CN114487129A - Flexible material damage identification method based on acoustic emission technology - Google Patents

Abstract

The invention relates to a damage identification method of a flexible material based on an acoustic emission technology, which comprises the steps of obtaining an acoustic emission signal of a flexible composite material to be detected; acquiring damage characteristic information of the acoustic emission signal, extracting characteristic parameters of the acoustic emission signal, and constructing characteristic vectors in different damage states; establishing a sample data set, randomly dividing a training set and a test set, and carrying out data normalization processing; and inputting the randomly divided sample data of the training set into an optimized extreme learning machine to obtain classified output results of different damage states. According to the method, the damage characteristic information of the acoustic emission signal is acquired in three levels of a time domain, a frequency domain and a transformation domain, the characteristic parameters of the damage signal are extracted by using principal component analysis dimensionality reduction processing, the damage characteristic vector is determined, the acoustic emission technology is combined with a limit learning machine algorithm based on particle swarm optimization, the hidden layer node of the limit learning machine is optimized, and effective and accurate damage identification can be carried out on the flexible material.

Description

Flexible material damage identification method based on acoustic emission technology

Technical Field

The invention belongs to the technical field of nondestructive testing, and particularly relates to a damage identification method of a flexible substrate based on an acoustic emission technology.

Background

The flexible base material has excellent performance characteristics, so the flexible base material is widely applied, however, the manufacturing process of the flexible base material is very complicated and is difficult to achieve precise control, so that the material quality is unstable, the flexible base material has certain randomness, and more or less internal defects such as pores, inclusions, cracks, looseness, poor bonding of fibers and a matrix interface, abrasion and the like exist. The creation, accumulation and propagation of cracks within the substrate exacerbate the dramatic loss of material strength and stiffness, greatly reducing the useful life of the structure, and sometimes potentially leading to catastrophic results. Moreover, the flexible base material is inevitably influenced by external action in engineering application, so that the damage degree of the flexible base material is timely and accurately monitored and evaluated, and measures are timely taken to ensure safety, and the method has important significance for ensuring that production and life of people are smoothly carried out.

At present, the acoustic emission technology is still one of the main methods for detecting damage of a related composite material, as a dynamic nondestructive detection technology, a received acoustic emission signal can directly reflect the induction, generation and development conditions of internal defects of a material or a component, and the acoustic emission technology has the advantages of real time, dynamic property, high sensitivity and the like, and is widely applied to monitoring the damage state of a composite material structure. Since the flexible substrate is a composite material having a rather complicated microstructure, there are various types of damage to the microstructure.

In the damage process, the damage characteristic of the composite material is the result of superposition of multiple damage sources, the propagation of acoustic emission waves is complex, the distortion of received acoustic emission signals is serious, and the waveforms of the acoustic emission signals are complex and have multiple interference sources, so that a more accurate signal processing algorithm and a more accurate signal classification algorithm are needed in the damage identification process. The most important thing in signal processing is to remove noise in the signal, extract characteristic parameters which are specific to the type of the detected defect, and obtain effective and accurate results through analyzing and classifying the characteristic parameters.

Disclosure of Invention

The invention aims to solve the technical problem of providing a damage identification method of a flexible base material based on an acoustic emission technology, which realizes damage identification of different damages and improves the accuracy of the damage identification by extracting damage characteristic parameters of an acoustic emission signal.

In order to solve the technical problems, the invention adopts the technical scheme that: a damage identification method of a flexible material based on an acoustic emission technology comprises the following steps:

step (1), acquiring an acoustic emission signal of a flexible substrate to be detected;

step (2), acquiring damage characteristic information of the acoustic emission signals, extracting characteristic parameters of the acoustic emission signals, and constructing characteristic vectors in different damage states;

step (3), establishing a sample data set, randomly dividing a training set and a test set, and carrying out data normalization processing;

step (4), optimizing the number of nodes of a hidden layer of the extreme learning machine by using a particle swarm algorithm;

and (5) inputting the randomly divided sample data of the training set into an optimized extreme learning machine to obtain classified output results of different damage states.

Further, the step (2) specifically includes the following steps:

s201, preparing flexible substrate samples in different damage states to be detected;

s202, acquiring acoustic emission signals generated by the flexible composite material under different damage states to serve as damage sample signals;

s203, extracting damage characteristic information of the damage sample signal to construct characteristic quantities in different damage states;

and S204, performing principal component analysis and dimension reduction processing on the characteristic quantities under different damage states through a data preprocessing process of normalization and regularization, and determining the characteristic vectors under different damage states.

Further, in step S203, the extracting of the damage characteristic information mainly includes time domain extracting, frequency domain extracting and transform domain extracting, and the signal self-characteristics with a large degree of discrimination, the FFT spectrum and the wavelet decomposition coefficients are selected to extract characteristic values, so as to construct a characteristic vector.

Further, step S203 specifically includes the following steps:

step a, after preprocessing of wavelet denoising and the like, extracting damage characteristic information is determined from two aspects, namely extracting characteristics including time domain peak values and crest coefficients of time domain of data on one hand; another aspect is features in two dimensions of the graph, including root mean square values and variances; respectively performing wavelet decomposition on the three types of data to extract characteristic parameters, and selecting a time domain peak value, a crest coefficient, a root mean square value and a variance as a first part of a characteristic vector;

step b, performing FFT transformation on the waveform data subjected to wavelet denoising, drawing an FFT spectrum of a damaged sample signal, screening out common characteristics under each defect condition, and selecting a spectrum curve peak value and a frequency domain peak coefficient as a second part of the characteristic vector determined by the common characteristics;

and c, performing wavelet decomposition on the acoustic emission signal data in different damage states by determining the optimal wavelet basis function and the number of decomposition layers, and then extracting corresponding wavelet coefficient energy characteristics and wavelet mean square error characteristics as a third part of the characteristic vector.

Further, in step S204, the principal component analysis and dimension reduction processing is principal component analysis performed on the acoustic emission signal characteristic parameters, and determines the damage characteristic vectors in different damage states, which specifically includes the following steps:

according to the principle that the mean square error of the acoustic emission signal characteristic parameters is minimized, observed by principal component analysis, the number of principal components in the acoustic emission characteristic parameters is determined, the main characteristic dimensions of the characteristic quantities under different damage states are extracted, the parameter dimension reduction of the characteristic quantities under different damage states is realized, and the characteristic vectors belonging to different damage states are constructed.

Further, the step (4) specifically includes the following steps:

s401, constructionELM model, initializing ELM model parameters, including hidden layer weightswAnd biasbObtaining the output layer weight by calculationβCompleting the construction of an ELM model to obtain a trained ELM model;

s402, inputting the training set and the test set obtained in the step (3) into an extreme learning machine, initializing parameters of a particle swarm algorithm, training and optimizing an ELM model, taking an output result of sample data of the test set as a difference value between an actual value and a predicted value output by the ELM to obtain a test error, wherein the percentage of the test error to the actual value is an ELM error rate, and taking the test error as fitness;

s403, obtaining an individual extreme value and a global extreme value through calculation, and continuously updating the position and the speed of the particles to realize the adjustment and optimization of the weight and the bias of the hidden layer of the extreme learning machine;

and S404, repeating the step S403 to continue training and learning, searching for a global optimal solution, updating the individual extreme value and the global extreme value by calculating the particle fitness until the ELM error rate is less than or equal to a set value and meeting an end condition to obtain the ELM optimal hidden layer node.

The acoustic emission signal acquired in the flexible substrate damage process is analyzed and processed in three angles of a time domain, a frequency domain and a transform domain by adopting a wavelet decomposition method, and the acoustic emission signal has good time-frequency local characteristics. The method comprises the steps of establishing an ELM-based damage identification model for training and testing, optimizing hidden layer nodes through a particle swarm algorithm, and realizing classification of different acoustic emission signals, so that the effectiveness of the flexible substrate damage identification method is verified. The method improves the accuracy of damage detection, more carefully identifies the damage of the flexible composite material, and more reliably evaluates the damage degree.

Compared with the prior art, the invention has the following advantages and effects:

the invention relates to a damage identification method of a flexible composite material based on an acoustic emission technology, which optimizes hidden layer nodes of an extreme learning machine by using an acoustic emission detection technology and the extreme learning machine and utilizing a particle swarm algorithm, improves the uncertainty of randomly acquired parameters of the extreme learning machine, and improves the accuracy of model classification.

The detection method provided by the invention can amplify local characteristics of the signal analysis signal during signal processing, has good time-frequency local characteristics, can obtain more signal characteristics than the traditional analysis method, is used for reducing the damage condition of the flexible composite material, and has small positioning error and high reduction precision.

According to the method, on the basis of extracting the characteristic vector by combining an acoustic emission means and a limit learning machine with a particle swarm optimization algorithm, after the acoustic emission signals of different damage states are subjected to wavelet threshold denoising pretreatment, the peak value, the crest coefficient, the energy of the wavelet coefficient, the mean square difference value and the like of each component are extracted, the characteristic vector of the damage signal is constructed from 3 levels of a time domain, a frequency domain and a transform domain, and the method is accurate and effective.

Drawings

FIG. 1 is a diagram of the method implementation of the present invention;

FIG. 2 is a schematic view of the acoustic emission detection principle of the present invention;

FIG. 3 is a schematic diagram of feature vector extraction according to the present invention;

FIG. 4 is a flow chart of a particle swarm optimization extreme learning machine;

FIG. 5 shows the ELM classification results of single random samples.

Detailed Description

The method for identifying the damage of the flexible material based on the acoustic emission technology mainly identifies different damages by using an extreme learning machine algorithm optimized by particle swarm through acoustic emission detection, and then analyzes and discusses a classification result. The extreme learning machine is a neural network algorithm with simple parameter setting and wide application, the weight between an input layer and a hidden layer and a threshold between neurons of the hidden layer are randomly given by the algorithm, and only the number of nodes of the hidden layer is required to be set to obtain a unique optimal solution without adjustment in the training process. However, the hidden layer parameters randomly generated by the extreme learning machine algorithm cause poor network generalization performance, and in order to improve the prediction accuracy, the number of hidden layer nodes needs to be increased. And too many hidden layer nodes increase the complexity of the network and easily cause the problem of overfitting. The Particle Swarm Optimization (PSO) is an optimization tool with global optimization capability based on swarm intelligence, swarm intelligence generated by cooperation and competition among the swarm guides optimization search, solutions in each generation of swarm have the advantages of self learning improvement and learning to other people, and can obtain better solutions, but local search capability of the solutions is weak. Aiming at the problems of the two algorithms, the particle swarm optimization algorithm is introduced into the extreme learning machine so as to achieve the purposes of global optimization and rapid convergence.

As shown in fig. 1, an exemplary embodiment of the present invention provides a damage identification method for a flexible material based on an acoustic emission technology, including the following steps:

and (1) acquiring an acoustic emission signal of the flexible substrate to be detected.

The acoustic emission detection principle is as shown in fig. 2, an acoustic emission sensor collects an acoustic emission signal of the flexible composite material, and the acoustic emission signal is processed by a signal amplifier, a signal conditioner and a signal collection processing system and then is input into a computer.

Step (2), acquiring damage characteristic information of the acoustic emission signal, extracting characteristic parameters of the acoustic emission signal, and constructing the acoustic emission signal Feature vectors in the same damage state.

Relatively specifically, the step (2) includes the steps of:

s201, preparing flexible base material samples in different damage states to be detected, wherein the flexible base material samples mainly comprise acoustic emission signals of a flexible substrate under the three conditions of no damage, holes and groove-shaped damage. Wherein the sizes of the two lesions are: size of hole damage:φ2mm × 1mm, trough lesion size: 4 mm. times.0.5 mm.

S202, collecting acoustic emission signals generated by the flexible composite material in different damage states as damage sample signals.

And S203, extracting damage characteristic information of the damage sample signal to construct characteristic quantities in different damage states.

And S204, performing principal component analysis and dimension reduction processing on the characteristic quantities under different damage states through a data preprocessing process of normalization and regularization, and determining the characteristic vectors under different damage states.

In step S203, the extraction of the acoustic emission characteristic parameters, that is, the damage characteristic information, mainly includes time domain extraction, frequency domain extraction, and transform domain extraction, as shown in fig. 3.

Considering that an acoustic emission signal generated in the damage process of the flexible composite material to be detected is a transient signal in a non-stationary state, a wavelet transformation means capable of performing multi-scale analysis is adopted for extracting a transformation domain, and the signal self characteristic with high discrimination, an FFT (fast Fourier transform) frequency spectrum and a wavelet decomposition coefficient are selected to extract a characteristic value so as to construct a characteristic vector, and the method specifically comprises the following steps:

step a, after the acquired acoustic emission signals are preprocessed through wavelet denoising and the like, the extraction of damage characteristic information can be determined from two aspects, namely, on one hand, the extraction of the time domain extraction characteristics of the data, including the time domain peak valuev1 and crest factorv2, etc. as the first and second feature amounts, respectively. Another aspect is a feature in two dimensions of a graph, including a root mean square valuev3 sum of variancev4, etc. as the third and fourth class characteristic quantities. Respectively carrying out wavelet decomposition on the three types of data to complete the extraction of characteristic parameters, and selecting time domain peak valuesv1. Crest factorv2 and root mean square valuev3 sum of variancev4 as the first part of the feature vector, as shown in table 1.

TABLE 1 time domain extraction of feature quantities

Step b, the screening process of extracting characteristic values from wavelet decomposition components is complex, and the data calculation amount is huge, so that the waveform data subjected to wavelet denoising is subjected to FFT (fast Fourier transform) conversion, the FFT frequency spectrum of the damaged sample signal is drawn, the common characteristic under each defect condition is screened out, and the peak value of a frequency spectrum curve is selectedv5 sum frequency crest factorvAnd 6 are respectively used as the fifth and sixth class feature quantities, and the second part of the feature vector is determined according to the fifth and sixth class feature quantities, as shown in table 2.

TABLE 2 FFT spectral feature quantity

Step c, completing wavelet denoising quality evaluation on the damage signal by adopting a composite evaluation index fusing multi-class characteristics, and determining the optimal wavelet basis function asdb5 wavelet bases and 3 decomposition layers corresponding to the wavelet bases are adopted for acoustic emission signal data in different damage statesdb5 wavelet basis is used for 3-layer wavelet decomposition, and then corresponding wavelet coefficient energy characteristics and wavelet mean square error characteristics are extracted to be used as seventh to fourteenth class characteristic quantities (v7~v14) And this is used as the third part of the feature vector, as shown in tables 3 and 4.

TABLE 3 wavelet coefficient energies: (v7-v10) Characteristic amount

Table 4 wavelet mean square error (v11-v14) Characteristic amount

The method for evaluating the wavelet denoising quality of the damage signal by fusing the composite evaluation index of the multi-class characteristics specifically comprises the following steps:

1) an alternative parameter is determined. Selecting the number of decomposition layers to be 2-8, and selecting wavelet basis functionsdb5~db13 and use ofrigrsureA threshold criterion and an adaptive threshold method based on Stein's unbiased risk estimation.

2) Calculating the root mean square error under each parameter condition (RMSE) And smoothness (r) According to formulas (1) and (2), the normalized root mean square error and the weight of smoothness are obtained;

wherein the content of the first and second substances,PRMSEis normalizedRoot mean square error;Pris the normalized smoothness;

is composed ofPRMSEStandard deviation of (d);

is composed ofPRMSEThe mean value of (a);

is composed ofPrStandard deviation of (d);

is composed ofPrThe mean value of (a);

is the weight of the normalized root mean square error;

is the weight of the normalized smoothness.

3) Calculating a composite index according to a formula (3) to obtainTAnd when the value is minimum, the corresponding optimal candidate parameter is obtained. The smaller the composite evaluation index is, the better the denoising effect on the noise-containing signal is. And when the composite index takes the minimum value, determining the optimal decomposition layer number and the wavelet basis function.

In the above steps, the principal component analysis and dimension reduction processing in step S204 is principal component analysis performed on the acoustic emission signal characteristic parameters, and determines the damage characteristic vectors in different damage states, specifically including the following steps:

determining the number of principal components in the acoustic emission characteristic parameter according to the principle of minimizing the mean square error of the acoustic emission signal characteristic parameter observed by principal component analysis, extracting the main characteristic dimensions of the characteristic quantity under different damage states, realizing the dimension reduction of the parameter of the characteristic quantity under different damage states, and reducing the characteristic quantity in the step S203Dimension processing to obtain damage characteristic vector

Is shown as

。

And (3) establishing a sample data set, randomly dividing the training set and the test set, and carrying out data normalization processing.

Randomly selected 150 sets of feature vectors as the classified training and prediction data set, with 123 sets of feature vectors as training samples and the remaining 27 sets as test samples.

And (4) optimizing the number of nodes of a hidden layer of the extreme learning machine by using a particle swarm algorithm.

As shown in fig. 4, the method specifically includes the following steps:

s401, constructing an ELM model, and initializing ELM model parameters including hidden layer weightwAnd biasbMaking an activation function of the required ELM, inputting the sample data of the training set into an ELM model, wherein the activation function adoptsSigmoidFunction of obtaining output layer weights by calculationβCompleting the construction of an ELM model to obtain a trained ELM model;

the ELM model is a three-layer neural network structure and comprises 1 input layer, 1 hidden layer and 1 output layer. Wherein, the input layer comprises 14 neuron nodes, and the output layer comprises 3 neuron nodes.

S402, inputting the training set and the test set obtained in the step (3) into an extreme learning machine, initializing parameters of a particle swarm algorithm, training and optimizing an ELM model, taking an output result of sample data of the test set as a difference value between an actual value and a predicted value output by the ELM to obtain a test error, wherein the percentage of the test error to the actual value is an ELM error rate, and taking the test error as fitness;

s403, obtaining an individual extreme value and a global extreme value through calculation, and continuously updating the position and the speed of the particles to realize the adjustment and optimization of the weight and the bias of the hidden layer of the extreme learning machine;

s404, repeating the step S403 to continue training and learning, searching for a global optimal solution, updating the individual extreme value and the global extreme value by calculating the particle fitness, ending the circulation when the ELM error rate is less than or equal to the set value of 0.11, and obtaining the ELM optimal hidden layer node of 23, wherein the recognition effect is the best when the classification accuracy is 92.6%, as shown in Table 5.

TABLE 5 Classification accuracy of different hidden layer node numbers

Step (5), inputting the randomly divided sample data of the training set into an optimized extreme learning machine to obtain different training sets And (5) classifying the damage state and outputting a result.

Inputting the randomly divided sample data of the training set into an optimized extreme learning machine, setting labels for three acoustic emission signals, setting the identification category of a non-damage signal to be 0, the identification category of a hole damage signal to be 1, and the identification category of a groove type damage signal to be 2, obtaining classification output results of different damage states, wherein the ELM classification result of a single random sample is shown in FIG. 5. The classification accuracy rate is the percentage of the correct classification result of the test samples to the number of the test samples, and the ELM model obtained through calculation has a high classification accuracy rate of 92.59%.

The above description is only of the preferred embodiments of the present invention, and it should be noted that: it will be apparent to those skilled in the art that various modifications and adaptations can be made without departing from the principles of the invention, and such modifications and adaptations are intended to be within the scope of the invention.

Claims (6)

1. A damage identification method of a flexible material based on an acoustic emission technology is characterized by comprising the following steps:

step (1), acquiring an acoustic emission signal of a flexible substrate to be detected;

step (2), acquiring damage characteristic information of the acoustic emission signals, extracting characteristic parameters of the acoustic emission signals, and constructing characteristic vectors in different damage states;

step (3), establishing a sample data set, randomly dividing a training set and a test set, and carrying out data normalization processing;

step (4), optimizing the number of nodes of a hidden layer of the extreme learning machine by using a particle swarm algorithm;

and (5) inputting the randomly divided sample data of the training set into an optimized extreme learning machine to obtain classified output results of different damage states.

2. The acoustic emission technology-based damage identification method for a flexible material according to claim 1, wherein: the step (2) specifically comprises the following steps:

s201, preparing flexible substrate samples in different damage states to be detected;

s202, acquiring acoustic emission signals generated by the flexible composite material under different damage states to serve as damage sample signals;

s203, extracting damage characteristic information of the damage sample signal to construct characteristic quantities in different damage states;

and S204, performing principal component analysis and dimension reduction processing on the characteristic quantities under different damage states through a data preprocessing process of normalization and regularization, and determining the characteristic vectors under different damage states.

3. The acoustic emission technology-based damage identification method for a flexible material according to claim 2, wherein: in step S203, the extraction of the damage characteristic information mainly includes time domain extraction, frequency domain extraction, and transform domain extraction, and selects the signal characteristics with a large degree of discrimination, the FFT spectrum, and the wavelet decomposition coefficients to extract characteristic values, and constructs a characteristic vector.

4. The acoustic emission technology-based damage identification method for a flexible material according to claim 3, wherein: step S203 specifically includes the following steps:

step a, after preprocessing of wavelet denoising and the like, extracting damage characteristic information is determined from two aspects, namely extracting characteristics including time domain peak values and crest coefficients of time domain of data on one hand; another aspect is features in two dimensions of the graph, including root mean square values and variances; respectively performing wavelet decomposition on the three types of data to extract characteristic parameters, and selecting a time domain peak value, a crest coefficient, a root mean square value and a variance as a first part of a characteristic vector;

step b, performing FFT transformation on the waveform data subjected to wavelet denoising, drawing an FFT spectrum of a damaged sample signal, screening out common characteristics under each defect condition, and selecting a spectrum curve peak value and a frequency domain peak coefficient as a second part of the characteristic vector determined by the common characteristics;

and c, performing wavelet decomposition on the acoustic emission signal data in different damage states by determining the optimal wavelet basis function and the number of decomposition layers, and then extracting corresponding wavelet coefficient energy characteristics and wavelet mean square error characteristics as a third part of the characteristic vector.

5. Method for the identification of damage to a flexible material based on acoustic emission technology according to claim 2, 3 or 4, characterized in that: in step S204, the principal component analysis dimension reduction processing is principal component analysis performed on the acoustic emission signal characteristic parameters, and determines the damage characteristic vectors in different damage states, specifically including the following steps:

according to the principle that the mean square error of the acoustic emission signal characteristic parameters is minimized, observed by principal component analysis, the number of principal components in the acoustic emission characteristic parameters is determined, the main characteristic dimensions of the characteristic quantities under different damage states are extracted, the parameter dimension reduction of the characteristic quantities under different damage states is realized, and the characteristic vectors belonging to different damage states are constructed.

6. The acoustic emission technology-based damage identification method for a flexible material according to claim 1, wherein the step (4) comprises the following steps:

s401, constructing an ELM model, and initializing ELM model parameters including hidden layer weightwAnd biasbObtaining the output layer weight by calculationβCompleting the construction of an ELM model to obtain a trained ELM model;

s402, inputting the training set and the test set obtained in the step (3) into an extreme learning machine, initializing parameters of a particle swarm algorithm, training and optimizing an ELM model, taking an output result of sample data of the test set as a difference value between an actual value and a predicted value output by the ELM to obtain a test error, wherein the percentage of the test error to the actual value is an ELM error rate, and taking the test error as fitness;

s403, obtaining an individual extreme value and a global extreme value through calculation, and continuously updating the position and the speed of the particles to realize the adjustment and optimization of the weight and the bias of the hidden layer of the extreme learning machine;

and S404, repeating the step S403 to continue training and learning, searching for a global optimal solution, updating the individual extreme value and the global extreme value by calculating the particle fitness until the ELM error rate is less than or equal to a set value and meeting an end condition to obtain the ELM optimal hidden layer node.

Images