Transformer fault voiceprint diagnosis method based on Transform algorithm

Abstract

The invention discloses a transformer fault voiceprint diagnosis method based on a Transform algorithm, and the method comprises the steps: installation of collection equipment: installing a sound collection device at the periphery of a transformer, and collection of sound signals: collecting the sound signals in the operation process of the transformer through the sound collection device, carrying out normalization, pre-emphasis, MRSVD noise reduction processing, filtering processing and signal feature extraction on the acquired sound signals; extracting a dominant frequency, a voiceprint component variation, a peak factor and a signal sound pressure level of the sound signals; according to the invention, the method can realize early warning and accurate diagnosis of faults of the transformer, and improves stable operation and reliability of the transformer. The method comprises the following steps of: acquiring a sound signal characteristic vector, caching the vectors of the four characteristics into a system, comparing the sound signal characteristic vector which is being diagnosed with the characteristic vector cached in the last time period, and calculating the variable quantity of each characteristic, and the like. The method can realize early warning and accurate diagnosis of the faults of the transformer, and improves stable operation and reliability of the transformer.

Description

Technical Field

[0001] The present invention relates to the technical field of power transformers, and in particular to a transformer fault soundprint diagnosis method based on a Transformer algorithm. Background Art

[0002] Power transformers (abbreviated as transformers) are very important electrical equipment in power systems. In the long-term operation of transformers, it is inevitable that the transformers will not be able to operate normally due to potential reasons such as discharge, overheating, loose components, etc. If early warning is not given in time, serious safety accidents and production accidents may occur. To ensure the stable and safe operation of transformers, diagnostic methods that can accurately identify the fault type are required.

[0003] At present, the diagnosis methods for transformers at home and abroad are mainly through the analysis of ultrasonic signals, oil and gas composition changes, infrared images, vibration signals, sound signals and other means of transformers. Among them, mechanical fault identification based on voiceprints has been widely studied in recent years due to its advantages such as high accuracy, timely detection and non-invasiveness. The application of voiceprint recognition in transformer detection is mainly based on the differences in sound characteristics of different fault types. When the transformer equipment in the substation is in operation, it is affected by various factors such as internal current and magnetic field to generate sound signals, and different operating states are manifested as different voiceprints. By analyzing the sound characteristics of the machine or equipment, and then performing feature extraction and deep neural network training, a large number of samples are autonomously learned, and the weights and bias parameters of the deep neural network are continuously improved, and the sound signal characteristics are autonomously obtained, so as to realize intelligent fault diagnosis.

[0004] At present, most of the voiceprint technologies used to diagnose transformer faults at home and abroad extract the time-frequency domain features of speech signals, such as short-time energy, wavelet packet decomposition, short-time Fourier, Mel-scale spectrum coefficients and other speech features, and then use support vector machine SVM, decision tree DT, multi-layer perceptron MLP, convolutional neural network CNN, residual network RESNET, long short-term memory LSTM and other classification models to classify and identify the features. Although certain effects have been achieved, there are still problems such as recognition accuracy bottlenecks, few recognition types, high resource usage, and complicated sample calibration. Summary of the invention

[0005] The purpose of this section is to summarize some aspects of the embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the specification abstract and the invention title of this application to avoid blurring the purpose of this section, the specification abstract and the invention title, and such simplifications or omissions cannot be used to limit the scope of the present invention.

[0006] Therefore, the purpose of the present invention is to provide a transformer fault voiceprint diagnosis method based on the Transformer algorithm, so as to achieve early warning and accurate diagnosis of transformer faults and improve the stable operation and reliability of the transformer.

[0007] To solve the above technical problems, according to one aspect of the present invention, the present invention provides the following technical solutions:

[0008] A transformer fault voiceprint diagnosis method based on Transformer algorithm, comprising:

[0009] S1. Installation of collection equipment: Install the sound collection device around the transformer;

[0010] S2. Sound signal collection: using a sound collection device to collect sound signals during the operation of the transformer, and adding the collected sound signals to the transformer sound sample library;

[0011] S3, sound signal preprocessing: normalize, pre-emphasize, perform MRSVD noise reduction processing, and filter processing on the collected sound signals;

[0012] S4, signal feature extraction: extract the main frequency of the sound signal, the change of the voiceprint component, the peak factor and the signal sound pressure level, and cache the vectors of the four features into the system;

[0013] S5, signal feature matching: compare the feature vector of the sound signal being diagnosed with the feature vector cached in the previous period, and calculate the change amount of each feature;

[0014] S6, abnormal sound judgment: if the characteristic change in step S5 exceeds the threshold, it is determined that there is an abnormal sound, and the next step of fault type identification is performed; otherwise, it is determined that the transformer sound is normal, and the label of the sample in the sample library is updated;

[0015] S7, if step S6 determines that the voice is a foreign voice, convert the voice signal pre-processed in step S3 into a Mel time-frequency spectrum;

[0016] S8, putting the time-frequency spectrum calculated in step S7 into the pre-trained Transformer model for recognition. If the recognition result is the set fault category, the fault category is output and the sound sample is added to the sample library for iterating the model. Otherwise, it is output as a normal signal and the label of the sample in the sample library is updated.

[0017] S9, fault type output: If in step S8, the model recognition result is the set fault category, then the fault category is output and the sample is added to the fault sample library for iterating the model; otherwise, the output is normal and the sample is added to the sample library to be reviewed.

[0018] As a preferred solution of the transformer fault voiceprint diagnosis method based on the Transformer algorithm described in the present invention, in step S1, the voiceprint collection microphone for collecting the transformer sound signal meets the following requirements: sensitivity greater than 10mv / Pa, frequency response range 20Hz~20kHz, sampling frequency 48000Hz, bit depth greater than 16bit, dynamic range 38dBA-110dB, and total harmonic distortion rate not greater than 3%.

[0019] As a preferred solution of the transformer fault soundprint diagnosis method based on the Transformer algorithm described in the present invention, in step S2, the real-time collected sound signal is intercepted according to a fixed length, and the entire sound signal greater than 1 s is selected for analysis.

[0020] As a preferred solution of the transformer fault soundprint diagnosis method based on the Transformer algorithm described in the present invention, in step S3, the specific process of signal preprocessing is as follows:

[0021] S301, data normalization: converting audio data of different amplitude ranges to a uniform range [-1, 1];

[0022] S302, pre-emphasis: Minimize the influence of the glottal pulse through a first-order high-pass filter, wherein the first-order high-pass filter is a first-order FIR filter with a coefficient of 0.9375.

[0023] S303, MRSVD noise reduction processing: MRSVD constructs the signal into a second-order matrix, decomposing it into two components each time, and then performs inverse operation on the extracted signal to reconstruct it, thereby achieving noise reduction on the original signal;

[0024] S304, filtering processing: using a Butterworth filter to filter out background noise below 20 Hz.

[0025] As a preferred solution of the transformer fault voiceprint diagnosis method based on the Transformer algorithm described in the present invention, in step S4, the voiceprint component change calculation formula is:

[0026]

[0027] Where D is the change in voiceprint components, f max is the upper limit frequency of the calculation, A1(n×10) is the n×10 Hz frequency component amplitude of the sound signal being diagnosed, and A2(n×10) is the n×10 Hz frequency component amplitude of the sound signal cached in the previous period.

[0028] As a preferred solution of the transformer fault soundprint diagnosis method based on the Transformer algorithm described in the present invention, in step S4, the peak factor calculation formula is:

[0029]

[0030] Among them, X peak is the peak value of the signal, X RMS is the RMS value of the signal.

[0031] As a preferred solution of the transformer fault voiceprint diagnosis method based on the Transformer algorithm described in the present invention, in step S5, the specific process of signal feature matching is as follows:

[0032] S501, if the signal main frequency is not within the preset range, it is recorded as 0, otherwise it is recorded as 1;

[0033] S502, if the change in the voiceprint component is greater than 10%, it is recorded as 0, otherwise it is recorded as 1;

[0034] S503, if the change in the peak factor between the currently diagnosed sound signal and the sound signal cached in the previous period is greater than 10%, it is recorded as 0, otherwise it is recorded as 1;

[0035] S504: If the total sound pressure level or the segmented sound pressure level exceeds the threshold, it is recorded as 1, otherwise it is recorded as 0;

[0036] S505. Add the results of steps S501-S504 to obtain a value T.

[0037] As a preferred solution of the transformer fault voiceprint diagnosis method based on the Transformer algorithm described in the present invention, in step S6, the abnormal sound judgment is as follows: if the value T calculated in step S5 is greater than 0, it is judged that there is an abnormal sound, otherwise it is output that there is no abnormal sound and the diagnosis of the next voice segment is entered.

[0038] As a preferred solution of the transformer fault voiceprint diagnosis method based on the Transformer algorithm described in the present invention, in step S7, when the speech signal preprocessed in step S3 is converted into a Mel time-frequency spectrum, the FFT window length is selected to be greater than 8192, the data of the mel filter group is selected to be greater than 128, and the size of the Mel spectrum is scaled to 224*244.

[0039] As a preferred solution of the transformer fault voiceprint diagnosis method based on the Transformer algorithm described in the present invention, in step S8, the Transformer model is trained as follows:

[0040] S801, constructing a model transformer sound sample library, where the sample library samples include sounds collected from different transformers, including fault samples, environmental noise samples, and samples of transformer equipment to be detected;

[0041] S802, data cleaning: filter out damaged samples and samples that are not transformer backgrounds, check the sample length of the sample set, and divide the samples into samples of fixed length;

[0042] S803, data labeling: labeling whether the sample contains fault sound or environmental sound;

[0043] S804, all labeled samples are combined into a standard sample library;

[0044] S805, extracting the mel spectrum of the samples from the standard sample library organized in S804, and scaling it to a size of 224*224 to form a mel spectrum set;

[0045] S806, model training: Build a Transformer-based swin-transformer model and use the Mel spectrogram sample set compiled in S804 for training.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] 1. By converting the sound signal into a Mel time-frequency spectrum image to replace the sound signal as the input of the model, the time domain and frequency domain characteristics of the sound signal are taken into account. At the same time, due to the different textures of different sounds on the Mel spectrum, the model has strong interpretability and reduces the granularity of annotation. A Transformer-based swin-transformer model is constructed, and the ambient sound and fault sound of the transformer are accurately identified by model tuning. The stationary signal is filtered through general features to reduce computing resources. The purpose of continuous optimization and iteration of the model is achieved by continuously expanding the training set.

[0048] 2. The present invention proposes a method of converting speech signals into Mel spectrograms and then using image classification to effectively reduce the timeliness of sound sample annotation.

[0049] 3. The swin-transformer model based on the Transformer model and its tuning method proposed in the present invention can effectively increase the recognition accuracy and generalization.

[0050] 4. The present invention uses the main frequency, voiceprint component change, peak factor and sound pressure level threshold characteristics, taking into account the time domain and frequency domain characteristics of speech for filtering, which can effectively reduce the system's resource occupancy and processing capacity. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the present invention will be described in detail below in combination with the accompanying drawings and detailed embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative labor. Among them:

[0052] Figure 1 This is a flow chart of a transformer fault voiceprint diagnosis method based on the Transformer algorithm of the present invention;

[0053] Figure 2 It is a schematic diagram of the installation of a sound collection device in a transformer fault voiceprint diagnosis method based on the Transformer algorithm of the present invention;

[0054] Figure 3 This is a model structure diagram of the Transformer in a transformer fault voiceprint diagnosis method based on the Transformer algorithm of the present invention. DETAILED DESCRIPTION

[0055] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0056] Secondly, the present invention is described in detail with reference to schematic diagrams. When describing the embodiments of the present invention in detail, for the sake of convenience, the cross-sectional diagrams showing the device structure will not be partially enlarged according to the general scale, and the schematic diagrams are only examples, which should not limit the scope of protection of the present invention. In addition, in actual production, the three-dimensional dimensions of length, width and depth should be included.

[0057] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0058] The present invention provides a transformer fault voiceprint diagnosis method based on the Transformer algorithm, which can realize early warning and accurate diagnosis of transformer faults and improve the stable operation and reliability of the transformer.

[0059] like Figure 1 As shown, this embodiment discloses a transformer fault voiceprint diagnosis method based on the Transformer algorithm, comprising the following steps:

[0060] Step S1, installation of acquisition equipment: Figure 2As shown, the sound collection device is installed near the transformer. In this embodiment, the non-contact device is installed on the cooler side and the oil pillow side; the sensor parameters of the device are sensitivity 10mv / Pa, frequency response range 20Hz~20kHz, sampling frequency 48000Hz, bit depth 16bit, dynamic range 38dBA-110dB;

[0061] Step S2, sound signal collection: Use a sound collection device to collect sound signals during the operation of the transformer. The device collects data in real time and uses data with a window length of 10s as a sample to be detected;

[0062] Step S3, sound signal preprocessing: normalize, pre-emphasize, perform MRSVD noise reduction processing, and filter processing on the collected sound signal, as follows:

[0063] Step S301, data normalization: converting audio data of different amplitude ranges to a unified range of [-1, 1];

[0064] Step S302, pre-emphasis: use a first-order FIR filter to process the signal with a coefficient of 0.9375.

[0065] Step S303: MRSVD noise reduction processing, the detailed decomposition process is as follows:

[0066] (1) For the original signal s(t) = (x1, x2, ..., x N ,)Construct the following matrix with 2 rows:

[0067]

[0068] (2) For H∈R (N-1)×2 Perform singular value decomposition:

[0069] H=U∧V T

[0070] Where U∈R (N-1)×(N-1) , V∈R 2×2 is an orthogonal matrix; Λ is a diagonal matrix; Λ=(diag(σ1,σ2),0) or its transpose, with a rank of 2.

[0071] (3) According to the singular value decomposition principle, the signal corresponding to the larger singular value is an approximate signal A dominated by the useful signal. j , the smaller singular value corresponds to the detail signal D dominated by noise j , and then continue to decompose the approximate signal;

[0072] (4) After multi-level decomposition of the signal, the signal features hidden under complex interference can be extracted, and finally noise reduction is achieved by inverse operation and reconstruction of the extracted signal. MRSVD does not need to determine the reconstruction order of effective singular values, and can extract useful detail features in a strong noise background more precisely than SVD.

[0073] Step S304, filtering: using a 5th order Butterworth filter to filter out background noise below 20 Hz;

[0074] Step S4, signal feature extraction: extract the main frequency of the sound signal, the change in voiceprint components, the peak factor and the signal sound pressure level, and cache the vectors of the four features into the system. Specifically:

[0075] S401, extracting the main frequency of the signal: performing FFT operation on the signal, and then taking the frequency point with the largest amplitude as a feature;

[0076] S402, extract the change in the signal voiceprint component, and the calculation formula is as follows:

[0077]

[0078] D is the change in the voiceprint component; f max is the upper limit frequency of the calculation, A1(n×10) is the amplitude of the n×10 Hz frequency component of the sound signal being diagnosed, and A2(n×10) is the amplitude of the n×10 Hz frequency component of the sound signal cached in the previous period;

[0079] S403, peak factor calculation, the calculation formula is as follows:

[0080]

[0081] Xpeak is the peak value of the signal, X RMS is the RMS value of the signal;

[0082] S404, signal sound pressure level calculation: calculate the total sound pressure level and segmented sound pressure level of the signal; in this embodiment, a MEMS microphone is used, which is different from an analog microphone, and its calculation formula is:

[0083] SPL = 20 × log 10 (RMS)+PAOP

[0084] Wherein, RMS is the root mean square value of the signal, and PAOP is the maximum sound pressure level of the digital microphone (in dB). In this embodiment, the segmentation method of the segmented sound pressure level is to segment the 10s signal into 0.2s segments and intercept them.

[0085] Step S5, signal feature matching: compare the feature vector of the sound signal being diagnosed with the feature vector cached in the previous period, and calculate the change amount of each feature, specifically:

[0086] S501: If the signal main frequency is not within the preset range, it is recorded as 0, otherwise it is recorded as 1. The following are common transformer main frequencies: 110kv transformer: 100Hz, 220kv transformer: 200Hz, 500kv transformer: 500Hz, 800kv transformer: 400Hz;

[0087] S502, if the change in the voiceprint component is greater than 10%, it is recorded as 0, otherwise it is recorded as 1;

[0088] S503, if the change in the peak factor between the currently diagnosed sound signal and the sound signal cached in the previous period is greater than 10%, it is recorded as 0, otherwise it is recorded as 1;

[0089] S504. If the total sound pressure level or the segmented sound pressure level exceeds the threshold, it is recorded as 1, otherwise it is recorded as 0. In this implementation case, the setting of the threshold is based on the actual transformer environment;

[0090] S505, adding the results of steps S501-S504 to obtain a value T;

[0091] Step S6, abnormal sound judgment: if the characteristic change in step S5 exceeds the threshold, it is determined that there is an abnormal sound, that is, the value of T is greater than 0, and the next step of fault type identification is performed; otherwise, it is determined that the transformer sound is normal, and the label of the sample in the sample library is updated;

[0092] Step S7: If the voice signal is judged as a foreign voice in step S6, the voice signal preprocessed in step S3 is converted into a Mel time-frequency spectrum. In this implementation case, the FFT window length is selected as 8192, and the data of the Mel filter group is selected as 256. In order to adapt to the input of the model, the size of the Mel spectrum needs to be scaled to 224*244;

[0093] Step S8: put the time-frequency spectrum calculated in step S7 into the pre-trained Transformer model for recognition. If the recognition result is the set fault category, the fault category is output and the sound sample is added to the sample library for iterating the model; otherwise, the normal signal is output and the label of the sample in the sample library is updated. The specific process of model training is as follows:

[0094] S801. Construct a sound sample library of the model transformer. To expand the scope of use of the model, the sample library should include sounds collected from different transformers, such as 110kv transformers, 220kv transformers, 500kv transformers, and 800kv transformers; include fault samples and environmental noise samples; and include samples of transformer equipment to be tested.

[0095] S802, data cleaning: filter out damaged samples and samples that are not transformer backgrounds; check the sample length of the sample set and divide the samples into samples of fixed length. In this implementation case, in order to obtain a higher frequency resolution of the Mel spectrum, 10s is used as a sample.

[0096] S803. Data labeling: Label whether the sample contains fault sounds or environmental sounds. In this implementation case, the main faults identified are abnormal noise from the cooler, DC bias magnetization, loose clamps, heavy overload, short-circuit impact, partial discharge, and loose core components. The main environmental sounds identified are rain, thunder, pedestrians talking, birds singing, whistles, and frogs.

[0097] S804, all labeled samples form a standard sample library;

[0098] S805, extract the mel spectrum of the sample from the standard sample library sorted in S804 (same as step S7), and scale it to a size of 224*224 to form a mel spectrum set;

[0099] S806, model training: Build a swin-transformer model based on Transformer, and use the Mel spectrogram sample set organized in S804 for training. The model architecture is shown in the attached Figure 3 , the specific process is as follows:

[0100] (1) Input an RGB image with a size of 224*224*3;

[0101] (2) The image is segmented and position encoded through the Patch Partition layer;

[0102] (3) The segmented data is subjected to feature mapping through the Linear Embedding layer;

[0103] (4) The feature mapped data is input into the Transformer block with improved self-attention calculation, and together with Linear Embedding, it is called the first stage;

[0104] (5) Unlike stage 1, stages 2-4 require Patch Merging for downsampling before inputting the model to produce a hierarchical representation;

[0105] (6) Finally, the data after stage 4 is classified through the output module (including a LayerNorm layer, an AdaptiveAvgPool1d layer and a fully connected layer) to obtain the classification result;

[0106] In this example, in order to solve the problem of sample imbalance, Focal-loss is used as the model loss function; cutout and mixup are used to enhance the data; and the cosine annealing decay learning rate scheduling strategy is used to tune the model;

[0107] S9, fault type output: If the model recognition result in step S8 is the set fault category, the fault category is output and the sample is added to the fault sample library for iterating the model; otherwise, the output is normal and the sample is added to the sample library to be reviewed.

[0108] Although the present invention has been described above with reference to the embodiments, various modifications may be made thereto and parts thereof may be replaced by equivalents without departing from the scope of the present invention. In particular, as long as there is no structural conflict, the various features in the embodiments disclosed in the present invention may be used in combination with each other in any manner, and the fact that these combinations are not exhaustively described in this specification is only for the sake of omitting space and saving resources. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A transformer fault voiceprint diagnosis method based on Transformer algorithm, characterized in that: include: S1. Installation of collection equipment: Install the sound collection device around the transformer; S2. Sound signal collection: using a sound collection device to collect sound signals during the operation of the transformer, and adding the collected sound signals to the transformer sound sample library; S3, sound signal preprocessing: normalize, pre-emphasize, perform MRSVD noise reduction processing, and filter processing on the collected sound signals; S4, signal feature extraction: extract the main frequency of the sound signal, the change of the voiceprint component, the peak factor and the signal sound pressure level, and cache the vectors of the four features into the system; S5, signal feature matching: compare the feature vector of the sound signal being diagnosed with the feature vector cached in the previous period, and calculate the change amount of each feature; S6, abnormal sound judgment: if the characteristic change in step S5 exceeds the threshold, it is determined that there is an abnormal sound, and the next step of fault type identification is performed; otherwise, it is determined that the transformer sound is normal, and the label of the sample in the sample library is updated; S7, if step S6 determines that the voice is a foreign voice, convert the voice signal pre-processed in step S3 into a Mel time-frequency spectrum; S8, putting the time-frequency spectrum calculated in step S7 into the pre-trained Transformer model for recognition. If the recognition result is the set fault category, the fault category is output and the sound sample is added to the sample library for iterating the model. Otherwise, it is output as a normal signal and the label of the sample in the sample library is updated. S9, fault type output: If in step S8, the model recognition result is the set fault category, then the fault category is output and the sample is added to the fault sample library for iterating the model; otherwise, the output is normal and the sample is added to the sample library to be reviewed.

2. According to claim 1, a transformer fault voiceprint diagnosis method based on Transformer algorithm is characterized in that: In step S1, the voiceprint collection microphone for collecting transformer sound signals meets the following requirements: sensitivity greater than 10mv / Pa, frequency response range 20Hz-20kHz, sampling frequency 48000Hz, bit depth greater than 16bit, dynamic range 38dBA-110dB, and total harmonic distortion rate no greater than 3%.

3. According to the transformer fault voiceprint diagnosis method based on Transformer algorithm in claim 1, it is characterized in that: In step S2, the sound signal collected in real time is intercepted according to a fixed length, and the entire sound signal longer than 1 s is selected for analysis.

4. According to the transformer fault voiceprint diagnosis method based on Transformer algorithm in claim 1, it is characterized in that: In step S3, the specific process of signal preprocessing is as follows: S301, data normalization: converting audio data of different amplitude ranges to a uniform range [-1, 1]; S302, pre-emphasis: Minimize the influence of the glottal pulse through a first-order high-pass filter, wherein the first-order high-pass filter is a first-order FIR filter with a coefficient of 0.9375. S303, MRSVD noise reduction processing: MRSVD constructs the signal into a second-order matrix, decomposing it into two components each time, and then performs inverse operation on the extracted signal to reconstruct it, thereby achieving noise reduction on the original signal; S304, filtering processing: using a Butterworth filter to filter out background noise below 20 Hz.

5. According to claim 1, a transformer fault voiceprint diagnosis method based on Transformer algorithm is characterized in that: In step S4, the voiceprint component change calculation formula is: Where D is the change in voiceprint components, f max is the upper limit frequency of the calculation, A1(n×10) is the n×10 Hz frequency component amplitude of the sound signal being diagnosed, and A2(n×10) is the n×10 Hz frequency component amplitude of the sound signal cached in the previous period.

6. According to claim 1, a transformer fault voiceprint diagnosis method based on Transformer algorithm is characterized in that: In step S4, the peak factor calculation formula is: Among them, X peak is the peak value of the signal, X RMS is the RMS value of the signal.

7. The transformer fault voiceprint diagnosis method based on Transformer algorithm according to claim 1 is characterized in that: In step S5, the specific process of matching signal features is as follows: S501, if the signal main frequency is not within the preset range, it is recorded as 0, otherwise it is recorded as 1; S502, if the change in the voiceprint component is greater than 10%, it is recorded as 0, otherwise it is recorded as 1; S503, if the change in the peak factor between the currently diagnosed sound signal and the sound signal cached in the previous period is greater than 10%, it is recorded as 0, otherwise it is recorded as 1; S504: If the total sound pressure level or the segmented sound pressure level exceeds the threshold, it is recorded as 1, otherwise it is recorded as 0; S505. Add the results of steps S501-S504 to obtain a value T.

8. According to the transformer fault voiceprint diagnosis method based on Transformer algorithm in claim 1, it is characterized in that: In step S6, the abnormal sound judgment is specifically as follows: if the value T calculated in step S5 is greater than 0, it is judged that there is an abnormal sound, otherwise it is output that there is no abnormal sound and the diagnosis of the next segment of speech is entered.

9. The transformer fault voiceprint diagnosis method based on Transformer algorithm according to claim 1 is characterized in that: In step S7, when the speech signal preprocessed in step S3 is converted into a Mel time-frequency spectrum, the FFT window length is selected to be greater than 8192, the data of the Mel filter group is selected to be greater than 128, and the size of the Mel spectrum is scaled to 224*244.

10. The transformer fault voiceprint diagnosis method based on Transformer algorithm according to claim 1 is characterized in that: In step S8, the Transformer model is trained as follows: S801, constructing a model transformer sound sample library, where the sample library samples include sounds collected from different transformers, including fault samples, environmental noise samples, and samples of transformer equipment to be detected; S802, data cleaning: filter out damaged samples and samples that are not transformer backgrounds, check the sample length of the sample set, and divide the samples into samples of fixed length; S803, data labeling: labeling whether the sample contains fault sound or environmental sound; S804, all labeled samples are combined into a standard sample library; S805, extracting the mel spectrum of the samples from the standard sample library organized in S804, and scaling it to a size of 224*224 to form a mel spectrum set; S806, model training: Build a Transformer-based swin-transformer model and use the Mel spectrogram sample set compiled in S804 for training.

Images