A transformer voiceprint anomaly detection method based on multi-band self-supervision

Abstract

The present invention relates to a transformer voiceprint anomaly detection method based on multi-band self-supervision, comprising: collecting and preprocessing transformer voiceprint data; extracting features from the collected voiceprint data; designing a self-supervised multi-layer fully connected neural network model, and self-supervising learning of transformer working voiceprint data features, and judging the differences between voiceprint data features; in self-supervised learning, when continuous voiceprint data features are input, the multi-layer fully connected neural network model outputs continuous voiceprint data anomaly monitoring difference results; using a template matching algorithm to perform recognition and detection of abnormal frequency feature vectors, separate and register abnormal frequency features, and realize generalized recognition and detection. The present invention retains the key features in the transformer voiceprint while compressing the parameters of non-critical frequency bands, reducing the feature vector dimension, reducing the calculation burden, speeding up system analysis, greatly saving the workload of professionals, and reducing the accuracy deviation caused by manual intervention.

Description

Technical Field

[0001] The present invention relates to the technical field of electric power monitoring, and in particular to a transformer voiceprint anomaly detection method based on multi-band self-supervision. Background Art

[0002] The safety and stability of the power system is the basis of all activities in contemporary society. It plays a vital role in the normal operation of cities and the security of residents' lives. With the development of the country and the progress of society, the power load is further divided into industrial power, residential power, municipal power, etc. The dependence of various social activities on the power system is further strengthened, and the demand for power consumption is further increased, which further improves the requirements for power supply quality and safety and reliability. In order to meet this severe challenge, the construction of smart grids is imminent, which is not only conducive to the improvement of the level of intelligence and user interaction, but also can alleviate energy pressure and effectively optimize the allocation of resources. With the gradual expansion of the transaction scale of the power market, the further development of the spot market, transmission and distribution prices, and incremental distribution, the provision of corresponding supporting facilities and auxiliary services to the market can bring corresponding economic effects and bring new vitality and vitality to the power industry.

[0003] As one of the important tasks of the power system, transformer anomaly detection is often used in real-time grid planning, fault analysis, real-time dispatching and other tasks. With the further opening of the power market and the further maturity of market management, transformer anomaly detection will play a greater role. The feature extraction process in the existing voiceprint anomaly detection technology usually uses Mel-frequency cepstral coefficients (MFCC) to extract voiceprint features to simulate the auditory characteristics of the human ear. However, this method is difficult to accurately represent the key frequency characteristics in the transformer voiceprint detection task, resulting in insufficient algorithm accuracy. In addition, due to the influence of factors such as load and voltage, the working voiceprint of each transformer has its own characteristics. The existing technology requires separate parameter debugging for each device. The detection accuracy is highly dependent on manual parameter settings, which consumes a lot of manpower costs. It is difficult to make timely adjustments when the working state of the transformer changes, resulting in low detection technology accuracy, poor robustness, and weak generalization ability. Summary of the invention

[0004] The purpose of the present invention is to provide a transformer voiceprint anomaly detection method based on multi-band self-supervision, which can reduce the dimension of feature vectors, reduce the computational burden, speed up system analysis, greatly save the workload of professionals, and reduce the accuracy deviation caused by manual intervention.

[0005] To achieve the above object, the present invention adopts the following technical solution: a transformer voiceprint anomaly detection method based on multi-band self-supervision, the method comprising the following steps in order:

[0006] (1) Collect and preprocess the voiceprint data of the transformer;

[0007] (2) Extracting features from the collected voiceprint data to obtain compression frequency features for the transformer;

[0008] (3) Design a self-supervised multi-layer fully connected neural network model, and self-supervise the learning of the working voiceprint data features of the transformer to determine the differences between the voiceprint data features;

[0009] (4) In self-supervised learning, when continuous voiceprint data features are input, the multi-layer fully connected neural network model outputs continuous voiceprint data anomaly monitoring difference results to obtain an abnormal frequency feature vector set;

[0010] (5) Use the template matching algorithm to identify and detect abnormal frequency feature vectors, separate and register the abnormal frequency features, and achieve generalized identification and detection.

[0011] The step (1) specifically refers to: performing segmentation, framing and windowing processing on the collected voiceprint data of the transformer, and dividing each input voiceprint data into a series of time domain spectrum data frames, a total of Z frames, recorded as:

[0012] W=[w1,w2,...,w Z ]

[0013] Use discrete Fourier transform to convert the time domain spectrum data frame into the frequency domain spectrum data frame, which is recorded as:

[0014] F=[f1,f2,...,f Z ].

[0015] The step (2) specifically comprises the following steps:

[0016] (2a) For the transformer soundprint frequency range [f low ,f high ], with 1000Hz as an interval, the frequency is divided into a set of frequency intervals {F1, F2, ..., F N};

[0017] (2b) For the F1 interval, the compression frequency c i With the original frequency f i The corresponding relationship is as follows:

[0018]

[0019] Among them, A1 is the frequency compensation coefficient, k1 is the compression ratio of the interval, and k1 can be 5, 10, and 20;

[0020] (2c) For intervals F except F1 n≠1, compression frequency c i With the original frequency f i The corresponding relationship is as follows:

[0021]

[0022] Among them, A n , A m is the compensation coefficient between the nth and mth intervals, k n , k m is the compression ratio between the nth and mth intervals, k m≠1 satisfy:

[0023]

[0024] Wherein, B is the exponential compensation coefficient;

[0025] (2d) The frequency multiple of 50 Hz in the F1 interval is denoted as F 50Hz ={50,100,...,950,1000}, the corresponding compressed frequency set is recorded as:

[0026] C 50Hz ={c 50Hz ,...,c 1000Hz}

[0027] The key frequency set of the compressed interval of the non-F1 interval is recorded as:

[0028] C m≠1 =∪ m≠1 C m

[0029] Among them, C m represents the compressed frequency set of the mth interval;

[0030] According to the calculation accuracy requirements, select according to frequency band I:

[0031]

[0032] Among them, D m represents the density compensation coefficient of the mth interval, and floor(·) represents the rounding down operation, thereby obtaining the corresponding compressed frequency set of all key frequencies:

[0033] C key =C 50Hz ∪C m≠1 ={c1,c2,...,c K}

[0034] Among them, K is the total number of key frequencies, and the corresponding original key frequencies are recorded as:

[0035] F key={f1 key ,f2 key ,...,f K key};

[0036] (2e) Taking each key frame as the interval center, the middle distance between the left and right adjacent key frames is selected as the interval boundary, and the maximum spectrum amplitude in the interval is used as the representative value of the key frame, which is expressed as:

[0037]

[0038] in, is the voiceprint frequency spectrum, from which the key frequency vector is obtained as the compression frequency feature for the transformer, and P is the number of feature vector dimensions, expressed as:

[0039] v key =[v1,...,v P ].

[0040] The step (3) specifically comprises the following steps:

[0041] (3a) Construct a multi-layer fully connected neural network model, where each layer is represented as: [E1, E2, E3, E4, E5, E6, E7, E8], where:

[0042] The input layer E1 contains P neurons, where P is the number of feature vector dimensions. During the training phase, 20% of the neurons are randomly discarded for training.

[0043] The hidden layer E2 contains 300 neurons, and 20% of the neurons are randomly discarded for training during the training phase;

[0044] The hidden layer E3 contains 64 neurons, and 10% of the neurons are randomly abandoned for training during the training phase;

[0045] The hidden layer E4 contains 32 neurons, and 10% of the neurons are randomly abandoned for training during the training phase;

[0046] The hidden layer E5 contains 32 neurons, and 10% of the neurons are randomly abandoned for training during the training phase;

[0047] The hidden layer E6 contains 64 neurons, and 10% of the neurons are randomly abandoned for training during the training phase;

[0048] The hidden layer E7 contains 300 neurons, and 20% of the neurons are randomly discarded for training during the training phase;

[0049] The output layer E8 contains P neurons, where P is the number of feature vector dimensions;

[0050] The output layer C is connected after the hidden layer E4, contains 32 neurons, and the output is 1-dimensional difference;

[0051] (3b) In the training phase, the input data is the input key frequency vector set V after frequency compression of the voiceprint data. input ;

[0052] The output data of the output layer E8 is the output key frequency vector set V output ;

[0053] During training, first set V input The loss function of the b-th batch of vectors during the training gradient backpropagation is expressed as follows:

[0054]

[0055]

[0056] Where N is the number of feature vectors in the b-th batch vector set, are the nth input and output feature vectors in the bth batch, respectively. Respectively The representative value of the pth key frame, P is the number of feature vector dimensions, α k is the dimension loss weight, β is the loss sensitivity coefficient, where α k For a finite nonlinear equalization function, use a logarithm-based function:

[0057] α k =αlog(k)+β

[0058] Among them, α and β are compensation coefficients;

[0059] The output layer C outputs the predicted difference s. The expected difference is 0 under self-supervision, and the cross entropy loss function is used for training.

[0060] (3c) In the verification phase, the input data is the input key frequency vector set V after the voiceprint data frequency compression input , the output is the difference of the output layer C, which is used as the abnormality monitoring result of the voiceprint data.

[0061] The step (4) specifically comprises the following steps:

[0062] (4a) Set an alarm trigger difference threshold τ s , using the binary method, if the threshold is exceeded, an abnormal monitoring alarm is output; if the threshold is not exceeded, it is considered that no abnormality has occurred at present;

[0063] (4b) Set an alarm trigger frequency threshold τ f, when the alarm trigger frequency exceeds the threshold within a period of time T, the difference threshold τ is automatically increased s , so that the trigger frequency is negatively correlated with the trigger threshold, expressed as:

[0064]

[0065] (4c) Set the maximum value of the alarm trigger difference As the maximum value of triggering difference adjustment, if the difference threshold reaches the maximum value, adaptive adjustment will no longer be performed even if the trigger frequency exceeds the threshold;

[0066] (4d) After adaptive adjustment, the adaptive difference result S = [s1, s2, ..., s N ], that is, the difference result of abnormal monitoring of voiceprint data;

[0067] (4e) According to the adaptive difference result, the frame position corresponding to the abnormal alarm is obtained, and the key frequency vector set V is input. input , get the abnormal frequency feature vector set V abn , realizing the time location and feature acquisition of abnormal voiceprints.

[0068] The step (5) specifically comprises the following steps:

[0069] (5a) The abnormal frequency feature vector set V abn Divided into 50Hz frequency multiple vectors and other frequency vector sets, denoted as and

[0070] (5b) Annotate the anomalies located by the voiceprint data and organize them into an anomaly standard annotation file;

[0071] (5c) and the abnormal standard file to form the first abnormal voiceprint compression frequency feature standard data set. and the abnormal standard file are merged into a second abnormal voiceprint compression frequency feature standard data set;

[0072] (5d) In the abnormal voiceprint registration stage, the first abnormal voiceprint compressed frequency feature standard data set is registered as the abnormal voiceprint intrinsic frequency feature dictionary, and the second abnormal voiceprint compressed frequency feature standard data set is registered as the abnormal voiceprint non-intrinsic frequency feature dictionary;

[0073] (5e) In the verification stage, the input is the verification frequency feature vector, which is divided into 50 Hz frequency multiple feature vectors and other frequency feature vectors, and compared with the abnormal voiceprint intrinsic frequency feature dictionary and abnormal voiceprint non-intrinsic frequency feature dictionary by template matching algorithm, and finally the intrinsic anomaly and non-intrinsic anomaly categories most similar to the current verification feature vector are obtained.

[0074] The template matching algorithm in step (5e) adopts any one of the dynamic time warping method, the cosine similarity method, and the mixed Gaussian model method.

[0075] It can be seen from the above technical scheme that the beneficial effects of the present invention are: first, the present invention retains the key features in the transformer voiceprint while compressing the parameters of non-critical frequency bands, reducing the dimension of the feature vector, reducing the calculation burden and speeding up the system analysis speed; second, the present invention does not require any manual parameter setting, and can automatically learn the current transformer working characteristics. When the working condition of the transformer changes, only simple operations are required to complete the reconfiguration of the system parameters, which greatly saves the workload of professionals and reduces the accuracy deviation caused by manual intervention; third, the present invention can automatically adjust the warning trigger threshold according to the number of warning triggers within a period of time, so as to achieve a significant reduction in the number of triggers while retaining the most important abnormal warning, and reduce the amount of false warnings caused by external factors such as the environment; fourth, the present invention is based on the existing template matching algorithm, by separating the transformer intrinsic spectrum characteristics and external spectrum characteristics, to achieve the evaluation of the transformer intrinsic working condition and generalized category fault detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 is a flow chart of the method of the present invention;

[0077] Figure 2 Schematic diagram of the structure of a self-supervised multi-layer fully connected neural network model;

[0078] Figure 3 This is the flow chart of the template matching algorithm. DETAILED DESCRIPTION

[0079] like Figure 1 As shown, a transformer voiceprint anomaly detection method based on multi-band self-supervision includes the following steps in sequence:

[0080] The method comprises the following steps in order:

[0081] (1) Collect and preprocess the voiceprint data of the transformer;

[0082] (2) Extracting features from the collected voiceprint data to obtain compression frequency features for the transformer;

[0083] (3) Design a self-supervised multi-layer fully connected neural network model, and self-supervise the learning of the working voiceprint data features of the transformer to determine the differences between the voiceprint data features;

[0084] (4) In self-supervised learning, when continuous voiceprint data features are input, the multi-layer fully connected neural network model outputs continuous voiceprint data anomaly monitoring difference results to obtain an abnormal frequency feature vector set;

[0085] (5) Use the template matching algorithm to identify and detect abnormal frequency feature vectors, separate and register the abnormal frequency features, and achieve generalized identification and detection.

[0086] The step (1) specifically refers to: performing segmentation, framing and windowing processing on the collected voiceprint data of the transformer, and dividing each input voiceprint data into a series of time domain spectrum data frames, a total of Z frames, recorded as:

[0087] W=[w1,w2,...,w Z ]

[0088] Use discrete Fourier transform to convert the time domain spectrum data frame into the frequency domain spectrum data frame, which is recorded as:

[0089] F=[f1,f2,...,f Z ].

[0090] The step (2) specifically comprises the following steps:

[0091] (2a) For the transformer soundprint frequency range [f low ,f high ], with 1000Hz as an interval, the frequency is divided into a set of frequency intervals {F1, F2, ..., F N};

[0092] (2b) For the F1 interval, the compression frequency c i With the original frequency f i The corresponding relationship is as follows:

[0093]

[0094] Among them, A1 is the frequency compensation coefficient, k1 is the compression ratio of the interval, and k1 can be 5, 10, and 20;

[0095] (2c) For intervals F except F1 n≠1 , compression frequency c i With the original frequency f i The corresponding relationship is as follows:

[0096]

[0097] Among them, A n , A m is the compensation coefficient between the nth and mth intervals, k n , k mis the compression ratio between the nth and mth intervals, k m≠1 satisfy:

[0098]

[0099] Wherein, B is the exponential compensation coefficient;

[0100] (2d) The frequency multiple of 50 Hz in the F1 interval is denoted as F 50Hz ={50,100,...,950,1000}, the corresponding compressed frequency set is recorded as:

[0101] C 50Hz ={c 50Hz ,...,c 1000Hz}

[0102] The key frequency set of the compressed interval of the non-F1 interval is recorded as:

[0103] C m≠1 =∪ m≠1 C m

[0104] Among them, C m represents the compressed frequency set of the mth interval;

[0105] According to the calculation accuracy requirements, select according to frequency band I:

[0106]

[0107] Among them, D m represents the density compensation coefficient of the mth interval, and floor(·) represents the rounding down operation, thereby obtaining the corresponding compressed frequency set of all key frequencies:

[0108] C key =C 50Hz ∪C m≠1 ={c1,c2,...,c K}

[0109] Among them, K is the total number of key frequencies, and the corresponding original key frequencies are recorded as:

[0110]

[0111] (2e) Taking each key frame as the interval center, the middle distance between the left and right adjacent key frames is selected as the interval boundary, and the maximum spectrum amplitude in the interval is used as the representative value of the key frame, which is expressed as:

[0112]

[0113] in, is the voiceprint frequency spectrum, from which the key frequency vector is obtained as the compression frequency feature for the transformer, and P is the number of feature vector dimensions, expressed as:

[0114] v key =[v1,...,v P ].

[0115] The step (3) specifically comprises the following steps:

[0116] (3a) Construct a multi-layer fully connected neural network model, such as Figure 2 As shown, each layer is represented as: [E1, E2, E3, E4, E5, E6, E7, E8], where:

[0117] The input layer E1 contains P neurons, where P is the number of feature vector dimensions. During the training phase, 20% of the neurons are randomly discarded for training.

[0118] The hidden layer E2 contains 300 neurons, and 20% of the neurons are randomly discarded for training during the training phase;

[0119] The hidden layer E3 contains 64 neurons, and 10% of the neurons are randomly abandoned for training during the training phase;

[0120] The hidden layer E4 contains 32 neurons, and 10% of the neurons are randomly abandoned for training during the training phase;

[0121] The hidden layer E5 contains 32 neurons, and 10% of the neurons are randomly abandoned for training during the training phase;

[0122] The hidden layer E6 contains 64 neurons, and 10% of the neurons are randomly discarded for training during the training phase;

[0123] The hidden layer E7 contains 300 neurons, and 20% of the neurons are randomly discarded for training during the training phase;

[0124] The output layer E8 contains P neurons, where P is the number of feature vector dimensions;

[0125] The output layer C is connected after the hidden layer E4, contains 32 neurons, and the output is 1-dimensional difference;

[0126] (3b) In the training phase, the input data is the input key frequency vector set V after frequency compression of the voiceprint data. input ;

[0127] The output data of the output layer E8 is the output key frequency vector set V output ;

[0128] During training, first set V inputThe loss function of the b-th batch of vectors during the training gradient backpropagation is expressed as follows:

[0129]

[0130]

[0131] Where N is the number of feature vectors in the b-th batch vector set, are the nth input and output feature vectors in the bth batch, respectively. Respectively The representative value of the pth key frame, P is the number of feature vector dimensions, α k is the dimension loss weight, β is the loss sensitivity coefficient, where α k For a finite nonlinear equalization function, use a logarithm-based function:

[0132] α k =αlog(k)+β

[0133] Among them, α and β are compensation coefficients;

[0134] The output layer C outputs the predicted difference s. The expected difference is 0 under self-supervision, and the cross entropy loss function is used for training.

[0135] (3c) In the verification phase, the input data is the input key frequency vector set V after the voiceprint data frequency compression input , the output is the difference of the output layer C, which is used as the abnormality monitoring result of the voiceprint data.

[0136] The step (4) specifically comprises the following steps:

[0137] (4a) Set an alarm trigger difference threshold τ s , using the binary method, if the threshold is exceeded, an abnormal monitoring alarm is output; if the threshold is not exceeded, it is considered that no abnormality has occurred at present;

[0138] (4b) Set an alarm trigger frequency threshold τ f , when the alarm trigger frequency exceeds the threshold within a period of time T, the difference threshold τ is automatically increased s , so that the trigger frequency is negatively correlated with the trigger threshold, expressed as:

[0139]

[0140] (4c) Set the maximum value of the alarm trigger difference As the maximum value of triggering difference adjustment, if the difference threshold reaches the maximum value, adaptive adjustment will no longer be performed even if the trigger frequency exceeds the threshold;

[0141] (4d) After adaptive adjustment, the adaptive difference result S = [s1, s2, ..., s N ], that is, the difference result of abnormal monitoring of voiceprint data;

[0142] (4e) According to the adaptive difference result, the frame position corresponding to the abnormal alarm is obtained, and the key frequency vector set V is input. input , get the abnormal frequency feature vector set V abn , realizing the time location and feature acquisition of abnormal voiceprints.

[0143] The step (5) specifically comprises the following steps:

[0144] (5a) The abnormal frequency feature vector set V abn Divided into 50Hz frequency multiple vectors and other frequency vector sets, denoted as and

[0145] (5b) Annotate the anomalies located by the voiceprint data and organize them into an anomaly standard annotation file;

[0146] (5c) and the abnormal standard file to form the first abnormal voiceprint compression frequency feature standard data set. and the abnormal standard file are merged into a second abnormal voiceprint compression frequency feature standard data set;

[0147] (5d) In the abnormal voiceprint registration stage, the first abnormal voiceprint compressed frequency feature standard data set is registered as the abnormal voiceprint intrinsic frequency feature dictionary, and the second abnormal voiceprint compressed frequency feature standard data set is registered as the abnormal voiceprint non-intrinsic frequency feature dictionary;

[0148] (5e) In the verification stage, the input is the verification frequency feature vector, which is divided into 50 Hz frequency doubling feature vectors and other frequency feature vectors, and compared with the abnormal voiceprint intrinsic frequency feature dictionary and abnormal voiceprint non-intrinsic frequency feature dictionary by template matching algorithm, such as Figure 3 As shown, the intrinsic anomaly and non-intrinsic anomaly categories that are most similar to the current verification feature vector are finally obtained. The template matching algorithm in step (5e) adopts any one of the dynamic time warping method, the cosine similarity method, and the mixed Gaussian model method.

[0149] In summary, the present invention retains the key features in the transformer voiceprint while compressing the parameters of non-critical frequency bands, reducing the dimension of the feature vector, reducing the calculation burden and speeding up the system analysis speed; the present invention does not require any manual parameter setting, can automatically learn the current transformer working characteristics, and when the transformer working condition changes, only simple operations are required to complete the reconfiguration of the system parameters, which greatly saves the workload of professionals and reduces the accuracy deviation caused by manual intervention.

Claims

1. A transformer voiceprint anomaly detection method based on multi-band self-supervision, characterized by: The method comprises the following steps in order: (1) Collect and preprocess the voiceprint data of the transformer; (2) Extracting features from the collected voiceprint data to obtain compression frequency features for the transformer; (3) Design a self-supervised multi-layer fully connected neural network model, and self-supervise the learning of the working voiceprint data features of the transformer to determine the differences between the voiceprint data features; (4) In self-supervised learning, when continuous voiceprint data features are input, the multi-layer fully connected neural network model outputs continuous voiceprint data anomaly monitoring difference results to obtain an abnormal frequency feature vector set; (5) Use the template matching algorithm to identify and detect abnormal frequency feature vectors, separate and register the abnormal frequency features, and achieve generalized identification and detection.

2. The transformer voiceprint anomaly detection method based on multi-band self-supervision according to claim 1 is characterized in that: The step (1) specifically refers to: performing segmentation, framing and windowing processing on the collected voiceprint data of the transformer, and dividing each input voiceprint data into a series of time domain spectrum data frames, a total of Z frames, recorded as: In=[in1,in2,...,in Z ] Use discrete Fourier transform to convert the time domain spectrum data frame into the frequency domain spectrum data frame, which is recorded as: F=[f1,f2,...,f Z ]。 3. The transformer voiceprint anomaly detection method based on multi-band self-supervision according to claim 1 is characterized in that: The step (2) specifically comprises the following steps: (2a) For the transformer soundprint frequency range [f low ,f high ], with 1000Hz as an interval, the frequency is divided into a set of frequency intervals {F1, F2, ..., F N }; (2b) For the F1 interval, the compression frequency c i With the original frequency f i The corresponding relationship is as follows: Among them, A1 is the frequency compensation coefficient, k1 is the compression ratio of the interval, and k1 can be 5, 10, and 20; (2c) For intervals F except F1 n≠1 , compression frequency c i With the original frequency f i The corresponding relationship is as follows: Among them, A n , A m is the compensation coefficient between the nth and mth intervals, k n , k m is the compression ratio between the nth and mth intervals, k m≠1 satisfy: Wherein, B is the exponential compensation coefficient; (2d) The frequency multiple of 50 Hz in the F1 interval is denoted as F 50Hz ={50,100,...,950,1000}, the corresponding compressed frequency set is recorded as: C 50Hz ={c 50Hz ,...,c 1000Hz } The key frequency set of the compressed interval of the non-F1 interval is recorded as: C m≠1 =∪ m≠1 C m Among them, C m represents the compressed frequency set of the mth interval; According to the calculation accuracy requirements, select according to frequency band I: Among them, D m represents the density compensation coefficient of the mth interval, and floor(·) represents the rounding down operation, thereby obtaining the corresponding compressed frequency set of all key frequencies: C key =C 50Hz ∪C m≠1 ={c1,c2,...,c K } Among them, K is the total number of key frequencies, and the corresponding original key frequencies are recorded as: F key ={f1 key ,f2 key ,...,f K key }; (2e) Taking each key frame as the interval center, the middle distance between the left and right adjacent key frames is selected as the interval boundary, and the maximum spectrum amplitude in the interval is used as the representative value of the key frame, which is expressed as: in, is the voiceprint frequency spectrum, from which the key frequency vector is obtained as the compression frequency feature for the transformer, and P is the number of feature vector dimensions, expressed as: v key =[v1,...,v P ]。 4. The transformer voiceprint anomaly detection method based on multi-band self-supervision according to claim 1 is characterized in that: The step (3) specifically comprises the following steps: (3a) Construct a multi-layer fully connected neural network model, where each layer is represented as: [E1, E2, E3, E4, E5, E6, E7, E8], where: The input layer E1 contains P neurons, where P is the number of feature vector dimensions. During the training phase, 20% of the neurons are randomly discarded for training. The hidden layer E2 contains 300 neurons, and 20% of the neurons are randomly discarded for training during the training phase; The hidden layer E3 contains 64 neurons, and 10% of the neurons are randomly abandoned for training during the training phase; The hidden layer E4 contains 32 neurons, and 10% of the neurons are randomly abandoned for training during the training phase; The hidden layer E5 contains 32 neurons, and 10% of the neurons are randomly discarded for training during the training phase; The hidden layer E6 contains 64 neurons, and 10% of the neurons are randomly discarded for training during the training phase; The hidden layer E7 contains 300 neurons, and 20% of the neurons are randomly discarded for training during the training phase; The output layer E8 contains P neurons, where P is the number of feature vector dimensions; The output layer C is connected after the hidden layer E4, contains 32 neurons, and the output is 1-dimensional difference; (3b) In the training phase, the input data is the input key frequency vector set V after frequency compression of the voiceprint data. input ; The output data of the output layer E8 is the output key frequency vector set V output ; During training, first V input The loss function of the b-th batch of vectors during the training gradient backpropagation is expressed as follows: Where N is the number of feature vectors in the b-th batch vector set, are the nth input and output feature vectors in the bth batch, respectively. Respectively The representative value of the pth key frame, P is the number of feature vector dimensions, α k is the dimension loss weight, β is the loss sensitivity coefficient, where α k For a finite nonlinear equalization function, use a logarithm-based function: a k =αlog(k)+β Among them, α and β are compensation coefficients; The output layer C outputs the predicted difference s. The expected difference is 0 under self-supervision, and the cross entropy loss function is used for training. (3c) In the verification phase, the input data is the input key frequency vector set V after the voiceprint data frequency compression input , the output is the difference of the output layer C, which is used as the abnormality monitoring result of the voiceprint data.

5. The transformer voiceprint anomaly detection method based on multi-band self-supervision according to claim 1 is characterized in that: The step (4) specifically comprises the following steps: (4a) Set an alarm trigger difference threshold τ s , using the binary method, if the threshold is exceeded, an abnormal monitoring alarm is output; if the threshold is not exceeded, it is considered that no abnormality has occurred at present; (4b) Set an alarm trigger frequency threshold τ f , when the alarm trigger frequency exceeds the threshold within a period of time T, the difference threshold τ is automatically increased s , so that the trigger frequency is negatively correlated with the trigger threshold, expressed as: (4c) Set the maximum value of the alarm trigger difference As the maximum value of triggering difference adjustment, if the difference threshold reaches the maximum value, adaptive adjustment will no longer be performed even if the trigger frequency exceeds the threshold; (4d) After adaptive adjustment, the adaptive difference result S = [s1, s2, ..., s N ], that is, the difference result of abnormal monitoring of voiceprint data; (4e) According to the adaptive difference result, the frame position corresponding to the abnormal alarm is obtained, and according to the input key frequency vector set V input , get the abnormal frequency feature vector set V abn , realizing the time location and feature acquisition of abnormal voiceprints.

6. The transformer voiceprint anomaly detection method based on multi-band self-supervision according to claim 1 is characterized in that: The step (5) specifically comprises the following steps: (5a) The abnormal frequency feature vector set V abn Divided into 50Hz frequency multiple vectors and other frequency vector sets, recorded as and (5b) Annotate the anomalies located by the voiceprint data and organize them into an anomaly standard annotation file; (5c) and the abnormal standard file to form the first abnormal voiceprint compression frequency feature standard data set. and the abnormal standard file are merged into a second abnormal voiceprint compression frequency feature standard data set; (5d) In the abnormal voiceprint registration stage, the first abnormal voiceprint compressed frequency feature standard data set is registered as the abnormal voiceprint intrinsic frequency feature dictionary, and the second abnormal voiceprint compressed frequency feature standard data set is registered as the abnormal voiceprint non-intrinsic frequency feature dictionary; (5e) In the verification stage, the input is the verification frequency feature vector, which is divided into 50 Hz frequency multiple feature vectors and other frequency feature vectors, and compared with the abnormal voiceprint intrinsic frequency feature dictionary and abnormal voiceprint non-intrinsic frequency feature dictionary by template matching algorithm, and finally the intrinsic anomaly and non-intrinsic anomaly categories most similar to the current verification feature vector are obtained.

7. The transformer voiceprint anomaly detection method based on multi-band self-supervision according to claim 6 is characterized in that: The template matching algorithm in the step (5e) adopts any one of the dynamic time warping method, the cosine similarity method, and the mixed Gaussian model method.

Images