Unsupervised detection method and device for abnormal sound of distribution network pole switch

Abstract

In the power distribution network pole switch abnormal sound unsupervised detection method and device provided in the application, the frequency domain features and time domain features corresponding to the audio signal of the power distribution network pole switch are determined; the robust principal component analysis method is used to perform dimension reduction processing on the frequency domain features and time domain features corresponding to the audio signal, and target time domain features and target frequency domain features are determined; the target time domain features and target frequency domain features are input into a multi-scale signal adjustment autoencoder model and an NKNN-based proximity model to obtain first detection results and second detection results; when the first detection results and the second detection results are inconsistent, the target time domain features and the target frequency domain features are input into a multivariate Gaussian mixture model to obtain third detection results; based on a voting strategy, the first detection results, the second detection results and the third detection results are voted, and the target detection results of the audio signal are determined according to the voting results. In this way, the detection accuracy of the abnormal sound of the power distribution network pole switch can be improved.

Description

Technical Field

[0001] This application relates to the field of power system technology, and in particular to an unsupervised detection method and device for abnormal sounds of pole-mounted switches in power distribution networks. Background Technology

[0002] The rapid development of industry has led to a continuous increase in my country's electricity demand, placing higher demands on power transmission. Regular operation, inspection, and maintenance of power distribution lines and equipment are fundamental to ensuring the normal and safe operation of the power distribution network. Power distribution operators need to conduct regular inspections to identify threats to the normal and safe transmission of electricity. These inspections mainly cover overhead lines, equipment and channels, cable lines and channels, substations, and switchyards. Audio monitoring methods offer advantages such as high efficiency and low cost. Sound-based fault detection and early warning can issue warnings in advance, allowing for timely maintenance. In the operation and maintenance of power distribution lines, the detection of pole-mounted switchgear includes checking for abnormal sounds. Therefore, utilizing audio signal acquisition technology, pattern recognition, machine learning, and other computer methods to monitor abnormal sounds from pole-mounted switches can save manpower and time, and make the maintenance process more efficient and convenient.

[0003] Current abnormal sound detection steps include sound data collection, data preprocessing, feature extraction, and supervised or unsupervised classification. There is no unified standard for which sound features to extract. Both time-domain and frequency-domain features reflect the characteristics of audio from different perspectives. However, excessively high feature dimensionality can lead to slower model execution and decreased accuracy. Therefore, it is necessary to reduce feature dimensionality while comprehensively considering the characteristics of all features. Consequently, in existing technologies, collecting abnormal sound samples from pole-mounted switches is difficult, and the sample size is small, resulting in low applicability of detection models designed based on supervised learning, ultimately leading to low accuracy in judging whether the sound is abnormal. Summary of the Invention

[0004] The purpose of this application is to at least solve one of the aforementioned technical defects, particularly the technical defect of low accuracy in judging whether a sound is abnormal in the prior art.

[0005] Firstly, this application provides an unsupervised detection method for abnormal sounds from pole-mounted switches in a power distribution network, the method comprising:

[0006] Determine the frequency domain and time domain characteristics of the audio signal corresponding to the switch on the distribution network pole;

[0007] Robust principal component analysis is used to perform feature dimensionality reduction on the frequency domain features and time domain features corresponding to the audio signal, and target time domain features and target frequency domain features are selected from the time domain features and frequency domain features after feature dimensionality reduction, respectively.

[0008] The target time-domain features and the target frequency-domain features are input into a pre-established multi-scale signal modulated autoencoder model and a pre-established NKNN-based proximity model to obtain a first detection result of the multi-scale signal modulated autoencoder model and a second detection result of the NKNN-based proximity model.

[0009] When the first detection result and the second detection result are inconsistent, the target time domain features and the target frequency domain features are input into a pre-established multivariate Gaussian mixture model to obtain the third detection result of the multivariate Gaussian mixture model;

[0010] Based on a pre-set voting strategy, the first detection result, the second detection result, and the third detection result are voted on, and the target detection result of the audio signal is determined according to the voting results.

[0011] In one embodiment, the step of determining the frequency domain characteristics and time domain characteristics corresponding to the audio signal of the pole-mounted switch in the distribution network includes:

[0012] Acquire the audio signal of the pole-mounted switch in the power distribution network;

[0013] The audio signal is denoised, and the denoised audio signal is then subjected to audio frame segmentation.

[0014] Windowing is applied to the audio signal after audio frame segmentation.

[0015] The frequency domain features and time domain features of the windowed audio signal are extracted respectively to obtain the frequency domain features and time domain features of the audio signal of the switch on the power distribution network pole.

[0016] In one embodiment, the step of using robust principal component analysis to perform feature dimensionality reduction processing on the frequency domain features and time domain features corresponding to the audio signal includes:

[0017] Construct the original frequency domain feature matrix of the frequency domain features corresponding to the audio signal, and construct the original time domain feature matrix of the time domain features corresponding to the audio signal;

[0018] The original feature matrix in the frequency domain is decomposed into a low-rank component and a sparse component in the frequency domain, and the original feature matrix in the time domain is decomposed into a low-rank component and a sparse component in the time domain.

[0019] Based on the frequency domain low-rank component and the frequency domain sparse component, the frequency domain features after dimensionality reduction are determined, and based on the time domain low-rank component and the time domain sparse component, the time domain features after dimensionality reduction are determined.

[0020] In one embodiment, the step of selecting the target time-domain feature and the target frequency-domain feature from the time-domain features and frequency-domain features after feature dimensionality reduction processing includes:

[0021] Determine multiple time-domain feature values ​​corresponding to the time-domain features after dimensionality reduction, and determine multiple frequency-domain feature values ​​corresponding to the frequency-domain features after dimensionality reduction;

[0022] The time-domain features corresponding to the first N time-domain feature values ​​after dimensionality reduction are selected as the target time-domain features, and the frequency-domain features corresponding to the first M frequency-domain feature values ​​after dimensionality reduction are selected as the target frequency-domain features, in descending order of their numerical values. Here, N and M are both positive integers greater than 1, and N and M are different.

[0023] In one embodiment, if the detection categories corresponding to the first detection result, the second detection result, and the third detection result are all audio normal or audio abnormal, then the step of voting on the first detection result, the second detection result, and the third detection result based on a pre-set voting strategy, and determining the target detection result of the audio signal according to the voting results, includes:

[0024] Based on the first detection result, the second detection result, and the third detection result, the number of votes corresponding to the detection category of normal audio and the number of votes corresponding to the detection category of abnormal audio are counted.

[0025] The detection category with the most votes is taken as the target detection result.

[0026] In one embodiment, if the first detection result, the second detection result, and the third detection result all include the probability corresponding to the detection category of normal audio and the probability corresponding to the detection category of abnormal audio, then the step of voting on the first detection result, the second detection result, and the third detection result based on a pre-set voting strategy, and determining the target detection result of the audio signal according to the voting result, includes:

[0027] The first confidence level of the multi-scale signal conditioning autoencoder model, the second confidence level of the NKNN-based proximity model, and the third confidence level of the multivariate Gaussian mixture model are determined respectively.

[0028] For the detection category of normal audio, the weighted probability corresponding to the detection category of normal audio is determined by weighted summation based on the first confidence level, the second confidence level, and the third confidence level, as well as the probability corresponding to the detection category in the first detection result, the probability corresponding to the detection category in the second detection result, and the probability corresponding to the detection category in the third detection result.

[0029] For the detection category of audio anomaly, the weighted probability corresponding to the detection category of audio anomaly is determined by weighted summation based on the first confidence level, the second confidence level, and the third confidence level, as well as the probability corresponding to the detection category in the first detection result, the probability corresponding to the detection category in the second detection result, and the probability corresponding to the detection category in the third detection result.

[0030] Based on the weighted probabilities corresponding to the detection categories of normal audio and abnormal audio, the detection category with the larger weighted probability is taken as the target detection result.

[0031] In one embodiment, the method further includes:

[0032] When the first detection result and the second detection result are consistent, the first detection result or the second detection result shall be taken as the target detection result of the audio signal.

[0033] Secondly, this application provides an unsupervised detection device for abnormal sounds of pole-mounted switches in power distribution networks, the device comprising:

[0034] The feature determination module is used to determine the frequency domain and time domain characteristics of the audio signal corresponding to the pole-mounted switch in the distribution network;

[0035] The target feature selection module is used to perform feature dimensionality reduction processing on the frequency domain features and time domain features corresponding to the audio signal using a robust principal component analysis method, and select target time domain features and target frequency domain features from the time domain features and frequency domain features after feature dimensionality reduction processing, respectively.

[0036] The detection result acquisition module is used to input the target time-domain features and the target frequency-domain features into a pre-established multi-scale signal modulated autoencoder model, and into a pre-established NKNN-based proximity model, to obtain the first detection result of the multi-scale signal modulated autoencoder model and the second detection result of the NKNN-based proximity model;

[0037] The target feature input module is used to input the target time-domain features and the target frequency-domain features into a pre-established multivariate Gaussian mixture model when the first detection result and the second detection result are inconsistent, so as to obtain the third detection result of the multivariate Gaussian mixture model;

[0038] The target detection result determination module is used to vote on the first detection result, the second detection result, and the third detection result based on a pre-set voting strategy, and determine the target detection result of the audio signal based on the voting result.

[0039] Thirdly, this application provides a storage medium storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps of the unsupervised detection method for abnormal sounds of pole-mounted switches in the power distribution network described in any of the above embodiments.

[0040] Fourthly, this application provides a computer device, including: one or more processors, and a memory;

[0041] The memory stores computer-readable instructions, which, when executed by the one or more processors, perform the steps of the unsupervised detection method for abnormal sounds of pole-mounted switches in the power distribution network as described in any of the above embodiments.

[0042] As can be seen from the above technical solutions, the embodiments of this application have the following advantages:

[0043] In the unsupervised detection method and apparatus for abnormal sounds from pole-mounted switches in the distribution network provided in this application, the frequency domain features and time domain features corresponding to the audio signal of the pole-mounted switch are first determined. Then, robust principal component analysis is used to perform feature dimensionality reduction processing on the frequency domain features and time domain features corresponding to the audio signal. Target time domain features and target frequency domain features are selected from the dimensionality-reduced time domain features and frequency domain features, respectively. The target time domain features and target frequency domain features are then input into a pre-established multi-scale signal conditioning autoencoder model and a pre-established NKNN-based proximity model to obtain a first detection result from the multi-scale signal conditioning autoencoder model and a second detection result from the NKNN-based proximity model. When the first and second detection results are inconsistent, the target time domain features and target frequency domain features are input into a pre-established multivariate Gaussian mixture model to obtain a third detection result from the multivariate Gaussian mixture model. Based on a pre-set voting strategy, the first, second, and third detection results are voted on, and the target detection result of the audio signal is determined according to the voting results. By employing robust principal component analysis, the dimensionality of the original time-domain and frequency-domain features can be reduced, thus mitigating problems such as slower model operation and decreased accuracy caused by excessively high feature dimensionality, thereby improving the model's efficiency and accuracy. By selecting target time-domain and target frequency-domain features, the most representative features can be chosen from multiple features, improving the model's discriminative ability. Using a multi-scale signal-adjusted autoencoder model and an NKNN-based proximity model, two detection results can be obtained. In cases where the first and second detection results are inconsistent, using a multivariate Gaussian mixture model can further improve the sound detection accuracy. By comprehensively evaluating multiple detection results based on a pre-defined voting strategy, the model's accuracy and robustness can be further improved. Thus, by fully considering techniques such as feature dimensionality reduction, multi-model fusion, and unsupervised learning, and combining them with a voting strategy, this approach exhibits high detection accuracy and applicability, and can be effectively applied to the unsupervised detection of abnormal sounds from pole-mounted switches in power distribution networks. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 A flowchart illustrating the unsupervised detection method for abnormal sounds from pole-mounted switches in a power distribution network provided in this application embodiment;

[0046] Figure 2A flowchart illustrating the steps for determining the frequency domain and time domain characteristics of the audio signal corresponding to a pole-mounted switch in a distribution network, as provided in an embodiment of this application.

[0047] Figure 3 A flowchart illustrating the feature dimensionality reduction processing steps provided in the embodiments of this application;

[0048] Figure 4 This is a schematic diagram of the unsupervised detection device for abnormal sounds of pole-mounted switches in a power distribution network provided in an embodiment of this application.

[0049] Figure 5 This is a schematic diagram of the internal structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0050] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0051] This application provides an unsupervised detection method for abnormal sounds from pole-mounted switches in a power distribution network. The following embodiments illustrate this method using a control device in a power distribution network as an example. It is understood that the control device can be any device with data processing capabilities, including but not limited to a single server, server cluster, personal laptop, desktop computer, etc. Figure 1 As shown, this application provides an unsupervised detection method for abnormal sounds from pole-mounted switches in a power distribution network. The method includes:

[0052] S101: Determine the frequency domain and time domain characteristics of the audio signal corresponding to the switch on the distribution network pole.

[0053] Among them, pole-mounted switches in power distribution networks refer to power switches installed on power distribution network poles. They are important equipment in the power system, mainly responsible for controlling and protecting the operation of power equipment. During the operation of the power system, the opening and closing operations of the switches generate specific audio signals. Frequency domain characteristics refer to the distribution and changes of the audio signal in the frequency domain. Time domain characteristics refer to the changes of the audio signal in the time domain. Since different characteristics can reflect different features of the audio signal from different perspectives, features in the time domain and frequency domain can be extracted separately. The extracted time domain features can include short-time average amplitude, short-time zero-crossing rate, short-time autocorrelation function, short-time average amplitude difference function, and linear predictive coding. Frequency domain features can include perceptual linear prediction, Mel-frequency cepstral coefficients, and power-regularized cepstral coefficients.

[0054] In this step, the audio signal from the switch on the distribution network pole can be collected first, and then the frequency domain characteristics and time domain characteristics corresponding to the audio signal can be determined. In one example, the audio signal can be directly analyzed in the time domain to extract time domain characteristics such as short-time average amplitude, short-time zero-crossing rate, short-time autocorrelation function, short-time average amplitude difference function, and linear predictive coding. The audio signal can then be converted from the time domain to the frequency domain using methods such as Fourier transform or short-time Fourier transform, and then frequency domain characteristics such as perceptual linear prediction, Mel-frequency cepstral coefficients, and power regularized cepstral coefficients can be extracted.

[0055] S102: Using robust principal component analysis, the frequency domain features and time domain features corresponding to the audio signal are subjected to feature dimensionality reduction processing, and target time domain features and target frequency domain features are selected from the time domain features and frequency domain features after feature dimensionality reduction processing, respectively.

[0056] Robust Principal Component Analysis (RPCA) is a method for high-dimensional data analysis, an improved version of the traditional Principal Component Analysis (PCA) method. While traditional PCA is sensitive to outliers, RPCA handles data containing outliers much better. Feature reduction is a method used to reduce the dimensionality of data, primarily aiming to reduce data complexity, improve data processing efficiency, and preserve the main features of the data as much as possible.

[0057] In this step, RPCA processing can be performed on the frequency and time domain features of the audio signal to extract the main features and remove the influence of outliers. Then, target time and frequency domain features can be selected from the dimensionality-reduced time and frequency domain features. These target time and frequency domain features are understood to be the features most relevant to the switching state. In one example, the method for selecting target features can be based on statistical analysis or a feature selection method based on machine learning.

[0058] S103: Input the target time-domain features and the target frequency-domain features into a pre-established multi-scale signal conditioning autoencoder model and into a pre-established NKNN-based proximity model to obtain the first detection result of the multi-scale signal conditioning autoencoder model and the second detection result of the NKNN-based proximity model.

[0059] Among them, the multi-scale signal conditioning autoencoder model is an unsupervised learning model based on deep learning, which can be used to process time-series data, such as audio signals. This model typically consists of an encoder and a decoder. The encoder encodes the input data into a low-dimensional latent space, and the decoder decodes the low-dimensional latent space back into the original data space. The decoded signal can then be compared with normal or abnormal audio signals to obtain the audio signal detection result. The NKNN (Nearest K-Nearest Neighbors) proximity model is a classification model. NKNN is an improved version of the K-Nearest Neighbors (KNN) algorithm, which considers not only the K nearest samples but also the proximity of each sample. The basic idea of ​​the NKNN model is that if most of a sample's K nearest neighbors belong to a certain class, then the sample is also likely to belong to that class. The first and second detection results refer to the conclusions drawn by the model after detecting the audio signal. In one example, the first and second detection results are of the same type. This can be a category label, which can be audio normal or audio abnormal, or an audio abnormality score or an audio abnormality index. It can also be multiple category labels and the probability corresponding to each category label, such as audio normal, audio abnormal, and fault degree.

[0060] In this step, the target time-domain features and target frequency-domain features are first input into a multi-scale signal conditioning autoencoder model to obtain the first detection result. Simultaneously, the target time-domain features and target frequency-domain features are input into a proximity model based on NKNN to obtain the second detection result. By comparing the first and second detection results, the state of the pole-mounted switch in the distribution network can be determined more accurately.

[0061] In one example, during the process of comparing the decoded signal with a normal or abnormal audio signal, the multi-scale signal conditioning autoencoder model can pre-set a threshold for normal audio. By comparing the decoded signal with the normal audio signal, the signal similarity can be obtained. If the signal similarity is greater than or equal to the threshold for normal audio, the detection result can be determined to be that the original audio signal is normal. If the signal similarity is less than the threshold for normal audio, the detection result can be determined to be that the original audio signal is abnormal.

[0062] S104: When the first detection result and the second detection result are inconsistent, the target time domain features and the target frequency domain features are input into the pre-established multivariate Gaussian mixture model to obtain the third detection result of the multivariate Gaussian mixture model.

[0063] The Gaussian Mixture Model (GMM) is a statistical model that assumes data is generated from a mixture of several Gaussian distributions. In the GMM, each Gaussian distribution corresponds to a category or a subset. The data generation process involves first selecting a Gaussian distribution from the mixture with a certain probability, and then sampling data from this Gaussian distribution.

[0064] In this step, if the first and second detection results are inconsistent, the target's time-domain features and frequency-domain features are input into a multivariate Gaussian mixture model to obtain a third detection result. It can be understood that the third detection result, the first detection result, and the second detection result are all of the same type.

[0065] S105: Based on a pre-set voting strategy, vote on the first detection result, the second detection result, and the third detection result, and determine the target detection result of the audio signal based on the voting results.

[0066] Voting strategy is a decision rule used to determine the final result based on multiple independent judgments. It can be used to integrate the results of multiple detectors to obtain a more reliable and accurate final detection result. The voting strategy can be set according to the specific circumstances; for example, it can be soft voting, hard voting, or a combination of both.

[0067] In this step, the first detection result, the second detection result, and the third detection result are voted on according to the voting strategy, so as to determine the target detection result of the audio signal from the first detection result, the second detection result, and the third detection result based on the voting results.

[0068] In the unsupervised detection method and apparatus for abnormal sounds from pole-mounted switches in the distribution network provided in this application, the frequency domain features and time domain features corresponding to the audio signal of the pole-mounted switch are first determined. Then, robust principal component analysis is used to perform feature dimensionality reduction processing on the frequency domain features and time domain features corresponding to the audio signal. Target time domain features and target frequency domain features are selected from the dimensionality-reduced time domain features and frequency domain features, respectively. The target time domain features and target frequency domain features are then input into a pre-established multi-scale signal conditioning autoencoder model and a pre-established NKNN-based proximity model to obtain a first detection result from the multi-scale signal conditioning autoencoder model and a second detection result from the NKNN-based proximity model. When the first and second detection results are inconsistent, the target time domain features and target frequency domain features are input into a pre-established multivariate Gaussian mixture model to obtain a third detection result from the multivariate Gaussian mixture model. Based on a pre-set voting strategy, the first, second, and third detection results are voted on, and the target detection result of the audio signal is determined according to the voting results. By employing robust principal component analysis, the dimensionality of the original time-domain and frequency-domain features can be reduced, thus mitigating problems such as slower model operation and decreased accuracy caused by excessively high feature dimensionality, thereby improving the model's efficiency and accuracy. By selecting target time-domain and frequency-domain features, the most representative features can be chosen from multiple features, improving the model's discriminative ability. Using a multi-scale signal-adjusted autoencoder model and an NKNN-based proximity model, two different detection results can be obtained. In cases where the first and second detection results are inconsistent, using a multivariate Gaussian mixture model can further improve the sound detection accuracy. By comprehensively evaluating multiple detection results based on a pre-defined voting strategy, the model's accuracy and robustness can be further improved. Thus, by fully considering techniques such as feature dimensionality reduction, multi-model fusion, and unsupervised learning, and combining them with a voting strategy, this approach exhibits high detection accuracy and applicability, and can be effectively applied to the unsupervised detection of abnormal sounds from pole-mounted switches in power distribution networks.

[0069] like Figure 2 As shown, in one embodiment, the step of determining the frequency domain characteristics and time domain characteristics corresponding to the audio signal of the switch on the distribution network pole includes:

[0070] S201: Obtain the audio signal of the switch on the power distribution pole;

[0071] S202: Denoise the audio signal and perform audio frame segmentation on the denoised audio signal;

[0072] S203: Window the audio signal after audio framing;

[0073] S204: Extract the frequency domain features and time domain features of the windowed audio signal respectively to obtain the frequency domain features and time domain features corresponding to the audio signal of the switch on the power distribution network pole.

[0074] Audio framing refers to dividing a long audio signal into short segments, each called a frame. This transforms a long audio signal into a short-term changing signal, facilitating subsequent processing. Common audio framing methods include overlapping windowing and non-overlapping windowing. Windowing multiplies each frame of the audio signal by a window function to reduce frequency domain leakage and increase smoothness. Commonly used window functions include the Hamming window, Heining window, and Blackman window. The windowed audio signal can then be used for frequency domain analysis and time domain analysis.

[0075] Specifically, the audio signal from the switch on the power distribution network pole is acquired, and then subjected to denoising, audio framing, and windowing processes sequentially. Frequency and time domain features are extracted from the windowed audio signal to obtain the corresponding frequency and time domain features of the audio signal from the switch on the power distribution network pole. In one example, a microphone or other recording device can be used to acquire the audio signal from the switch on the power distribution network pole. In another example, denoising algorithms based on wavelet transform or spectral subtraction can be used to denoise the audio signal.

[0076] In one example, the raw audio signal acquired from the pole-mounted switch contains environmental noise. The DeepXi algorithm, based on the minimum mean square error method using logarithmic spectral analysis, is used to denoise the audio signal. The denoised audio signal is then framed using a sliding window method to improve audio resolution. Each frame of the audio signal is then windowed using a Hamming window, thus completing the audio signal preprocessing steps.

[0077] In this embodiment, by extracting the frequency domain and time domain features of the audio signal from the pole-mounted switch in the distribution network, various feature information related to the audio signal can be obtained, thereby improving the accuracy and reliability of audio signal diagnosis.

[0078] like Figure 3 As shown, in one embodiment, the step of using robust principal component analysis to perform feature dimensionality reduction processing on the frequency domain features and time domain features corresponding to the audio signal includes:

[0079] S301: Construct the original frequency domain feature matrix of the frequency domain features corresponding to the audio signal, and construct the original time domain feature matrix of the time domain features corresponding to the audio signal;

[0080] S302: Decompose the original frequency domain feature matrix into a low-rank frequency domain component and a sparse frequency domain component, and decompose the original time domain feature matrix into a low-rank time domain component and a sparse time domain component.

[0081] S303: Determine the frequency domain features after dimensionality reduction based on the frequency domain low-rank component and the frequency domain sparse component, and determine the time domain features after dimensionality reduction based on the time domain low-rank component and the time domain sparse component.

[0082] The frequency domain original feature matrix is ​​a matrix containing frequency domain features, with each row representing the frequency domain feature vector of a sample. The time domain original feature matrix is ​​a matrix containing time domain features, with each row representing the time domain feature vector of a sample. The frequency domain low-rank component refers to a low-rank matrix that can represent the frequency domain original feature matrix well. The frequency domain sparse component refers to a sparse matrix that can represent noise or anomaly information in the frequency domain original feature matrix. The time domain low-rank component refers to a low-rank matrix that can represent the time domain original feature matrix well. The time domain sparse component refers to a sparse matrix that can represent noise or anomaly information in the time domain original feature matrix.

[0083] Specifically, firstly, the original frequency domain feature matrix of the audio signal's frequency domain features and the original time domain feature matrix of the audio signal's time domain features are constructed. Next, the original frequency domain feature matrix is ​​decomposed to obtain low-rank frequency domain components and sparse frequency domain components. The low-rank frequency domain components retain the main information in the original frequency domain feature matrix, while the sparse frequency domain components have higher noise suppression and anomaly detection capabilities. Similarly, the original time domain feature matrix is ​​decomposed to obtain low-rank time domain components and sparse time domain components. The low-rank time domain components retain the main information in the original time domain feature matrix, while the sparse time domain components have higher noise suppression and anomaly detection capabilities. Based on the low-rank frequency domain components and sparse time domain components, the dimensionality-reduced frequency domain features and the dimensionality-reduced time domain features can be determined.

[0084] In one example, the original frequency domain feature matrix X is decomposed into a low-rank frequency domain component L and a sparse frequency domain component S:

[0085] X = L + S

[0086] RPCA achieves more accurate dimensionality reduction by solving a convex matrix factorization optimization problem. The objective function of this convex problem is:

[0087]

[0088] In the formula, the parameters λ and 1-norm are parameters of the interference term. Generally, a lower rank in L means a larger error, i.e., a larger sum of non-zero elements in S. The objective function aims to achieve optimization by finding a balance between these two factors.

[0089] In this embodiment, the frequency and time domain features of audio signals typically have high dimensionality. High dimensionality leads to increased computational complexity and reduces algorithm efficiency. Robust principal component analysis (BPM) is used for feature dimensionality reduction, which reduces the feature dimensionality, simplifying data analysis and processing, and improving algorithm efficiency. Audio signals contain rich information in both the frequency and time domains, but some of this information may be redundant or unimportant. BPM extracts the most representative feature information, removing redundant and unimportant information, thereby improving the algorithm's accuracy.

[0090] In one embodiment, the step of selecting the target time-domain feature and the target frequency-domain feature from the time-domain features and frequency-domain features after feature dimensionality reduction processing includes:

[0091] Determine multiple time-domain feature values ​​corresponding to the time-domain features after dimensionality reduction, and determine multiple frequency-domain feature values ​​corresponding to the frequency-domain features after dimensionality reduction;

[0092] The time-domain features corresponding to the first N time-domain feature values ​​after dimensionality reduction are selected as the target time-domain features, and the frequency-domain features corresponding to the first M frequency-domain feature values ​​after dimensionality reduction are selected as the target frequency-domain features, in descending order of their numerical values. Here, N and M are both positive integers greater than 1, and N and M are different.

[0093] Specifically, first, we determine multiple time-domain feature values ​​corresponding to the dimensionality-reduced time-domain features, and multiple frequency-domain feature values ​​corresponding to the dimensionality-reduced frequency-domain features. We then sort the time-domain feature values ​​in descending order of value and select the top N time-domain feature values ​​corresponding to the dimensionality-reduced time-domain features as the target time-domain features. Similarly, we sort the frequency-domain feature values ​​in descending order of value and select the top M frequency-domain feature values ​​corresponding to the dimensionality-reduced frequency-domain features as the target frequency-domain features. The specific values ​​of N and M can be determined based on the actual situation. Both N and M are positive integers greater than 1, and N and M cannot be equal to ensure that the selected time-domain features and frequency-domain features are not completely identical.

[0094] In this embodiment, by selecting the top N time-domain feature values ​​and the top M frequency-domain feature values, the most representative features can be screened out, thereby improving the representativeness of the features. The most representative time-domain and frequency-domain features better reflect the speech information of the audio signal, enabling the algorithm to more accurately identify and classify different speech signals, thus improving the accuracy and discriminative power of the algorithm.

[0095] In one embodiment, if the detection categories corresponding to the first detection result, the second detection result, and the third detection result are all audio normal or audio abnormal, then the step of voting on the first detection result, the second detection result, and the third detection result based on a pre-set voting strategy, and determining the target detection result of the audio signal according to the voting results, includes:

[0096] Based on the first detection result, the second detection result, and the third detection result, the number of votes corresponding to the detection category of normal audio and the number of votes corresponding to the detection category of abnormal audio are counted.

[0097] The detection category with the most votes is taken as the target detection result.

[0098] Specifically, based on pre-defined classification criteria, the first, second, and third detection results are categorized into two types: normal audio and abnormal audio. A voting system is then implemented for each category, specifically counting the number of votes for the "normal audio" category and the "abnormal audio" category. The category with the most votes is selected as the target detection result. Furthermore, if the number of votes for "normal audio" and "abnormal audio" is equal, the final result can be determined according to pre-defined rules, such as defaulting to "abnormal audio" as the target detection result or conducting another vote.

[0099] In this embodiment, voting among multiple detection results can effectively reduce the false positive rate of a single detection result and improve the accuracy of the final detection result. Determining the target detection result by statistically analyzing the detection category with the most votes can further reduce the probability of incorrect judgments.

[0100] In one embodiment, if the first detection result, the second detection result, and the third detection result all include the probability corresponding to the detection category of normal audio and the probability corresponding to the detection category of abnormal audio, then the step of voting on the first detection result, the second detection result, and the third detection result based on a pre-set voting strategy, and determining the target detection result of the audio signal according to the voting result, includes:

[0101] The first confidence level of the multi-scale signal conditioning autoencoder model, the second confidence level of the NKNN-based proximity model, and the third confidence level of the multivariate Gaussian mixture model are determined respectively.

[0102] For the detection category of normal audio, the weighted probability corresponding to the detection category of normal audio is determined by weighted summation based on the first confidence level, the second confidence level, and the third confidence level, as well as the probability corresponding to the detection category in the first detection result, the probability corresponding to the detection category in the second detection result, and the probability corresponding to the detection category in the third detection result.

[0103] For the detection category of audio anomaly, the weighted probability corresponding to the detection category of audio anomaly is determined by weighted summation based on the first confidence level, the second confidence level, and the third confidence level, as well as the probability corresponding to the detection category in the first detection result, the probability corresponding to the detection category in the second detection result, and the probability corresponding to the detection category in the third detection result.

[0104] Based on the weighted probabilities corresponding to the detection categories of normal audio and abnormal audio, the detection category with the larger weighted probability is taken as the target detection result.

[0105] Specifically, firstly, for the multi-scale signal conditioning autoencoder model, the NKNN-based proximity model, and the multivariate Gaussian mixture model, the first confidence level of the multi-scale signal conditioning autoencoder model, the second confidence level of the NKNN-based proximity model, and the third confidence level of the multivariate Gaussian mixture model are determined, respectively. For the detection category of normal audio, the weighted probability corresponding to the detection category of normal audio is calculated by weighted summation based on the first, second, and third confidence levels, as well as the probabilities corresponding to this detection category in the first, second, and third detection results. For the detection category of abnormal audio, the weighted probability corresponding to the detection category of abnormal audio is also calculated by weighted summation based on the first, second, and third confidence levels, as well as the probabilities corresponding to this detection category in the first, second, and third detection results. Based on the weighted probabilities corresponding to the detection categories of normal and abnormal audio, the detection category with the larger weighted probability is selected as the final target detection result.

[0106] In one example, the first confidence level of the multi-scale signal conditioning autoencoder model is set to 0.8, the second confidence level of the NKNN-based proximity model is set to 0.6, and the third confidence level of the multivariate Gaussian mixture model is set to 0.7. Assume the probability of the first detection result being normal audio is 0.8 and the probability of the second detection result being abnormal audio is 0.2; the probability of the third detection result being normal audio is 0.6 and the probability of the second detection result being abnormal audio is 0.4; and the probability of the fourth detection result being normal audio is 0.7 and the probability of the third detection result being abnormal audio is 0.3. The weighted probabilities of the normal audio detection category and the abnormal audio detection category are calculated using a weighted summation method. The weighted probability for the normal audio detection category is: 0.8*0.8 + 0.6*0.6 + 0.7*0.7 = 1.49, and the weighted probability for the abnormal audio detection category is: 0.8*0.2 + 0.6*0.4 + 0.7*0.3 = 0.61. The weighted probabilities of the detection categories for normal audio and abnormal audio are compared, and the detection category with the larger weighted probability is selected as the final target detection result, that is, the final target detection result is normal audio.

[0107] In this embodiment, each model has its unique characteristics and performance. By combining the confidence levels of multiple models, the normal or abnormal state of the audio can be evaluated more comprehensively, reducing the subjectivity and limitations of a single model. Each model provides a probability value for its corresponding detection category. By weighted summing of the probability information from different models, the probability of the audio being normal or abnormal can be estimated more accurately, reducing the possibility of misjudgment.

[0108] In one embodiment, the method further includes:

[0109] When the first detection result and the second detection result are consistent, the first detection result or the second detection result shall be taken as the target detection result of the audio signal.

[0110] Specifically, when the first and second detection results are consistent, either one can be used as the target detection result for the audio signal. This simplifies the judgment process and reduces computational load and complexity.

[0111] The following describes the unsupervised detection device for abnormal sounds of pole-mounted switches in a power distribution network provided in the embodiments of this application. The unsupervised detection device for abnormal sounds of pole-mounted switches in a power distribution network described below can be referred to in correspondence with the unsupervised detection method for abnormal sounds of pole-mounted switches in a power distribution network described above. Figure 4 As shown, this application provides an unsupervised detection device for abnormal sounds of pole-mounted switches in power distribution networks. The device includes:

[0112] The feature determination module 401 is used to determine the frequency domain characteristics and time domain characteristics of the audio signal corresponding to the pole-mounted switch in the distribution network;

[0113] The target feature selection module 402 is used to perform feature dimensionality reduction processing on the frequency domain features and time domain features corresponding to the audio signal using a robust principal component analysis method, and select target time domain features and target frequency domain features from the time domain features and frequency domain features after feature dimensionality reduction processing, respectively.

[0114] The detection result acquisition module 403 is used to input the target time domain features and the target frequency domain features into a pre-established multi-scale signal modulated autoencoder model, and into a pre-established NKNN-based proximity model, to obtain the first detection result of the multi-scale signal modulated autoencoder model and the second detection result of the NKNN-based proximity model.

[0115] The target feature input module 404 is used to input the target time domain features and the target frequency domain features into a pre-established multivariate Gaussian mixture model when the first detection result and the second detection result are inconsistent, so as to obtain the third detection result of the multivariate Gaussian mixture model;

[0116] The first target detection result determination module 405 is used to vote on the first detection result, the second detection result and the third detection result based on a pre-set voting strategy, and determine the target detection result of the audio signal according to the voting result.

[0117] In one embodiment, the feature determination module 401 includes:

[0118] An audio signal acquisition unit is used to acquire the audio signal of the pole-mounted switch in the power distribution network;

[0119] The first audio signal preprocessing unit is used to denoise the audio signal and perform audio frame segmentation on the denoised audio signal.

[0120] The second audio signal preprocessing unit is used to perform windowing processing on the audio signal after audio frame processing;

[0121] The feature determination unit is used to extract the frequency domain features and time domain features of the windowed audio signal, respectively, to obtain the frequency domain features and time domain features corresponding to the audio signal of the switch on the power distribution network pole.

[0122] In one embodiment, the target feature selection module 402 includes:

[0123] The feature matrix construction unit is used to construct the original frequency domain feature matrix of the frequency domain features corresponding to the audio signal, and to construct the original time domain feature matrix of the time domain features corresponding to the audio signal.

[0124] The feature matrix decomposition unit is used to decompose the original frequency domain feature matrix into a low-rank frequency domain component and a sparse frequency domain component, and to decompose the original time domain feature matrix into a low-rank time domain component and a sparse time domain component.

[0125] The dimension reduction feature determination unit is used to determine the dimension-reduced frequency domain features based on the frequency domain low-rank component and the frequency domain sparse component, and to determine the dimension-reduced time domain features based on the time domain low-rank component and the time domain sparse component.

[0126] In one embodiment, the target feature selection module 402 includes:

[0127] The eigenvalue determination unit is used to determine multiple time-domain eigenvalues ​​corresponding to the time-domain features after dimensionality reduction, and to determine multiple frequency-domain eigenvalues ​​corresponding to the frequency-domain features after dimensionality reduction.

[0128] The target feature selection unit is used to select the dimensionality-reduced time-domain features corresponding to the first N time-domain feature values ​​as the target time-domain features in descending order of the time-domain feature values, and to select the dimensionality-reduced frequency-domain features corresponding to the first M frequency-domain feature values ​​as the target frequency-domain features in descending order of the frequency-domain feature values, where N and M are both positive integers greater than 1, and N and M are different.

[0129] In one embodiment, if the detection categories corresponding to the first detection result, the second detection result, and the third detection result are all audio normal or audio abnormal, then the first target detection result determination module 405 includes:

[0130] The voting count unit is used to count the number of votes corresponding to the detection category of normal audio and the number of votes corresponding to the detection category of abnormal audio based on the first detection result, the second detection result and the third detection result;

[0131] The first target detection result determination unit is used to determine the detection category with the most votes as the target detection result.

[0132] In one embodiment, if the first detection result, the second detection result, and the third detection result all include the probability corresponding to the detection category of normal audio and the probability corresponding to the detection category of abnormal audio, then the first target detection result determination module 405 includes:

[0133] The confidence determination unit is used to determine the first confidence of the multi-scale signal conditioning autoencoder model, the second confidence of the NKNN-based proximity model, and the third confidence of the multivariate Gaussian mixture model, respectively.

[0134] The first weighted summation unit is used to determine the weighted probability corresponding to the detection category of normal audio by weighted summation based on the first confidence level, the second confidence level, the third confidence level, and the probability corresponding to the detection category in the first detection result, the probability corresponding to the detection category in the second detection result, and the probability corresponding to the detection category in the third detection result.

[0135] The second weighted summation unit is used to determine the weighted probability corresponding to the detection category of audio anomaly by weighted summation based on the first confidence level, the second confidence level, the third confidence level, and the probability corresponding to the detection category in the first detection result, the probability corresponding to the detection category in the second detection result, and the probability corresponding to the detection category in the third detection result.

[0136] The second target detection result determination unit is used to determine the target detection result based on the weighted probability corresponding to the detection category of normal audio and the weighted probability corresponding to the detection category of abnormal audio.

[0137] In one embodiment, the apparatus further includes:

[0138] The second target detection result determination module is used to take the first detection result or the second detection result as the target detection result of the audio signal when the first detection result and the second detection result are consistent.

[0139] In one embodiment, this application also provides a storage medium storing computer-readable instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of the unsupervised detection method for abnormal sounds of pole-mounted switches in a power distribution network as described in any of the above embodiments.

[0140] In one embodiment, this application also provides a computer device storing computer-readable instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of the unsupervised detection method for abnormal sounds of pole-mounted switches in a power distribution network as described in any of the above embodiments.

[0141] Indicatively, such as Figure 5 As shown, Figure 5 This is a schematic diagram of the internal structure of a computer device 500 provided in an embodiment of this application. The computer device 500 can be provided as a server. (Refer to...) Figure 5The computer device 500 includes a processing component 502, which further includes one or more processors, and memory resources represented by memory 501 for storing instructions, such as application programs, that can be executed by the processing component 502. The application programs stored in memory 501 may include one or more modules, each corresponding to a set of instructions. Furthermore, the processing component 502 is configured to execute instructions to perform the unsupervised detection method for abnormal sounds of pole-mounted switches in the power distribution network according to any of the above embodiments.

[0142] The computer device 500 may also include a power supply component 503 configured to perform power management of the computer device 500, a wired or wireless network interface 504 configured to connect the computer device 500 to a network, and an input / output (I / O) interface 505. The computer device 500 may operate on an operating system stored in memory 501, such as Windows Server™, Mac OS X™, Unix™, Linux™, Free BSD™, or similar.

[0143] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0144] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. In this document, "a," "an," "the," "the," and "its" may also include plural forms unless the context clearly indicates otherwise. "Multiple" refers to at least two, such as 2, 3, 5, or 8, etc. "And / or" includes any and all combinations of the related listed items.

[0145] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0146] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An unsupervised detection method for abnormal sounds from pole-mounted switches in a power distribution network, characterized in that, The method includes: Determine the frequency domain and time domain characteristics of the audio signal corresponding to the switch on the distribution network pole; Robust principal component analysis is used to perform feature dimensionality reduction on the frequency domain features and time domain features corresponding to the audio signal, and target time domain features and target frequency domain features are selected from the time domain features and frequency domain features after feature dimensionality reduction, respectively. The target time-domain features and the target frequency-domain features are input into a pre-established multi-scale signal modulated autoencoder model and a pre-established NKNN-based proximity model to obtain a first detection result of the multi-scale signal modulated autoencoder model and a second detection result of the NKNN-based proximity model. When the first detection result and the second detection result are inconsistent, the target time domain features and the target frequency domain features are input into a pre-established multivariate Gaussian mixture model to obtain the third detection result of the multivariate Gaussian mixture model; Based on a pre-set voting strategy, the first detection result, the second detection result, and the third detection result are voted on, and the target detection result of the audio signal is determined according to the voting results.

2. The unsupervised detection method for abnormal sounds of pole-mounted switches in a power distribution network according to claim 1, characterized in that, The steps for determining the frequency domain and time domain characteristics of the audio signal corresponding to the switch on the distribution network pole include: Acquire the audio signal of the pole-mounted switch in the power distribution network; The audio signal is denoised, and the denoised audio signal is then subjected to audio frame segmentation. Windowing is applied to the audio signal after audio frame segmentation. The frequency domain features and time domain features of the windowed audio signal are extracted respectively to obtain the frequency domain features and time domain features of the audio signal of the switch on the power distribution network pole.

3. The unsupervised detection method for abnormal sounds of pole-mounted switches in a power distribution network according to claim 1, characterized in that, The step of performing feature dimensionality reduction processing on the frequency domain and time domain features corresponding to the audio signal using robust principal component analysis includes: Construct the original frequency domain feature matrix of the frequency domain features corresponding to the audio signal, and construct the original time domain feature matrix of the time domain features corresponding to the audio signal; The original feature matrix in the frequency domain is decomposed into a low-rank component and a sparse component in the frequency domain, and the original feature matrix in the time domain is decomposed into a low-rank component and a sparse component in the time domain. Based on the frequency domain low-rank component and the frequency domain sparse component, the frequency domain features after dimensionality reduction are determined, and based on the time domain low-rank component and the time domain sparse component, the time domain features after dimensionality reduction are determined.

4. The unsupervised detection method for abnormal sounds of pole-mounted switches in a power distribution network according to claim 1, characterized in that, The step of selecting the target time-domain feature and the target frequency-domain feature from the time-domain features and frequency-domain features after feature dimensionality reduction processing includes: Determine multiple time-domain feature values ​​corresponding to the time-domain features after dimensionality reduction, and determine multiple frequency-domain feature values ​​corresponding to the frequency-domain features after dimensionality reduction; The time-domain features corresponding to the first N time-domain feature values ​​after dimensionality reduction are selected as the target time-domain features, and the frequency-domain features corresponding to the first M frequency-domain feature values ​​after dimensionality reduction are selected as the target frequency-domain features, in descending order of their numerical values. Here, N and M are both positive integers greater than 1, and N and M are different.

5. The unsupervised detection method for abnormal sounds of pole-mounted switches in a power distribution network according to claim 1, characterized in that, If the detection categories corresponding to the first detection result, the second detection result, and the third detection result are all audio normal or audio abnormal, then the step of voting on the first detection result, the second detection result, and the third detection result based on a pre-set voting strategy, and determining the target detection result of the audio signal according to the voting results, includes: Based on the first detection result, the second detection result, and the third detection result, the number of votes corresponding to the detection category of normal audio and the number of votes corresponding to the detection category of abnormal audio are counted. The detection category with the most votes is taken as the target detection result.

6. The unsupervised detection method for abnormal sounds of pole-mounted switches in a power distribution network according to claim 1, characterized in that, If the first detection result, the second detection result, and the third detection result all include the probability corresponding to the detection category of normal audio and the probability corresponding to the detection category of abnormal audio, then the step of voting on the first detection result, the second detection result, and the third detection result based on a pre-set voting strategy, and determining the target detection result of the audio signal according to the voting result, includes: The first confidence level of the multi-scale signal conditioning autoencoder model, the second confidence level of the NKNN-based proximity model, and the third confidence level of the multivariate Gaussian mixture model are determined respectively. For the detection category of normal audio, the weighted probability corresponding to the detection category of normal audio is determined by weighted summation based on the first confidence level, the second confidence level, and the third confidence level, as well as the probability corresponding to the detection category in the first detection result, the probability corresponding to the detection category in the second detection result, and the probability corresponding to the detection category in the third detection result. For the detection category of audio anomaly, the weighted probability corresponding to the detection category of audio anomaly is determined by weighted summation based on the first confidence level, the second confidence level, and the third confidence level, as well as the probability corresponding to the detection category in the first detection result, the probability corresponding to the detection category in the second detection result, and the probability corresponding to the detection category in the third detection result. Based on the weighted probabilities corresponding to the detection categories of normal audio and abnormal audio, the detection category with the larger weighted probability is taken as the target detection result.

7. The unsupervised detection method for abnormal sounds of pole-mounted switches in a power distribution network according to any one of claims 1 to 6, characterized in that, The method further includes: When the first detection result and the second detection result are consistent, the first detection result or the second detection result shall be taken as the target detection result of the audio signal.

8. An unsupervised detection device for abnormal sound of pole-mounted switches in a power distribution network, characterized in that, The device includes: The feature determination module is used to determine the frequency domain and time domain characteristics of the audio signal corresponding to the pole-mounted switch in the distribution network; The target feature selection module is used to perform feature dimensionality reduction processing on the frequency domain features and time domain features corresponding to the audio signal using a robust principal component analysis method, and select target time domain features and target frequency domain features from the time domain features and frequency domain features after feature dimensionality reduction processing, respectively. The detection result acquisition module is used to input the target time-domain features and the target frequency-domain features into a pre-established multi-scale signal modulated autoencoder model, and into a pre-established NKNN-based proximity model, to obtain the first detection result of the multi-scale signal modulated autoencoder model and the second detection result of the NKNN-based proximity model; The target feature input module is used to input the target time-domain features and the target frequency-domain features into a pre-established multivariate Gaussian mixture model when the first detection result and the second detection result are inconsistent, so as to obtain the third detection result of the multivariate Gaussian mixture model; The target detection result determination module is used to vote on the first detection result, the second detection result, and the third detection result based on a pre-set voting strategy, and determine the target detection result of the audio signal based on the voting result.

9. A storage medium, characterized in that: The storage medium stores computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps of the unsupervised detection method for abnormal sounds of pole-mounted switches in a power distribution network as described in any one of claims 1 to 7.

10. A computer device, characterized in that, include: One or more processors, and memory; The memory stores computer-readable instructions, which, when executed by the one or more processors, perform the steps of the unsupervised detection method for abnormal sounds of pole-mounted switches in a power distribution network as described in any one of claims 1 to 7.

Images