Transformer state identification method and device, electronic equipment and storage medium

Abstract

The invention provides a transformer state identification method and device, electronic equipment and a storage medium, and relates to the technical field of electric power, and the method comprises the steps: obtaining to-be-tested audio data of a transformer; inputting the to-be-detected audio data into the voiceprint extraction model to obtain a target voiceprint output by the voiceprint extraction model; the voiceprint extraction model is obtained by training first sample audio data of stateless tags of a first number of first sample transformers; and identifying the state of the transformer based on the target voiceprint and the at least one registered voiceprint. The method comprises the following steps: training based on a large amount of first sample audio data without state labels to obtain a voiceprint extraction model, determining at least one registered voiceprint based on a small amount of second sample audio data with state labels through the voiceprint extraction model, and further realizing state recognition of the transformer in combination with the registered voiceprint. Under the condition that only a small amount of sample audio data with state labels are needed, the accuracy of transformer state identification is improved.

Description

Technical Field

[0001] The present invention relates to the field of electric power technology, and in particular to a transformer state identification method, device, electronic equipment and storage medium. Background Art

[0002] In energy scenarios such as electricity, with the increasing demand for high-voltage and ultra-high-voltage power grid construction, higher and higher requirements are placed on the reliability of power systems, especially transformers (if a failure occurs, it will cause economic losses and even cause large-scale social shutdowns). Therefore, early warning of transformer failures is becoming increasingly important.

[0003] In related technologies, such as Figure 1 As shown, usually based on the transformer audio data with normal state labels and the transformer audio data with fault state labels, the deep neural network model is supervisedly trained to finally obtain a transformer state recognition model. The audio data to be tested of the transformer to be tested is input into the trained transformer state recognition model to obtain the posterior probability of each state category, and the state category corresponding to the maximum posterior probability is determined as the state of the transformer to be tested.

[0004] However, in the above-mentioned related technologies, when the transformer state recognition model is trained in a supervised manner, a large amount of labeled transformer audio data is required as training data. If the training data is insufficient, the accuracy of transformer state recognition will be reduced. Summary of the invention

[0005] The present invention provides a transformer state recognition method, device, electronic device and storage medium, which are used to solve the defect in the prior art that the accuracy of transformer state recognition will be reduced if the training data is insufficient.

[0006] The present invention provides a transformer state identification method, comprising the following steps.

[0007] Obtain the audio data to be tested of the transformer; Inputting the audio data to be tested into a voiceprint extraction model to obtain a target voiceprint output by the voiceprint extraction model; the voiceprint extraction model is trained based on the first sample audio data without state labels of the first sample transformer of the first number; The state of the transformer is identified based on the target voiceprint and at least one registered voiceprint; the at least one registered voiceprint is a voiceprint obtained by inputting a second number of second sample audio data including a state label of a second sample transformer into the voiceprint extraction model, and the second number is smaller than the first number.

[0008] According to a transformer state identification method provided by the present invention, the state of the transformer is identified based on the target voiceprint and at least one registered voiceprint, including: For each of the registered voiceprints, determining the similarity between the target voiceprint and the registered voiceprint; The state of the transformer is determined based on the target state label of the second sample audio data corresponding to the registered voiceprint with the maximum similarity.

[0009] According to a transformer state recognition method provided by the present invention, the voiceprint extraction model is trained based on the following method: Obtaining first sample audio data of the stateless tag of the first sample transformer; Based on the first sample audio data, determining a training data set, the training data set including sample audio data of at least two preset sampling rates corresponding to the first sample audio data; For each of the sample audio data of the preset sampling rate in the training data set, feature extraction is performed on the sample audio data of the preset sampling rate to obtain sample audio features of the sample audio data of the preset sampling rate; Inputting the sample audio features of each of the preset sampling rates into an initial encoder to obtain the prediction coding features corresponding to each of the preset sampling rates output by the initial encoder; Determining first loss information based on predictive coding features of different preset sampling rates corresponding to the first sample audio data; Based on the first loss information, the voiceprint extraction model is determined.

[0010] According to a transformer state identification method provided by the present invention, the voiceprint extraction model is determined based on the first loss information, including: Inputting the predicted coding features corresponding to each of the preset sampling rates into an initial decoder to obtain the predicted decoding features corresponding to each of the preset sampling rates output by the initial decoder; Determining second loss information based on predicted decoding features and sample audio features at the same preset sampling rate corresponding to the first sample audio data; Based on the second loss information and the first loss information, the network parameters of the initial encoder are adjusted to obtain the voiceprint extraction model.

[0011] According to a transformer state identification method provided by the present invention, the feature extraction of the sample audio data of the preset sampling rate to obtain the sample audio features of the sample audio data of the preset sampling rate includes: Based on at least two types of filters, feature extraction is performed on the sample audio data of the preset sampling rate, and for each type of the filters, sample audio features of the sample audio data of the preset sampling rate corresponding to the filters are obtained; the filters include at least two types of equal-pitch Mel filters, equal-pitch frequency filters, and equal-pitch anti-Mel filters; The step of inputting the sample audio features of each of the preset sampling rates into an initial encoder to obtain the prediction coding features corresponding to each of the preset sampling rates output by the initial encoder includes: The sample audio features of the preset sampling rates corresponding to all the filters are concatenated to obtain a combined audio feature; The combined audio feature is input into the initial encoder to obtain the prediction coding feature corresponding to each of the preset sampling rates corresponding to each of the filters output by the initial encoder.

[0012] According to a transformer state recognition method provided by the present invention, determining a training data set based on the first sample audio data includes: Acquire third sample audio data of stateless tags of other devices except the first sample transformer; The training data set is determined based on the first sample audio data and the third sample audio data, wherein the training data set also includes sample audio data of at least two preset sampling rates corresponding to the third sample audio data.

[0013] According to a transformer state identification method provided by the present invention, the sampling rates of the first sample audio data and the third sample audio data are both greater than the maximum preset sampling rate among all preset sampling rates; The determining the training data set based on the first sample audio data and the third sample audio data includes: Downsampling the first sample audio data to obtain sample audio data of at least two preset sampling rates corresponding to the first sample audio data; Downsampling the third sample audio data to obtain sample audio data of at least two preset sampling rates corresponding to the third sample audio data; The training data set is determined based on sample audio data of at least two preset sampling rates corresponding to the first sample audio data and sample audio data of at least two preset sampling rates corresponding to the third sample audio data.

[0014] The present invention also provides a transformer state identification device, comprising: An acquisition unit, used for acquiring audio data to be tested of the transformer; An extraction unit, configured to input the audio data to be tested into a voiceprint extraction model to obtain a target voiceprint output by the voiceprint extraction model; the voiceprint extraction model is trained based on the first sample audio data without state labels of the first sample transformer of the first number; An identification unit is used to identify the state of the transformer based on the target voiceprint and at least one registered voiceprint; the at least one registered voiceprint is a voiceprint obtained by inputting a second number of second sample audio data including a state label of a second sample transformer into the voiceprint extraction model, and the second number is less than the first number.

[0015] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, any of the transformer state identification methods described above is implemented.

[0016] The present invention also provides a non-transitory computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the transformer state identification method described in any one of the above is implemented.

[0017] The present invention also provides a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the transformer state identification method as described above is implemented.

[0018] The transformer state identification method, device, electronic device and storage medium provided by the present invention input the acquired audio data of the transformer to be tested into the voiceprint extraction model to obtain the target voiceprint output by the voiceprint extraction model, and identify the state of the transformer based on the target voiceprint and at least one registered voiceprint, wherein the voiceprint extraction model is trained based on a first number of first sample audio data without state labels, and the registered voiceprint is a voiceprint obtained by inputting a second number of second sample audio data including state labels into the voiceprint extraction model. It can be seen that the present invention first obtains a voiceprint extraction model based on a large number of first sample audio data without state labels, and then determines at least one registered voiceprint based on a small number of second sample audio data with state labels through the voiceprint extraction model, and then realizes the state identification of the transformer in combination with the registered voiceprint, which improves the accuracy of transformer state identification when only a small number of sample audio data with state labels are required. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0020] Figure 1 It is a flow chart of transformer status identification in related technology.

[0021] Figure 2 It is a flow chart of a transformer state identification method provided in an embodiment of the present invention.

[0022] Figure 3 It is a spectrum diagram of the equal-height Mel filter provided by an embodiment of the present invention.

[0023] Figure 4 This is one of the flow charts of the training method of the voiceprint extraction model provided in the embodiment of the present invention.

[0024] Figure 5 This is the second flow chart of the training method of the voiceprint extraction model provided in the embodiment of the present invention.

[0025] Figure 6 It is a structural diagram of a network model provided by an embodiment of the present invention.

[0026] Figure 7 It is a structural schematic diagram of a transformer state identification device provided in an embodiment of the present invention.

[0027] Figure 8 It is a schematic diagram of the physical structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0028] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0029] At present, industrial intelligent manufacturing has entered an era of rapid development. With the continuous development of science and technology and automation technology, industries such as manufacturing and energy are gradually moving towards the field of automation and intelligence. Therefore, the normal operation of industrial equipment has become increasingly important. Once a sudden failure of industrial equipment occurs, it will not only cause huge losses to the industrial manufacturing of a region, but also have great secondary safety hazards, causing major economic losses, and may even lead to catastrophic casualties or adverse social impacts. In energy scenarios such as electricity, with the increase in demand for high-voltage and ultra-high-voltage power grid construction, higher and higher requirements are placed on the reliability of power systems, especially transformers (if a failure occurs, it will cause economic losses and even cause large-scale social shutdowns). Therefore, early warning of transformer failures is becoming more and more important.

[0030] At present, the fault state of the transformer can be detected by machine recognition of the sound emitted when the transformer fails, that is, through state recognition technology. The fault state of the transformer usually includes DC bias magnetization, short circuit impact, partial discharge and heavy overload. The specific detection principle of transformer fault state recognition is: the transformer is subjected to electrical, magnetic, mechanical and other stresses to produce vibrations, and the formed mechanical waves are converted into audio signals after being transmitted through solid-liquid-gas media. The audio signal is captured by the acoustic sensor (microphone) to obtain audio data. The audio data contains a large amount of time-frequency domain feature information. When the transformer is abnormal, the sound also changes. Combined with artificial intelligence methods such as deep learning, the state of the transformer can be effectively identified.

[0031] However, in the related art, the transformer state recognition model is trained in a supervised manner, which requires a large amount of labeled transformer audio data as training data. If the training data is insufficient, the accuracy of transformer state recognition will be reduced.

[0032] Based on this, the present invention proposes a transformer state recognition method, which first obtains a voiceprint extraction model based on a large amount of first sample audio data without state labels, and then determines at least one registered voiceprint based on a small amount of second sample audio data with state labels through the voiceprint extraction model, and then realizes transformer state recognition in combination with the registered voiceprint. In the case where only a small amount of sample audio data with state labels is required, the accuracy of transformer state recognition is improved.

[0033] Combine the following Figure 2-Figure 6 The transformer state identification method of the present invention is described. The execution subject of the transformer state identification method can be an electronic device such as a terminal, a computer or a server, or a transformer state identification device set in the electronic device, and the transformer state identification device can be implemented by software, hardware or a combination of the two.

[0034] Figure 2 is a flow chart of a transformer state identification method provided by an embodiment of the present invention, such as Figure 2 As shown, the transformer state identification method includes the following steps: Step 201: Obtain audio data to be tested of the transformer.

[0035] For example, when the transformer is working normally or abnormally, the transformer is subjected to electrical, magnetic, mechanical and other stresses to generate vibrations. The formed mechanical waves are converted into audio signals after being transmitted through the solid-liquid-gas medium. The audio signals are captured by the acoustic sensor (microphone), thereby obtaining the audio data to be tested of the transformer.

[0036] Step 202: input the audio data to be tested into a voiceprint extraction model to obtain a target voiceprint output by the voiceprint extraction model; the voiceprint extraction model is trained based on the first sample audio data without state labels of the first number of first sample transformers.

[0037] For example, when the audio data to be tested of the transformer is obtained, the audio data to be tested is downsampled to at least two preset sampling rates according to the actual sampling rate, and at least two types of filters are used to perform feature extraction on the downsampled audio data to be tested of the transformer to obtain audio features to be tested corresponding to the downsampled audio data to be tested of each filter, and the audio features to be tested corresponding to the downsampled audio data to be tested of each filter are spliced ​​to obtain spliced ​​features to be tested, and the spliced ​​features to be tested are input into a pre-trained voiceprint extraction model, and the input spliced ​​features to be tested are analyzed by the voiceprint extraction model to finally obtain a target voiceprint corresponding to the spliced ​​features to be tested output by the voiceprint extraction model.

[0038] It should be noted that when the actual sampling rate of the audio data to be tested of the transformer is the same as the preset sampling rate, downsampling is not required.

[0039] It should be noted that when the at least two types of filters include two types of filters, the two types of filters include at least two of equal-height Mel filters, equal-height frequency filters, and equal-height inverse Mel filters. Preferably, the at least two types of filters include three types of filters, and the three types of filters are equal-height Mel filters, equal-height frequency filters, and equal-height inverse Mel filters. Among them, the equal-height frequency filters are equally spaced triangular filters, and the equal-height inverse Mel filters have the characteristics of dense filters at high frequencies and sparse filters at low frequencies. The calculation formulas for the Mel frequency of the equal-height Mel filters and the calculation formulas for the inverse Mel frequency of the equal-height inverse Mel filters can be expressed by the following formulas (1) and (2), respectively: f_mel = 2595 log10(1 + f / 700) (1) f_imel = 700 (10 ^ (f / 2595) - 1) (2) Among them, f_mel represents the Mel frequency of the equal-height Mel filter, f represents the normal frequency corresponding to the equal-height frequency filter, and f_imel represents the inverse Mel frequency of the equal-height inverse Mel filter. Figure 3 is a spectrum diagram of an equal-height Mel filter provided by an embodiment of the present invention, such as Figure 3 As shown, the features of the audio data to be tested of the transformer can be extracted based on the frequency spectrum of the equal-height Mel filter.

[0040] Step 203: Identify the state of the transformer based on the target voiceprint and at least one registered voiceprint; the at least one registered voiceprint is a voiceprint obtained by inputting a second number of second sample audio data including a state label of a second sample transformer into the voiceprint extraction model, and the second number is smaller than the first number.

[0041] For example, after the voiceprint extraction model training is completed, the second number of second sample audio data including the state label is downsampled to at least two preset sampling rates according to the actual sampling rate, for example, downsampled to 32k and / or 16k, so as to enrich the content information of the sample audio data, for example: the second sample audio data of 48k is downsampled to 32k and / or 16k, and the second sample audio data of 20k is only downsampled to 16k. Then all the downsampled second sample audio data are input into the voiceprint extraction model, and the voiceprint is extracted by inference of the voiceprint extraction model, and the extracted voiceprint is called the registered voiceprint, and the registered voiceprint is stored corresponding to the second sample audio data including the state label.

[0042] When the target voiceprint corresponding to the splicing feature to be tested of the transformer is obtained, the target voiceprint is compared with each pre-registered voiceprint to determine the target state label of the second sample audio data corresponding to the registered voiceprint closest to the target voiceprint, and the state represented by the target state label is determined as the state of the transformer. For example, if the state represented by the target state label is DC bias magnetization, the state of the transformer is determined to be DC bias magnetization, if the state represented by the target state label is short-circuit shock, the state of the transformer is determined to be short-circuit shock, if the state represented by the target state label is partial discharge, the state of the transformer is determined to be partial discharge, and if the state represented by the target state label is heavy overload, the state of the transformer is determined to be heavy overload. Of course, when the state represented by the target state label is normal, the state of the transformer is also normal, indicating that the transformer is not faulty.

[0043] It should be noted that the second number of second sample audio data including status labels are determined based on the audio signals collected when the transformer has all types of faults currently occurring. When new faults are discovered subsequently, it is necessary to determine new sample audio data based on the new audio signals collected when the transformer has a new fault. Then, the new sample audio features corresponding to the new sample audio data are input into the voiceprint extraction model to obtain a new registered voiceprint. The new registered voiceprint is stored in correspondence with the new sample audio data including the status label, and the registered voiceprint is continuously improved to further improve the accuracy of transformer status recognition.

[0044] The transformer state identification method provided by the present invention inputs the obtained audio data of the transformer to be tested into the voiceprint extraction model to obtain the target voiceprint output by the voiceprint extraction model, and identifies the state of the transformer based on the target voiceprint and at least one registered voiceprint, wherein the voiceprint extraction model is obtained by training based on a first number of first sample audio data without state labels, and the registered voiceprint is a voiceprint obtained by inputting a second number of second sample audio data including state labels into the voiceprint extraction model. It can be seen that the present invention first obtains a voiceprint extraction model based on a large number of first sample audio data without state labels, and then determines at least one registered voiceprint based on a small number of second sample audio data with state labels through the voiceprint extraction model, and then realizes the state identification of the transformer in combination with the registered voiceprint, which improves the accuracy of transformer state identification when only a small number of sample audio data with state labels are required.

[0045] In one embodiment, the above step 203 identifies the state of the transformer based on the target voiceprint and at least one registered voiceprint, which can be specifically implemented in the following manner: For each of the registered voiceprints, the similarity between the target voiceprint and the registered voiceprint is determined; based on the target state label of the second sample audio data corresponding to the registered voiceprint with the maximum similarity, the state of the transformer is determined.

[0046] For example, for each registered voiceprint, the cosine distance between the target voiceprint and the registered voiceprint is calculated, the cosine distance is used to characterize the similarity between the target voiceprint and the registered voiceprint, the registered voiceprint corresponding to the maximum cosine distance is determined as the voiceprint closest to the target voiceprint, the target state label of the second sample audio data corresponding to the registered voiceprint with the maximum cosine distance is obtained, and the state represented by the target state label is determined as the state of the transformer.

[0047] In this embodiment, the state of the transformer can be determined based on the similarity between the target voiceprint and each registered voiceprint, thereby improving the accuracy of transformer state recognition.

[0048] In one embodiment, Figure 4 is one of the flow charts of the training method of the voiceprint extraction model provided by the embodiment of the present invention, such as Figure 4 As shown, the voiceprint extraction model is trained based on the following method: Step 401: Obtain first sample audio data of the stateless tag of the first sample transformer.

[0049] For example, a large amount of stateless label audio data and some open source collection audio data when the first sample transformer is running are collected, and all the collected stateless label audio data are used as first sample audio data, and the sampling rate of the first sample audio data is greater than the preset sampling rate.

[0050] Step 402: Determine a training data set based on the first sample audio data, wherein the training data set includes sample audio data of at least two preset sampling rates corresponding to the first sample audio data.

[0051] For example, when all the first sample audio data are obtained, for each first sample audio data, the first sample audio data can be downsampled to obtain sample audio data of at least two preset sampling rates corresponding to the first sample audio data. According to the same method, sample audio data of at least two preset sampling rates corresponding to each first sample audio data can be obtained, and the set of sample audio data of at least two preset sampling rates corresponding to all the first sample audio data is determined as a training data set. Taking two preset sampling rates as an example, 32k and 16k respectively, the training data set includes 32k sample audio data and 16k sample audio data corresponding to all the first sample audio data.

[0052] Step 403: For each of the sample audio data at the preset sampling rate in the training data set, feature extraction is performed on the sample audio data at the preset sampling rate to obtain sample audio features of the sample audio data at the preset sampling rate.

[0053] For example, for the sample audio data of each preset sampling rate in the training data set, feature extraction is performed on the sample audio data of the preset sampling rate, so as to obtain the sample audio features of the sample audio data of each preset sampling rate in the training data set, and then the sample audio features of 32k sample audio data and the sample audio features of 16k sample audio data of the same first sample audio data are bound to ensure that the sample audio features of 32k and the sample audio features of 16k in each batch exist and correspond one to one.

[0054] It should be noted that when there is sample audio data in the training data set whose length is less than the preset length, the sample audio data less than the preset length needs to be padded to the preset length by zero padding.

[0055] Step 404: Input the sample audio features of each of the preset sampling rates into an initial encoder to obtain the prediction coding features corresponding to each of the preset sampling rates output by the initial encoder.

[0056] For example, the network model uses a masked autoencoder (MAE) to mask the features by 70%. The initial encoder can use a Transformer structure to input the sample audio features of each preset sampling rate into the initial encoder, and the initial encoder encodes and analyzes the sample audio features of each preset sampling rate, and finally outputs the predicted coding features corresponding to each preset sampling rate.

[0057] Step 405: Determine first loss information based on predictive coding features of different preset sampling rates corresponding to the first sample audio data.

[0058] For example, after obtaining the predicted coding features corresponding to all preset sampling rates in the training data set, for each first sample audio data corresponding to a different preset sampling rate, the first loss information is calculated using the mean square error (MSE) based on the predicted coding features of the different preset sampling rates corresponding to the first sample audio data. The first loss information is the distance error of the predicted coding features of the different preset sampling rates corresponding to the first sample audio data. The first loss information can also be called contrast loss, which aims to bring the representation vectors of samples with different sampling rates closer.

[0059] Step 406: Determine the voiceprint extraction model based on the first loss information.

[0060] For example, when the first loss information is calculated, the distance error representing the first loss information is compared with the preset error. When the distance error representing the first loss information is greater than the preset error, the network parameters of the initial encoder are adjusted, and the process is continuously iterated until the distance error representing the first loss information is less than or equal to the preset error. The final initial encoder is determined as the voiceprint extraction model.

[0061] In this embodiment, based on the predicted coding features of different preset sampling rates corresponding to the first sample audio data, MSE loss is used to calculate the first loss information, and the characterization vectors of samples with different sampling rates are approximated based on the first loss information. The final voiceprint extraction model has the ability to adapt to the input of audio data with different sampling rates, and can perform voiceprint similarity comparison between audio data with different sampling rates.

[0062] In one embodiment, Figure 5 This is the second flow chart of the training method of the voiceprint extraction model provided by the embodiment of the present invention. Figure 5 As shown, the above step 406 determines the voiceprint extraction model based on the first loss information, which can be specifically implemented by the following steps: Step 4061: input the predicted coding features corresponding to each of the preset sampling rates into an initial decoder to obtain the predicted decoding features corresponding to each of the preset sampling rates output by the initial decoder.

[0063] Among them, the initial decoder can adopt a convolutional neural network (CNN) structure.

[0064] For example, when the predicted coding features corresponding to each preset sampling rate output by the initial encoder are obtained, the predicted coding features corresponding to each preset sampling rate are input into the initial decoder, and the predicted coding features corresponding to each preset sampling rate are decoded by the initial decoder, and finally the predicted decoding features corresponding to each preset sampling rate are output.

[0065] Step 4062: Determine second loss information based on the predicted decoding features and sample audio features of the same preset sampling rate corresponding to the first sample audio data.

[0066] For example, when the predicted decoding features corresponding to each preset sampling rate output by the initial decoder are obtained, for each preset sampling rate, the second loss information is calculated using MSE based on the predicted decoding features of the preset sampling rate corresponding to the first sample audio data and the sample audio features of the preset sampling rate. The second loss information may also be called reconstruction loss. For example, the second loss information is calculated based on the difference between the predicted decoding features of 32k corresponding to the first sample audio data and the input sample audio features of 32k, and the difference between the predicted decoding features of 16k corresponding to the first sample audio data and the input sample audio features of 16k.

[0067] Step 4063: Based on the second loss information and the first loss information, adjust the network parameters of the initial encoder to obtain the voiceprint extraction model.

[0068] For example, when the first loss information and the second loss information are obtained, the network parameters of the initial encoder are adjusted based on the first loss information and the second loss information, and the network parameters of the initial decoder are adjusted based on the second loss information, and the iteration is continued until the convergence condition is reached. Finally, the trained initial encoder is used as a voiceprint extractor, and the feature input is no longer masked, that is, the final trained initial encoder is determined as the voiceprint extraction model. Figure 6 is a schematic diagram of the structure of the network model provided by the embodiment of the present invention, such as Figure 6 As shown, the network model includes an initial encoder Encoder and an initial decoder Decoder, the sample audio features of each preset sampling rate are input into the initial encoder to obtain the predicted coding features corresponding to each preset sampling rate output by the initial encoder, the predicted coding features corresponding to each preset sampling rate output by the initial encoder are input into the initial decoder to obtain the predicted decoding features corresponding to each preset sampling rate output by the initial decoder, the first loss information is determined based on the predicted coding features of different preset sampling rates corresponding to the first sample audio data, the second loss information is determined based on the predicted decoding features and the sample audio features of the same preset sampling rate corresponding to the first sample audio data, the network parameters of the initial encoder are adjusted based on the second loss information and the first loss information to obtain the voiceprint extraction model.

[0069] In this embodiment, on the basis of the first loss information, the second loss information is determined based on the predicted decoding features and sample audio features of the same preset sampling rate corresponding to the first sample audio data, and self-supervised training of the voiceprint extraction model is implemented based on the first loss information and the second loss information, thereby improving the robustness and accuracy of the voiceprint extraction model.

[0070] In one embodiment, in step 403, feature extraction is performed on the sample audio data of the preset sampling rate to obtain the sample audio features of the sample audio data of the preset sampling rate, which can be specifically implemented in the following manner: Based on at least two types of filters, feature extraction is performed on the sample audio data of the preset sampling rate, and for each type of filter, sample audio features of the sample audio data of the preset sampling rate corresponding to the filter are obtained; the filters include at least two types of equal-pitch Mel filters, equal-pitch frequency filters and equal-pitch anti-Mel filters.

[0071] Accordingly, the above step 404 inputs the sample audio features of each of the preset sampling rates into the initial encoder to obtain the prediction coding features corresponding to each of the preset sampling rates output by the initial encoder, which can be specifically implemented in the following manner: The sample audio features of the preset sampling rates corresponding to all the filters are spliced ​​to obtain a combined audio feature; the combined audio feature is input into the initial encoder to obtain the predicted coding features corresponding to the preset sampling rates corresponding to the filters output by the initial encoder.

[0072] For example, taking three types of filters as an example, the three types of filters are equal-pitch Mel filters, equal-pitch frequency filters and equal-pitch anti-Mel filters. For each type of filter, based on the filter, feature extraction is performed on the sample audio data of the preset sampling rate to obtain the sample audio features of the sample audio data of the preset sampling rate corresponding to the filter. The sample audio features can be filter bank features. The sample audio features output by the three types of filters are spliced, and the feature splicing is 3 T The combined audio features of F, wherein 3 represents the channel dimension, T represents the time dimension, and F represents the frequency dimension, are input into the initial encoder to obtain the prediction coding features corresponding to each preset sampling rate corresponding to each filter output by the initial encoder.

[0073] In this embodiment, the traditional filterbank feature uses an equal-height Mel filter, where the filters are dense at low frequencies and sparse at high frequencies, which corresponds to the objective rule that the higher the frequency, the slower the human ear. However, in addition to low-frequency signals, transformers also have high-frequency signals such as partial discharges. Only filterbank features that fit the characteristics of the human ear cannot obtain information about high-frequency signals. Therefore, the present invention uses at least two types of filters, namely, equal-height Mel filters, equal-height frequency filters, and equal-height anti-Mel filters, to extract sample audio features to obtain a combination of multiple frequency scale features, thereby making the information of the combined audio features input to the initial encoder more complete, reducing information loss, and further improving the accuracy of the voiceprint extraction model obtained by the final training.

[0074] In one embodiment, the above step 402 determines the training data set based on the first sample audio data, which can be specifically implemented in the following manner: Obtain stateless label-free third sample audio data of other devices except the first sample transformer; determine the training data set based on the first sample audio data and the third sample audio data, wherein the training data set also includes sample audio data of at least two preset sampling rates corresponding to the third sample audio data.

[0075] Optionally, sampling rates of the first sample audio data and the third sample audio data are both greater than a maximum preset sampling rate among all preset sampling rates.

[0076] Downsample the first sample audio data to obtain sample audio data of at least two preset sampling rates corresponding to the first sample audio data; downsample the third sample audio data to obtain sample audio data of at least two preset sampling rates corresponding to the third sample audio data; determine the training data set based on the sample audio data of at least two preset sampling rates corresponding to the first sample audio data and the sample audio data of at least two preset sampling rates corresponding to the third sample audio data.

[0077] Among them, other equipment may be electrical equipment and / or industrial equipment, the electrical equipment may be, for example, a switch cabinet and a reactor, etc., and the industrial equipment may be, for example, a motor and an air compressor, etc.

[0078] For example, when all the first sample audio data are obtained, the first sample audio data can be downsampled for each first sample audio data to obtain sample audio data of at least two preset sampling rates corresponding to the first sample audio data. According to the same method, sample audio data of at least two preset sampling rates corresponding to each first sample audio data can be obtained.

[0079] It is also possible to obtain third sample audio data of stateless tags of other devices except the first sample transformer, that is, to obtain third sample audio data of stateless tags of electrical equipment and / or industrial equipment. The sampling rate of the third sample audio data is greater than the preset sampling rate. For each third sample audio data, the third sample audio data can be downsampled to obtain sample audio data of at least two preset sampling rates corresponding to the third sample audio data. According to the same method, sample audio data of at least two preset sampling rates corresponding to each third sample audio data can be obtained.

[0080] A set of sample audio data of at least two preset sampling rates corresponding to all first sample audio data and a set of sample audio data of at least two preset sampling rates corresponding to all third sample audio data are determined as a training data set. Taking two preset sampling rates as an example, 32k and 16k respectively, the training data set includes 32k sample audio data and 16k sample audio data corresponding to all first sample audio data and 32k sample audio data and 16k sample audio data corresponding to all third sample audio data.

[0081] In this embodiment, a training data set is determined based on sample audio data of at least two preset sampling rates corresponding to the first sample audio data, and sample audio data of at least two preset sampling rates corresponding to the third sample audio data, so that the training data set includes sample audio data of other devices besides the transformer, enriching the data source of the training data set, so that the voiceprint extraction model trained based on the training data set can adapt to unknown audio data to be tested, thereby improving the generalization of the voiceprint extraction model; in addition, the sample audio data is downsampled to at least two preset sampling rates according to the actual sampling rate, so that the retrieval of the closest voiceprint is more accurate.

[0082] The transformer state identification device provided by the present invention is described below. The transformer state identification device described below and the transformer state identification method described above can be referred to each other.

[0083] Figure 7 is a schematic diagram of the structure of a transformer state identification device provided by an embodiment of the present invention, such as Figure 7 As shown, the transformer state identification device 700 includes an acquisition unit 701, an extraction unit 702 and an identification unit 703, wherein: An acquisition unit 701 is used to acquire audio data to be tested of the transformer; An extraction unit 702 is used to input the audio data to be tested into a voiceprint extraction model to obtain a target voiceprint output by the voiceprint extraction model; the voiceprint extraction model is trained based on the first sample audio data without state labels of the first number of first sample transformers; The identification unit 703 is used to identify the state of the transformer based on the target voiceprint and at least one registered voiceprint; the at least one registered voiceprint is a voiceprint obtained by inputting a second number of second sample audio data including a state label of a second sample transformer into the voiceprint extraction model, and the second number is less than the first number.

[0084] The transformer state identification device provided by the present invention inputs the obtained audio data of the transformer to be tested into the voiceprint extraction model to obtain the target voiceprint output by the voiceprint extraction model, and identifies the state of the transformer based on the target voiceprint and at least one registered voiceprint, wherein the voiceprint extraction model is obtained by training based on a first number of first sample audio data without state labels, and the registered voiceprint is a voiceprint obtained by inputting a second number of second sample audio data including state labels into the voiceprint extraction model. It can be seen that the present invention first obtains a voiceprint extraction model based on a large number of first sample audio data without state labels, and then determines at least one registered voiceprint based on a small number of second sample audio data with state labels through the voiceprint extraction model, and then realizes the state identification of the transformer in combination with the registered voiceprint, which improves the accuracy of transformer state identification when only a small number of sample audio data with state labels are required.

[0085] Based on any of the above embodiments, the identification unit 703 is specifically used for: For each of the registered voiceprints, determining the similarity between the target voiceprint and the registered voiceprint; The state of the transformer is determined based on the target state label of the second sample audio data corresponding to the registered voiceprint with the maximum similarity.

[0086] Based on any of the above embodiments, the voiceprint extraction model is trained based on the following method: Obtaining first sample audio data of the stateless tag of the first sample transformer; Based on the first sample audio data, determining a training data set, the training data set including sample audio data of at least two preset sampling rates corresponding to the first sample audio data; For each of the sample audio data of the preset sampling rate in the training data set, feature extraction is performed on the sample audio data of the preset sampling rate to obtain sample audio features of the sample audio data of the preset sampling rate; Inputting the sample audio features of each of the preset sampling rates into an initial encoder to obtain the prediction coding features corresponding to each of the preset sampling rates output by the initial encoder; Determining first loss information based on predictive coding features of different preset sampling rates corresponding to the first sample audio data; Based on the first loss information, the voiceprint extraction model is determined.

[0087] Based on any of the foregoing embodiments, determining the voiceprint extraction model based on the first loss information includes: Inputting the predicted coding features corresponding to each of the preset sampling rates into an initial decoder to obtain the predicted decoding features corresponding to each of the preset sampling rates output by the initial decoder; Determining second loss information based on the predicted decoding features and the sample audio features at the same preset sampling rate corresponding to the first sample audio data; Based on the second loss information and the first loss information, the network parameters of the initial encoder are adjusted to obtain the voiceprint extraction model.

[0088] Based on any of the above embodiments, the step of extracting features from the sample audio data of the preset sampling rate to obtain sample audio features of the sample audio data of the preset sampling rate includes: Based on at least two types of filters, feature extraction is performed on the sample audio data of the preset sampling rate, and for each type of the filters, sample audio features of the sample audio data of the preset sampling rate corresponding to the filters are obtained; the filters include at least two types of equal-pitch Mel filters, equal-pitch frequency filters, and equal-pitch anti-Mel filters; The step of inputting the sample audio features of each of the preset sampling rates into an initial encoder to obtain the prediction coding features corresponding to each of the preset sampling rates output by the initial encoder includes: splicing the sample audio features of the preset sampling rates corresponding to all the filters to obtain a combined audio feature; The combined audio feature is input into the initial encoder to obtain the prediction coding feature corresponding to each of the preset sampling rates corresponding to each of the filters output by the initial encoder.

[0089] Based on any of the foregoing embodiments, determining a training data set based on the first sample audio data includes: Acquire third sample audio data of stateless tags of other devices except the first sample transformer; The training data set is determined based on the first sample audio data and the third sample audio data, wherein the training data set also includes sample audio data of at least two preset sampling rates corresponding to the third sample audio data.

[0090] Based on any of the above embodiments, the sampling rates of the first sample audio data and the third sample audio data are both greater than the maximum preset sampling rate among all preset sampling rates; The determining the training data set based on the first sample audio data and the third sample audio data includes: Downsampling the first sample audio data to obtain sample audio data of at least two preset sampling rates corresponding to the first sample audio data; Downsampling the third sample audio data to obtain sample audio data of at least two preset sampling rates corresponding to the third sample audio data; The training data set is determined based on sample audio data of at least two preset sampling rates corresponding to the first sample audio data and sample audio data of at least two preset sampling rates corresponding to the third sample audio data.

[0091] Figure 8 is a schematic diagram of the physical structure of an electronic device provided by an embodiment of the present invention, such as Figure 8 As shown, the electronic device may include: a processor 810, a communications interface 820, a memory 830 and a communication bus 840, wherein the processor 810, the communications interface 820 and the memory 830 communicate with each other via the communication bus 840. The processor 810 may call the logic instructions in the memory 830 to execute the transformer state identification method, which includes: obtaining the audio data to be tested of the transformer; Inputting the audio data to be tested into a voiceprint extraction model to obtain a target voiceprint output by the voiceprint extraction model; the voiceprint extraction model is trained based on the first sample audio data without state labels of the first sample transformer of the first number; The state of the transformer is identified based on the target voiceprint and at least one registered voiceprint; the at least one registered voiceprint is a voiceprint obtained by inputting a second number of second sample audio data including a state label of a second sample transformer into the voiceprint extraction model, and the second number is smaller than the first number.

[0092] In addition, the logic instructions in the above-mentioned memory 830 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0093] On the other hand, the present invention further provides a computer program product, the computer program product includes a computer program, the computer program can be stored in a non-transitory computer-readable storage medium, when the computer program is executed by a processor, the computer can execute the transformer state identification method provided by the above methods, the method includes: obtaining audio data to be tested of the transformer; Inputting the audio data to be tested into a voiceprint extraction model to obtain a target voiceprint output by the voiceprint extraction model; the voiceprint extraction model is trained based on the first sample audio data without state labels of the first sample transformer of the first number; The state of the transformer is identified based on the target voiceprint and at least one registered voiceprint; the at least one registered voiceprint is a voiceprint obtained by inputting a second number of second sample audio data including a state label of a second sample transformer into the voiceprint extraction model, and the second number is smaller than the first number.

[0094] In another aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the method for identifying a transformer state provided by the above methods is implemented, and the method comprises: obtaining audio data to be tested of the transformer; Inputting the audio data to be tested into a voiceprint extraction model to obtain a target voiceprint output by the voiceprint extraction model; the voiceprint extraction model is trained based on the first sample audio data without state labels of the first sample transformer of the first number; The state of the transformer is identified based on the target voiceprint and at least one registered voiceprint; the at least one registered voiceprint is a voiceprint obtained by inputting a second number of second sample audio data including a state label of a second sample transformer into the voiceprint extraction model, and the second number is smaller than the first number.

[0095] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0096] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0097] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A transformer state identification method, characterized in that: include: Obtain the audio data to be tested of the transformer; Inputting the audio data to be tested into a voiceprint extraction model to obtain a target voiceprint output by the voiceprint extraction model; The voiceprint extraction model is trained based on the first sample audio data without state labels of the first sample transformer of the first number; Based on the target voiceprint and at least one registered voiceprint, identifying the state of the transformer; The at least one registered voiceprint is a voiceprint obtained by inputting a second number of second sample audio data including state labels of a second sample transformer into the voiceprint extraction model, and the second number is smaller than the first number.

2. The transformer state identification method according to claim 1, characterized in that: The identifying the state of the transformer based on the target voiceprint and at least one registered voiceprint includes: For each of the registered voiceprints, determining the similarity between the target voiceprint and the registered voiceprint; The state of the transformer is determined based on the target state label of the second sample audio data corresponding to the registered voiceprint with the maximum similarity.

3. The transformer state identification method according to claim 1, characterized in that: The voiceprint extraction model is trained based on the following method: Obtaining first sample audio data of the stateless tag of the first sample transformer; Based on the first sample audio data, determining a training data set, the training data set including sample audio data of at least two preset sampling rates corresponding to the first sample audio data; For each of the sample audio data of the preset sampling rate in the training data set, feature extraction is performed on the sample audio data of the preset sampling rate to obtain sample audio features of the sample audio data of the preset sampling rate; Inputting the sample audio features of each of the preset sampling rates into an initial encoder to obtain the prediction coding features corresponding to each of the preset sampling rates output by the initial encoder; Determining first loss information based on predictive coding features of different preset sampling rates corresponding to the first sample audio data; Based on the first loss information, the voiceprint extraction model is determined.

4. The transformer state identification method according to claim 3, characterized in that: The determining the voiceprint extraction model based on the first loss information includes: Inputting the predicted coding features corresponding to each of the preset sampling rates into an initial decoder to obtain the predicted decoding features corresponding to each of the preset sampling rates output by the initial decoder; Determining second loss information based on predicted decoding features and sample audio features at the same preset sampling rate corresponding to the first sample audio data; Based on the second loss information and the first loss information, the network parameters of the initial encoder are adjusted to obtain the voiceprint extraction model.

5. The transformer state identification method according to claim 3, characterized in that: The extracting features of the sample audio data of the preset sampling rate to obtain the sample audio features of the sample audio data of the preset sampling rate includes: Based on at least two types of filters, feature extraction is performed on the sample audio data of the preset sampling rate, and for each type of the filters, sample audio features of the sample audio data of the preset sampling rate corresponding to the filters are obtained; the filters include at least two types of equal-pitch Mel filters, equal-pitch frequency filters, and equal-pitch anti-Mel filters; The step of inputting the sample audio features of each of the preset sampling rates into an initial encoder to obtain the prediction coding features corresponding to each of the preset sampling rates output by the initial encoder includes: splicing the sample audio features of the preset sampling rates corresponding to all the filters to obtain a combined audio feature; The combined audio feature is input into the initial encoder to obtain the prediction coding feature corresponding to each of the preset sampling rates corresponding to each of the filters output by the initial encoder.

6. The transformer state identification method according to any one of claims 3 to 5, characterized in that: The determining of a training data set based on the first sample audio data includes: Acquire third sample audio data of stateless tags of other devices except the first sample transformer; The training data set is determined based on the first sample audio data and the third sample audio data, wherein the training data set also includes sample audio data of at least two preset sampling rates corresponding to the third sample audio data.

7. The transformer state identification method according to claim 6, characterized in that: The sampling rates of the first sample audio data and the third sample audio data are both greater than the maximum preset sampling rate among all preset sampling rates; The determining the training data set based on the first sample audio data and the third sample audio data includes: Downsampling the first sample audio data to obtain sample audio data of at least two preset sampling rates corresponding to the first sample audio data; Downsampling the third sample audio data to obtain sample audio data of at least two preset sampling rates corresponding to the third sample audio data; The training data set is determined based on sample audio data of at least two preset sampling rates corresponding to the first sample audio data and sample audio data of at least two preset sampling rates corresponding to the third sample audio data.

8. A transformer state identification device, characterized in that: include: An acquisition unit, used for acquiring audio data to be tested of the transformer; An extraction unit, used for inputting the audio data to be tested into a voiceprint extraction model to obtain a target voiceprint output by the voiceprint extraction model; The voiceprint extraction model is trained based on the first sample audio data without state labels of the first sample transformer of the first number; an identification unit, configured to identify a state of the transformer based on the target voiceprint and at least one registered voiceprint; The at least one registered voiceprint is a voiceprint obtained by inputting a second number of second sample audio data including state labels of a second sample transformer into the voiceprint extraction model, and the second number is smaller than the first number.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the transformer state identification method according to any one of claims 1 to 7 is implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the transformer state identification method according to any one of claims 1 to 7 is implemented.

Images