CN115846852A - FSW welding monitoring method based on DS evidence theory and neural network - Google Patents

Abstract

The invention discloses a FSW welding monitoring method based on DS evidence theory and neural network, comprising the following steps: S1: establishing an acoustic emission acquisition system to collect acoustic emission signals; S2: extracting and analyzing the acoustic emission signals, and making a data set; S3: Establish a backbone neural network model; S4: Establish a multi-task classification network structure; S5: Combine DS evidence theory to establish a multi-task learning discriminant index; S6: Iterate the backbone neural network and multi-task classification network, if the evaluation result reaches the optimal If the condition is optimal or the maximum number of iterations is reached, the training is stopped and the weight of the multi-task classification network is saved to complete the identification and classification of the FSW acoustic emission welding defect signal. The combination of layers solves the shortcomings of multiple single tasks that consume training costs, and identifies and analyzes defect parts, achieving the purpose of accurately locating defect locations and defect types, and improving recognition efficiency.

Description

A FSW Welding Monitoring Method Based on DS Evidence Theory and Neural Network

technical field

The invention relates to the technical field of friction stir welding, in particular to a FSW micro-welding monitoring method based on DS evidence theory and neural network.

Background technique

Friction stir welding (FSW) is the most widely used welding method in the rail transit equipment manufacturing industry. Welding defects are potential damage methods for workpiece fatigue fracture. Fatigue fractures caused by welding defects are highly concealed, and once they occur, they will lead to catastrophic accidents and cause serious economic losses. Acoustic emission testing is a new method of dynamic non-destructive testing based on the internal structure feedback of the workpiece, which can directly detect and judge the internal defects of the workpiece, and reflect the welding status of FSW welding plates in real time. Therefore, analyzing the acoustic emission signal manifestations of defects during welding, and accurately identifying and locating defect parts is of great significance to ensure the safe operation of rail vehicles.

In signal analysis and feature extraction, most of them use one method for feature extraction, and use a variety of neural networks to identify and classify after feature extraction; in the existing signal analysis methods, a single signal analysis method only analyzes the signal from a certain angle. After the analysis, even if the analysis effect is obvious, it is also one-sided; in the mainstream classification neural network, the SoftMax layer uses the maximum membership function for classification. This method is easy to cause the problem of loss of classification results. When the model is built, the application of multiple backbone networks increases the training cost, affects the recognition efficiency, and consumes training costs.

Contents of the invention

In order to solve the problems of low efficiency and high cost of existing signal feature extraction and identification, the present invention provides a FSW welding monitoring method based on DS evidence theory and neural network.

A kind of FSW welding monitoring method based on DS evidence theory and neural network of the present invention comprises the following steps:

S1: Establish an acoustic emission acquisition system to collect acoustic emission signals;

S2: extract and analyze the acoustic emission signal, and make a data set;

S3: Input the data set in S2, and establish the backbone neural network model;

S4: Input the results in S3, and establish a multi-task classification network structure;

S5: Combined with DS evidence theory, establish multi-task learning discriminant indicators;

S6: Iterate the backbone neural network and multi-task classification network. If the evaluation result reaches the optimal condition or reaches the maximum number of iterations, stop the training and save the weight of the multi-task classification network to complete the identification and classification of FSW acoustic emission welding defect signals .

Preferably, in said S1, an acoustic emission sensor of model R15A, a collection card of model PCIE-1816H, and a sampling rate of 100 kHz are used to collect acoustic signals.

Preferably, said S2 includes the following steps:

S2-1: Use the Mel spectrum to analyze the filtered data data, the expression is as follows:

mel=f _mel (data)

In the formula: mel is the eigenvector after the Mel spectrum transformation, f _mel is the Mel frequency, and the expression of the Mel frequency curve is as follows:

Where: f is the original frequency.

S2-2: Use short-time Fourier transform to analyze the filtered data data, the expression is as follows:

spec=F(data)

In the formula: spec is the feature vector after the short-time Fourier transform, and the definition expression of the short-time Fourier transform is as follows:

In the formula: t is the time domain, ω is the frequency domain, i is the imaginary number unit, and e is the base of the natural logarithm;

S2-3: Use wavelet transform to analyze the filtered data data, the expression is as follows:

cwt=cwt(data)

Among them, the wavelet transform definition expression of f(t) is as follows:

In the formula: α is the scale factor, τ is the time shift factor, ψ(t) is the wavelet basis;

S2-4: 32 eigenvalues were extracted from S2-1 to S2-3 respectively, a total of 96 eigenvalues were extracted, and a data set was made with a Hamming window of 30 ms as a unit.

Preferably, the backbone neural network model in S3 includes: the first layer is a one-dimensional convolutional layer with an output dimension of 2900*100; the second layer is a one-dimensional convolutional layer with an output dimension of 2800*100; the third The layer is a pooling layer with an output dimension of 1400*100; the fourth layer is a one-dimensional convolutional layer with an output dimension of 1300*160; the fifth layer is a one-dimensional convolutional layer with an output dimension of 1200*160; the sixth layer It is a pooling layer with an output dimension of 600*160; the seventh layer is a dropout layer with an output dimension of 600*160; the eighth layer is a fully connected layer with an output dimension of 300*160.

Preferably, the pooling layer of the third layer is a maximum pooling layer, and the pooling layer of the sixth layer is an average pooling layer,

Preferably, in said S4, the multi-task classification network structure includes: the first layer is a fully connected layer with an output dimension of 200*160; the second layer is a one-dimensional convolutional layer with an output dimension of 100*80; the third layer It is a one-dimensional convolutional layer with an output dimension of 50*20; the fourth layer is a pooling layer with an output dimension of 25*1; the fifth layer is a fine-tuned SoftMax layer with an output dimension of 5*1.

Preferably, the first fully connected layer in S4 is connected to the backbone neural network model in S3 and the time-frequency classification layer, and the input dimension is the output dimension of the fully connected layer in S3.

Preferably, the fifth fine-tuned SoftMax layer in S4 divides the FSW welding process into five categories: noise area, stirring head pressing area, stirring head lifting area, normal welding area and defect area, and the SoftMax classification principle is expressed The formula is as follows:

In the formula: P is the probability; e is the natural logarithm; v is the output vector; v _i is the value of the i-th category in v; k is the number of multiple outputs or categories of the neural network; v _j is the j-th in v The value of the category, i indicates the category that needs to be calculated currently, the calculation result is between 0 and 1, and the sum of the SoftMax values of all categories is 1.

Preferably, in said S5, the definition expression of DS evidence theory is as follows:

DS_result=DS(result _spec , result _mel , result _cwt )

In the formula: DS_result is the synthesis result of DS evidence theory; result _spec is the probability distribution of short-time Fourier feature vector identified by neural network; result _mel is the probability distribution of Mel spectral feature vector identified by neural network; result _cwt is wavelet Transform the probability distribution of the feature vector identified by the neural network;

By calculating the cross-entropy function between the output of DS evidence theory and the true value label, the expression of the multi-task learning discriminant index is as follows:

In the formula: H(p, q) is the loss function value of the cross-entropy function; p is the truth label; q is the synthesis result of DS evidence theory; x _i is each probability corresponding to the probability vector; n represents the probability vector corresponding to the welding category .

Preferably, in S6, the evaluation result reaching the optimal condition means that the difference between the cross-entropy functions obtained by two iterations in S5 is less than 0.001, and the maximum number of iterations is 1000. If one of the conditions is met, the training is stopped and the result is saved.

A FSW welding monitoring method based on DS evidence theory and neural network of the present invention takes FSW acoustic emission signal as the research object, conducts defect welding test, constructs a backbone neural network model data set with three time-frequency domain analysis algorithms, and combines the three The monitoring results of this time-frequency method are regarded as multi-task and a multi-task classification network structure is constructed. Combined with DS evidence theory, the SoftMax layer and cross-entropy loss function constitute a multi-task neural network evaluation index, which makes the backbone neural network and multi-task classification network converge quickly. Complete the identification and analysis of FSW welding signals to achieve the optimal FSW monitoring model.

Description of drawings

Fig. 1 is a kind of flow chart of the FSW welding monitoring method based on DS evidence theory and neural network of the present invention;

Fig. 2 is a structural diagram of the backbone neural network model of the present invention;

Fig. 3 is a structure diagram of the multi-task classification network of the present invention.

Detailed ways

A kind of neural network-based stochastic resonance FSW tiny feature defect feature extraction method of the present invention, shown in Figure 1, concrete steps are as follows:

S1: Establish an acoustic emission acquisition system to collect acoustic emission signals;

S2: extract and analyze the acoustic emission signal, and make a data set;

S3: Input the data set in S2, and establish the backbone neural network model;

S4: Input the results in S3, and establish a multi-task classification network structure;

S5: Combined with DS evidence theory, establish multi-task learning discriminant indicators;

S6: Iterate the backbone neural network and multi-task classification network. If the evaluation result reaches the optimal condition or reaches the maximum number of iterations, stop the training and save the weight of the multi-task classification network to complete the identification and classification of FSW acoustic emission welding defect signals .

In S1, the acoustic emission sensor model is R15A, the acquisition card model is PCIE-1816H, and the sampling rate is 100kHz to collect the acoustic signal.

In S2, Mel spectrum, short-time Fourier and wavelet transform are used to extract the features of the acoustic emission signal, and the steps are as follows:

S2-1: Use the Mel spectrum to analyze the filtered data data, the expression is as follows:

mel=f _mel (data)

In the formula: mel is the eigenvector after the Mel spectrum transformation, f _mel is the Mel frequency, and the expression of the Mel frequency curve is as follows:

Where: f is the original frequency.

S2-2: Use short-time Fourier transform to analyze the filtered data data, the expression is as follows:

spec=F(data)

In the formula: spec is the feature vector after the short-time Fourier transform, and the definition expression of the short-time Fourier transform is as follows:

In the formula: t is the time domain, ω is the frequency domain, i is the imaginary number unit, and e is the base of the natural logarithm;

S2-3: Use wavelet transform to analyze the filtered data data, the expression is as follows:

cwt=cwt(data)

Among them, the wavelet transform definition expression of f(t) is as follows:

In the formula: α is the scale factor, τ is the time shift factor, ψ(t) is the wavelet basis;

S2-4: 32 eigenvalues were extracted from S2-1 to S2-3 respectively, a total of 96 eigenvalues were extracted, and a data set was made with a Hamming window of 30 ms as a unit.

As shown in Figure 2, the backbone neural network model is established by combining the one-dimensional convolutional network and the fully connected network in S3. The backbone neural network model includes: the input dimension of the first layer is 3000*96, which is made by extracting feature values in S2 The data set, 3000 is the data of 30ms length, 96 is the feature vector of 30ms data, the first layer: convolutional layer, a one-dimensional convolutional neural network layer with a convolution kernel size of 100, defining 100 filters, activation function It is RELU, the output dimension is 2900*100, and acts on the lower layer; the second layer: convolutional layer, a one-dimensional convolutional neural network layer with a convolution kernel size of 100, defining 100 filters, the activation function is RELU, and the output The dimension is 2800*100, and acts on the lower layer; the third layer: pooling layer, uses the maximum pooling layer to reduce the complexity of the output and prevent data overfitting, the output dimension is 1400*100, and acts on the lower layer; Four layers: convolutional layer, a one-dimensional convolutional neural network layer with a convolution kernel size of 100, 160 filters are defined, the activation function is RELU, the output dimension is 1300*160, and it acts on the lower layer; the fifth layer: volume Product layer, a one-dimensional convolutional neural network layer with a convolution kernel size of 100, defining 160 filters, an activation function of RELU, an output dimension of 1200*160, and acting on the lower layer; the sixth layer: pooling layer, using The average pooling layer is used to further avoid overfitting, the output dimension is 600*160, and it acts on the lower layer; the seventh layer: the Dropout layer selects the Dropout layer with a ratio of 0.5, making the network less sensitive to small changes in the data , the output dimension is 600*160, and acts on the lower layer; the eighth layer: fully connected layer, the output dimension is 300*160.

As shown in Figure 3, the three time-frequency classification networks in S4 include short-time Fourier classification layer, Mel spectrum classification layer and wavelet classification layer. The three time-frequency classification network structures are all represented by the same classification network structure, collectively referred to as It is a multi-task classification network structure, including: the first layer: fully connected layer, which connects the backbone neural network model in S3 and three time-frequency classification layers, the input dimension is the output dimension of the fully connected layer in S3 300*160, and the output The dimension is 200*160; the second layer: convolutional layer, a one-dimensional convolutional neural network layer with a convolution kernel size of 100, defines 80 filters, the activation function is RELU, and the output dimension is 100*80, and acts on The lower layer; the third layer: convolutional layer, a one-dimensional convolutional neural network layer with a convolution kernel size of 50, defining 20 filter layers, the activation function is RELU, and the output dimension is 50*20, and acts on the lower layer; the fourth Layer: pooling layer, using the maximum pooling layer to reduce the complexity of the output and prevent data overfitting, the output dimension is 25*1; the fifth layer: the fine-tuned SoftMax layer, which divides the FSW welding process into five categories: Noise area, stirring head down pressure area, stirring head lifting area, normal welding area and defect area, the output dimension is 5*1.

The expression of SoftMax classification principle is as follows:

In the formula: P is the probability; e is the natural logarithm; v is the output vector; v _i is the value of the i-th category in v; k is the number of multiple outputs or categories of the neural network; v _j is the j-th in v The value of the category, i indicates the category that needs to be calculated currently, the calculation result is between 0 and 1, and the sum of the SoftMax values of all categories is 1.

In S5, the definition expression of DS evidence theory is as follows:

DS_result=DS(result _spec , result _mel , result _cwt )

In the formula: DS_result is the result of DS evidence theory synthesis; result _spec is the probability distribution output by the short-time Fourier feature vector after passing through the multi-task classification network structure; result _mel is the Mel spectral feature vector after passing through the multi-task classification network structure The output probability distribution; result _cwt is the probability distribution output by the wavelet transform feature vector after passing through the multi-task classification network structure. The three time-frequency algorithms respectively output five probability distributions through the fine-tuned SoftMax layer, and finally synthesized by DS evidence theory Output a unified result.

By calculating the cross-entropy function between the output of DS evidence theory and the true value label, the expression of the multi-task learning discriminant index is as follows:

In the formula: H(p, q) is the loss function value of the cross-entropy function; p is the truth label; q is the synthesis result of DS evidence theory; x _i is each probability corresponding to the probability vector; n represents the probability vector corresponding to the welding category , is the result synthesized by DS evidence theory.

pTrue value labels are assigned manually, including assigning unused welding process and defect learning labels according to the one-hot coding principle. As shown in Table 1, the welding process and defects are divided into five categories, which are welding start area (I) , Stable welding zone (II), transition zone (III), welding defect zone (IV), welding termination zone (V).

Table 1 Different welding areas and defect learning labels

In S6, the backpropagation algorithm is used to iterate the backbone neural network and the multi-task classification network. If the evaluation result reaches the optimal condition, that is, the difference between the cross-entropy function of the two iterations is less than 0.001 or the maximum number of iterations is 1000, which satisfies the One condition stops the training and saves the weights and parameters of the multi-task classification network to complete the identification and classification of FSW acoustic emission welding defect signals.

The backpropagation steps are as follows:

For neural network-dataset x vectors are:

x=(x ₁ , x ₂ , x ₃ , . . . , x _k )

After the activation function, the r vector is calculated:

r = ReLU(x)

After the r vector is calculated by the neural network, the deviation e is obtained:

e=forward(r)

Then find the partial derivative of each x _i for the deviation e as the gradient of the deviation e at the point x _i :

Update network weights:

w is the weight of each layer of neurons in the neural network, and LR is the learning rate.

The present invention first collects the acoustic emission signal of the FSW welding process through the acoustic emission sensor, and uses three kinds of signal time-frequency feature extraction methods (short-time Fourier, Mel spectrum, wavelet transform analysis) to extract the acoustic emission signal. Each method extracts 32 eigenvalues, and then use the one-dimensional convolutional backbone network model, combined with the SoftMax layer to design the short-time Fourier classification layer, the Mel spectrum classification layer, and the wavelet classification layer to construct a multi-task classification network structure, and extract the three time-frequency methods into A total of 96 eigenvalues are used as the input of the backbone neural network model for training, and the output value after training is used as the input of the multi-task classification network, and the back-propagation algorithm is used to iterate the backbone neural network model and the multi-task classification network, and pass DS evidence theory and cross-entropy loss function establish multi-task learning discriminant index evaluation. If the evaluation conditions are met, the training can be stopped and the backbone neural network and multi-task classification structure model weights and parameters can be saved to complete the identification and classification of FSW acoustic emission welding defect signals.

The present invention extracts multiple algorithms for the acoustic signal, uses the combination of DS evidence theory and SoftMax layer, solves the shortcomings of multiple single tasks that consume training costs, identifies and analyzes defect parts, and achieves accurate positioning of defect positions and defect types The purpose is to improve the recognition efficiency.

The present invention has been described by means of embodiments, and those skilled in the art will appreciate that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, the features and examples may be modified to adapt a particular situation and material to the teachings of the invention without departing from the spirit and scope of the invention. Therefore, the present invention is not limited by the specific embodiments disclosed here, and all embodiments falling within the scope of the claims of the present application belong to the protection scope of the present invention.

Claims (10)

1. A FSW welding monitoring method based on DS evidence theory and neural network is characterized by comprising the following steps:

s1: establishing an acoustic emission acquisition system, and acquiring an acoustic emission signal;

s2: extracting and analyzing the acoustic emission signals, and making a data set;

s3: inputting the data set in the S2, and establishing a backbone neural network model;

s4: inputting the result in the S3 and establishing a multi-task classification network structure;

s5: establishing a multitask learning discrimination index by combining a DS evidence theory;

s6: and iterating the main neural network and the multi-task classification network, stopping training and storing the weight of the multi-task classification network if the evaluation result reaches the optimal condition or the maximum iteration number, and finishing the identification and classification of the FSW acoustic emission welding defect signal.

2. The FSW welding monitoring method based on the DS evidence theory and the neural network as claimed in claim 1, wherein in S1, an acoustic emission sensor of a type R15A and a collection card of a type PCIE-1816H are used, and an acoustic signal is collected at a sampling rate of 100 kHz.

3. The FSW welding monitoring method based on DS evidence theory and neural network as claimed in claim 1, wherein the step of S2 comprises the following steps:

s2-1: and analyzing the filtered data by using a Mel frequency spectrum, wherein the expression is as follows:

mel＝f _mmel (data)

in the formula: mel is a feature vector after Mel frequency spectrum transformation, f _mel Is a Mel frequency curveThe line expression is as follows:

in the formula: f is the original frequency.

S2-2: analyzing the filtered data by using short-time Fourier transform, wherein the expression is as follows:

spec＝F(data)

in the formula: spec is a feature vector after short-time Fourier transform, and the short-time Fourier transform defines the expression as follows:

in the formula: t is a time domain, omega is a frequency domain, i is an imaginary number unit, and e is the base of a natural logarithm;

s2-3: analyzing the filtered data by using wavelet transformation, wherein the expression is as follows:

cwt＝cwt(data)

wherein, the wavelet transform definition expression of f (t) is as follows:

in the formula: alpha is a scale factor, tau is a time shift factor, psi (t) is a wavelet basis;

s2-4: and respectively extracting 32 characteristic values from S2-1 to S2-3, extracting 96 characteristic values in total, and preparing a data set by taking a Hamming window of 30ms as a unit.

4. The DS evidence theory and neural network-based FSW welding monitoring method as claimed in claim 1, wherein the trunk neural network model in S3 comprises: the first layer is a one-dimensional convolution layer with an output dimension of 2900 x 100; the second layer is a one-dimensional convolution layer with an output dimension of 2800 x 100; the third layer is a pooling layer, and the output dimension is 1400 × 100; the fourth layer is a one-dimensional convolution layer, and the output dimension is 1300 x 160; the fifth layer is a one-dimensional convolution layer with an output dimension of 1200 x 160; the sixth layer is a pooling layer with an output dimension of 600 × 160; the seventh layer is a Dropout layer with an output dimension of 600 × 160; the eighth layer is a fully connected layer with an output dimension of 300 x 160.

5. The FSW welding monitoring method based on DS evidence theory and neural network as claimed in claim 4, wherein the pooling layer of the third layer is a maximum pooling layer, and the pooling layer of the sixth layer is an average pooling layer.

6. The FSW welding monitoring method based on DS evidence theory and neural network as claimed in claim 1, wherein in S4, the multitask classification network structure comprises: the first layer is a fully connected layer with an output dimension of 200 × 160; the second layer is a one-dimensional convolution layer, and the output dimension is 100 x 80; the third layer is a one-dimensional convolution layer, and the output dimension is 50 x 20; the fourth layer is a pooling layer with an output dimension of 25 x 1; the fifth layer is the trimmed SoftMax layer with an output dimension of 5 x 1.

7. The FSW welding monitoring method based on the DS evidence theory and the neural network as claimed in claim 6, wherein the first fully connected layer in S4 is a connection between the main neural network model in S3 and the time-frequency classification layer, and the input dimension is the output dimension of the fully connected layer in S3.

8. The DS evidence theory and neural network-based FSW welding monitoring method as claimed in claim 6, wherein the fifth layer of fine-tuned SoftMax layer in S4 classifies FSW welding processes into five categories: the noise area, the stirring head lower pressing area, the stirring head lifting area, the normal welding area and the defect area, and the expression of the SoftMax classification principle is as follows:

in the formula: p is the probability; e is a natural logarithm; v is the output vector；v _i Is the value of the ith category in v; k represents a number of outputs or classes of the neural network; v. of _j For the value of the jth category in v, i represents the category that needs to be calculated currently, the calculation result is between 0 and 1, and the SoftMax values of all categories sum to 1.

9. The FSW welding monitoring method based on DS evidence theory and neural network as claimed in claim 1, wherein in S5, DS evidence theory defines the following expression:

DS_result＝DS(result _spec ，result _mel ，result _cwt )

in the formula: DS _ result is the result of DS evidence theory synthesis; result _spec The probability distribution of the short-time Fourier feature vector through neural network identification is obtained; result _mel Probability distribution of the Mel frequency spectrum characteristic vector through neural network identification; result _cwt Probability distribution of wavelet transformation characteristic vectors identified by a neural network;

the cross entropy function calculation is carried out on DS evidence theory output results and truth labels, namely the multi-task learning discriminant index expression is as follows:

in the formula: h (p, q) is a loss function value of the cross entropy function; p is a truth label; q is a DS evidence theory synthesis result; x is the number of _i Is the corresponding probability vector per probability; n represents a probability vector corresponding to the welding category.

10. The FSW welding monitoring method based on the DS evidence theory and the neural network as claimed in claim 1, wherein in S6, the evaluation result reaching the optimal condition means that a difference value of cross entropy functions obtained by two iterations in S5 is less than 0.001, the maximum number of iterations is 1000, and training is stopped and the result is saved when one of the conditions is met.

Images