Power transformer fault identification method

Abstract

The invention discloses a power transformer fault identification method comprising the following steps: collecting voiceprint sample data of various faults of a power transformer, and expanding and superposing environmental noise; performing feature extraction on the data superposed with the environmental noise to obtain voiceprint signal features, and converting the voiceprint signal features into feature vectors; the CNN-LSTM network model is optimized by using an improved fishing optimization algorithm, the optimized model is trained by using a characteristic vector, and the trained model is used for power transformer fault diagnosis; according to the method, the learning rate and the batch size of the CNN-LSTM network model are optimized by adopting the improved fishing optimization algorithm, the problem that hyper-parameters are difficult to accurately select is solved, and the condition in the power transformer can be accurately output in real time by diagnosing the characteristic quantity.

Description

Technical Field

[0001] The invention relates to the technical field of power equipment detection, and in particular to a method for identifying power transformer faults. Background Art

[0002] As a high-value key equipment of the power grid, power transformers usually need to carry out necessary maintenance work during their entire life cycle to ensure that the equipment is in a healthy operating state. With the development and progress of the power industry, the condition maintenance method of power transformers has gradually shifted from periodic power outage detection to non-power outage detection.

[0003] During the operation of the power transformer, the core, winding and other equipment will vibrate. The generated mechanical waves are transmitted to the fluid medium and appear as sound waves. The sound waves are received by the collection sensor to obtain sound signals. The characteristic information contained in the sound signal is the acoustic fingerprint, referred to as voiceprint.

[0004] The soundprint signal generated by the power transformer contains a large amount of equipment status information, so the transformer DC bias, core loosening, winding loosening, winding deformation and other mechanical abnormalities can be detected through the soundprint. At the same time, the sensor for collecting sound signals has the characteristics of non-contact, which does not affect the normal operation of the power transformer.

[0005] At present, the commonly used non-stop detection methods such as oil chromatography and ultrasound are difficult to timely identify abnormal conditions such as DC bias, loose core, loose winding, and winding deformation of transformers. If the above abnormal conditions are not handled in time, they may gradually develop and cause serious power transformer defects. Summary of the invention

[0006] In view of the deficiencies in the prior art, the present invention provides a method for identifying faults in a power transformer, the purpose of which is to solve the problems in the background technology.

[0007] To achieve the above object, the present invention provides the following technical solution: a method for identifying a power transformer fault, comprising the following steps:

[0008] Step S1: Collect voiceprint sample data of various faults of power transformers, expand the collected voiceprint sample data using a sample augmentation method, and superimpose environmental noise on the expanded voiceprint sample data;

[0009] Step S2: extracting features from the voiceprint sample data after superimposing the environmental noise to obtain voiceprint signal features, and converting the voiceprint signal features into feature vectors;

[0010] Step S3: construct a CNN-LSTM network model, optimize the learning rate and batch size of the CNN-LSTM network model as the goal of improving the fishing optimization algorithm, retain the optimal parameters of the CNN-LSTM network model, and obtain the CFOA-CNN-LSTM network model;

[0011] Step S4: using the feature vector in step S2 to train the CFOA-CNN-LSTM network model to obtain a power transformer fault recognition model;

[0012] Step S5: inputting the real-time collected power transformer voiceprint signal data into the power transformer fault identification model to realize power transformer fault diagnosis;

[0013] In step S3, the specific process of optimizing the learning rate and batch size of the CNN-LSTM network model as the target of improving the fishing optimization algorithm is as follows:

[0014] Step S301: using the learning rate and batch size of the CNN-LSTM network model as the number of fishermen in the improved fishing optimization algorithm to initialize the population;

[0015] Step S302: simulating an exploration phase, which is divided into an independent search mode and a group fishing mode;

[0016] Step S3021: simulating an independent search mode, and introducing a Cauchy mutation operator into the independent search mode for improvement;

[0017] Step S3022: simulating a group fishing mode;

[0018] Step S303: simulation development stage;

[0019] Step S304: Determine whether the current number of evaluations exceeds the maximum number of evaluations. If so, determine the current global optimal position as the optimal learning rate and batch size for the CNN-LSTM network model. If not, return to step S302 to continue optimizing.

[0020] Furthermore, the exploration phase and development phase of the improved fishing optimization algorithm are composed of EFs indicates the current evaluation number, and MaxEFs indicates the maximum evaluation number; when When is the exploration stage, the improved fishing optimization algorithm performs global search; when When entering the development stage, the fishing optimization algorithm is improved for local search.

[0021] Furthermore, the learning rate and batch size of the CNN-LSTM network model are used as the number of fishermen in the improved fishing optimization algorithm for population initialization, which is expressed as:

[0022] Fisher i，j =(ub j -lb j )*r+lb j ;

[0023] In the formula, Fisher i，j represents the position of the i-th fisherman in the j-th dimension; ub j and lb j They represent the upper and lower bounds of the j dimension respectively; r represents a random number between (0,1).

[0024] Furthermore, the transition between the independent search mode and the group fishing mode is determined by the capture rate α; the capture rate α is expressed as:

[0025]

[0026] A random number p∈(0, 1) is used for simulation. When p<α, an independent search mode is adopted. When p≥α, a group fishing mode is adopted.

[0027] Further, the independent search mode is expressed as:

[0028]

[0029] Where Exp is the empirical analysis value obtained by fishermen using fishermen in other locations other than themselves as reference objects, and its value range is (-1, 1); fit max and fit min are the worst and best fitness values ​​after the t-th position update, respectively; T represents the number of iterations of the fisherman’s position; represents the position of the i-th fisherman after T iterations in the j-th dimension; represents the position of the i-th fisherman after T+1 iterations in the j-th dimension; R s is a random number between (0,1); Dis is the Euclidean distance between the fisherman and the reference object; s is a d-dimensional random unit vector; R is the exploration range, and R≤Dis; fit i represents the fitness value of the i-th fisherman after the T-th iteration; fit pos represents the fitness value of the fisherman at any position pos after the Tth iteration; Represents the position of the fisherman at any position pos after T iterations in the j dimension.

[0030] Furthermore, the Cauchy mutation operator is introduced into the independent search mode for improvement, and the Cauchy probability density function f(x; x0, γ) is expressed as:

[0031]

[0032] Where x is a random variable; x0 is the location parameter; γ is the scale parameter, γ>0;

[0033] Substituting the Cauchy probability density function into the independent search mode, it is expressed as:

[0034]

[0035] In the formula, To increase the position of the i-th fisherman in the j-dimension after the T+1-th iteration after the Cauchy mutation perturbation; is the position of the i-th fisherman in the j-dimension after the T+1-th iteration; cauchy(0,1) is the random phasor that satisfies the Cauchy distribution cauchy in the interval (0,1).

[0036] Furthermore, the group fishing mode is simulated, which is expressed as:

[0037]

[0038] In the formula, c is a group of 3-4 fishermen; Centre c The target point surrounded by group c; represents the position of the i-th fisherman in group c in the j dimension after T+1 updates; represents the position of the i-th fisherman in group c in the j dimension after T updates; r2 is the speed at which the fisherman approaches the center, with a value range of (0,1); r3 is the offset of the movement, with a value range of (-1,1), which gradually decreases with the increase of EFs; represents the position of each fisherman in group c after T iterations; mean represents the mean of the parameter, that is, the center position of group c.

[0039] Furthermore, the simulation development phase is expressed as:

[0040]

[0041] Where, σ is the variance; is the position of the i-th fisherman after T+1 iterations; Gbest is the global optimal position; GD is the Gaussian distribution function; r4 is a random number in the range of {1,2,3}; Fisher represents the average value matrix of each dimension of the fisherman center.

[0042] Furthermore, the specific process of step S4 is: dividing the feature vector in step S2 into a training set, a validation set and a test set in proportion, using the training set to train the CFOA-CNN-LSTM network model, using the test set to test the CFOA-CNN-LSTM network model, and using the validation set to verify the CFOA-CNN-LSTM network model to obtain a power transformer fault identification model.

[0043] Further, the specific process of step S1 is as follows: step S101: by deploying directional sound collection sensors around the power transformer, the voiceprint sample data of the power transformer in the 220kV substation when it is normal and faulty, that is, the voiceprint sample data of various faults of the power transformer are obtained; wherein, the deployment mode of the sound collection sensor is pole-type, which is arranged on four sides of the power transformer respectively; the sampling frequency of the sound collection sensor is 48kHz, the frequency response is 20Hz~20kHz, and the audio dynamic range is 30dbA~120dbA;

[0044] Step S102: using a sample augmentation method to achieve data expansion by adjusting parameters of the voiceprint sample data;

[0045] Step S103: superimposing environmental noise on the expanded voiceprint sample data.

[0046] Compared with the existing technology, the present invention has the following beneficial effects:

[0047] (1) The present invention solves the problem of the difficulty in accurately selecting hyperparameters by using an improved fishing optimization algorithm to optimize the learning rate and batch size of the CNN-LSTM network model. By diagnosing the characteristic quantity, the internal situation of the power transformer can be output in real time and accurately.

[0048] (2) The present invention increases the diversity of the fisher population and improves the algorithm's optimization capability by introducing the Cauchy mutation operator into the independent search mode of the fishing optimization algorithm.

[0049] (3) The present invention effectively expands the sample size and improves the generalization ability of the model by augmenting the voiceprint sample data and superimposing environmental noise, so that the recognition accuracy of the model is higher under noise interference.

[0050] (4) By combining the CNN neural network with the LSTM neural network, the present invention can effectively process the spatiotemporal dependencies in the voiceprint signal and significantly improve the effect of transformer fault signal feature extraction. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a flow chart of the steps of the present invention.

[0052] Figure 2 This is a comparison chart of diagnostic accuracy before and after optimization of the CNN-LSTM network model. DETAILED DESCRIPTION

[0053] like Figure 1 As shown, the present invention provides a technical solution: a method for identifying a power transformer fault, comprising the following steps:

[0054] Step S1: Collect voiceprint sample data of various faults of power transformers, expand the collected voiceprint sample data using a sample augmentation method, and superimpose environmental noise on the expanded voiceprint sample data.

[0055] Step S2: extracting features from the voiceprint sample data after superimposing the environmental noise to obtain voiceprint signal features, and converting the voiceprint signal features into feature vectors.

[0056] Step S3: construct a CNN-LSTM network model, and optimize the learning rate (LearningRate) and batch size (batch size) of the CNN-LSTM network model as the goal of improving the fishing optimization algorithm, retain the optimal parameters of the CNN-LSTM network model, and obtain the CFOA-CNN-LSTM network model.

[0057] Step S4: Use the feature vector in step S2 to train the CFOA-CNN-LSTM network model to obtain a power transformer fault identification model.

[0058] Step S5: inputting the real-time collected power transformer voiceprint signal data into the power transformer fault identification model to realize power transformer fault diagnosis.

[0059] The specific process of step S1 is as follows:

[0060] Step S101: Deploy directional sound collection sensors around the power transformer to obtain voiceprint sample data of the power transformer in the 220kV substation when it is normal and faulty, that is, voiceprint sample data of various faults of the power transformer; wherein the deployment method of the sound collection sensor is pole-type, which is arranged on the four sides of the power transformer; the sampling frequency of the sound collection sensor is 48kHz, the frequency response is 20Hz~20kHz, and the audio dynamic range is 30dbA~120dbA.

[0061] Step S102: adopting the sample augmentation method to achieve data expansion by adjusting the sampling rate, pitch, amplitude and other parameters of the voiceprint sample data.

[0062] Step S103: superimpose environmental noise on the expanded voiceprint sample data; when actually collecting voiceprint sample data, accidental noises such as bird calls and whistles are usually not encountered, but considering the generalization ability of the training model, it is necessary to incorporate these transient noise factors into the training samples; since environmental noise is additive noise, audio fusion can be achieved by amplitude superposition.

[0063] Among them, feature extraction is performed on the voiceprint sample data after superimposing the environmental noise to obtain the voiceprint signal feature, and the voiceprint signal feature is converted into a feature vector. The Mel cepstral coefficient is used as the voiceprint signal feature, and the feature dimension is set to 64.

[0064] Among them, in step S3, the CNN-LSTM network model is built based on the Keras open source framework and is divided into two parts. The first part uses the strong feature extraction capability of the convolutional neural network to design a two-layer one-dimensional convolution structure, with the convolution kernel (filters) set to 8 and 10, the convolution size (kernel_size) set to 25, the padding (padding) set to same, the input size (input_shape) set to 64×1, and the activation function (activation) selects the relu function; in the second part, in order to improve the convergence speed, a Dropout layer is added to sparse the network parameters, and the sparse coefficient is set to 0.25, and then a long short-term memory network (LSTM) is added to enhance the memory ability of the voiceprint signal features, and finally a Dense layer is added to compress the number of network nodes to 2.

[0065] Among them, in step S3, the specific process of optimizing the learning rate (Learning Rate) and batch size (batchsize) of the CNN-LSTM network model as the target of improving the fishing optimization algorithm is as follows:

[0066] Step S301: The learning rate and batch size of the CNN-LSTM network model are used as the number of fishermen in the improved fishing optimization algorithm to initialize the population, which is expressed as:

[0067] Fisher i，j =(ub j -lb j )*r+lb j ;

[0068] In the formula, Fisher i，j represents the position of the i-th fisherman in the j-th dimension; ub j and lb j They represent the upper and lower bounds of the j dimension respectively; r represents a random number between (0,1).

[0069] The improved fishing optimization algorithm is divided into exploration stage and development stage. When is the exploration stage, the improved fishing optimization algorithm performs global search, EFs represents the current evaluation number, MaxEFs represents the maximum evaluation number; when When entering the development stage, the fishing optimization algorithm is improved for local search.

[0070] Step S302: Simulate the exploration phase: At the beginning of fishing, fish are most abundant and the catch rate is the highest, but as the fishermen's catch gradually decreases, the catch rate decreases; the exploration phase is divided into independent search mode and group fishing mode, and the conversion between independent search mode and group fishing mode is determined by the capture rate α; the capture rate α is expressed as:

[0071]

[0072] When the capture rate α is high, fishermen adopt an independent search mode; when the capture rate α decreases, fishermen adopt a group fishing mode; using a random number p∈(0, 1) for simulation, when p<α, an independent search mode is adopted, and when p≥α, a group fishing mode is adopted.

[0073] Step S3021: simulating an independent search mode;

[0074] Fishermen adjust their exploration direction and position according to the catch of other fishermen. When their own catch is good, they conduct reverse search, which is expressed as:

[0075]

[0076] Where Exp is the empirical analysis value obtained by fishermen using fishermen in other locations other than themselves as reference objects, and its value range is (-1, 1); fit max and fit min are the worst and best fitness values ​​after the t-th position update, respectively; T represents the number of iterations of the fisherman’s position; represents the position of the i-th fisherman after T iterations in the j-th dimension; represents the position of the i-th fisherman after T+1 iterations in the j-th dimension; R s is a random number between (0,1); Dis is the Euclidean distance between the fisherman and the reference object; s is a d-dimensional random unit vector used to determine the main travel direction and distance; R is the exploration range, which varies with the absolute value of Exp and the number of current evaluations EFs, and R≤Dis; fit i represents the fitness value of the i-th fisherman after the T-th iteration; fit pos represents the fitness value of the fisherman at any position pos after the Tth iteration; Represents the position of the fisherman at any position pos after T iterations in the j dimension.

[0077] Since the independent search mode is realized by random numbers, it is easy to cause the improved fishing optimization algorithm to miss some better solutions and reduce the convergence accuracy. In order to enhance the thoroughness of the search and eliminate the negative impact of the local optimal solution, the Cauchy mutation operator is introduced into the independent search mode to increase the diversity of the fisher population and improve the algorithm's optimization ability.

[0078] The Cauchy probability density function f(x; x0, γ) is expressed as:

[0079]

[0080] Where x is a random variable; x0 is the location parameter; γ is the scale parameter, γ>0.

[0081] Substituting the Cauchy probability density function into the independent search mode, it is expressed as:

[0082]

[0083] In the formula, To increase the position of the i-th fisherman in the j-dimension after the T+1-th iteration after the Cauchy mutation perturbation; is the position of the i-th fisherman in the j-dimension after the T+1-th iteration; cauchy(0,1) is a random phasor that satisfies the Cauchy distribution in the interval (0,1).

[0084] Step S3022: simulating a group fishing mode;

[0085] As fish are continuously caught and fishermen gradually gather, the movement and deviation of fishermen decrease, and fishermen begin to form groups of 3-4 people at random and cooperate with each other to fish, which can be expressed as:

[0086]

[0087] In the formula, c is a group of 3-4 fishermen; Centre c The target point surrounded by group c; represents the position of the i-th fisherman in group c in the j dimension after T+1 updates; represents the position of the i-th fisherman in group c in the j dimension after T updates; r2 is the speed at which the fisherman approaches the center, with a value range of (0,1); r3 is the offset of the movement, with a value range of (-1,1), which gradually decreases with the increase of EFs; represents the position of each fisherman in group c after T iterations; mean represents the mean of the parameter, that is, the center position of group c.

[0088] Step S303: simulation development stage;

[0089] After the above stages, some fish escaped and were not caught by fishermen. Therefore, in the development stage, the distribution of fishermen is as follows: with the fish school as the center, the concentration gradually becomes sparse from the middle to the periphery, and the distribution range gradually decreases outward; fishermen in the center catch fish schools, while fishermen in the periphery catch escaped fish; fishermen conduct strategic fishing and work together to improve the fishing rate; Gaussian distribution is used to simulate the distribution of fishermen, which is expressed as:

[0090]

[0091] Where, σ is the variance; is the position of the i-th fisherman after T+1 iterations; Gbest is the global optimal position; GD is the Gaussian distribution function; r4 is a random number in the range of {1,2,3}, which is used to distribute the fishermen in three ranges; Fisher represents the mean value matrix of each dimension of the fisherman center.

[0092] Step S304: Determine whether the current number of evaluations exceeds the maximum number of evaluations. If so, determine the current global optimal position as the optimal learning rate and batch size for the CNN-LSTM network model. If not, return to step S302 to continue optimizing.

[0093] The specific process of step S4 is: divide the feature vector in step S2 into a training set, a validation set and a test set in a ratio of 70:15:15, use the training set to train the CFOA-CNN-LSTM network model, use the test set to test the CFOA-CNN-LSTM network model, and use the validation set to verify the CFOA-CNN-LSTM network model to obtain a power transformer fault identification model.

[0094] like Figure 2 As shown in the figure, the diagnostic accuracy of the CNN-LSTM network model optimized by the improved fishing optimization algorithm is significantly better than that of the original CNN-LSTM network model.

[0095] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying a power transformer fault, characterized in that: The steps include: Step S1: Collect voiceprint sample data of various faults of power transformers, expand the collected voiceprint sample data using a sample augmentation method, and superimpose environmental noise on the expanded voiceprint sample data; Step S2: extracting features from the voiceprint sample data after superimposing the environmental noise to obtain voiceprint signal features, and converting the voiceprint signal features into feature vectors; Step S3: construct a CNN-LSTM network model, optimize the learning rate and batch size of the CNN-LSTM network model as the goal of improving the fishing optimization algorithm, retain the optimal parameters of the CNN-LSTM network model, and obtain the CFOA-CNN-LSTM network model; Step S4: using the feature vector in step S2 to train the CFOA-CNN-LSTM network model to obtain a power transformer fault recognition model; Step S5: inputting the real-time collected power transformer voiceprint signal data into the power transformer fault identification model to realize power transformer fault diagnosis; In step S3, the specific process of optimizing the learning rate and batch size of the CNN-LSTM network model as the target of improving the fishing optimization algorithm is as follows: Step S301: using the learning rate and batch size of the CNN-LSTM network model as the number of fishermen in the improved fishing optimization algorithm to initialize the population; Step S302: simulating an exploration phase, which is divided into an independent search mode and a group fishing mode; Step S3021: simulating an independent search mode, and introducing a Cauchy mutation operator into the independent search mode for improvement; Step S3022: simulating a group fishing mode; Step S303: simulation development stage; Step S304: Determine whether the current number of evaluations exceeds the maximum number of evaluations. If so, determine the current global optimal position as the optimal learning rate and batch size for the CNN-LSTM network model. If not, return to step S302 to continue optimizing.

2. A method for identifying a power transformer fault according to claim 1, characterized in that: The exploration phase and development phase of the improved fishing optimization algorithm are composed of EFs indicates the current evaluation number, and MaxEFs indicates the maximum evaluation number; when When is the exploration stage, the improved fishing optimization algorithm performs global search; when When entering the development stage, the fishing optimization algorithm is improved for local search.

3. A method for identifying a power transformer fault according to claim 2, characterized in that: The learning rate and batch size of the CNN-LSTM network model are used as the number of fishermen in the improved fishing optimization algorithm to initialize the population, expressed as: Fisher i，j =(ub j -lb j )*r+lb j ; In the formula, Fisher i，j represents the position of the i-th fisherman in the j-th dimension; ub j and lb j They represent the upper and lower bounds of the j dimension respectively; r represents a random number between (0,1).

4. A method for identifying a power transformer fault according to claim 3, characterized in that: The transition between the independent search mode and the group fishing mode is determined by the capture rate α; the capture rate α is expressed as: A random number p∈(0, 1) is used for simulation. When p<α, an independent search mode is adopted. When p≥α, a group fishing mode is adopted.

5. A method for identifying a power transformer fault according to claim 4, characterized in that: The independent search mode is expressed as: Where Exp is the empirical analysis value obtained by fishermen using fishermen in other locations other than themselves as reference objects, and its value range is (-1, 1); fit max and fit min are the worst and best fitness values ​​after the t-th position update, respectively; T represents the number of iterations of the fisherman’s position; represents the position of the i-th fisherman after T iterations in the j-th dimension; represents the position of the i-th fisherman after T+1 iterations in the j-th dimension; R s is a random number between (0,1); Dis is the Euclidean distance between the fisherman and the reference object; s is a d-dimensional random unit vector; R is the exploration range, and R≤Dis; fit i represents the fitness value of the i-th fisherman after the T-th iteration; fit pos represents the fitness value of the fisherman at any position pos after the Tth iteration; Represents the position of the fisherman at any position pos after T iterations in the j dimension.

6. A method for identifying a power transformer fault according to claim 5, characterized in that: The Cauchy mutation operator is introduced into the independent search mode for improvement. The Cauchy probability density function f(x; x0, γ) is expressed as: Where x is a random variable; x0 is the location parameter; γ is the scale parameter, γ>0; Substituting the Cauchy probability density function into the independent search mode, it is expressed as: In the formula, To increase the position of the i-th fisherman in the j-dimension after the T+1-th iteration after the Cauchy mutation perturbation; is the position of the i-th fisherman in the j-dimension after the T+1-th iteration; cauchy(0,1) is the random phasor that satisfies the Cauchy distribution cauchy in the interval (0,1).

7. A method for identifying a power transformer fault according to claim 6, characterized in that: Simulate the group fishing mode, expressed as: In the formula, c is a group of 3-4 fishermen; Centre c The target point surrounded by group c; represents the position of the i-th fisherman in group c in the j dimension after T+1 updates; represents the position of the i-th fisherman in group c in the j dimension after T updates; r2 is the speed at which the fisherman approaches the center, with a value range of (0,1); r3 is the offset of the movement, with a value range of (-1,1), which gradually decreases with the increase of EFs; represents the position of each fisherman in group c after T iterations; mean represents the mean of the parameter, that is, the center position of group c.

8. A method for identifying a power transformer fault according to claim 7, characterized in that: Simulate the development phase, expressed as: Where, σ is the variance; is the position of the i-th fisherman after T+1 iterations; Gbest is the global optimal position; GD is the Gaussian distribution function; r4 is a random number in the range of {1,2,3}; Fisher represents the average value matrix of each dimension of the fisherman center.

9. A method for identifying a power transformer fault according to claim 8, characterized in that: The specific process of step S4 is: divide the feature vector in step S2 into a training set, a validation set and a test set in proportion, use the training set to train the CFOA-CNN-LSTM network model, use the test set to test the CFOA-CNN-LSTM network model, and use the validation set to verify the CFOA-CNN-LSTM network model to obtain a power transformer fault identification model.

10. A method for identifying power transformer faults according to claim 9, characterized in that: The specific process of step S1 is as follows: Step S101: by deploying directional sound collection sensors around the power transformer, the soundprint sample data of the power transformer in the 220kV substation when it is normal and faulty, that is, the soundprint sample data of various faults of the power transformer are obtained; wherein, the deployment mode of the sound collection sensor is pole-type, which is arranged on the four sides of the power transformer respectively; the sampling frequency of the sound collection sensor is 48kHz, the frequency response is 20Hz~20kHz, and the audio dynamic range is 30dbA~120dbA; Step S102: using a sample augmentation method to achieve data expansion by adjusting parameters of the voiceprint sample data; Step S103: superimposing environmental noise on the expanded voiceprint sample data.

Images