A Neural Network Identification Method for Porcelain Insulator Damage Based on Genetic Algorithm Feature Parameter Optimization
The acoustic signals are collected through the force hammer-microphone system and the characteristic parameters are optimized using genetic algorithms. Combined with the BP neural network, the subjectivity and equipment limitations of the existing porcelain insulator detection methods are solved, and efficient and accurate porcelain insulator damage recognition is achieved.
Patent Information
- Application Number
- CN202110617866.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-06-01
- Publication Date
- 2024-12-13
- Estimated Expiration
- 2041-06-01
AI Technical Summary
The existing porcelain insulator detection methods have problems such as strong subjectivity, large equipment and inconvenience, and high environmental requirements, which leads to insufficient detection efficiency and accuracy, making it difficult to effectively identify porcelain insulator damage.
The force hammer-microphone system is used to collect acoustic signals, extract the time domain, shape and wavelet energy characteristic parameters, and optimize the characteristic parameters through genetic algorithms, and combine them with BP neural network to identify them to achieve efficient detection of porcelain insulator damage.
Optimizing the feature parameter input BP neural network through genetic algorithms improves the accuracy and efficiency of porcelain insulator damage detection, reduces the influence of human and environmental factors, and achieves efficient damage recognition.
Smart Images
Figure CN113361383B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a porcelain insulator damage identification method, in particular to a porcelain insulator damage neural network identification method based on genetic algorithm characteristic parameter optimization. The method is suitable for porcelain insulator damage identification and belongs to the field of non-destructive testing. Background Art
[0002] Porcelain insulators are important components of the national power grid system, playing the role of insulating and supporting conductors. Porcelain insulator specimens are mainly composed of quartz particles distributed in a glass ground state. Due to the difference in expansion coefficients between the matrix and the quartz particles, porcelain insulators are prone to microcracks when in service. However, ceramics are a typical brittle material, and cracks that appear in service are very easy to expand, which in turn causes very serious consequences. Therefore, the detection of porcelain insulators is particularly important.
[0003] Commonly used porcelain insulator inspection methods include visual inspection, ultrasonic inspection, infrared imaging inspection and ultraviolet corona imaging inspection, among which visual inspection is the most commonly used inspection method when porcelain insulators are formed, that is, inspectors use measuring tools, combined with product inspection standards and inspection experience of inspectors, to inspect and judge the quality status of the surface of porcelain insulator workpieces. It is easy to operate and flexible. However, due to the strong subjectivity of the visual method, it is easy to miss inspections and misjudgments. Ultrasonic inspection is also a commonly used porcelain insulator inspection method. Ultrasonic inspection of porcelain insulators is divided into creeping wave inspection method and phased array inspection method. The principle of creeping wave inspection method is that when the ultrasonic incident angle is close to the first critical angle, the surface of the inspected workpiece will generate creeping waves that propagate along the surface of the workpiece. If a defect is found, a defect echo will be generated to detect defects in porcelain insulators. However, the ultrasonic wave is greatly attenuated when propagating inside the workpiece, and with the presence of clutter, the effect of on-site inspection is poor. , which has great limitations; the phased array method has solved the disadvantages of the creeping wave method very well, but the ultrasonic detection method also has limitations because the detection equipment is too large and not easy to carry, and both methods require power off for detection; infrared imaging and ultraviolet corona imaging are both methods that can detect damage without power off, but both have high requirements for the environment. If the experimental environment does not meet the detection conditions, the damage detection rate is low; the detection conditions of the above detection methods are relatively harsh. In order to improve this situation and ensure the detection rate of porcelain insulator damage, a new porcelain insulator detection method needs to be studied.
[0004] Using the change of structural parameters to detect the damage of porcelain insulators is an effective detection method. Among them, the change of structural damage state will directly cause the change of received acoustic signal, and the acoustic signal contains a lot of damage information, which can be characterized by characteristic parameters. Therefore, this paper uses a hammer-microphone system to excite porcelain insulator specimens with different damage states, and extracts three types of characteristic parameters from the received acoustic signal, namely time domain, shape, and wavelet energy characteristic parameters. Afterwards, the acoustic signal is identified by the BP neural network method, thereby achieving the purpose of porcelain insulator damage detection.
[0005] If the three types of characteristic parameters are directly input into the neural network for training, it may cause information redundancy, increase the amount of calculation, and affect the recognition effect of the neural network. Therefore, this paper uses genetic algorithm to further optimize the characteristic parameters, find the most suitable parameters for damage characterization from the three types of characteristic parameters, and form an eigenvalue matrix to input into the neural network structure, so as to realize the detection of porcelain insulator damage efficiently and intelligently. Summary of the invention
[0006] The design of the present invention is to provide a method for identifying porcelain insulator damage, in particular, a method for identifying porcelain insulator damage by a neural network based on optimization of characteristic parameters of a genetic algorithm. The method obtains three types of characteristic parameters of a received acoustic signal: time domain characteristic parameters, frequency domain characteristic parameters, and wavelet energy characteristic parameters. The classifier uses a BP neural network to optimize the BP neural network input parameters and the acoustic signal characteristic parameters, respectively. The characteristic parameters are optimized using a genetic algorithm, and the optimized recognition model is used to identify the porcelain insulator acquisition signal.
[0007] The present invention proposes a neural network identification method for porcelain insulator damage based on genetic algorithm feature parameter optimization, and its basic principle is as follows:
[0008] During the service life of porcelain insulators, the upper and lower parts are more likely to be damaged due to the uneven force on the main body. The acoustic signals of three porcelain insulator samples were collected using a hammer-microphone system. Since the acoustic signals contain a large amount of characteristic information about the porcelain insulator specimens, three types of characteristic parameters were extracted from the signals.
[0009] (1) Time domain characteristic parameters
[0010] The received signal of the porcelain insulator contains a lot of characteristic information. Different characteristic parameters represent different characteristic information, and the time domain characteristic parameters can be directly extracted from the time domain signal.
[0011]
[0012]
[0013] Table 1
[0014] (2) Shape characteristic parameters
[0015] The shape characteristic parameters of the signal also contain a lot of characteristic information and can be used to characterize the damage of porcelain insulators.
[0016] Shape feature parameters can also be directly extracted from the signal and can be effectively used to express the characteristics of the signal.
[0017]
[0018] Table 2
[0019] (3) Wavelet energy characteristic parameters
[0020] The received signal of the porcelain insulator is a typical non-stationary random signal. If we want to extract the energy distribution of the signal at different scales, we need to extract the eigenvalues related to the wavelet energy entropy. The wavelet entropy theory is a theory that combines the wavelet analysis method with the information entropy theory, and can effectively characterize the energy distribution of the signal in the time domain and frequency domain.
[0021] Performing 3-layer wavelet packet transform on each signal, the signal is divided into 8 frequency bands, and 8 wavelet energy ratios, 8 wavelet scale entropies, wavelet energy entropy and wavelet singular spectrum entropy of each frequency band can be obtained, a total of 18 eigenvalues.
[0022]
[0023]
[0024] Table 3
[0025] The above 32 eigenvalues that can be extracted from a single signal are optimized using a genetic algorithm. A genetic algorithm is a probabilistic random search optimization algorithm that simulates natural evolution phenomena such as natural selection, hybridization, and mutation of biological populations to evolve excellent individuals. In the process of feature selection, it can be used as a search strategy for feature subsets (the overall design framework of genetic search feature reduction is as follows Figure 4 Its basic principle is as follows:
[0026] Genetic algorithm is a probabilistic random search optimization algorithm. Its basic operation process is as follows:
[0027] The process of genetic algorithm consists of six steps, namely chromosome encoding, population initialization, individual fitness evaluation, selection, crossover and mutation. This study uses binary arrays to represent the chromosomes of individual individuals, and the encoding length is the chromosome length; when the population is initialized, the encoding of each individual in the initial group will be generated in a random manner; when evaluating individual fitness, the fitness function is the only basis for the transmission or elimination of individual individuals. The higher the individual fitness value, the greater the possibility of being inherited to the next generation, and vice versa, it is more likely to be eliminated; when performing selection operations, the elimination of individuals will be determined based on the fitness value; when performing crossover operations, two mutually crossed genes will exchange chromosomes at a certain position in the chromosome, thereby generating two new individuals, which is the main way to generate new individuals; when performing mutation operations, in general binary-coded individuals, the number at a certain position is changed, which has the effect of maintaining population diversity.
[0028] Specifically, in this algorithm combining genetic algorithm and neural network, for chromosome coding, the chromosome coding in the individual is "0" or "1", which represents whether the characteristic value participates in the subsequent neural network training. For fitness evaluation, the quality of individuals in the population is positively correlated with the level of fitness, but the quality of the target is negatively correlated with the level of mean square error, so the fitness function takes the opposite of the mean square error value between the expected output of the neural network and the actual output.
[0029] Assume that a trained neural network corresponding to an individual predicts N test samples, and the actual output and predicted output of the test samples are a 1 ,a 2 ...a N and b 1 ,b 2 ,...,b N , then the mean square error MSE can be expressed as follows:
[0030]
[0031] Where: a j is the actual output, b j is the predicted output, N is the number of samples. The fitness function O is shown in formula (2).
[0032] O=-MSE (2)
[0033] The target value is normalized as the individual fitness function. When the genetic algorithm optimization is completed, the current chromosome number sequence is taken and decoded and converted into an eigenvalue matrix for subsequent porcelain insulator damage identification research.
[0034] The BP neural network is used to identify porcelain insulator damage. The BP network can approximate and store a large number of input-output nonlinear mapping relationships. Its training rule is to use the steepest descent method to continuously iterate and update the network weights and thresholds through back propagation to minimize the network's root mean square error. The BP neural network can be used to identify porcelain insulator damage efficiently and accurately, achieving the purpose of porcelain insulator damage detection.
[0035] The invention proposes a neural network identification method for porcelain insulator damage based on genetic algorithm characteristic parameter optimization, which is implemented by the following steps:
[0036] Step 1: Build the experimental system.
[0037] according to Figure 1 Build a porcelain insulator detection system.
[0038] Step 2: Acoustic signal collection.
[0039] The hammer excitation signal and the microphone receiving signal are collected synchronously to collect the acoustic signals of three types of porcelain insulator specimens, namely "porcelain insulator without defects", "porcelain insulator with defects on the upper part" and "porcelain insulator with defects on the lower part". The porcelain insulator training signals and test signals are collected successively.
[0040] Step 3: Acoustic signal preprocessing.
[0041] The signal collected every minute is truncated into several porcelain insulator knocking signals, and the knocking signals are subjected to Fourier transform operation to obtain the time domain signal and spectrum diagram.
[0042] Step 4: Extraction of acoustic signal feature parameters.
[0043] The time domain feature parameters, shape feature parameters and wavelet energy feature parameters are extracted respectively.
[0044] Step 5: Optimization of BP neural network.
[0045] The 32-dimensional full feature parameter matrix is used to optimize the BP neural network input parameters, including the "transfer function from input layer to hidden layer", "transfer function from hidden layer to output layer" and "neural network learning rate".
[0046] Step 6: Optimization of feature parameters
[0047] Optimizing Neural Network Inputs Using Genetic Algorithms:
[0048] (1) Set the population size to 50, the crossover probability to 0.9, the mutation probability to 0.02, and the iteration termination condition to 300 times;
[0049] (2) Binary encoding is performed on the chromosome. A single signal can extract 32-dimensional feature parameters. The length of the chromosome is 32. "0" means that the corresponding feature parameter does not participate in training, and "1" means that it participates in training.
[0050] (3) Using random coding to initialize the population;
[0051] (4) Calculate the fitness function, decode the chromosome and generate the corresponding feature matrix, input the neural network training, calculate the mean square error between the actual output and the predicted output, take the inverse of the mean square error and normalize it as the individual fitness function;
[0052] (5) Calculate the fitness of each individual, determine the inherited individuals and the eliminated individuals, and perform transmission, crossover and mutation operations; repeat steps (4) and (5) before reaching the iteration termination condition; according to the chromosome encoding of the final result, take the eigenvalue corresponding to "1" to form an eigenvalue matrix to complete the optimization process.
[0053] Step 7: Construction of BP neural network
[0054] (1) Input the optimized training samples into the neural network, set the number of input nodes to 16, the number of output nodes to 3, the number of hidden layer nodes to 11, the input layer-hidden layer transfer function to tansig, and the hidden layer-output layer transfer function to logsig;
[0055] (2) Train the BP neural network, iteratively solve the optimal network weights and threshold matrix, and save the training results.
[0056] Step 8: Assessment of service status of porcelain insulators.
[0057] The eigenvalues of the test samples are extracted, and according to the eigenvalue matrix of the optimization results, they are input into the BP neural network model, and the matching degree between the test samples and various training samples is calculated respectively, and the samples are classified into the category with the highest matching degree.
[0058] The present invention has the following advantages:
[0059] 1) The genetic algorithm is used to optimize the extracted characteristic parameters of the three types of acoustic signals, and the characteristic parameter matrix that is most suitable for characterizing the damage state of porcelain insulators is selected in an unsupervised manner.
[0060] 2) The service quality of porcelain insulators is evaluated based on BP neural network, which avoids the influence of human and environmental factors on defect identification and realizes the detection of porcelain insulator damage. BRIEF DESCRIPTION OF THE DRAWINGS
[0061] Figure 1 Schematic diagram of porcelain insulator experimental system
[0062] Figure 2(a)-Figure 2(c) Schematic diagram of collecting acoustic signals
[0063] Figure 3 The influence of learning rate on the performance of BP neural network
[0064] Figure 4 Genetic Algorithm Flowchart
[0065] Figure 5 Genetic algorithm iteration process and results
[0066] Figure 6(a)-Figure 6(d) Distribution of some input feature values of neural network
[0067] Figure 7 BP neural network training performance graph
[0068] Figure 8(a)-Figure 8(b) Test sample recognition results
[0069] Fig. 9 Flowchart of this method DETAILED DESCRIPTION
[0070] The present invention will be further described below in conjunction with specific experiments:
[0071] The implementation process of this experiment includes the following steps:
[0072] Step 1: Build the experimental system.
[0073] Build a force hammer-microphone system: connect the modal force hammer to the force sensor, connect the microphone to the signal amplifier, and connect the amplifier and force sensor to the input 1 interface and input 2 interface of the HS3 acquisition board respectively. Connect the output interface to the PC end to synchronously collect the force hammer excitation signal and the microphone receiving signal.
[0074] Step 2: Acoustic signal collection.
[0075] Use a hammer to stimulate porcelain insulators with different damage states, and use a microphone to receive the acoustic signal. (The schematic diagram of the porcelain insulator experimental system is shown in Figure 1 The sampling frequency is selected as 30kHz, and the HS3 acquisition board is used to collect and store the signal in real time. The signal storage is in minutes, and the acoustic signal of each minute is stored in different Excel tables. The experiment selects 3 different types of porcelain insulator workpieces (complete, upper defect and lower defect), and collects training samples and test samples. The training samples will be divided into two parts for parameter optimization. The process of parameter optimization is to randomly extract 50 groups of signals from the training samples each time as optimized test samples, and the rest as optimized training samples. The operation is repeated 5 times to obtain the 5 average recognition rates as a measure of parameter optimization.
[0076] Step 3: Acoustic signal preprocessing.
[0077] The acoustic signal collected in step 2 is exported, and the signal is divided into several acoustic signals based on the signal interception method of amplitude. (The schematic diagram of collecting acoustic signals is shown in Figure 2(a)-Figure 2(c) (shown)
[0078] Step 4: Extraction of acoustic signal feature parameters.
[0079] Three types of characteristic parameters are extracted from each signal, including 7 time domain characteristic parameters, 7 shape characteristic parameters and 18 wavelet energy characteristic parameters, and the extracted characteristic parameters are normalized.
[0080] Step 5: Parameter optimization of BP neural network.
[0081] 32-dimensional feature parameters are extracted from the collected acoustic signal training samples and test samples to form a training eigenvalue matrix and a test eigenvalue matrix. The training sample eigenvalue matrix is directly used as input to optimize the BP neural network structure.
[0082] (1) Optimize the input layer-hidden layer transfer function and the hidden layer-output layer transfer function, and select the transfer function combination with the best neural network recognition efficiency: "tansig-logsig";
[0083] (2) Change the network learning rate, study the effect of the learning rate on the neural network recognition effect, and finally determine the network learning rate parameter to be 0.005. (The effect of learning rate on the neural network recognition efficiency is as follows Figure 3 (shown)
[0084] Step 6: Optimization of feature parameters
[0085] Genetic algorithm is used to optimize the 32-dimensional feature parameters (the genetic algorithm flow chart is as follows Figure 4 As shown in FIG. 2 , 20 signals are selected from the three types of signals, a total of 60 groups of signals are selected as training samples, and the test samples are consistent with those described above. The optimization process is divided into the following steps:
[0086] (1) Set the genetic algorithm parameters, the iteration termination condition is 300 times, the population size is 50, the individual chromosome length is the same as the eigenvalue dimension, which is set to 32, the crossover probability is 0.8, and the mutation probability is 0.02;
[0087] (2) Individual initialization coding, using binary coding, each dimension of feature parameters has two possibilities, participating in neural network training is coded as "1", and not participating in training is coded as "0";
[0088] (3) Individual decoding, generating the characteristic parameter matrix corresponding to each chromosome, inputting them into the BP neural network for training, and inputting the test signal to obtain the mean square error between the predicted output and the actual output. The inverse of the mean square error is taken and normalized as the fitness function of the genetic algorithm. Since there is randomness in neural network training, the mean square error is the result of 10 repeated trainings.
[0089] (4) Retain the individuals with the highest fitness in each generation, and perform natural selection, crossover, and mutation operations on the population to obtain a new population;
[0090] (5) If the termination condition is not met, steps (3) and (4) are repeated. If the termination condition is met, the final result is selected for chromosome decoding to obtain the optimized eigenvalue combination and conduct subsequent research.
[0091] (Genetic algorithm optimization process is as follows Figure 5 (shown)
[0092] Step 7: Iterative training of BP neural network
[0093] (1) Input the training samples into the neural network and set the initial value parameters of the neural network, where the input layer is set to 16, the output layer is set to 3, the hidden layer neurons are set to 11 according to the formula, the hidden layer transfer function selects the tansig transfer function, the output layer transfer function selects the logsig transfer function, the training algorithm selects the gradient descent adaptive learning training function, the maximum number of training times is set to 5000, the training accuracy is 0.001, and the learning rate is 0.005; (The distribution of some characteristic parameters is shown in Figure 6(a)-Figure 6(d) As shown: Figure 6 (a) is the skewness parameter, Figure 6 (b) is the wavelet scale entropy, Figure 6 (c)
[0094] is the margin factor, and Figure 6(d) is the kurtosis factor)
[0095] (2) Train the neural network with eigenvalue input, determine the final weight and threshold matrix, and optimize the optimal neural network structure parameters. (The neural network performance curve is shown in Figure 7 (shown)
[0096] Step 8: Evaluation of damage status of porcelain insulators.
[0097] The optimized test sample eigenvalue matrix is input into the BP neural network trained in step seven to identify porcelain insulator damage. The test sample is identified by the neural network, and the matching degree between the test sample and various training samples is calculated to identify the damage state of the test sample. In this experiment, the recognition rate is 91.7% when the full eigenvalue matrix before optimization is used as input. In comparison, the optimized 16-dimensional eigenvalue matrix is used as the test sample, with a total of 48 test samples, of which 47 test samples have accurate porcelain insulator defect recognition, and 1 test sample has a false detection. The detection rate of the upper defective test sample and the non-defective test sample is 100%, and the lower defective test sample has 1 false detection. The recognition rate of the optimized eigenvalue extraction method is 97.9%. (The neural network recognition results are shown in Figure 2. Figure 8(a)-Figure 8(b) As shown, Figure 8(a) is the result diagram before optimization, and Figure 8(b) is the result diagram after optimization)
[0098] The implementation flow chart of this method is as follows Fig. 9 shown.
[0099] The above is a typical application of the present invention, and the application of the present invention is not limited thereto.
Claims
1. A neural network identification method for porcelain insulator damage based on genetic algorithm characteristic parameter optimization, characterized in that: This is accomplished by the following steps: Step 1: Build the experimental system; Build a porcelain insulator detection system; Step 2: Acoustic signal collection; The hammer excitation signal and the microphone receiving signal are synchronously collected to collect the acoustic signals of three types of porcelain insulator specimens, namely "porcelain insulator without defects", "porcelain insulator with defects on the upper part", and "porcelain insulator with defects on the lower part". The porcelain insulator training signals and test signals are collected successively; Step 3: acoustic signal preprocessing; The signal collected every minute is truncated into several porcelain insulator knocking signals, and the knocking signals are subjected to Fourier transform operation to obtain time domain signals and spectrum diagrams; Step 4: Extraction of acoustic signal characteristic parameters; Extract time domain characteristic parameters, shape characteristic parameters and wavelet energy characteristic parameters respectively; Step 5: Optimization of BP neural network; Use the 32-dimensional full-feature parameter matrix to optimize the BP neural network input parameters, including "input layer-hidden layer transfer function", "hidden layer-output layer transfer function" and "neural network learning rate"; Step 6: Optimization of feature parameters; Optimizing Neural Network Inputs Using Genetic Algorithms: (1) Set the population size to 50, the crossover probability to 0.9, the mutation probability to 0.02, and the iteration termination condition to 300 times; (2) Binary encoding is performed on the chromosome. A single signal is used to extract 32-dimensional feature parameters. The length of the chromosome is 32. "0" means that the corresponding feature parameter does not participate in training, and "1" means that it participates in training. (3) Using random coding to initialize the population; (4) Calculate the fitness function, decode the chromosome and generate the corresponding feature matrix, input the neural network training, calculate the mean square error between the actual output and the predicted output, take the inverse of the mean square error and normalize it as the individual fitness function; (5) Calculate the fitness of each individual, determine the inherited individuals and the eliminated individuals, and perform transmission, crossover and mutation operations; repeat steps (4) and (5) before reaching the iteration termination condition; according to the chromosome encoding of the final result, take the eigenvalues relative to "1" to form an eigenvalue matrix to complete the optimization process; Step 7: Construction of BP neural network (1) Input the optimized training samples into the neural network, set the number of input nodes to 16, the number of output nodes to 3, the number of hidden layer nodes to 11, the input layer-hidden layer transfer function to tansig, and the hidden layer-output layer transfer function to logsig; (2) Train the BP neural network, iteratively solve the optimal network weights and threshold matrix, and save the training results; Step 8: Evaluation of service status of porcelain insulators; Extract the eigenvalues of the test samples, input the eigenvalue matrix of the optimization results into the BP neural network model, calculate the matching degree between the test samples and various training samples, and classify them into the category with the highest matching degree; During the service life of porcelain insulators, the upper and lower parts of the porcelain insulators are more likely to be damaged due to the uneven force on the main body; The hammer-microphone acquisition system is used to collect the acoustic signals of Sanli porcelain insulator samples. Since the acoustic signals contain a lot of characteristic information about the porcelain insulator workpiece, three types of characteristic parameters are extracted from the signals; (1) Time domain characteristic parameters The received signal of the porcelain insulator contains a lot of characteristic information; different characteristic parameters represent different characteristic information. The following time domain characteristic parameters are extracted from the time domain signal; (2) Shape characteristic parameters The shape characteristic parameters of the signal also contain a lot of characteristic information, which can be used to characterize the damage of porcelain insulators; The shape feature parameters are as follows: S1 peak factor S2 shape parameter S3 Kurtosis S4 Pulse Factor S5 skewness parameter S6 Margin Factor S7 Kurtosis Factor (3) Wavelet energy characteristic parameters The received signal of porcelain insulator is a typical non-stationary random signal. If we want to extract the energy distribution of the signal at different scales, we need to extract the eigenvalues related to the wavelet energy entropy. Wavelet entropy theory is a theory that is a combination of wavelet analysis and information entropy theory. It can effectively characterize the energy distribution of signals in the time domain and frequency domain. The collected signal is transformed into 8 frequency bands by three layers of wavelet packet transform. The wavelet energy entropy is calculated based on the wavelet packet coefficients of the third layer. For a single sound signal, 8 wavelet energy ratios, 8 wavelet scale entropies, 1 wavelet energy entropy and 1 wavelet singular spectrum entropy are obtained, for a total of 18 eigenvalues.
2. The method for identifying porcelain insulator damage by neural network based on genetic algorithm characteristic parameter optimization according to claim 1 is characterized in that: The process of genetic algorithm consists of six steps, namely chromosome encoding, population initialization, individual fitness evaluation, selection, crossover and mutation; this study uses binary arrays to represent the chromosomes of individual individuals, and the encoding length is the chromosome length; when the population is initialized, the encoding of each individual in the initial group will be generated in a random manner; when evaluating individual fitness, the fitness function is the only basis for the transmission or elimination of individual individuals. The higher the individual fitness value, the greater the possibility of being inherited to the next generation, and vice versa, it is more likely to be eliminated; when performing selection operations, the elimination of individuals will be determined based on the fitness value; when performing crossover operations, two mutually crossed genes will exchange chromosomes at a certain position in the chromosome, thereby generating two new individuals, which is the main way to generate new individuals; when binary-coded individuals perform mutation operations, the population diversity is maintained by changing the number at a certain position; The genetic algorithm parameter optimization method is as follows: In the algorithm combining the genetic algorithm with the neural network, for chromosome coding, the chromosome coding in the individual is "0" or "1", which represents whether the characteristic value participates in the subsequent neural network training; for fitness evaluation, the quality of the individual in the population is positively correlated with the fitness level, but the quality of the target is negatively correlated with the mean square error value, so the fitness function takes the opposite of the mean square error value between the expected output of the neural network and the actual output; Suppose a trained neural network corresponding to an individual predicts N test samples, and the actual output and predicted output of the test samples are a1, a2...a N and b1,b2,...,b N , then the mean square error MSE is expressed as follows: Where: a j is the actual output, b j is the predicted output, N is the number of samples; the fitness function O is shown in formula (2); O=-MSE(2) The target value is normalized as the individual fitness function. When the genetic algorithm optimization is completed, the current chromosome number sequence is taken and decoded and converted into an eigenvalue matrix.
3. The method for identifying porcelain insulator damage by neural network based on genetic algorithm characteristic parameter optimization according to claim 1 is characterized in that: Step 1: Build the experimental system; Build a force hammer-microphone system: connect the modal force hammer to the force sensor, connect the microphone to the signal amplifier, and connect the amplifier and force sensor to the input 1 interface and input 2 interface of the HS3 acquisition board respectively. Connect the output interface to the PC end to synchronously collect the force hammer excitation signal and the microphone receiving signal; Step 2: Acoustic signal collection; A hammer is used to stimulate porcelain insulators in different damage states, and a microphone is used to receive acoustic signals. The sampling frequency is selected as 30kHz. The HS3 acquisition board is used to collect and store the signals in real time. The signal storage is in units of minutes, and the acoustic signals per minute are stored in different Excel tables. Three different types of porcelain insulator workpieces are selected for the experiment, namely complete, upper defect and lower defect. Training samples and test samples are collected. The training samples will be divided into two parts for parameter optimization. The process of parameter optimization is to randomly extract multiple groups of signals from the training samples each time as optimized test samples, and the rest as optimized training samples. The operation is repeated 5 times, and the average recognition rate of 5 times is obtained as the measure of parameter optimization. Step 3: acoustic signal preprocessing; Exporting the acoustic signal collected in step 2, and dividing the signal into a plurality of acoustic signals based on a signal interception method of amplitude; Step 4: Extraction of acoustic signal characteristic parameters; Extract three types of characteristic parameters from each signal, including 7 time domain characteristic parameters, 7 shape characteristic parameters and 18 wavelet energy characteristic parameters, and normalize the extracted characteristic parameters; Step 5: Parameter optimization of BP neural network; Extract 32-dimensional feature parameters from the collected acoustic signal training samples and test samples respectively to form a training eigenvalue matrix and a test eigenvalue matrix. Use the training sample eigenvalue matrix directly as input to optimize the BP neural network structure. Optimize the input layer-hidden layer transfer function and the hidden layer-output layer transfer function, and select the transfer function combination with the best neural network recognition efficiency: "tansig-logsig"; Change the network learning rate, study the effect of the learning rate on the neural network recognition effect, and finally determine the network learning rate parameter to be 0.005; Step 6: Optimization of feature parameters The 32-dimensional feature parameters are optimized using a genetic algorithm. 20 signals are selected from each of the three types of signals, a total of 60 groups of signals are used as training samples, which are divided into the following steps: (1) Set the genetic algorithm parameters, the iteration termination condition is 300 times, the population size is 50, the individual chromosome length is the same as the characteristic dimension, which is set to 32, the crossover probability is 0.8, and the mutation probability is 0.02; (2) Individual initialization coding, using binary coding, each dimension of feature parameters has two possibilities, participating in neural network training is coded as "1", and not participating in training is coded as "0"; (3) Individual decoding, generating the special positive parameter matrix corresponding to each chromosome, inputting them into the BP neural network for training, and inputting the test signal to obtain the mean square error between the predicted output and the actual output. The inverse of the mean square error is taken and normalized as the fitness function of the genetic algorithm. Since there is randomness in neural network training, the mean square error is the result of 10 repeated trainings. (4) Retain the individuals with the highest fitness in each generation, and perform natural selection, crossover, and mutation operations on the population to obtain a new population; (5) If the termination condition is not met, repeat steps (3) and (4). If the termination condition is met, select the final result for chromosome decoding; Step 7: Iterative training of BP neural network (1) Input the training samples into the neural network and set the initial value parameters of the neural network. The input layer is set to 16, the output layer is set to 3, the hidden layer neurons are set to 11 according to the formula, the hidden layer transfer function selects the tansig transfer function, and the output layer transfer function selects the logsig transfer function. The training algorithm selects the gradient descent adaptive learning training function, sets the maximum number of training times to 5000, the training accuracy to 0.001, and the learning rate to 0.005; (2) Train the neural network with eigenvalue input, determine the final weight and threshold matrix, and output the optimal neural network structure parameters.