A method for diagnosing transformer core loose fault soundprint

Abstract

The invention discloses a transformer core loose fault soundprint diagnosis method, comprising the following steps: using wavelet transform to convert collected soundprint time series data into a soundprint feature spectrum; establishing a sensor dynamic response data fusion algorithm based on an entropy weight method, and fusing the soundprint feature spectrum in real time according to the relative importance of the soundprint time series data to obtain a wavelet dynamic fusion soundprint feature spectrum; using an improved crown porcupine optimization algorithm to optimize and train an improved convolutional neural network; importing the wavelet dynamic fusion soundprint feature spectrum into the trained improved convolutional neural network for feature extraction and recognition to obtain a final diagnosis result; the invention can enhance the thoroughness of algorithm search and eliminate the negative influence of local optimal solutions by introducing a random walk strategy and a Cauchy mutation operator into the first defense strategy and the second defense strategy of the crown porcupine algorithm respectively.

Description

Technical Field

[0001] The invention relates to the technical field of fault diagnosis of electric power equipment, and in particular to a transformer core loose fault soundprint diagnosis method. Background Art

[0002] Transformers are key equipment in power systems and play a key role in power systems. Loose core is one of the main causes of faults such as overheating and insulation damage. During the operation of the transformer, the continuous sound signal generated by the loose core contains rich information and reflects the status of the transformer to a large extent. At the same time, the sensor for collecting sound signals is non-contact, flexible to install, and does not affect the normal operation of the transformer.

[0003] The existing transformer core loosening fault diagnosis based on voiceprint has achieved certain results, but there are still problems such as difficulty in processing large-scale non-stationary signals and unclear characteristics of slight loose sound signals. In addition, the existing methods only process single sensor signals and cannot grasp the overall changes of the equipment. Summary of the invention

[0004] In view of the deficiencies in the prior art, the present invention provides a transformer core loose fault soundprint diagnosis method, which aims to solve the problems in the background technology.

[0005] To achieve the above object, the present invention provides the following technical solution: a transformer core loose fault soundprint diagnosis method, comprising the following steps:

[0006] Step S1: using multiple sensors to collect voiceprint time series data of the transformer, and using wavelet transform to convert the collected voiceprint time series data into a voiceprint feature spectrum;

[0007] Step S2: Establish a sensor dynamic response data fusion algorithm based on the entropy weight method, and fuse the voiceprint feature map in real time according to the relative importance of the voiceprint time series data to obtain a wavelet dynamic fusion voiceprint feature map; use the wavelet dynamic fusion voiceprint feature map to construct a sample set;

[0008] Step S3: Build an improved ShufflenetV2 convolutional neural network and determine the activation function and classifier;

[0009] Step S4: optimizing the improved ShufflenetV2 convolutional neural network using the improved crown porcupine optimization algorithm;

[0010] Step S5: using the sample set to train the optimized improved ShufflenetV2 convolutional neural network to obtain the optimal improved ShufflenetV2 convolutional neural network;

[0011] Step S6: import the wavelet dynamic fusion voiceprint feature map into the optimal improved ShufflenetV2 convolutional neural network for feature extraction and recognition to obtain the final diagnosis result;

[0012] The specific process of step S4 is:

[0013] Step S401: using the convolution kernel weights of the improved ShufflenetV2 convolutional neural network as the initial population position of the improved crested porcupine optimization algorithm to perform parameter optimization;

[0014] Step S402: Circular population reduction;

[0015] Select the best crested porcupine individuals from the current crested porcupine population and reintroduce them into the crested porcupine population;

[0016] Step S403: simulating the first defense strategy and introducing the Cauchy mutation operator for improvement;

[0017] Step S404: simulating the second defense strategy and introducing a random walk strategy for improvement;

[0018] Step S405: simulating the third defense strategy;

[0019] Step S406: simulating the fourth defense strategy;

[0020] Step S407: determine whether the current iteration number reaches the preset maximum iteration number, if not, continue the iteration, otherwise stop the iteration, and output the optimal crested porcupine individual position of the current iteration as the optimal solution for parameter optimization.

[0021] Furthermore, the initial population of the improved crown porcupine optimization algorithm is expressed as:

[0022] ;

[0023] In the formula, It indicates the number of crested porcupine individuals, i.e. the size of the crested porcupine population; represents the i-th candidate solution in the search space; and are the lower and upper limits of the search range respectively; A random number between 0 and 1.

[0024] Furthermore, the best crested porcupine individuals are selected from the current crested porcupine population and reintroduced into the crested porcupine population, which is expressed as:

[0025] ;

[0026] In the formula, Indicates the total number of iterations; Indicates the current iteration number; Indicates the maximum value of the total number of iterations; Represents the remainder or modulo operator; Represents the minimum number of crested porcupine individuals in a newly generated crested porcupine population.

[0027] Furthermore, the first defense strategy is simulated, which is expressed as:

[0028] ;

[0029] In the formula, Indicates After the iteration The location of each crested porcupine individual; Indicates After the iteration The location of each crested porcupine individual; Indicates the position of the current best crested porcupine individual; Represents the vector generated between the current best crested porcupine individual and the crested porcupine individual randomly selected from the current crested porcupine population; Represents random numbers based on normal distribution; Represents a random value in the interval [0,1];

[0030] It is expressed as:

[0031] ;

[0032] In the formula, represents the position of a crested porcupine individual randomly selected from the current crested porcupine population; is a random number between [1,N];

[0033] Introducing the Cauchy mutation operator into the first defense strategy; Cauchy probability density function It is expressed as:

[0034] ;

[0035] In the formula, is a random variable; is a positional parameter; is the scale parameter, ;

[0036] Will Bring it into the first defense strategy, expressed as:

[0037] ;

[0038] In the formula, To increase the Cauchy probability density function, After the iteration The location of each crested porcupine individual; for The interval satisfies the Cauchy distribution A random vector of .

[0039] Furthermore, the second defense strategy is simulated, which is expressed as:

[0040] ;

[0041] In the formula, is a random value generated between 0 and 1; represents a vector; Indicates the current position of the predator; and represents the positions of two crested porcupine individuals randomly selected from the current crested porcupine population, and are two random integers between [1,N];

[0042] The random walk strategy is introduced in the second defense strategy, which is expressed as:

[0043] ;

[0044] In the formula, is the set of steps of random walk; is the cumulative sum of walking steps; Indicates the number of walking steps; For the Number of walking steps; represents a random function, ; Represents a variable that contains a randomly generated number between 1 and 0.

[0045] Furthermore, the third defense strategy is simulated, which is expressed as:

[0046] ;

[0047] In the formula, is a parameter used to control the search direction; represents the defense factor; is the odor diffusion factor; represents the position of a crested porcupine individual randomly selected from the current crested porcupine population, is a random integer between [1,N];

[0048] It is expressed as:

[0049] ;

[0050] It is expressed as:

[0051] ;

[0052] It is expressed as:

[0053] ;

[0054] In the formula, Indicates The first iteration The objective function value of each crested porcupine individual; Represents a minimum value that avoids division by zero; represents a vector containing randomly generated values ​​between 0 and 1; represents a variable containing a randomly generated number between 1 and 0; Represents a vector, which is used to simulate the three situations in which the second defense strategy exists.

[0055] Furthermore, the fourth defense strategy is simulated, which is expressed as:

[0056] ;

[0057] In the formula, is the convergence speed factor; and All are random values ​​in the interval [0,1]; Is the impact The average force of crested porcupine individuals of the predator; represents a random vector;

[0058] It is expressed as:

[0059] ;

[0060] ;

[0061] ;

[0062] ;

[0063] In the formula, It is The first iteration The mass of the predator; It is Crested porcupine individuals are in The final speed during the iteration; It is Crested porcupine individuals The initial velocity at the iteration; is a randomly generated vector between 0 and 1; represents a natural constant; represents the objective function value of a crested porcupine individual randomly selected from the current crested porcupine population; represents the position of a crested porcupine individual randomly selected from the current crested porcupine population;

[0064] for ,delete The denominator of is expressed as:

[0065] .

[0066] Furthermore, the second defense strategy has three situations:

[0067] The first case: When = 0, the crested porcupine individual stops diffusing the scent, the predator stops moving, and the distance between the predator and the crested porcupine individual remains unchanged;

[0068] The second situation: When =1, the predator is nearby and the crested porcupine individual emits odor;

[0069] The third situation: When When it is a combination of 0 and 1, the predator keeps a safe distance from the crested porcupine individual, and the crested porcupine individual does not release the scent.

[0070] Furthermore, the specific process of step S2 is as follows:

[0071] Step S201: The corresponding sensor The dimensional voiceprint feature maps are spliced ​​to form a matrix :

[0072] ;

[0073] In the formula, Indicated by The sensor collects Dimensional voiceprint feature map;

[0074] Step S202: Matrix Standardize and get the standard matrix :

[0075] ;

[0076] In the formula, Expressing the The sensor receives the The result after standardization of the dimensional voiceprint feature map; Indicated by The sensor collects Dimensional voiceprint feature map, , ;

[0077] No. Information entropy of sensors It is expressed as:

[0078] ;

[0079] In the formula, ,and ; Represents the total dimension of the voiceprint feature map; express The weight of

[0080] according to Calculate the information entropy weight value corresponding to each sensor :

[0081] ;

[0082] In the formula, Indicates The information entropy of each sensor is ;

[0083] Step S203: Wavelet coefficients of the sensors The weight value corresponding to each sensor is calculated by the entropy weight method Fusion is performed to obtain the fused wavelet coefficient matrix , draw the fusion wavelet coefficient matrix to obtain the multi-sensor fusion voiceprint feature map MSF-CWT, that is, the wavelet dynamic fusion voiceprint feature map; among them, The wavelet coefficients of the sensors are based on the wavelet coefficient matrix get; A vector representing the scale factor; Represents the translation factor vector.

[0084] Furthermore, in step S3, the activation function of the improved ShufflenetV2 convolutional neural network adopts the MobieNetV3 model. The activation function is expressed as:

[0085] ;

[0086] In the formula, for Improvement of the function to limit its value to a maximum value; To improve the input value of ShufflenetV2 convolutional neural network;

[0087] The improved ShufflenetV2 convolutional neural network classifier is The layer is improved and expressed as:

[0088] ;

[0089] In the formula, for Dimensions of layers; for The elements, ; for The input of the layer.

[0090] Compared with the existing technology, the present invention has the following beneficial effects:

[0091] (1) The present invention optimizes the convolution kernel of the ShufflenetV2 convolutional neural network by using an improved crown porcupine algorithm, so that the neural network can more efficiently capture key information in the input data, thereby enhancing the feature extraction capability, recognition accuracy and efficiency of the neural network.

[0092] (2) The present invention can enhance the thoroughness of algorithm search and eliminate the negative influence of local optimal solutions by introducing random walk strategy and Cauchy mutation operator into the first defense strategy and the second defense strategy of the crested porcupine algorithm respectively. The random walk strategy can help the algorithm locate the optimal solution more accurately during the search process, thereby improving the convergence accuracy. The introduction of Cauchy mutation method can help the algorithm jump out of the local optimal solution during the search process, thereby enhancing the global search capability of the algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] Figure 1 It is a flow chart of the steps of the present invention.

[0094] Figure 2 This is a comparison chart of diagnostic accuracy before and after optimization of the ShufflenetV2 convolutional neural network. DETAILED DESCRIPTION

[0095] like Figure 1 As shown, the present invention provides a technical solution: a transformer core loose fault soundprint diagnosis method, comprising the following steps:

[0096] Step S1: Use multiple sensors to collect the voiceprint time series data of the transformer, and use wavelet transform (CWT) to convert the collected voiceprint time series data into a voiceprint feature map.

[0097] Step S2: Establish a sensor dynamic response data fusion algorithm based on the entropy weight method, and fuse the voiceprint feature map in real time according to the relative importance of the voiceprint time series data to obtain a wavelet dynamic fusion voiceprint feature map; use the wavelet dynamic fusion voiceprint feature map to construct a sample set.

[0098] Step S3: Build an improved ShufflenetV2 convolutional neural network and determine the activation function and classifier.

[0099] Step S4: Using the improved crown porcupine optimization algorithm to optimize the improved ShufflenetV2 convolutional neural network.

[0100] Step S5: Use the sample set to train the optimized improved ShufflenetV2 convolutional neural network to obtain the optimal improved ShufflenetV2 convolutional neural network.

[0101] Step S6: Import the wavelet dynamic fusion voiceprint feature map into the optimal improved ShufflenetV2 convolutional neural network for feature extraction and recognition to obtain the final diagnosis result.

[0102] The specific process of step S1 is as follows:

[0103] Step S1.1: Use multiple sensors to collect sound signals of the transformer to obtain transformer voiceprint time series data.

[0104] Step S1.2: After the transformer voiceprint time series data is framed and windowed, continuous wavelet transform is performed; the wavelet transform is expressed as:

[0105] (1);

[0106] In the formula, Represents input data (voiceprint time series data after framing and windowing); represents the mother wavelet; is the scale factor, which is used to determine the size of the time-frequency window; is the translation factor, which is used to determine the position of the time-frequency window; Represents the wavelet transform coefficient, which is used to reflect the local characteristics of the signal at different scales and positions; Indicates the time variable during the integration process The differential quantity used in the integration.

[0107] The specific process of step S1.2 is:

[0108] Determine the parameters of wavelet transform, including wavelet basis function, scale factor, wavelet center frequency, translation factor, window function parameters, signal sampling rate, and then substitute and the parameters of wavelet transform, we get The wavelet coefficient matrix :

[0109] (2);

[0110] In the formula, A vector representing the scale factor; represents the translation factor vector; is the length of the scale factor; is the length of the translation factor; Represents the wavelet coefficients at each scale.

[0111] The specific process of step S2 is as follows:

[0112] Step S201: The corresponding sensor The dimensional voiceprint feature maps are spliced ​​to form a matrix :

[0113] (3);

[0114] In the formula, Indicated by The sensor collects Dimensional voiceprint feature map.

[0115] Step S202: Matrix Standardize and get the standard matrix :

[0116] (4);

[0117] In the formula, Indicates The sensor receives the The result after standardization of the dimensional voiceprint feature map; Indicated by The sensor collects Dimensional voiceprint feature map, , .

[0118] No. Information entropy of sensors It is expressed as:

[0119] (5);

[0120] In the formula, ,and ; Represents the total dimension of the voiceprint feature map; express The weight occupied.

[0121] according to Calculate the information entropy weight value corresponding to each sensor :

[0122] (6);

[0123] In the formula, Indicates The information entropy of each sensor is .

[0124] Step S203: Wavelet coefficients of the sensor The weight value corresponding to each sensor is calculated by the entropy weight method Fusion is performed to obtain the fused wavelet coefficient matrix , draw the fusion wavelet coefficient matrix to obtain the multi-sensor fusion voiceprint feature map MSF-CWT, that is, the wavelet dynamic fusion voiceprint feature map; among them, The wavelet coefficients of the sensors are based on the wavelet coefficient matrix get.

[0125] Among them, in step S3, the activation function of the improved ShufflenetV2 convolutional neural network adopts the MobieNetV3 model The activation function solves the problem of neuron "necrosis" and is expressed as:

[0126] (7);

[0127] In the formula, for Improvement of the function to limit its value to a maximum value; To improve the input value of ShufflenetV2 convolutional neural network.

[0128] The improved ShufflenetV2 convolutional neural network classifier is The layer is improved and expressed as:

[0129] (8);

[0130] In the formula, for Dimensions of layers; for The elements, ; for The input of the layer; Represents a natural constant.

[0131] The specific process of step S4 is as follows:

[0132] Step S401: The convolution kernel weights of the improved ShufflenetV2 convolutional neural network are used as the initial population position of the improved crown porcupine optimization algorithm. The initial population is expressed as:

[0133] (9);

[0134] In the formula, represents the number of crested porcupine individuals (crested porcupine population size); represents the i-th candidate solution in the search space; and are the lower and upper limits of the search range respectively; A random number between 0 and 1.

[0135] Step S402: Circular population reduction;

[0136] In the iterative process of the improved crown porcupine optimization algorithm, the best crown porcupine individuals are selected from the current crown porcupine population and reintroduced into the crown porcupine population to accelerate the convergence speed, improve diversity, and avoid falling into the local minimum, which is expressed as:

[0137] (10);

[0138] In the formula, Indicates the total number of iterations; Indicates the current iteration number; Indicates the maximum value of the total number of iterations; Represents the remainder or modulo operator; Indicates the minimum number of crested porcupine individuals in the newly generated crested porcupine population. The crested porcupine population size cannot be less than .

[0139] Step S403: simulating the first defense strategy;

[0140] When the crested porcupine becomes aware of a predator, it raises and flaps its spines, at which point the predator moves toward or away from the crested porcupine, as indicated by:

[0141] (11);

[0142] In the formula, Indicates After the iteration The location of each crested porcupine individual; Indicates After the iteration The location of each crested porcupine individual; Indicates the position of the current best crested porcupine individual; Represents the vector generated between the current best crested porcupine individual and the crested porcupine individual randomly selected from the current crested porcupine population, used to represent the predator in the The position at the iteration; Represents random numbers based on normal distribution; Represents a random value in the interval [0,1].

[0143] It is expressed as:

[0144] (12);

[0145] In the formula, represents the position of a crested porcupine individual randomly selected from the current crested porcupine population; is a random number between [1,N].

[0146] When the crested porcupine optimization algorithm adopts the first defense strategy, it generates random numbers, which will cause it 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 first defense strategy to increase the diversity of the crested porcupine population, improve the global optimization ability of the algorithm, and enhance the robustness of the algorithm. The Cauchy probability density function It is expressed as:

[0147] (13);

[0148] In the formula, is a random variable; is the location parameter, indicating the peak of the distribution; is the scale parameter, controlling the width of the distribution, .

[0149] Will Bring it into the first defense strategy, expressed as:

[0150] (14);

[0151] In the formula, To increase the Cauchy probability density function, After the iteration The location of each crested porcupine individual; for The interval satisfies the Cauchy distribution A random vector of .

[0152] Step S404: simulating the second defense strategy;

[0153] The crested porcupine makes noise and threatens the predator. When the predator approaches the crested porcupine, the crested porcupine makes more noise, which is represented by:

[0154] (15);

[0155] In the formula, is a random value generated between 0 and 1; represents a vector; Indicates the current position of the predator; and represents the positions of two crested porcupine individuals randomly selected from the current crested porcupine population, and are two random integers between [1,N].

[0156] In order to further improve the adaptive optimization ability and global optimization ability of the crown porcupine optimization algorithm, the random walk strategy is introduced in the second defense strategy, which is expressed as:

[0157] (16);

[0158] In the formula, is the set of steps of random walk; is the cumulative sum of walking steps; Indicates the number of walking steps; For the Number of walking steps; represents a random function, ; Represents a variable that contains a randomly generated number between 1 and 0.

[0159] Since the feasible range of the random walk strategy has boundaries, it cannot be directly used to update the position of the crested porcupine individual. In order to ensure random walk within the feasible range, the position of the crested porcupine individual needs to be normalized, which can be expressed as:

[0160] (17);

[0161] In the formula, For the Crested porcupine The position in the iteration, and Respectively Minimum and maximum values ​​of the dimensional random walk variable; and Respectively The random walk variable is The minimum and maximum values ​​of the iterations.

[0162] Step S405: simulating the third defense strategy;

[0163] Crested porcupines secrete a foul odor that spreads around them, keeping predators away, as shown by:

[0164] (18);

[0165] In the formula, is a parameter used to control the search direction; represents the defense factor; is the odor diffusion factor; represents the position of a crested porcupine individual randomly selected from the current crested porcupine population, is a random integer between [1,N].

[0166] It is expressed as:

[0167] (19);

[0168] It is expressed as:

[0169] (20);

[0170] It is expressed as:

[0171] (twenty one);

[0172] In the formula, Indicates The first iteration The objective function value of each crested porcupine individual; Represents a minimum value that avoids division by zero; represents a vector containing randomly generated values ​​between 0 and 1; represents a variable containing a randomly generated number between 1 and 0; Represents a vector, which is used to simulate three possible situations of the second defense strategy:

[0173] The first case: When =0, the crested porcupine stops diffusing its scent, and the predator stops moving because of fear. At this time, the distance between the predator and the crested porcupine remains unchanged.

[0174] The second situation: When =1, the predator is nearby and the crested porcupine will emit a noticeable scent.

[0175] The third situation: When When it is a combination of 0 and 1, the predator keeps a safe distance from the crested porcupine and the crested porcupine does not release the scent.

[0176] Step S406: simulating the fourth defense strategy;

[0177] When the crested porcupine is within the predator's prey range and the predator attacks the crested porcupine, the crested porcupine counterattacks the predator. This process can be simulated by a one-dimensional inelastic collision, expressed as:

[0178] (twenty two);

[0179] In the formula, is the convergence speed factor; and All are random values ​​in the interval [0,1]; Is the impact The average force of crested porcupine individuals of the predator; represents a random vector.

[0180] It is expressed as:

[0181] (twenty three);

[0182] (twenty four);

[0183] (25);

[0184] (26);

[0185] In the formula, It is The iteration The mass of the predator; It is Crested porcupine individuals are The final speed at which to iterate, and select random solutions from it for allocation based on the current population; It is Crested porcupine individuals The initial velocity at the iteration; is a randomly generated vector between 0 and 1; represents the objective function value of a crested porcupine individual randomly selected from the current crested porcupine population; Represents the position of a crested porcupine individual randomly selected from the current crested porcupine population.

[0186] for , the average force of the crested porcupine individual is calculated based on the current iteration number as the denominator, and the current iteration number increases linearly during the optimization process; therefore, the influence of the average force of the crested porcupine individual is gradually minimized. In fact, the minimum value of this factor has an adverse effect on the performance of the crested porcupine optimization algorithm, because they may not be able to find better solutions in various areas around the current best solution; therefore, removing The denominator of is expressed as:

[0187] (27).

[0188] This approach will help create a wide range of values ​​within the search space, allowing for a more comprehensive examination of the area around the best solution to date.

[0189] Step S407: determine whether the current iteration number reaches the preset maximum iteration number, if not, continue the iteration, otherwise stop the iteration, and output the optimal crested porcupine individual position of the current iteration as the optimal solution for parameter optimization.

[0190] The specific process of step S5 is as follows:

[0191] The wavelet dynamic fusion voiceprint feature map in the sample set is input into the optimized improved ShufflenetV2 convolutional neural network, and the prediction result is output. The prediction result is compared with the actual result to calculate the error. When the error reaches the predetermined convergence adjustment, the model parameters of the optimized improved ShufflenetV2 convolutional neural network are saved to obtain the optimal improved ShufflenetV2 convolutional neural network.

[0192] like Figure 2 As shown, the original ShufflenetV2 convolutional neural network and the improved ShufflenetV2 convolutional neural network optimized by the improved crown porcupine optimization algorithm in the present invention are run in a computer programming language (python), and the looseness of the transformer core is identified, and the identification results are compared. Obviously, the diagnostic accuracy of the improved ShufflenetV2 convolutional neural network optimized by the improved crown porcupine optimization algorithm in the present invention is significantly better than that of the original ShufflenetV2 convolutional neural network.

[0193] 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 transformer core loose fault soundprint diagnosis method, characterized in that: The steps include: Step S1: using multiple sensors to collect voiceprint time series data of the transformer, and using wavelet transform to convert the collected voiceprint time series data into a voiceprint feature spectrum; Step S2: Establish a sensor dynamic response data fusion algorithm based on the entropy weight method, and fuse the voiceprint feature map in real time according to the relative importance of the voiceprint time series data to obtain a wavelet dynamic fusion voiceprint feature map; use the wavelet dynamic fusion voiceprint feature map to construct a sample set; Step S3: Build an improved ShufflenetV2 convolutional neural network and determine the activation function and classifier; Step S4: optimizing the improved ShufflenetV2 convolutional neural network using the improved crown porcupine optimization algorithm; Step S5: using the sample set to train the optimized improved ShufflenetV2 convolutional neural network to obtain the optimal improved ShufflenetV2 convolutional neural network; Step S6: import the wavelet dynamic fusion voiceprint feature map into the optimal improved ShufflenetV2 convolutional neural network for feature extraction and recognition to obtain the final diagnosis result; The specific process of step S4 is: Step S401: using the convolution kernel weights of the improved ShufflenetV2 convolutional neural network as the initial population position of the improved crested porcupine optimization algorithm to perform parameter optimization; Step S402: Circular population reduction; Select the best crested porcupine individuals from the current crested porcupine population and reintroduce them into the crested porcupine population; Step S403: simulating the first defense strategy and introducing the Cauchy mutation operator for improvement; Step S404: simulating the second defense strategy and introducing a random walk strategy for improvement; Step S405: simulating the third defense strategy; Step S406: simulating the fourth defense strategy; Step S407: determine whether the current iteration number reaches the preset maximum iteration number, if not, continue the iteration, otherwise stop the iteration, and output the optimal crested porcupine individual position of the current iteration as the optimal solution for parameter optimization.

2. The transformer core loosening fault soundprint diagnosis method according to claim 1 is characterized by: The initial population of the improved crown porcupine optimization algorithm is expressed as: ； In the formula, It indicates the number of crested porcupine individuals, i.e. the size of the crested porcupine population; represents the i-th candidate solution in the search space; and are the lower and upper limits of the search range respectively; A random number between 0 and 1.

3. The transformer core loosening fault soundprint diagnosis method according to claim 2 is characterized by: Select the best crested porcupine individuals from the current crested porcupine population and reintroduce them into the crested porcupine population, expressed as: ； In the formula, Indicates the total number of iterations; Indicates the current iteration number; Indicates the maximum value of the total number of iterations; Represents the remainder or modulo operator; Represents the minimum number of crested porcupine individuals in a newly generated crested porcupine population.

4. The transformer core loosening fault soundprint diagnosis method according to claim 3 is characterized by: Simulate the first defense strategy, expressed as: ； In the formula, Indicates After the iteration The location of each crested porcupine individual; Indicates After the iteration The location of each crested porcupine individual; Indicates the position of the current best crested porcupine individual; Represents the vector generated between the current best crested porcupine individual and the crested porcupine individual randomly selected from the current crested porcupine population; Represents random numbers based on normal distribution; Represents a random value in the interval [0,1]; It is expressed as: ； In the formula, represents the position of a crested porcupine individual randomly selected from the current crested porcupine population; is a random number between [1,N]; Introducing the Cauchy mutation operator into the first defense strategy; Cauchy probability density function It is expressed as: ； In the formula, is a random variable; is a positional parameter; is the scale parameter, ; Will Bring it into the first defense strategy, expressed as: ； In the formula, To increase the Cauchy probability density function, After the iteration The location of each crested porcupine individual; for The interval satisfies the Cauchy distribution A random vector of .

5. The transformer core loose fault soundprint diagnosis method according to claim 4 is characterized by: Simulate the second defense strategy, expressed as: ； In the formula, is a random value generated between 0 and 1; represents a vector; Indicates the current position of the predator; and represents the positions of two crested porcupine individuals randomly selected from the current crested porcupine population, and are two random integers between [1,N]; The random walk strategy is introduced in the second defense strategy, which is expressed as: ； In the formula, is the set of steps of random walk; is the cumulative sum of walking steps; Indicates the number of walking steps; For the Number of walking steps; represents a random function, ; Represents a variable that contains a randomly generated number between 1 and 0.

6. The transformer core loose fault soundprint diagnosis method according to claim 5 is characterized by: Simulate the third defense strategy, expressed as: ； In the formula, is a parameter used to control the search direction; represents the defense factor; is the odor diffusion factor; represents the position of a crested porcupine individual randomly selected from the current crested porcupine population, is a random integer between [1,N]; It is expressed as: ； It is expressed as: ； It is expressed as: ； In the formula, Indicates The iteration The objective function value of each crested porcupine individual; Represents a minimum value that avoids division by zero; represents a vector containing randomly generated values ​​between 0 and 1; represents a variable containing a randomly generated number between 1 and 0; Represents a vector, which is used to simulate the three situations in which the second defense strategy exists.

7. The transformer core loose fault soundprint diagnosis method according to claim 6 is characterized by: Simulate the fourth defense strategy, expressed as: ； In the formula, is the convergence speed factor; and All are random values ​​in the interval [0,1]; Is the impact The average force of crested porcupine individuals of the predator; represents a random vector; It is expressed as: ； ； ； ； In the formula, It is The iteration The mass of the predator; It is Crested porcupine individuals are in The final speed during the iteration; It is Crested porcupine individuals The initial velocity at the iteration; is a randomly generated vector between 0 and 1; represents a natural constant; represents the objective function value of a crested porcupine individual randomly selected from the current crested porcupine population; represents the position of a crested porcupine individual randomly selected from the current crested porcupine population; for ,delete The denominator of is expressed as: 。 8. The transformer core loosening fault soundprint diagnosis method according to claim 7 is characterized by: The second defense strategy exists in three situations: The first case: When = 0, the crested porcupine individual stops diffusing the scent, the predator stops moving, and the distance between the predator and the crested porcupine individual remains unchanged; The second situation: When =1, the predator is nearby and the crested porcupine individual emits odor; The third situation: When When it is a combination of 0 and 1, the predator keeps a safe distance from the crested porcupine individual, and the crested porcupine individual does not release the scent.

9. The transformer core loose fault soundprint diagnosis method according to claim 8, characterized in that: The specific process of step S2 is: Step S201: The corresponding sensor The dimensional voiceprint feature maps are spliced ​​to form a matrix : ； In the formula, Indicated by The sensor collects Dimensional voiceprint feature map; Step S202: Matrix Standardize and get the standard matrix : ； In the formula, Indicates The sensor receives the The result after standardization of the dimensional voiceprint feature map; Indicated by The sensor collects Dimensional voiceprint feature map, , ; No. Information entropy of sensors It is expressed as: ； In the formula, ,and ; Represents the total dimension of the voiceprint feature map; express The weight of according to Calculate the information entropy weight value corresponding to each sensor : ； In the formula, Indicates The information entropy of each sensor is ; Step S203: Wavelet coefficients of the sensor The weight value corresponding to each sensor is calculated by the entropy weight method Fusion is performed to obtain the fused wavelet coefficient matrix , draw the fusion wavelet coefficient matrix to obtain the multi-sensor fusion voiceprint feature map MSF-CWT, that is, the wavelet dynamic fusion voiceprint feature map; among them, The wavelet coefficients of the sensors are based on the wavelet coefficient matrix get; A vector representing the scale factor; Represents the translation factor vector.

10. The transformer core loose fault soundprint diagnosis method according to claim 9, characterized in that: In step S3, the activation function of the improved ShufflenetV2 convolutional neural network adopts the MobieNetV3 model The activation function is expressed as: ； In the formula, for Improvement of the function to limit its value to a maximum value; To improve the input value of ShufflenetV2 convolutional neural network; The improved ShufflenetV2 convolutional neural network classifier is The layer is improved and expressed as: ； In the formula, for Dimensions of layers; for The elements, ; for The input of the layer.

Images