold can eliminate false judgments of anomaly detection, and the sliding window method can more accurately locate the anomaly occurrence time period, so the obtained anomaly detection result is more consistent with the actual anomaly of the substation to be detected.

[0124] Compared to existing technologies, this application first acquires historical voiceprint feature data of multiple substations, extracts local time-frequency features based on a two-dimensional convolutional network to obtain the first voiceprint feature, and then extracts features based on a bidirectional recurrent neural network to obtain the second voiceprint feature. This can more accurately extract the time-frequency feature part of the voiceprint feature. Then, based on the second voiceprint feature, the preset initial voiceprint detection model is trained iteratively multiple times based on contrastive learning and perceptron classification to obtain the voiceprint detection model. By using the more accurate second voiceprint feature as the training set, and combining contrastive learning and perceptron classification to optimize the model parameters, the three work together to improve the accuracy of the final output voiceprint detection model. This improves the accuracy of anomaly detection when performing anomaly detection on the substation to be detected based on real-time voiceprint features.

[0125] Corresponding to the aforementioned method, please refer to Figure 2. This application provides a power equipment anomaly detection device based on voiceprint, including a time-frequency feature acquisition module 210, a voiceprint feature extraction module 220, a detection model training module 230, and an equipment anomaly detection module 240.

[0126] The time-frequency feature acquisition module 210 is used to acquire historical voiceprint feature data of multiple power equipment, and extract local time-frequency features of the historical voiceprint feature data based on a two-dimensional convolutional network to obtain the first voiceprint feature.

[0127] The voiceprint feature extraction module 220 is used to extract features from the first voiceprint feature based on a bidirectional recurrent neural network to obtain the second voiceprint feature;

[0128] The detection model training module 230 is used to perform multiple iterations of training on a preset initial voiceprint detection model based on the second voiceprint feature. During the iteration, the model parameters of the initial voiceprint detection model are jointly optimized based on contrastive learning and perceptron classification until the model parameters meet the preset values, and the training is completed to obtain the voiceprint detection model.

[0129] The equipment anomaly detection module 240 is used to acquire the real-time voiceprint features of the substation equipment to be detected, and input the real-time voiceprint features into the voiceprint detection model to obtain the anomaly detection result of the substation equipment to be detected.

[0130] In some embodiments of this application, the time-frequency feature acquisition module 210 includes a voiceprint feature acquisition submodule; the voiceprint feature acquisition submodule includes an audio acquisition and segmentation unit and a sequence time-frequency analysis unit;

[0131] The audio acquisition and segmentation unit is used to acquire historical audio data from multiple power equipment and segment the historical audio data into multiple sample audio sequences according to a preset sequence length.

[0132] The sequence time-frequency analysis unit is used to perform adaptive time-frequency analysis on the multiple sample audio sequences to obtain historical voiceprint feature data corresponding to the multiple power equipment.

[0133] In some embodiments of this application, the time-frequency feature acquisition module 210 includes a time-frequency feature extraction submodule; the time-frequency feature extraction submodule includes a voiceprint feature enhancement unit and a time-frequency feature extraction unit;

[0134] The voiceprint feature enhancement unit is used to perform data enhancement on the historical voiceprint feature data based on a mask to obtain enhanced voiceprint feature data.

[0135] The time-frequency feature extraction unit is used to downsample the enhanced voiceprint feature data based on a two-dimensional convolutional network to extract local time-frequency features and obtain the first voiceprint feature.

[0136] In some embodiments of this application, the detection model training module 230 includes a voiceprint feature pooling submodule and a detection model training submodule;

[0137] The voiceprint feature pooling submodule is used to perform average pooling on the second voiceprint feature to obtain the third voiceprint feature.

[0138] The detection model training submodule is used to perform multiple iterations of training on a preset initial voiceprint detection model based on the third voiceprint feature. During iteration, the joint total loss is calculated based on contrastive learning and perceptron classification, and the model parameters are optimized according to the joint total loss until the model parameters meet the preset values, thus completing the training and obtaining the voiceprint detection model. The joint total loss includes contrastive loss based on contrastive learning and cross-entropy loss based on perceptron classification.

[0139] The joint total loss is as follows:

[0140] L=α*L contrast +(1-α)*L ce ;

[0141] Among them, L,L contrast， L ce These are the joint total loss, the contrastive loss, and the cross-entropy loss, respectively, with α being the loss balance parameter.

[0142] The comparison loss is specifically as follows:

[0143] Among them, H c H is the third voiceprint feature based on the mapping output of a multilayer perceptron. + H - These are the positive and negative samples of the voiceprint features, respectively. sim(a,b) is the similarity between input a and input b, and τ is the temperature parameter.

[0144] In some embodiments of this application, the device anomaly detection module 240 includes an anomaly score acquisition submodule, an anomaly time period location submodule, and an anomaly result acquisition submodule;

[0145] The anomaly score acquisition submodule is used to input the real-time voiceprint features into the voiceprint detection model to obtain the anomaly detection score of the substation to be detected.

[0146] The abnormal time period location submodule is used to locate the abnormal time period of the substation to be detected based on the abnormal detection score and the sliding window method when the abnormal detection score is greater than the preset abnormal threshold.

[0147] The abnormal result acquisition submodule is used to obtain the abnormal detection result of the substation to be detected based on the abnormal detection score and the abnormal occurrence time period.

[0148] This application first acquires historical voiceprint feature data of multiple substations, and extracts local time-frequency features based on a two-dimensional convolutional network to obtain the first voiceprint feature. Then, it extracts features based on a bidirectional recurrent neural network to obtain the second voiceprint feature, which can more accurately extract the time-frequency feature part of the voiceprint feature. Based on the second voiceprint feature, the preset initial voiceprint detection model is trained iteratively multiple times based on contrastive learning and perceptron classification to obtain the voiceprint detection model. By using the more accurate second voiceprint feature as the training set, and combining contrastive learning and perceptron classification to optimize the model parameters, the three work together to improve the accuracy of the final output voiceprint detection model. This improves the accuracy of anomaly detection when performing anomaly detection on the substation to be detected based on real-time voiceprint features.

[0149] It should be understood that the apparatus provided in the embodiments of this application corresponds to the aforementioned method. The voiceprint-based transformer equipment anomaly detection apparatus provided in the embodiments of this application can implement the voiceprint-based transformer equipment anomaly detection method provided in any embodiment of this application.

[0150] Adaptively, embodiments of this application also provide a computer device and a computer-readable storage medium.

[0151] The computer device includes: a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor;

[0152] The processor executes the computer program to implement the voiceprint-based abnormal detection method for power equipment according to this application.

[0153] The computer-readable storage medium stores multiple instructions adapted for loading by a processor to execute a voiceprint-based abnormality detection method for power equipment according to this application.

[0154] The above description represents some embodiments of this application, providing a further detailed explanation of the purpose, technical solution, and beneficial effects of this application. It should be understood that the above-described embodiments of this application should not be construed as limiting this application. In particular, any changes, modifications, equivalent substitutions, and variations made by those skilled in the art within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for detecting anomalies in power equipment based on voiceprint, characterized in that, include: Historical voiceprint feature data of multiple power equipment are acquired, and local time-frequency features of the historical voiceprint feature data are extracted based on a two-dimensional convolutional network to obtain the first voiceprint feature. Based on a bidirectional recurrent neural network, feature extraction is performed on the first voiceprint feature to obtain the second voiceprint feature; Based on the second voiceprint feature, the preset initial voiceprint detection model is trained iteratively multiple times. During the iteration, the model parameters of the initial voiceprint detection model are jointly optimized based on contrastive learning and perceptron classification until the model parameters meet the preset values, and the training is completed to obtain the voiceprint detection model. The real-time voiceprint features of the substation to be detected are obtained, and the real-time voiceprint features are input into the voiceprint detection model to obtain the abnormal detection results of the substation to be detected.

2. The method for detecting abnormalities in power equipment based on voiceprint according to claim 1, characterized in that, The acquisition of historical voiceprint feature data from multiple power equipment specifically includes: Acquire historical audio data from multiple power equipment and divide the historical audio data into multiple sample audio sequences according to a preset sequence length; Adaptive time-frequency analysis is performed on the multiple sample audio sequences to obtain historical voiceprint feature data corresponding to the multiple power equipment.

3. The method for detecting abnormalities in power equipment based on voiceprint according to claim 1, characterized in that, The first voiceprint feature is obtained by extracting local time-frequency features from the historical voiceprint feature data based on a two-dimensional convolutional network, specifically including: The historical voiceprint feature data is augmented based on a mask to obtain enhanced voiceprint feature data. Based on a two-dimensional convolutional network, the enhanced voiceprint feature data is downsampled to extract local time-frequency features, thus obtaining the first voiceprint feature.

4. The method for detecting abnormalities in power equipment based on voiceprint according to claim 1, characterized in that, The step involves iteratively training a preset initial voiceprint detection model based on the second voiceprint feature. During each iteration, the model parameters of the initial voiceprint detection model are jointly optimized based on contrastive learning and perceptron classification until the model parameters meet preset values, thus completing the training and obtaining the voiceprint detection model. Specifically, this includes: The second voiceprint feature is averaged and pooled to obtain the third voiceprint feature; Based on the third voiceprint feature, the preset initial voiceprint detection model is trained iteratively multiple times. During the iteration, the joint total loss is calculated based on contrastive learning and perceptron classification, and the model parameters are optimized according to the joint total loss until the model parameters meet the preset value, thus completing the training and obtaining the voiceprint detection model. The joint total loss includes contrastive loss based on contrastive learning and cross-entropy loss based on perceptron classification. The joint total loss is as follows: L=α*L contrast +(1-a)*L ce ; Among them, L,L contrast ,L ce These are the joint total loss, the contrastive loss, and the cross-entropy loss, respectively, with α being the loss balance parameter. The comparison loss is specifically as follows: Among them, H c H is the third voiceprint feature based on the mapping output of a multilayer perceptron. + H - These are the positive and negative samples of the voiceprint features, respectively. sim(a,b) is the similarity between input a and input b, and τ is the temperature parameter.

5. The method for detecting abnormalities in power equipment based on voiceprint according to claim 4, characterized in that, The step of inputting the real-time voiceprint features into the voiceprint detection model to obtain the anomaly detection result of the substation to be detected specifically includes: The real-time voiceprint features are input into the voiceprint detection model to obtain the anomaly detection score of the substation to be detected. When the anomaly detection score is greater than the preset anomaly threshold, the anomaly occurrence time of the substation to be detected is located based on the anomaly detection score and the sliding window method. Based on the anomaly detection score and the time period in which the anomaly occurred, the anomaly detection result of the substation to be tested is obtained.

6. A device for detecting abnormalities in power equipment based on voiceprint, characterized in that, It includes a time-frequency feature acquisition module, a voiceprint feature extraction module, a detection model training module, and a device anomaly detection module; The time-frequency feature acquisition module is used to acquire historical voiceprint feature data of multiple power equipment, and extract local time-frequency features of the historical voiceprint feature data based on a two-dimensional convolutional network to obtain the first voiceprint feature. The voiceprint feature extraction module is used to extract features from the first voiceprint feature based on a bidirectional recurrent neural network to obtain the second voiceprint feature; The detection model training module is used to perform multiple iterations of training on the preset initial voiceprint detection model based on the second voiceprint feature. During the iteration, the model parameters of the initial voiceprint detection model are jointly optimized based on contrastive learning and perceptron classification until the model parameters meet the preset values, and the training is completed to obtain the voiceprint detection model. The equipment anomaly detection module is used to acquire the real-time voiceprint features of the substation equipment to be detected, and input the real-time voiceprint features into the voiceprint detection model to obtain the anomaly detection result of the substation equipment to be detected.

7. The abnormal detection device for substation equipment based on voiceprint according to claim 6, characterized in that, The time-frequency feature acquisition module includes a voiceprint feature acquisition submodule; the voiceprint feature acquisition submodule includes an audio acquisition and segmentation unit and a sequence time-frequency analysis unit; The audio acquisition and segmentation unit is used to acquire historical audio data from multiple power equipment and segment the historical audio data into multiple sample audio sequences according to a preset sequence length. The sequence time-frequency analysis unit is used to perform adaptive time-frequency analysis on the multiple sample audio sequences to obtain historical voiceprint feature data corresponding to the multiple power equipment.

8. The voiceprint-based transformer anomaly detection device according to claim 6, characterized in that, The time-frequency feature acquisition module includes a time-frequency feature extraction submodule; the time-frequency feature extraction submodule includes a voiceprint feature enhancement unit and a time-frequency feature extraction unit; The voiceprint feature enhancement unit is used to perform data enhancement on the historical voiceprint feature data based on a mask to obtain enhanced voiceprint feature data. The time-frequency feature extraction unit is used to downsample the enhanced voiceprint feature data based on a two-dimensional convolutional network to extract local time-frequency features and obtain the first voiceprint feature.

9. The abnormal detection device for substation equipment based on voiceprint according to claim 6, characterized in that, The detection model training module includes a voiceprint feature pooling submodule and a detection model training submodule; The voiceprint feature pooling submodule is used to perform average pooling on the second voiceprint feature to obtain the third voiceprint feature. The detection model training submodule is used to perform multiple iterations of training on a preset initial voiceprint detection model based on the third voiceprint feature. During iteration, the joint total loss is calculated based on contrastive learning and perceptron classification, and the model parameters are optimized according to the joint total loss until the model parameters meet the preset values, thus completing the training and obtaining the voiceprint detection model. The joint total loss includes contrastive loss based on contrastive learning and cross-entropy loss based on perceptron classification. The joint total loss is as follows: L=α*L contrast +(1-a)*L ce ; Among them, L,L contrast ,L ce These are the joint total loss, the contrastive loss, and the cross-entropy loss, respectively, with α being the loss balance parameter. The comparison loss is specifically as follows: Among them, H c H is the third voiceprint feature based on the mapping output of a multilayer perceptron. + H - These are the positive and negative samples of the voiceprint features, respectively. sim(a,b) is the similarity between input a and input b, and τ is the temperature parameter.

10. The voiceprint-based transformer anomaly detection device according to claim 6, characterized in that, The equipment anomaly detection module includes an anomaly score acquisition submodule, an anomaly time period location submodule, and an anomaly result acquisition submodule; The anomaly score acquisition submodule is used to input the real-time voiceprint features into the voiceprint detection model to obtain the anomaly detection score of the substation to be detected. The abnormal time period location submodule is used to locate the abnormal time period of the substation to be detected based on the abnormal detection score and the sliding window method when the abnormal detection score is greater than the preset abnormal threshold. The abnormal result acquisition submodule is used to obtain the abnormal detection result of the substation to be detected based on the abnormal detection score and the abnormal occurrence time period.

Images