A switch cabinet partial discharge audio recognition method based on deep meta learning

Abstract

This paper presents a deep meta-learning-based audio recognition method for partial discharge in switchgear, belonging to the field of audio signal processing. Addressing the practical problems of large data differences and decreased cross-domain performance in current switchgear partial discharge audio recognition methods, a target recognition method based on deep meta-learning is proposed, which can effectively improve the audio recognition performance of switchgear partial discharge under complex conditions. The method includes the following steps: 1) preprocessing method for audio signals collected from simulated partial discharge phenomena in high-voltage switchgear; 2) deep model design; 3) application of deep meta-learning strategy. Based on deep learning and meta-learning, its main features are strong practicality and portability, enabling it to meet the cross-domain generalization requirements of audio recognition in real-world scenarios and improve the accuracy and applicability of switchgear partial discharge recognition methods.

Description

Technical Field

[0001] This invention belongs to the field of audio signal processing, specifically relating to a method for partial discharge audio recognition of switchgear based on deep meta-learning. Background Technology

[0002] High reliability of power equipment is crucial for the safety of ultra-large-scale power transmission and distribution and the power grid. High-voltage switchgear, mainly composed of high-voltage switches, is widely used in every aspect of the power grid, and its safety and reliability directly affect the safety of the power grid and the overall power system. The main reason for high-voltage switchgear failures is the lack of targeted and effective monitoring methods. Traditional fault detection in high-voltage switchgear requires power outages or relies on manual inspection, resulting in problems such as failure to handle faults in a timely manner and difficulty in detection. The main faults in high-voltage switchgear are insulation faults, and before insulation faults occur, phenomena such as partial discharge exist within the switchgear. Therefore, the operating status parameters of the high-voltage switchgear equipment can be obtained by real-time detection of partial discharge phenomena, thereby timely detection of potential insulation faults. Based on this, academia and industry have successively proposed methods such as electrical pulse detection, ultra-high frequency detection, electromagnetic wave detection, optical detection, ultrasonic detection, and audio analysis (Wang Qiang. Research and Sensor Design of Intelligent Online Monitoring System for Partial Discharge of Switchgear [D]. Xihua University, 2013) to achieve the goal of real-time detection of partial discharge phenomena in switchgear. Among them, non-invasive detection methods can achieve online detection without affecting the operating status of the switchgear, which is more in line with practical application needs.

[0003] Audio analysis, as a superior non-intrusive method for non-electrical quantity detection (Tang Yunhui, Hu Xilin. Research on partial discharge detection algorithm for switchgear based on acoustic signal analysis [J]. Electronic Devices, 2018: 644-648), is simple in principle, highly practical, and applicable to various devices. Therefore, audio analysis has multiple advantages in switchgear partial discharge detection and has been widely applied in substation monitoring (Zhou Ling. Research on partial discharge fault identification algorithm based on ultrasonic signals [D]., 2013). Currently, the main feature of switchgear partial discharge detection research based on audio analysis is the Mel Frequency Cepstrum Coefficient (MFCC) feature, and the pattern recognition algorithms used are mainly traditional neural networks and support vector machines. However, due to differences in acquisition distance, acquisition equipment, and switchgear model, audio signals generated by the same type of partial discharge in switchgear may also differ. Therefore, switchgear partial discharge detection requires higher generalization ability from the model. Against this backdrop, this invention proposes a method for partial discharge audio recognition of switchgear based on deep meta-learning, which can make full use of multi-source domain data with differences and effectively improve the audio recognition performance of partial discharge detection models under conditions of large data differences and cross-domain performance degradation. Summary of the Invention

[0004] The purpose of this invention is to address the practical problems of partial discharge audio recognition methods for switchgear, such as large data differences and reduced cross-domain performance. It provides a deep meta-learning-based method for partial discharge audio recognition in switchgear, improving the audio recognition performance under conditions of large data differences and reduced cross-domain performance, thus meeting practical application needs and having greater practical application value.

[0005] This invention includes the following steps:

[0006] 1) Preprocessing of audio signals collected from simulated partial discharge phenomena in high-voltage switchgear;

[0007] 2) Design of the depth model;

[0008] 3) Application of deep meta-learning strategies.

[0009] In step 1), the audio signal preprocessing of the simulated high-voltage switchgear partial discharge phenomenon is performed. To obtain the audio signal generated by the partial discharge of the switchgear, a partial discharge simulation model is used to collect audio signals in a real switchgear, recording the audio signals generated by three types of partial discharge: tip, air gap, and suspension. The specific steps are as follows:

[0010] (1) In order to reduce the impact of random noise, large data volume of a single audio file and invalid audio frames, the original audio signal is preprocessed to filter out unimportant information and background noise, and audio signal pre-emphasis, audio framing, endpoint detection, etc. are performed.

[0011] The pre-emphasis processed audio signal has a flatter spectrum, making it easier to analyze the spectrum and channel parameters; furthermore, pre-emphasis effectively suppresses random noise and eliminates DC drift when calculating the short-time energy of the audio signal; the transfer function of a first-order high-pass digital filter used for pre-emphasis processing is:

[0012] H(z) = 1 - αz -1

[0013] Where α is the pre-emphasis coefficient, usually taken as 0.9 < α < 1.0; z is the input of the filter.

[0014] A single audio file contains a huge amount of data, making it difficult to process. To facilitate calculation, considering the periodicity and short-term stability of partial discharge phenomena in switchgear, the entire audio segment can be divided into frames.

[0015] Since some signals in the entire audio signal are not valid information, it is necessary to distinguish between valid and invalid audio frames through endpoint detection. Valid audio frames are extracted using a dual-threshold endpoint detection method, and short-time energy and zero-crossing rate are combined for endpoint detection to obtain a higher quality valid audio signal.

[0016] Short-time energy, i.e., the audio signal x of the i-th frame of the n-th sample. i The energy E of (n) i The calculation formula is:

[0017]

[0018] Zero crossing rate, i.e., the audio signal x of the i-th frame i (n) The number of times Z crosses the 0 axis (time axis) i The calculation formula is:

[0019]

[0020] Where sgn[·] is the sign function, i.e.:

[0021]

[0022] (2) To improve the recognition accuracy of audio signals, it is necessary to extract effective audio features; Mel Frequency Cepstrum Coefficient (MFCC) is selected as the classification feature; the calculation of MFCC includes Fast Fourier Transform (FFT), Mel filtering, and Discrete Cosine Transform (DCT), and the calculation process is as follows:

[0023] (a) Fast Fourier Transform: The input of the FFT is a frame of audio signal x(n); the output is a complex number X(k) containing N frequency bands, representing the amplitude and phase of a certain frequency component in the original signal. Its calculation formula is:

[0024]

[0025] (b) Mel Filtering: The input is the output of the FFT, containing the energy information of the frame signal in each frequency band; it is passed through a Mel filter bank H(k) and logarithmic calculation is performed to obtain the logarithmic energy spectrum s(m); where H(k) consists of M triangular filters, and the center frequency of the m-th triangular filter is f(m), then the frequency response of each triangular filter is:

[0026]

[0027] in,

[0028] The logarithmic energy spectrum s(m) is:

[0029]

[0030] (c) Discrete Cosine Transform (DCT): The Discrete Cosine Transform is a transformation similar to the Discrete Fourier Transform. By performing the Discrete Cosine Transform on the logarithmic energy spectrum, the Lth order MFCC can be obtained.

[0031]

[0032] In step 2), the design of the depth model involves the following specific steps:

[0033] To better describe the relationship between the centroid and the feature distribution, a cosine similarity-based classification network framework is used, introducing learnable cross-domain classification centroid parameters to guide the optimization process of the feature distribution, thereby reducing the dependence of the classification centroid on the source domain data.

[0034] The classification model consists of a feature extractor. and task classifier Composition, in which For input space, For tag space, Corresponding to a low-dimensional feature space; for feature extractors Use a deep convolutional neural network and perform normalization on the network's output; for the classifier Use normalized feature vectors as classifiers Input, Normalized classification weights Composed of temperature coefficient T, obtained from cosine similarity, classifier for:

[0035]

[0036] The prediction result is Classification weight It can be viewed as the centroid of the learnable categories, possessing domain independence and trainability; in subsequent training and optimization processes, It can be used as a generalized classification boundary learned by the model to guide the model's feature alignment.

[0037] In step 3), the application of the deep meta-learning strategy is carried out in the following specific steps:

[0038] To address the audio recognition problem of partial discharge audio recognition methods for switchgear under conditions of large data differences and degraded performance across domains, a meta-learning strategy is introduced into deep learning. The specific steps are as follows:

[0039] (1) In order to address the problem that meta-learning-based domain generalization methods have limited meta-training tasks and insufficient generalization experience, a model-independent Mixed Episodic Training (MET) approach is proposed. This approach expands the limited set of meta-training tasks for multi-domain generalization to single-domain generalization into a more complex set of meta-training tasks that includes various generalization scenarios such as multi-domain generalization to single-domain, multi-domain generalization to multi-domain, and multi-domain adaptation. The Mixed Episodic Training approach increases the model's knowledge accumulation on generalization tasks and improves the model's generalization ability.

[0040] (2) To fully leverage the cross-domain generalization centroid The guiding role is to update the cross-domain generalization centroid separately during the meta-training process, by constraining the cross-domain generalization centroid. The distribution relationship between the feature extractor output and the parameter update process is optimized; in the internal optimization stage of the meta-training process, the parameter φ of the feature extractor is fixed, and the classifier parameter θ corresponding to the generalized classification centroid is optimized first.

[0041] (3) In order to make full use of the supervision information of the source domain data and improve the performance of the domain generalization method based on meta-learning, a loss function based on semantic alignment and classification boundary alignment is introduced. and By using fine-grained alignment methods to constrain the semantic feature space and the classification boundary of the classifier, the two loss functions effectively utilize the label information of the samples to effectively extract and optimize domain-independent and class-related classification features, thereby improving the generalization ability of the model.

[0042] This invention is based on deep learning and meta-learning. Its main features are strong practicality and portability. It can meet the cross-domain generalization requirements of audio recognition in real-world scenarios and improve the accuracy and applicability of the partial discharge recognition method for switchgear.

[0043] Compared with the prior art, the innovative aspects of this invention include:

[0044] 1) Applying deep meta-learning to the recognition of partial discharge audio across domains with large data differences. This invention seeks a suitable solution for the recognition of partial discharge audio across domains with large data differences from multiple aspects, including original audio signal processing, deep model design, and the application of deep meta-learning strategies, and has achieved good results.

[0045] 2) The confused scenario training mode and the parameter update method based on cross-domain generalization centroid are used to improve the performance of partial discharge audio recognition in cross-domain data with large differences. Traditional meta-learning-based domain generalization methods have limited exposure to meta-training tasks and insufficient generalization experience. The confused scenario training mode increases the model's knowledge accumulation on the generalization task and improves the model's generalization ability. Traditional centroid classification networks based on statistical features ignore the marginal distribution of features. The parameter update method based on cross-domain generalization centroid can make full use of the label information of each feature, and also allows the classifier parameters and feature extractor parameters to participate in the semantic alignment process simultaneously. This fully leverages the guiding role of the cross-domain generalization centroid, optimizes the semantic feature space, improves the generalization ability of extracted features, and achieves the effect of improving the performance of partial discharge audio recognition in cross-domain data with large differences. Attached Figure Description

[0046] Figure 1 This is the original audio signal for different discharge types in the partial discharge detection dataset for switchgear.

[0047] Figure 2 This dataset contains MFCC features for different discharge types in a switchgear partial discharge detection dataset.

[0048] Figure 3 This is a deep meta-learning audio classification model based on a confused context training mode.

[0049] Figure 4t-SNE analysis on audio datasets for partial discharge anomaly detection in switchgear. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of the present invention clearer, the following embodiments will be used to further illustrate the present invention in conjunction with the accompanying drawings.

[0051] The embodiments of the present invention include the following steps:

[0052] 1) Preprocessing of audio signals collected from simulated partial discharge phenomena in high-voltage switchgear;

[0053] To obtain the audio signals generated by partial discharge in switchgear, this invention employs a partial discharge simulation model to collect audio signals in a real switchgear. Due to variations in background noise, acquisition equipment, and acquisition location during the experiment, audio data from different batches exhibit domain differences, consistent with actual application scenarios. This invention uses audio data from six batches of recordings, documenting audio signals generated by three types of partial discharge: tip, air gap, and suspension. The original audio signals are shown below. Figure 1 As shown.

[0054] Step 1: In order to reduce the impact of random noise, large data size of a single audio file and invalid audio frames, the original audio signal is preprocessed to filter out unimportant information and background noise, and audio signal pre-emphasis, audio framing, endpoint detection, etc.

[0055] The pre-emphasis processed audio signal has a flatter spectrum, making it easier to analyze the spectrum and channel parameters. Furthermore, pre-emphasis effectively suppresses random noise and eliminates DC drift when calculating the short-time energy of audio signals (Du Shibin. Research on Fault Monitoring Algorithm for Electrical Equipment Based on Audio Features [D]. Jinan: Shandong University, 2014). The transfer function of a first-order high-pass digital filter used for pre-emphasis processing is:

[0056] H(z) = 1 - αz -1

[0057] Where α is the pre-emphasis coefficient, which is usually taken as 0.9 < α < 1.0, and here we set α = 0.97.

[0058] A single audio file contains a massive amount of data, making processing difficult. To facilitate calculations, and considering the periodicity and short-term stability of partial discharge phenomena in switchgear, the entire audio segment can be divided into frames. In this invention, the frame length is set to 5 power frequency cycles, and the frame shift is set to 2 power frequency cycles.

[0059] Since some signals in the entire audio signal are not valid information, endpoint detection is needed to distinguish between valid and invalid audio frames. In this invention, a dual-threshold endpoint detection method is used to extract valid audio frames, and short-time energy and zero-crossing rate are combined for endpoint detection to obtain a higher quality valid audio signal.

[0060] Short-time energy, i.e., the audio signal x of the i-th frame of the n-th sample. i The energy E of (n) i The calculation formula is:

[0061]

[0062] Zero crossing rate, i.e., the audio signal x of the i-th frame i (n) The number of times Z crosses the 0 axis (time axis) i The calculation formula is:

[0063]

[0064] Where sgn[·] is the sign function, i.e.:

[0065]

[0066] Step 2: To improve the recognition accuracy of audio signals, it is necessary to extract effective audio features. This invention selects MFCC as the classification feature. MFCC, as an excellent audio feature, can not only be used for audio feature extraction but also reduce the computational dimensionality, and is widely used in audio feature extraction tasks. The obtained MFCC is as follows: Figure 2 As shown. Calculating MFCC involves the following steps: Fast Fourier Transform (FFT), Mel filtering, and Discrete Cosine Transform (DCT). The calculation process is as follows:

[0067] (1) Fast Fourier Transform. The input of the FFT is a frame of audio signal x(n); the output is a complex number X(k) containing N frequency bands, representing the amplitude and phase of a certain frequency component in the original signal. Its calculation formula is:

[0068]

[0069] (2) Mel Filtering. The input is the output of the FFT, which contains the energy information of the frame signal in each frequency band. This is passed through a Mel filter bank H(k) and logarithmic calculation is performed to obtain the logarithmic energy spectrum s(m) (Wang Shijia. Research on Sound Event Recognition Based on Deep Learning [D]. Southeast University, 2018). H(k) consists of M triangular filters. Let the center frequency of the m-th triangular filter be f(m), then the frequency response of each triangular filter is:

[0070]

[0071] in,

[0072] The logarithmic energy spectrum s(m) is:

[0073]

[0074] (3) Discrete Cosine Transform (DCT). The Discrete Cosine Transform is a transformation similar to the Discrete Fourier Transform. The Lth order MFCC can be obtained by performing a Discrete Cosine Transform on the logarithmic energy spectrum.

[0075]

[0076] 2) Design of the depth model;

[0077] To better describe the relationship between the centroid and feature distribution, this invention attempts to introduce a learnable centroid into the model to guide subsequent feature optimization. This invention uses a cosine similarity-based classification network framework, introducing learnable cross-domain classification centroid parameters to guide the feature distribution optimization process. Compared to domain statistical classification centroids obtained from statistical domain feature distributions, using learnable cross-domain classification centroids reduces the dependence of the classification centroid on the source domain data to some extent.

[0078] In this invention, the classification model is composed of a feature extractor. and task classifier Composition, in which For input space, For tag space, This corresponds to a low-dimensional feature space. For feature extractors... This invention uses a deep convolutional neural network and performs normalization on the network's output. For the classifier... Use normalized feature vectors as classifiers Input, Normalized classification weights Composed of temperature coefficient T, obtained from cosine similarity, classifier for:

[0079]

[0080] The final prediction result is Classification weight It can be viewed as the centroid of the learnable categories, possessing domain independence and trainability. In subsequent training and optimization processes, It can be used as a generalized classification boundary learned by the model to guide the model's feature alignment.

[0081] 3) Application of deep meta-learning strategies.

[0082] To address the audio recognition problem of partial discharge audio recognition methods for switchgear under conditions of large data differences and degraded cross-domain performance, this invention introduces a meta-learning strategy into deep learning. The specific steps are as follows:

[0083] Step 1: To address the limitations of meta-learning-based domain generalization methods in terms of the limited meta-training tasks they encounter and insufficient generalization experience, this invention proposes a model-independent Mixed Episodic Training (MET) approach. This expands the limited set of meta-training tasks for multi-domain to single-domain generalization into a more complex set encompassing various generalization scenarios, including multi-domain to single-domain, multi-domain to multi-domain, and multi-domain adaptation. The Mixed Episodic Training approach increases the model's knowledge accumulation regarding generalization tasks, thereby improving its generalization ability.

[0084] Step 2: In order to fully leverage the cross-domain generalization centroid Guided by the principle of cross-domain generalization centroid, this invention updates the cross-domain generalization centroid separately during the meta-training process, by constraining the cross-domain generalization centroid. The parameter update process is optimized by considering the distribution relationship between the feature extractor output and the parameter parameters. During the internal optimization phase of the meta-training process, the feature extractor parameters φ are fixed, and the classifier parameters θ corresponding to the generalized classification centroid are optimized first.

[0085] Step 3: To fully utilize the supervised information from the source domain data and improve the performance of the meta-learning-based domain generalization method, this invention introduces a loss function based on semantic alignment and classification boundary alignment. and Fine-grained alignment methods are used to constrain the semantic feature space and the classification boundary of the classifier. Two loss functions effectively utilize the sample label information, efficiently extracting and optimizing domain-independent and class-related classification features, further improving the model's generalization ability.

[0086] The following combination Figure 3The application of deep meta-learning strategies will be described in further detail. The main idea of ​​the deep meta-learning method for confusing context training modes in this invention is as follows: Figure 3 As shown. In the confused scenario training mode part, K-1 source domains are randomly selected from K source domains to form the meta-training set. Extract data from all source domains as the meta-test set. The target domain is obfuscated by randomly assigning a percentage of each domain to a value between (0, 1). When the percentage of the Kth source domain is 1, the meta-test set becomes the target domain. In the domain generalization model section based on semantic alignment and classification boundary alignment, It is the supervised loss function in the meta-training process. It is a semantic alignment loss function based on cross-domain generalization centroid. It is further divided into and These are, respectively, the domain-independent semantic alignment loss function and the domain-dependent semantic alignment loss function in the meta-testing process. It is also the classifier alignment loss function.

[0087] The first step is to design a confusion scenario training mode based on meta-learning.

[0088] Inspired by Model-Agnostic Meta-learning (MAML) algorithms (Chelsea Finn, Pieter Abbeel, Sergey Levine. Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks[C]. Proceedings of the 34th International Conference on Machine Learning. PMLR 70, 2017:1126-1135), this invention proposes the MET approach, which allows the proposed method to access more varied domain distribution differences and further improves the model's generalization ability to unknown domains. Unlike previous meta-learning-based domain generalization tasks, the MET approach uses different meta-task sets during meta-training and meta-testing. To increase the number of meta-tasks, this invention randomly selects K-1 source domains from K source domains as the meta-training set. Randomly set the domain proportion and extract data from all source domains as the meta-test set. By extracting data of different proportions as the meta-test set and expanding the meta-task to a set of tasks generalized from multiple domains to a single domain and from multiple domains to multiple domains, the finite number of tasks K is expanded to an infinite number, thereby increasing the complexity of the tasks.

[0089] The second step is to optimize the feature extractor F and the task classifier. The update process of parameters φ and θ.

[0090] In the domain generalization task of this invention In, any subtask Both contain two datasets: one for meta-training and one for training. and used for meta-testing During the internal optimization phase of the meta-training process, the parameters φ of the feature extractor F are fixed, and the classifier parameters θ corresponding to the generalized classification centroid are optimized first. The classifier parameters θ are based on a task-specific supervised loss function. Updated

[0091]

[0092] Where α is the meta-training learning rate hyperparameter.

[0093] After obtaining the parameters and initial parameter set (φ,θ,θ′) from the meta-training process, a meta-testing process is performed to enable the model to obtain the desired parameters. Effective features. Calculate based on the updated parameters (φ,θ′). Forced alignment and The characteristics between them. The update process in the meta-test phase is as follows:

[0094]

[0095] Where β is the meta-test learning rate hyperparameter.

[0096] The third step is to design a loss function based on semantic alignment and classification boundary alignment.

[0097] Based on the generalized centroid Samples will be uniformly clustered near the centroids of the same class, and far from the centroids of other classes. Inspired by the Center Loss function (Wen Y, Zhang K, Li Z, et al. A discriminative feature learning approach for deep face recognition[C]. Proceedings of the European Conference on Computer Vision. Springer, Cham, 2016:499-515), this invention introduces cosine distance to measure the distance between sample features and centroids. The cosine value between them is expected to be close to 1, while the outlier features are... The cosine value between them should be as small as possible. To better distinguish easily mixed categories, this invention only constrains the heterogeneous feature pairs with the largest cosine value, rather than constraining all heterogeneous feature pairs. Loss function for semantic feature alignment. as follows:

[0098]

[0099] Inspired by metric learning and centroid networks (Zhao H, Des Combes RT, Zhang K, et al. On learning invariant representations for domain adaptation[C]. International Conference on Machine Learning.PMLR,2019:7523-7532), this invention aligns the classification boundaries of various source domains by restricting the alignment of the output distributions of different classifiers (i.e., centroids) in each domain. Specifically, given a set of K-1 source domains... From The samples can be processed by K-1 classifiers (i.e., in Single source domain D k The centroid obtained above Obtain the corresponding category score. The formula is as follows:

[0100]

[0101]

[0102] in, Those belonging to class C belong to D. k Samples of the domain The total number.

[0103] To obtain the distribution differences of different classification scores, this invention utilizes KL divergence to evaluate the pairwise distance between the classification score distributions. (Classifier alignment loss) The definition of is:

[0104]

[0105] Where (a,b)∈{s1,s2,……,s K-1},

[0106] because The discussion focuses on the feature distribution of a single domain, therefore It is named the domain-dependent semantic alignment loss function. Overall, From semantic alignment loss Alignment loss with classifier composition:

[0107]

[0108] Here, γ1 and γ2 are parameters that balance the semantic alignment and classifier alignment losses.

[0109] For ease of description, the method proposed in this invention is denoted as FAME. The comparison method, DeepAll, is a standard supervised learning method. Its training set consists of all source domains, and it is trained using the same model framework as the FAME method of this invention.

[0110] The experimental results of this invention are as follows: Figure 4 As shown, this visualization was obtained by analyzing the feature spaces of the FAME and DeepAll methods using the T-distributed stochastic neighbor embedding (t-SNE) algorithm. The classification result comes from a generalization task with six batches of data as the target domain. Figure 4 Figures (a) and (b) in the middle show the domain distribution and category distribution of DeepAll's output features, respectively. Figure 4 Figures (c) and (d) in the middle represent the domain distribution and class distribution of the output features of FAME, respectively. By comparison... Figure 4 As shown in Figures (a) and (c), the FAME method exhibits more pronounced clustering of the target domain features, effectively aligning the features of the multi-source domain and the target domain and reducing domain differences; (Comparison) Figure 4 As shown in Figures (b) and (d), the FAME method effectively aggregates samples of the same category from multiple source domains, with significant inter-class distances and smaller intra-class distances. The results of t-SNE validate the effectiveness of the FAME method.

[0111] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included in the scope of protection of the present invention.

[0112] References:

[0113] [1] Wang Qiang. Research on Intelligent Online Monitoring System for Partial Discharge of Switchgear and Sensor Design [D]. Xihua University, 2013.

[0114] [2] Tang Yunhui, Hu Xilin. Research on partial discharge detection algorithm of switchgear based on acoustic signal analysis [J]. Electronic Devices, 2018: 644-648.

[0115] [3] Zhou Ling. Research on partial discharge fault identification algorithm based on ultrasonic signal [D]. 2013.

[0116] [4] Du Shibin. Research on fault monitoring algorithm for electrical equipment based on audio features [D]. Jinan: Shandong University, 2014.

[0117] [5] Wang Shijia. Research on sound event recognition based on deep learning [D]. Southeast University, 2018.

[0118] [6] Chelsea Finn, Pieter Abbeel, Sergey Levine. Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks[C]. Proceedings of the 34thInternational Conference on Machine Learning. PMLR 70, 2017: 1126-1135.

[0119] [7]Wen Y, Zhang K, Li Z, et al. A discriminative feature learning approach for deep face recognition [C]. Proceedings of the European Conference on Computer Vision. Springer, Cham, 2016: 499-515.

[0120] [8] Zhao H, Des Combes RT, Zhang K, et al. On learning invariant representations for domain adaptation [C]. International Conference on Machine Learning. PMLR, 2019: 7523-7532.

Claims

1. A switch cabinet partial discharge audio recognition method based on deep meta-learning, characterized by The method comprises the following steps: 1) Preprocessing of audio signals collected from simulation of partial discharge phenomena of high-voltage switch cabinets: audio signals are collected from real switch cabinets by using a partial discharge simulation model, and audio signals generated by three types of partial discharge, i.e., tip, air gap and suspension, are recorded; The preprocessing of the audio signals collected from the simulation of the partial discharge phenomena of the high-voltage switch cabinet comprises the following steps: (1) Preprocessing of the original audio signals to filter out unimportant information and background noise, and performing pre-emphasis, audio framing and endpoint detection on the audio signals; A first-order high-pass digital filter is used for pre-emphasis processing, and its transfer function is: H(z) = 1 - az -1 Wherein, α is the pre-emphasis coefficient, 0.9 < α < 1.0; z is the input of the filter; Short-time energy and zero-crossing rate are used in combination for endpoint detection to obtain high-quality effective audio signals; (2) Selecting mel-frequency cepstral coefficient (MFCC) as the classification feature; the calculation of MFCC includes fast Fourier transform (FFT), mel filtering and discrete cosine transform (DCT), and the specific calculation process is as follows: (a) Fast Fourier transform: the input of FFT is a frame of audio signals x(n); the output is a complex number X(k) containing N frequency bands, representing the amplitude and phase of a certain frequency component in the original signal, and its calculation formula is: (b) Mel filtering: the input is the output result of FFT, containing the energy information of the frame signal in each frequency band; it is passed through a mel filter bank H(k) and logarithm calculation to obtain a log energy spectrum s(m); wherein H(k) is composed of M triangular filters, and the center frequency of the mth triangular filter is f(m), so the frequency response of each triangular filter is: wherein, The log energy spectrum s(m) is: (c) Discrete cosine transform: the log energy spectrum is subjected to discrete cosine transform, i.e., L-order MFCC is obtained: 2) Design of a deep model, and the specific steps are as follows: A learnable cross-domain classification centroid parameter is introduced to guide the optimization process of feature distribution, so as to reduce the dependence of the classification centroid on the source domain data; The classification model consists of a feature extractor and a task classifier wherein is the input space, is the label space, corresponds to a low-dimensional feature space; for the feature extractor a deep convolutional neural network is used and a normalization is performed on the output of the network; for the classifier the normalized feature vector is used as input, the classifier consists of normalized classification weights and a temperature coefficient T, resulting from a cosine similarity, the classifier is: The prediction result is where the classification weight The classification centroid is regarded as a learnable category, has domain independence and trainability; in the subsequent training optimization process, The generalization classification boundary learned by the model is used to guide the feature alignment of the model; 3) Application of a deep meta-learning strategy.

2. The switch cabinet partial discharge audio recognition method based on deep meta-learning according to claim 1, characterized in that In step 1) part (1), the short-time energy, i.e. the energy E i (n) of the i-th frame of the audio signal x i (n) is calculated as follows: The zero-crossing rate, i.e. the number of zero-crossings Z of the i-th frame of audio signals x i (n) the number of crossings Z through the 0-axis i , the 0-axis being the time axis, the calculation formula being: Wherein, sgn[·] is a sign function, i.e.:

3. The switch cabinet partial discharge audio recognition method based on deep meta-learning according to claim 1, characterized in that In step 3), the application of the deep meta-learning strategy comprises the following steps: (1) The meta-training task set of limited multi-domain generalization to single domain is expanded to a more complex meta-training task set including multi-domain generalization to single domain, multi-domain generalization to multi-domain and multi-domain adaptation; the confusion scenario training mode increases the knowledge accumulation of the model on the generalization task and improves the generalization ability of the model; (2) updating the cross-domain generalization centroid separately in the meta-training process, optimizing the parameter updating process by constraining the distribution relationship between the cross-domain generalization centroid and the feature extractor output; in the internal optimization stage of the meta-training process, fixing the parameters of the feature extractor, and preferentially optimizing the classifier parameters corresponding to the generalized classification centroid (3) Introducing loss functions based on semantic alignment and classification boundary alignment and By fine-grained alignment, the semantic feature space and the classification boundary of the classifier are respectively constrained; two loss functions use the label information of the sample to extract and optimize the domain-independent and class-dependent classification features to improve the generalization ability of the model.

Images