[go: up one dir, main page]

CN113361383A - Porcelain insulator damage neural network identification method based on genetic algorithm characteristic parameter optimization - Google Patents

Porcelain insulator damage neural network identification method based on genetic algorithm characteristic parameter optimization Download PDF

Info

Publication number
CN113361383A
CN113361383A CN202110617866.7A CN202110617866A CN113361383A CN 113361383 A CN113361383 A CN 113361383A CN 202110617866 A CN202110617866 A CN 202110617866A CN 113361383 A CN113361383 A CN 113361383A
Authority
CN
China
Prior art keywords
neural network
signal
training
porcelain insulator
individual
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202110617866.7A
Other languages
Chinese (zh)
Other versions
CN113361383B (en
Inventor
焦敬品
赵博生
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing University of Technology
Original Assignee
Beijing University of Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Beijing University of Technology filed Critical Beijing University of Technology
Priority to CN202110617866.7A priority Critical patent/CN113361383B/en
Publication of CN113361383A publication Critical patent/CN113361383A/en
Application granted granted Critical
Publication of CN113361383B publication Critical patent/CN113361383B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/086—Learning methods using evolutionary algorithms, e.g. genetic algorithms or genetic programming
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/02—Preprocessing
    • G06F2218/04—Denoising
    • G06F2218/06—Denoising by applying a scale-space analysis, e.g. using wavelet analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/08—Feature extraction

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Data Mining & Analysis (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Signal Processing (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Physiology (AREA)
  • Investigating Or Analyzing Materials By The Use Of Ultrasonic Waves (AREA)

Abstract

A porcelain insulator damage neural network identification method based on genetic algorithm characteristic parameter optimization belongs to the field of nondestructive testing. The method collects the acoustic signals of three different types of porcelain insulator test pieces, intercepts the collected acoustic signals by using an amplitude-based method, extracts three different types of characteristic parameters which are 32 time domain, shape and wavelet energy characteristic parameters respectively, and optimizes BP neural network input parameters including an input-hidden layer transfer function, a hidden-output layer transfer function and a learning rate. Optimizing the characteristic value matrix by using a genetic algorithm, and determining the optimized input characteristic value of the neural network; constructing the characteristic parameters of the test sample according to the optimized result, and inputting the characteristic parameters into the trained BP neural network model for identification; the high-efficiency recognition of the damage state of the porcelain insulator is realized by calculating the matching degree of the test sample and various different porcelain insulator signal training samples.

Description

Porcelain insulator damage neural network identification method based on genetic algorithm characteristic parameter optimization
Technical Field
The invention relates to a method for identifying damage of a porcelain insulator, in particular to a neural network identification method for damage of the porcelain insulator based on genetic algorithm characteristic parameter optimization. The method is suitable for damage identification of the porcelain insulator and belongs to the field of nondestructive testing.
Background
The porcelain insulator is an important component device in a national power grid system and plays a role in insulating and supporting a wire. The porcelain insulator test piece mainly comprises quartz particles distributed in a glass ground state. The ceramic insulator is very easy to generate micro cracks in a service state due to the difference of expansion coefficients of the matrix and the quartz particles, but the ceramic is a typical brittle material, and the cracks generated in service are very easy to expand, so that very serious results are caused, and the ceramic insulator is particularly important for detection of the ceramic insulator.
The common porcelain insulator detection means comprises visual detection, ultrasonic detection, infrared imaging detection, ultraviolet corona imaging detection and the like, wherein the visual detection is the most common detection mode during porcelain insulator forming, namely, detection personnel detect and judge the quality condition of the surface of a porcelain insulator workpiece by means of a measuring tool and in combination with the detection standard of a product and the detection experience of the detection personnel. The operation is simple and convenient, the flexibility is strong, however, the conditions of missed detection, erroneous judgment and the like are easy to occur due to the strong subjectivity of the visual method; the ultrasonic detection method is also a commonly used porcelain insulator detection method, and is divided into a creeping wave detection method and a phased array detection method, wherein the creeping wave detection method is based on the principle that when the incident angle of ultrasonic waves is close to a first critical angle, creeping waves propagating along the surface of a workpiece are generated on the surface of the workpiece to be detected, and if the defects occur, defect echoes are generated, so that the detection of the defects of the porcelain insulator is realized, but the propagation attenuation of the ultrasonic waves in the workpiece is extremely large, and along with the existence of clutter, the effect during field detection is poor, and the detection is very limited; the phased array method well solves the disadvantages of the creeping wave method, but the ultrasonic detection method has limitations because the detection equipment is too large and is inconvenient to carry, and the two methods both need to be powered off for detection; the infrared imaging method and the ultraviolet corona imaging method belong to methods for detecting damage without power failure, but the requirements of the infrared imaging method and the ultraviolet corona imaging method on the environment are high, and if the experimental environment does not meet the detection condition, the damage detection rate is low; the detection conditions of the detection means are relatively harsh, and in order to improve the condition and ensure the detection rate of the damage of the porcelain insulator, a novel porcelain insulator detection method needs to be researched.
The method for detecting the damage of the porcelain insulator by using the change of the structural parameters is an effective detection means. The change of the structural damage state can directly cause the change of the received sound signal, more damage information is stored in the sound signal, and the information carried in the sound signal can be represented through characteristic parameters. Therefore, the porcelain insulator test pieces in different damage states are excited through a force hammer-microphone system, three types of characteristic parameters, namely time domain, shape and wavelet energy characteristic parameters, in received sound signals are extracted, and then the sound signals are identified through a BP neural network method, so that the purpose of detecting the damage of the porcelain insulator is achieved.
If the three types of characteristic parameters are directly input into the neural network for training, information redundancy is possibly caused, the identification effect of the neural network is influenced while the calculated amount is increased, so that the characteristic parameters are further optimized by adopting a genetic algorithm, the parameters most suitable for damage characterization are searched from the three types of characteristic parameters, a characteristic value matrix is formed and input into the neural network structure, and the damage detection of the porcelain insulator is efficiently and intelligently realized.
Disclosure of Invention
The invention aims to provide 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 obtains three types of characteristic parameters of the received acoustic signal: the classifier adopts a BP neural network to respectively optimize input parameters of the BP neural network and characteristic parameters of acoustic signals, wherein the characteristic parameters are optimized by a genetic algorithm, and the ceramic insulator collected signals are identified by using an optimized identification model.
The invention provides a method for identifying a damaged neural network of a porcelain insulator based on genetic algorithm characteristic parameter optimization, which has the following basic principle:
during the service period of the porcelain insulator, the upper part and the lower part of the porcelain insulator are easy to damage due to unbalanced stress of the main body. The method is characterized in that a force hammer-microphone system is used for collecting acoustic signals of three porcelain insulator samples, and three types of characteristic parameters are extracted from the signals due to the fact that a large amount of characteristic information about a porcelain insulator test piece is stored in the acoustic signals.
(1) Time domain feature parameters
The received signal of the porcelain insulator contains a large amount of characteristic information. Different characteristic parameters represent different characteristic information, and the time domain characteristic parameters can be directly extracted through the time domain signals.
Figure BDA0003094094170000021
Figure BDA0003094094170000031
TABLE 1
(2) Shape feature parameter
The shape characteristic parameters of the signals also contain a large amount of characteristic information, and the shape characteristic parameters can be used for representing the damage of the porcelain insulator.
The shape characteristic parameters can also be directly extracted from the signals, and can be effectively used for characteristic expression of the signals.
Figure BDA0003094094170000032
TABLE 2
(3) Wavelet energy characteristic parameter
The received signal of the porcelain insulator is a typical non-stationary random signal, and if the energy distribution of the signal on different scales is to be extracted, the characteristic value related to the wavelet energy entropy is to be extracted. The wavelet entropy theory is a theory generated by combining a wavelet analysis method and an information entropy theory, and can effectively represent the energy distribution of signals on a time domain and a frequency domain.
And (3) performing 3-layer wavelet packet transformation on each signal, dividing the signal into 8 frequency bands, and respectively obtaining 8 wavelet energy ratios, 8 wavelet scale entropies, and 18 characteristic values of the wavelet energy entropies and the wavelet singular spectrum entropies of each frequency band.
Figure BDA0003094094170000033
Figure BDA0003094094170000041
TABLE 3
The 32 characteristic values which can be extracted from a single signal are optimized by using a genetic algorithm. The genetic algorithm is a probabilistic random search optimization algorithm, and excellent individuals are obtained through evolution by simulating natural evolution phenomena such as natural selection, hybridization, variation and the like of biological populations. It can be used as a search strategy for feature subset in the feature selection process (the general design framework flow of genetic search feature reduction is shown in fig. 8). Its basic principle is as follows:
the genetic algorithm is a probabilistic random search optimization algorithm, and the basic operation process is as follows:
the genetic algorithm process comprises six steps, namely chromosome coding, population initialization, individual fitness evaluation, selection, crossing and mutation. In the research, binary digit groups are used for representing chromosomes of a single body, and the coding length is the chromosome length; at population initialization, each individual code in the initial population will be generated in a random manner; when the individual fitness evaluation is carried out, the fitness function is used as the only basis for single body transfer or elimination, the higher the individual fitness value is, the higher the possibility of being inherited to the next generation is, otherwise, the more easy the individual fitness function is to be eliminated; when the selection operation is carried out, whether the individual is eliminated or not is determined according to the fitness value; when the crossover operation is carried out, the two genes which are crossed with each other can carry out chromosome interchange at a certain position in a chromosome, so that two new individuals are generated, and the method is a main mode for generating the new individuals; when mutation operation is performed, the number of a certain position in a general binary coding individual is changed, and the function of keeping population diversity is achieved.
Specifically, in the algorithm combining the genetic algorithm and the neural network, for the chromosome code, the chromosome code in the individual is '0' or '1', namely representing whether the characteristic value participates in the subsequent neural network training or not. For the fitness evaluation, the quality of individuals in the population is in positive correlation with the fitness, but the quality of the target is in negative correlation with the variance value, so that the fitness function takes the inverse number of the variance value of the predicted output and the actual output of the neural network.
Setting a trained neural network corresponding to an individual to predict N test samples, wherein the actual output and the predicted output of the test samples are respectively a1,a2...aNAnd b1,b2,...,bNThen the mean square error value MSE can be expressed as follows:
Figure BDA0003094094170000051
in the formula: a isjFor actual output, bjFor prediction output, N is the number of samples. The fitness function O is shown in equation (2).
O＝-MSE (2)
And (4) carrying out normalization operation on the target value, taking the current chromosome number sequence as an individual fitness function, decoding and converting the current chromosome number sequence into a characteristic value matrix when the genetic algorithm is optimized, and carrying out subsequent porcelain insulator damage identification research.
And the BP neural network is adopted to identify the damage of the porcelain insulator, and the BP network can approximate and store a large number of input-output nonlinear mapping relations. The training rule is to use the steepest descent method to continuously and iteratively update the weight and the threshold value of the network through back propagation so as to minimize the root mean square value of the error of the network. The damage of the porcelain insulator is identified through the BP neural network, the damage can be identified efficiently and accurately, and the purpose of detecting the damage of the porcelain insulator is achieved.
The invention provides a porcelain insulator damage neural network identification method based on genetic algorithm characteristic parameter optimization, which is realized by the following steps:
the method comprises the following steps: and (5) establishing an experimental system.
A porcelain insulator detection system is set up according to figure 1.
Step two: and collecting the acoustic signals.
The method comprises the steps of synchronously acquiring a force hammer excitation signal and a receiving signal of a microphone, acquiring acoustic signals of three types of porcelain insulator test pieces, namely a 'non-defective porcelain insulator', 'porcelain insulator with defect on the upper part thereof and' porcelain insulator with defect on the lower part thereof, and sequentially acquiring a porcelain insulator training signal and a testing signal.
Step three: and (4) preprocessing the acoustic signal.
And cutting the signals collected within every minute into a plurality of porcelain insulator knocking signals, and respectively carrying out Fourier transform operation on the knocking signals to obtain time domain signals and frequency spectrum diagrams.
Step four: and extracting the characteristic parameters of the acoustic signal.
And respectively extracting time domain characteristic parameters, shape characteristic parameters and wavelet energy characteristic parameters.
Step five: and (5) optimizing the BP neural network.
And optimizing BP neural network input parameters by using a 32-dimensional full-feature parameter matrix, wherein the BP neural network input parameters comprise an input layer-hidden layer transfer function, a hidden layer-output layer transfer function and a neural network learning rate.
Step six: preference of characteristic parameters
Optimizing neural network input using a genetic algorithm:
(1) setting the population quantity to be 50, the cross probability to be 0.9, the variation probability to be 0.02, and the iteration termination condition to be 300 times;
(2) binary coding is carried out on the chromosome, 32-dimensional characteristic parameters can be extracted from a single signal, the length of the chromosome is 32, 0 represents that the corresponding characteristic parameters do not participate in training, and 1 represents that the chromosome participates in training;
(3) initializing a population by adopting a random coding mode;
(4) calculating a fitness function, decoding a chromosome and generating a corresponding characteristic matrix, inputting the characteristic matrix into a neural network for training, calculating a mean square difference value between actual output and predicted output, and taking the opposite number of the mean square difference and normalizing the result to be used as an individual fitness function;
(5) calculating the fitness of each individual, determining genetic individuals and eliminated individuals, and performing transmission, crossing and mutation operations; repeating the steps (4) and (5) before the iteration termination condition is reached; and according to the chromosome coding of the final result, taking the eigenvalue corresponding to the '1' to form an eigenvalue matrix, and finishing the optimization process.
Step seven: construction of BP neural network
(1) Inputting the optimized training sample into a neural network, setting the number of input nodes of the network to be 16, the number of output nodes to be 3, the number of nodes of a hidden layer to be 11, and setting an input layer-hidden layer transfer function to be tansig and a hidden layer-output layer transfer function to be logsig;
(2) and training the BP neural network, iteratively solving the optimal network weight and the threshold matrix, and storing the training result.
Step eight: and (4) evaluating the service state of the porcelain insulator.
And extracting characteristic values of the test samples, inputting the characteristic values into a BP neural net model according to an optimized result characteristic value matrix, respectively calculating the matching degrees of the test samples and various training samples, and classifying the test samples into the class with the highest matching degree.
The invention has the following advantages:
1) and optimizing the extracted three types of acoustic signal characteristic parameters by using a genetic algorithm, and selecting a characteristic parameter matrix which is most suitable for representing the damage state of the porcelain insulator in an unsupervised and optimized manner.
2) The service state quality of the porcelain insulator is evaluated based on the BP neural network, the influence of human and environmental factors on defect identification is avoided, and the damage detection of the porcelain insulator is realized.
Drawings
FIG. 1 is a schematic diagram of a porcelain insulator experimental system
FIG. 2 schematic diagram of acoustic signal acquisition
FIG. 3 Effect of learning Rate on BP neural network Performance
FIG. 4 flow chart of genetic algorithm
FIG. 5 genetic Algorithm iterative Process and results
FIG. 6 distribution of input feature values of a neural network portion
FIG. 7 BP neural network training performance graph
FIG. 8 test sample identification results
FIG. 9 is a flow chart of the method
Detailed Description
The invention is further illustrated below with reference to specific experiments:
the experiment implementation process comprises the following steps:
the method comprises the following steps: and (5) establishing an experimental system.
Constructing a force hammer-microphone system: the modal force hammer is connected with the force sensor, the microphone is connected with the signal amplifier, the amplifier and the force sensor are respectively connected with an input 1 interface and an input 2 interface of the HS3 acquisition board card, an output interface is connected with a PC (personal computer) end, and a force hammer excitation signal and a microphone receiving signal are synchronously acquired.
Step two: and collecting the acoustic signals.
The method comprises the steps of exciting porcelain insulators in different damage states by using a force hammer, receiving acoustic signals by using a microphone, (a schematic diagram of a porcelain insulator experimental system is shown in figure 1) selecting a sampling frequency to be 30kHz, collecting and storing the signals in real time by using an HS3 collection board card, storing the signals in minutes, storing the acoustic signals of each minute in different excel tables, selecting 3 types of different porcelain insulator workpieces (complete, upper-defect and lower-defect) in an experiment, collecting a training sample and a test sample, wherein the training sample is divided into two parts for parameter optimization, the parameter optimization process is to randomly extract 50 groups of signals from the training sample every time as an optimized test sample, the rest parts are used as optimized training samples, and repeating 5 times of operation to obtain 5 times of average recognition rate as a measurement for parameter optimization.
Step three: and (4) preprocessing the acoustic signal.
And D, deriving the acoustic signals collected in the step II, and dividing the signals into a plurality of acoustic signals by using an amplitude-based signal interception method. (the schematic diagram of the collected sound signal is shown in FIG. 2)
Step four: and extracting the characteristic parameters of the acoustic signal.
Extracting three types of characteristic parameters in each signal, namely 7 time domain characteristic parameters, 7 shape characteristic parameters and 18 wavelet energy characteristic parameters, and performing normalization processing on the extracted characteristic parameters.
Step five: and optimizing parameters of the BP neural network.
Respectively extracting 32-dimensional characteristic parameters from the collected acoustic signal training sample and the collected acoustic signal testing sample to form a training characteristic value matrix and a testing characteristic value matrix, and directly taking the training sample characteristic value matrix as input to optimize the BP neural network structure.
(1) Optimizing the input layer-hidden layer transfer function and the hidden layer-output layer transfer function, and selecting the transfer function combination with the best neural network identification efficiency: "tansig-logsig";
(2) changing the network learning rate, researching the influence of the learning rate on the neural network identification effect, and finally determining the network learning rate parameter to be 0.005. (the influence of learning rate on neural network recognition efficiency is shown in FIG. 3)
Step six: preference of characteristic parameters
The 32-dimensional characteristic parameters are optimized by using a genetic algorithm, wherein a flow chart of the genetic algorithm is shown in fig. 4, 20 signals are respectively selected from three types of signals, 60 groups of signals are used as training samples, and a test sample is consistent with the above description, wherein the optimization process comprises the following steps:
(1) setting genetic algorithm parameters, wherein the iteration termination condition is 300 times, the population scale is 50, the individual chromosome length is the same as the dimension of the characteristic value, the individual chromosome length is set to be 32, the cross probability is 0.8, and the variation probability is 0.02;
(2) the individual initialization coding adopts binary coding, each dimension characteristic parameter has two possibilities, the coding is '1' when participating in the neural network training, and the coding is '0' when not participating in the training;
(3) decoding individuals, generating characteristic parameter matrixes corresponding to each chromosome, inputting the characteristic parameter matrixes into a BP neural network for training respectively, inputting test signals to obtain a mean square deviation value of prediction output and actual output, taking the opposite number of the mean square deviation value and normalizing the number of the mean square deviation value as a fitness function of a genetic algorithm, wherein the mean square deviation value is a result of 10 times of repeated training due to the fact that the neural network training has contingency;
(4) reserving the individuals with the maximum fitness in each generation, and performing natural selection, crossing and variation operations on the population to obtain a new population;
(5) if the end condition is not reached, repeating the steps (3) and (4), and if the end condition is reached, selecting the final result to perform chromosome decoding to obtain the optimized characteristic value combination, and performing subsequent research. (genetic Algorithm optimization Process is shown in FIG. 5)
Step seven: iterative training of BP neural networks
(1) Inputting a training sample into a neural network, setting initial value parameters of the neural network, wherein an input layer is set to be 16, an output layer is set to be 3, a hidden layer neuron is set to be 11 according to a formula, a tan sig transfer function is selected as a hidden layer transfer function, a logsig transfer function is selected as an output layer transfer function, a gradient descent self-adaptive learning training function is selected as a training algorithm, the maximum training frequency is set to be 5000 times, the training precision is 0.001, and the learning rate is 0.005; (the distribution of some characteristic parameters is shown in FIGS. 6(a) - (d) (a) is skewness parameter, (b) is wavelet scale entropy, (c) is margin factor, and (d) is kurtosis factor)
(2) And training the neural network input by the characteristic value, determining a final weight and a threshold matrix, and optimizing the optimal neural network structural parameters. (neural network Performance Curve is shown in FIG. 7)
Step eight: and (4) evaluating the damage state of the porcelain insulator.
Inputting the optimized characteristic value matrix of the test sample into the BP neural network trained in the seventh step, identifying the damage of the porcelain insulator, identifying the test sample through the neural network, calculating the matching degree of the test sample and various different training samples, and identifying the damage state of the test sample, wherein in the experiment, the identification rate is 91.7% when the full characteristic value matrix before optimization is used as input, compared with the experiment that the optimized 16-dimensional characteristic value matrix is used as the test sample and has 48 test samples, the porcelain insulator defect identification of 47 test samples is accurate, 1 test sample has false detection, the detection rate of the test sample with the defect on the upper part and the test sample without the defect is 100%, and the detection rate of the test sample with the defect on the lower part is 1 false detection, so the identification rate of the optimized characteristic value extraction method is 97.9%. (the neural network recognition results are shown in FIG. 8, in which (a) is the result graph before optimization and (b) is the result graph after optimization)
The flow chart of the method is shown in fig. 9.
The above is a typical application of the present invention, and the application of the present invention is not limited thereto.

Claims (4)

1.一种基于遗传算法特征参数优化的瓷绝缘子损伤神经网络识别方法，其特征在于，通过以下步骤实现的：1. a kind of porcelain insulator damage neural network identification method based on genetic algorithm characteristic parameter optimization, is characterized in that, realizes through the following steps: 步骤一：实验系统搭建；Step 1: Build the experimental system; 搭建瓷绝缘子检测系统；Build a detection system for porcelain insulators; 步骤二：声信号采集；Step 2: Acoustic signal acquisition; 对力锤激励信号和麦克风的接收信号进行同步采集，采集三类瓷绝缘子试件的声信号，分别是“无缺陷瓷绝缘子”、“上部含缺陷的瓷绝缘子”、“下部含缺陷的瓷绝缘子”，先后采集瓷绝缘子训练信号以及测试信号；The excitation signal of the hammer and the received signal of the microphone are collected synchronously, and the acoustic signals of three types of porcelain insulator specimens are collected, namely "defective porcelain insulator", "defective porcelain insulator in the upper part", and "porcelain insulator with defects in the lower part". ”, successively collect the training signal and the test signal of the porcelain insulator; 步骤三：声信号预处理；Step 3: Acoustic signal preprocessing; 将每分钟内采集的信号截断为若干瓷绝缘子敲击信号，分别对敲击信号进行傅里叶变换操作，得到时域信号和频谱图；The signal collected in every minute is truncated into several ceramic insulator knocking signals, and the Fourier transform operation is performed on the knocking signals respectively to obtain the time domain signal and the frequency spectrum; 步骤四：声信号特征参数提取；Step 4: extraction of acoustic signal feature parameters; 分别提取时域特征参数、形状特征参数及小波能量特征参数；Extract time domain feature parameters, shape feature parameters and wavelet energy feature parameters respectively; 步骤五：BP神经网络的优化；Step 5: Optimization of BP neural network; 使用32维全特征参数矩阵对BP神经网络输入参数进行优化，包括“输入层-隐含层的传递函数”、“隐含层-输出层的传递函数”以及“神经网络学习率”；Use 32-dimensional full feature parameter matrix to optimize the input parameters of BP neural network, including "input layer-hidden layer transfer function", "hidden layer-output layer transfer function" and "neural network learning rate"; 步骤六：特征参数的优选Step 6: Optimization of Feature Parameters 利用遗传算法优化神经网络输入：Optimizing neural network input using genetic algorithm: (1)设定种群数量50，交叉概率为0.9，变异概率为0.02，迭代终止条件300次；(1) Set the population size to 50, the crossover probability to be 0.9, the mutation probability to be 0.02, and the iteration termination condition to be 300 times; (2)对染色体进行二进制编码，单个信号可提取32维特征参数，则染色体的长度为32，“0”代表对应的特征参数不参与训练，“1”代表参与训练；(2) Binary coding is performed on the chromosome, and a single signal can extract 32-dimensional characteristic parameters, then the length of the chromosome is 32, "0" means that the corresponding characteristic parameter does not participate in the training, and "1" means that it participates in the training; (3)采用随机编码的方式进行种群初始化；(3) Use random coding to initialize the population; (4)计算适应度函数，对染色体解码并生成对应的特征矩阵，输入神经网络训练，计算实际输出与预测输出的均方差值，取均方差的相反数并归一化，作为个体适应度函数；(4) Calculate the fitness function, decode the chromosomes and generate the corresponding feature matrix, input the neural network training, calculate the mean square error value of the actual output and the predicted output, take the inverse of the mean square error and normalize it as the individual fitness function; (5)计算每个个体适应度，并决定遗传个体和淘汰个体，并进行传递、交叉和变异操作；在达到迭代终止条件之前，重复第(4)、(5)步；根据最终结果的染色体编码，取与“1”相对应的特征值组成特征值矩阵，完成优化过程；(5) Calculate the fitness of each individual, and determine the genetic individual and the eliminated individual, and perform transmission, crossover and mutation operations; before reaching the iteration termination condition, repeat steps (4) and (5); according to the chromosome of the final result Coding, take the eigenvalue corresponding to "1" to form an eigenvalue matrix to complete the optimization process; 步骤七：BP神经网络的构造Step 7: Construction of BP Neural Network (1)将优化后的训练样本输入到神经网络，设置网络的输入节点个数为16、输出节点个数为3、隐含层节点为11，输入层-隐含层传递函数为tansig、隐含层-输出层传递函数为logsig；(1) Input the optimized training samples into the neural network, set the number of input nodes of the network 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 be tansig, and the hidden layer to be 11. The transfer function between the containing layer and the output layer is logsig; (2)对BP神经网络进行训练，迭代求解最佳网络权值和阈值矩阵，保存训练结果；(2) Train the BP neural network, iteratively solve the optimal network weight and threshold matrix, and save the training results; 步骤八：瓷绝缘子服役状态评定；Step 8: Assessment of service status of porcelain insulators; 对测试样本进行特征值的提取，按照优化结果特征值矩阵，输入BP神经挽网络模型，分别计算出测试样本与各类训练样本的匹配度，并归类于匹配度最高的一类。Extract the eigenvalues of the test samples, input the BP neural pull network model according to the eigenvalue matrix of the optimization results, calculate the matching degree of the test samples and various training samples respectively, and classify them into the category with the highest matching degree. 2.根据权利要求1所述的一种基于遗传算法特征参数优化的瓷绝缘子损伤神经网络识别方法，其特征在于：2. a kind of porcelain insulator damage neural network identification method based on genetic algorithm characteristic parameter optimization according to claim 1, is characterized in that: 瓷绝缘子在服役期间，由于主体受力不均衡，瓷绝缘子上部和下部较容易产生损伤；使用力锤-麦克风采集系统对三种瓷绝缘子样品的声信号进行采集，由于声信号中蕴藏着大量的关于瓷绝缘子工件的特征信息，对信号进行三类特征参数的提取；During the service period of the porcelain insulator, due to the unbalanced force on the main body, the upper and lower parts of the porcelain insulator are more likely to be damaged. Regarding the characteristic information of the porcelain insulator workpiece, three types of characteristic parameters are extracted for the signal; (1)时域特征参数(1) Time-domain characteristic parameters 瓷绝缘子的接收信号中包含了大量特征信息；不同的特征参数代表了不同的特征信息，通过时域信号直接可以提取出如下时域特征参数；The received signal of the porcelain insulator contains a lot of characteristic information; different characteristic parameters represent different characteristic information, and the following time domain characteristic parameters can be directly extracted from the time domain signal;
Figure FDA0003094094160000021
Figure FDA0003094094160000021
Figure FDA0003094094160000031
Figure FDA0003094094160000031
 表1Table 1 (2)形状特征参数(2) Shape feature parameters 信号的形状特征参数也蕴藏着大量的特征信息，可用于表征瓷绝缘子损伤；形状特征参数具体如下；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 characteristic parameters are as follows;
Figure FDA0003094094160000032
Figure FDA0003094094160000032
 表2Table 2 (3)小波能量特征参数(3) Wavelet energy characteristic parameters 瓷绝缘子的接收信号是一种典型的非平稳随机信号，若要提取信号在不同尺度上的能量分布，则要对小波能量熵相关的特征值进行提取；小波熵理论是小波分析法和信息熵理论相结合而生的理论，能够对信号在时域和频域上的能量分布进行有效的表征；The received signal of the porcelain insulator is a typical non-stationary random signal. To extract the energy distribution of the signal on different scales, it is necessary to extract the eigenvalues related to the wavelet energy entropy; the wavelet entropy theory is the wavelet analysis method and information entropy. The theory combined with the theory can effectively characterize the energy distribution of the signal in the time domain and frequency domain;
Figure FDA0003094094160000033
Figure FDA0003094094160000033
Figure FDA0003094094160000041
Figure FDA0003094094160000041
 表3table 3 对采集信号进行3层小波包变换，将信号分成8个频段，由第3层的小波包系数进行小波能量熵的计算：对于单个声信号，分别可得到小波能量比8个，小波尺度熵8个，小波能量熵1个和小波奇异谱熵1个，共18个特征值。Perform 3-layer wavelet packet transformation on the collected signal, divide the signal into 8 frequency bands, and calculate the wavelet energy entropy by the wavelet packet coefficient of the third layer: For a single acoustic signal, 8 wavelet energy ratios and 8 wavelet scale entropy can be obtained respectively. There are 1 wavelet energy entropy and 1 wavelet singular spectrum entropy, a total of 18 eigenvalues. 3.根据权利要求2所述的一种基于遗传算法特征参数优化的瓷绝缘子损伤神经网络识别方法，遗传算法的过程由六步骤构成，分别是染色体编码、种群初始化、个体适应度评价、选择、交叉和变异；本次研究使用二进制数组表示单个体的染色体，编码长度即为染色体长度；在种群初始化时，初始群体中的每个个体编码将以随机的方式生成；在进行个体适应度评价时，适应度函数作为单个体传递或淘汰的唯一依据，个体适应度值越高，被遗传到下一代的可能性就越大，反之，则更容易被淘汰；在进行选择操作时，会根据适应度值决定个体的淘汰与否；在进行交叉操作时，两互相交叉的基因会在染色体中的某个位置进行染色体互换，从而产生两个新的个体，是生成新个体的主要方式；在进行变异操作时，一般的二进制编码个体中，使某个位置的数字发生变化，有保持种群多样性的作用；3. a kind of porcelain insulator damage neural network identification method based on genetic algorithm characteristic parameter optimization according to claim 2, the process of genetic algorithm consists of six steps, namely chromosome coding, population initialization, individual fitness evaluation, selection, Crossover and mutation; this study uses a binary array to represent the chromosome of a single individual, and the coding length is the length of the chromosome; when the population is initialized, the code of each individual in the initial population will be generated in a random manner; when evaluating the individual fitness , the fitness function is used as the only basis for the transmission or elimination of a single individual. The higher the individual fitness value, the greater the possibility of being inherited to the next generation, otherwise, the easier it is to be eliminated; The degree value determines whether the individual is eliminated or not; during the crossover operation, the two crossed genes will perform chromosome exchange at a certain position in the chromosome, thereby generating two new individuals, which is the main way to generate new individuals; When performing mutation operation, in a general binary coded individual, the number of a certain position is changed, which has the effect of maintaining the diversity of the population; 其特征在于遗传算法参数优化方法具体如下：It is characterized in that the genetic algorithm parameter optimization method is as follows: 本次遗传算法与神经网络相结合的算法中，对于染色体编码，个体中的染色体编码为“0”或“1”，即代表了特征值参与后续神经网络训练与否；对于适应度评价，种群中个体的好坏与适应度的高低为正相关，但目标的好坏与均方差值的高低成负相关，所以适应度函数则取神经网络预计输出与实际输出的均方差值的相反数；In this algorithm combining the genetic algorithm and the neural network, for the chromosome coding, the chromosome coding in the individual is "0" or "1", which means whether the eigenvalue participates in the subsequent neural network training or not; for the fitness evaluation, the population The quality of the individual is positively correlated with the level of fitness, but the quality of the target is negatively correlated with the level of the mean square error, so the fitness function takes the opposite of the mean square error between the expected output of the neural network and the actual output. number; 设一个个体对应的训练好的神经网络对N个测试样本进行预测，测试样本的实际输出和预测输出分别为a1,a2...aN和b1,b2,...,bN，则均方差值MSE则可如下表示：Suppose a trained neural network corresponding to an individual predicts N test samples, and the actual output and predicted output of the test sample are a 1 , a 2 ... a N and b 1 , b 2 , ..., b respectively N , then the mean square error value MSE can be expressed as follows:
Figure FDA0003094094160000051
Figure FDA0003094094160000051
 式中：aj为实际输出，bj为预测输出，N为样本个数；适应度函数O如公式(2)所示；In the formula: a j is the actual output, b j is the predicted output, and N is the number of samples; the fitness function O is shown in formula (2); O＝-MSE (2)O=-MSE (2) 对目标值进行归一化操作，作为个体适应度函数，遗传算法优化完成时，取当前染色体编号序列，并解码转化成为特征值矩阵。The target value is normalized as an individual fitness function. When the genetic algorithm optimization is completed, the current chromosome number sequence is taken and converted into an eigenvalue matrix by decoding. 4.根据权利要求1所述的一种基于遗传算法特征参数优化的瓷绝缘子损伤神经网络识别方法，其特征在于：4. a kind of porcelain insulator damage neural network identification method based on genetic algorithm characteristic parameter optimization according to claim 1, is characterized in that: 步骤一：实验系统搭建；Step 1: Build the experimental system; 搭建力锤-麦克风系统：将模态力锤与力传感器相连接，将麦克风与信号放大器相连接，并将放大器和力传感器分别接在HS3采集板卡的输入1接口和输入2接口，输出接口连接PC端，对力锤激励信号和麦克风的接收信号进行同步采集；Build the hammer-microphone system: connect the modal hammer to the force sensor, connect the microphone to the signal amplifier, and connect the amplifier and the force sensor to the input 1 interface and input 2 interface of the HS3 acquisition board respectively, and the output interface Connect to the PC terminal to synchronously collect the excitation signal of the hammer and the received signal of the microphone; 步骤二：声信号采集；Step 2: Acoustic signal acquisition; 使用力锤激励不同损伤状态的瓷绝缘子，并使用麦克风进行声信号的接收，采样频率选择为30kHz，使用HS3采集板卡对信号进行实时采集并存储，信号存储以分钟为单位，将每分钟的声信号存储在不同的excel表格中，实验选取3类不同的瓷绝缘子工件即完整、上缺陷和下缺陷，采集训练样本和测试样本两部分，其中训练样本将分为两部分进行参数优化，参数优化的过程是从训练样本中每次随机抽取多组信号作为优化测试样本，其余作为优化训练样本，并重复5次操作，得到5次平均识别率作为参数优化的量度；Use a force hammer to excite porcelain insulators with different damage states, and use a microphone 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 minutes. The acoustic signals are stored in different excel sheets. The experiment selects 3 different types of porcelain insulator workpieces, namely complete, upper and lower defects, and collects training samples and test samples. The training samples will be divided into two parts for parameter optimization. The optimization process is to randomly select multiple groups of signals from the training samples as the optimization test samples, and the rest are used as the optimization training samples, and repeat the operation 5 times to obtain the average recognition rate of 5 times as the measure of parameter optimization; 步骤三：声信号预处理；Step 3: Acoustic signal preprocessing; 导出步骤二中采集的声信号，基于幅值的信号截取方法将信号分为若干个声信号；The acoustic signal collected in step 2 is derived, and the signal is divided into several acoustic signals based on the amplitude-based signal interception method; 步骤四：声信号特征参数提取；Step 4: extraction of acoustic signal feature parameters; 提取各信号中的三类特征参数，分别为时域特征参数7个、形状特征参数7个和小波能量特征参数18个，并将提取的特征参数进行归一化处理；Extract three types of characteristic parameters in each signal, namely 7 time domain characteristic parameters, 7 shape characteristic parameters and 18 wavelet energy characteristic parameters, and normalize the extracted characteristic parameters; 步骤五：BP神经网络的参数优化；Step 5: Parameter optimization of BP neural network; 对采集的声信号训练样本和测试样本分别提取32维特征参数，组成训练特征值矩阵和测试特征值矩阵，将训练样本特征值矩阵直接作为输入，进行BP神经网络结构的优化；32-dimensional feature parameters are extracted from the collected acoustic signal training samples and test samples respectively to form a training eigenvalue matrix and a test eigenvalue matrix, and the training sample eigenvalue matrix is directly used as input to optimize the BP neural network structure; 优化输入层-隐含层传递函数和隐含层-输出层传递函数，选择神经网络识别效率最好的传递函数组合：“tansig-logsig”；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 recognition efficiency of the neural network: "tansig-logsig"; 改变网络学习率，研究学习率对神经网络识别效果的影响，并最终确定网络学习率参数为0.005；Change the network learning rate, study the influence of the learning rate on the recognition effect of the neural network, and finally determine the network learning rate parameter to be 0.005; 步骤六：特征参数的优选Step 6: Optimization of Feature Parameters 利用遗传算法对32维特征参数进行优选，从三类信号中分别选择20个信号，共60组信号作为训练样本，分为以下几个步骤：The genetic algorithm is used to optimize the 32-dimensional feature parameters, and 20 signals are selected from the three types of signals, and a total of 60 groups of signals are used as training samples, which are divided into the following steps: (1)设定遗传算法参数，迭代终止条件为300次，种群规模50，个体染色体长度与特征值维数相同，设置为32，交叉概率0.8，变异概率0.02；(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, set to 32, the crossover probability is 0.8, and the mutation probability is 0.02; (2)个体初始化编码，采用二进制编码，每一维特征参数皆有两种可能，参与神经网络训练则编码为“1”，不参与训练则编码为“0”；(2) Individual initialization coding adopts binary coding, and each dimension feature parameter has two possibilities. If participating in neural network training, it will be coded as "1", and if not participating in training, it will be coded as "0"; (3)个体解码，生成各染色体对应的特征参数矩阵，分别输入BP神经网络进行训练，并将测试信号输入，得到预测输出与实际输出的均方差值，取均方差值的相反数并归一化，作为遗传算法的适应度函数，由于神经网络训练存在偶然性，所以均方差值为10次重复训练的结果；(3) Individual decoding, generating the characteristic parameter matrix corresponding to each chromosome, respectively inputting the BP neural network for training, and inputting the test signal to obtain the mean square error value of the predicted output and the actual output, taking the opposite of the mean square error value and calculating Normalization, as the fitness function of the genetic algorithm, due to the contingency in the training of the neural network, the mean square error is the result of 10 repeated trainings; (4)保留每一代中适应度最大的个体，并对种群做自然选择、交叉和变异操作，得到新的种群；(4) Retain the individual with the greatest fitness in each generation, and perform natural selection, crossover and mutation operations on the population to obtain a new population; (5)若未达到终止条件，则反复(3)、(4)步操作，若达到终止条件，则选择最终结果进行染色体解码；(5) If the termination condition is not reached, repeat the operations of (3) and (4), if the termination condition is reached, select the final result to perform chromosome decoding; 步骤七：BP神经网络的迭代训练Step 7: Iterative training of BP neural network (1)将训练样本输入到神经网络中，并设置神经网络的初始值参数，其中输入层设置为16，输出层设置为3，隐含层神经元根据公式设置为11，隐含层传递函数选择tansig传递函数，输出层传递函数选择logsig传递函数，训练算法选择梯度下降自适应学习训练函数，设定最大训练次数为5000次，训练精度为0.001，学习率为0.005；(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 neurons of the hidden layer are set to 11 according to the formula, and the transfer function of the hidden layer is set to 11. Select the tansig transfer function, the output layer transfer function select the logsig transfer function, the training algorithm select the gradient descent adaptive learning training function, set the maximum training times to 5000 times, the training accuracy to 0.001, and the learning rate to 0.005; (2)对特征值输入的神经网络进行训练，确定最终的权值与阈值矩阵，输出出最佳的神经网络结构参数。(2) Train the neural network input with the eigenvalues, determine the final weight and threshold matrix, and output the optimal neural network structure parameters.
CN202110617866.7A 2021-06-01 2021-06-01 A Neural Network Identification Method for Porcelain Insulator Damage Based on Genetic Algorithm Feature Parameter Optimization Active CN113361383B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110617866.7A CN113361383B (en) 2021-06-01 2021-06-01 A Neural Network Identification Method for Porcelain Insulator Damage Based on Genetic Algorithm Feature Parameter Optimization

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110617866.7A CN113361383B (en) 2021-06-01 2021-06-01 A Neural Network Identification Method for Porcelain Insulator Damage Based on Genetic Algorithm Feature Parameter Optimization

Publications (2)

Publication Number Publication Date
CN113361383A true CN113361383A (en) 2021-09-07
CN113361383B CN113361383B (en) 2024-12-13

Family

ID=77531604

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110617866.7A Active CN113361383B (en) 2021-06-01 2021-06-01 A Neural Network Identification Method for Porcelain Insulator Damage Based on Genetic Algorithm Feature Parameter Optimization

Country Status (1)

Country Link
CN (1) CN113361383B (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115659136A (en) * 2022-12-09 2023-01-31 中铁第一勘察设计院集团有限公司 Wireless interference signal waveform identification method based on neural network
CN115762685A (en) * 2022-12-02 2023-03-07 浙江大学 Sound-transmitting layer for reducing acoustic transmission loss of dissimilar material interface and optimization method

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108665095A (en) * 2018-04-27 2018-10-16 东华大学 Short term power prediction technique based on genetic algorithm optimization Elman neural networks
CN111259750A (en) * 2020-01-10 2020-06-09 西北工业大学 Underwater sound target identification method for optimizing BP neural network based on genetic algorithm
CN111598231A (en) * 2020-03-23 2020-08-28 天津职业技术师范大学(中国职业培训指导教师进修中心) Prediction of cable joint conductor temperature based on GA-optimized serial BP network
CN112307927A (en) * 2020-10-26 2021-02-02 重庆邮电大学 Research on recognition of MPSK signal in non-cooperative communication based on BP network

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108665095A (en) * 2018-04-27 2018-10-16 东华大学 Short term power prediction technique based on genetic algorithm optimization Elman neural networks
CN111259750A (en) * 2020-01-10 2020-06-09 西北工业大学 Underwater sound target identification method for optimizing BP neural network based on genetic algorithm
CN111598231A (en) * 2020-03-23 2020-08-28 天津职业技术师范大学(中国职业培训指导教师进修中心) Prediction of cable joint conductor temperature based on GA-optimized serial BP network
CN112307927A (en) * 2020-10-26 2021-02-02 重庆邮电大学 Research on recognition of MPSK signal in non-cooperative communication based on BP network

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115762685A (en) * 2022-12-02 2023-03-07 浙江大学 Sound-transmitting layer for reducing acoustic transmission loss of dissimilar material interface and optimization method
CN115762685B (en) * 2022-12-02 2023-07-28 浙江大学 A sound-permeable layer and its optimization method for reducing the acoustic transmission loss at the interface of dissimilar materials
CN115659136A (en) * 2022-12-09 2023-01-31 中铁第一勘察设计院集团有限公司 Wireless interference signal waveform identification method based on neural network

Also Published As

Publication number Publication date
CN113361383B (en) 2024-12-13

Similar Documents

Publication Publication Date Title
CN103077402B (en) Partial discharge of transformer mode identification method based on singular value decomposition algorithm
CN111160520A (en) BP neural network wind speed prediction method based on genetic algorithm optimization
CN107144643B (en) A kind of damnification recognition method of Lamb wave monitoring signals statistical parameter
CN102809718A (en) Ultra-high-frequency partial discharge signal identification method for gas insulated switchgear (GIS)
CN109142514B (en) Defect detection device and method based on pulse eddy current array
CN113361383B (en) A Neural Network Identification Method for Porcelain Insulator Damage Based on Genetic Algorithm Feature Parameter Optimization
CN103018338A (en) A Nondestructive Testing Method for Concrete Based on Acoustic Emission and Neural Network
CN111916089B (en) Hail detection method and device based on acoustic signal feature analysis
CN110555247A (en) structure damage early warning method based on multipoint sensor data and BilSTM
CN115165164B (en) Concrete compressive stress identification method and system based on deep learning admittance characteristics
CN112114047B (en) GAs-liquid flow parameter detection method based on acoustic emission-GA-BP neural network
CN107121501A (en) A kind of turbine rotor defect classification method
CN114137079A (en) Ultrasonic guided wave nondestructive testing method based on combination of deep learning and Duffing system
CN114487129A (en) Damage identification method for flexible materials based on acoustic emission technology
CN110568082A (en) A Discrimination Method for Cable Broken Wire Based on Acoustic Emission Signal
CN110688983A (en) Microseismic signal identification method based on multi-mode optimization and ensemble learning
Yang et al. Crack detection in carbide anvil using acoustic signal and deep learning with particle swarm optimisation
CN117635461B (en) A modeling method for predicting the firmness of vegetables and fruits
CN114595733B (en) Bridge cable broken wire signal recognition method and system based on long short-term memory network
CN115452957B (en) Small sample metal damage identification method based on attention prototype network
CN111272875A (en) A non-destructive testing method for apple brittleness based on vibration and sound signals
CN112651537A (en) Photovoltaic power generation ultra-short term power prediction method and system
CN114217179A (en) High-voltage switch cabinet partial discharge monitoring method
CN112528849A (en) Structure health monitoring method based on inner product matrix and deep learning
CN115586254B (en) Method and system for identifying metal material based on convolutional neural network

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant