A neural network-based stochastic resonance FSW micro-defect feature extraction method

Abstract

The application discloses a kind of random resonance FSW micro feature extraction methods based on neural network, comprising the following steps: S1: establishing acoustic emission acquisition system;S2: extracting acoustic emission signal feature;S3: establishing neural network model, and iteration optimization random resonance system structure parameter;S4: construct a kind of neural network loss function based on random resonance;S5: update iteration neural network parameter, determine whether optimal random resonance effect is reached;S6: extract and analyze acoustic emission signal micro feature;The present application takes FSW welding defect as the target of analysis, combines acoustic emission signal and the feature analysis of extraction, designs a kind of neural network loss function based on random resonance, and using neural network model more effectively finds the structure parameter of random resonance system, so that random resonance system and acoustic emission signal reach maximum resonance, effectively carry out feature extraction to the micro feature of welding defect position.

Description

A neural network-based method for extracting minute defects in stochastic resonant FSWs Technical Field

[0001] This invention relates to the field of friction stir welding technology, specifically to a method for extracting minute defect features in stochastic resonant FSW based on neural networks. Background Technology

[0002] Acoustic emission testing in friction stir welding is a novel method for dynamic non-destructive testing based on feedback from the internal structure of the workpiece. It can directly detect and judge internal defects in the workpiece and reflect the current situation of acoustic emission signal analysis in friction stir welding. Currently, the analysis of acoustic emission signals mainly uses traditional signal analysis methods (time domain, frequency domain, time-frequency domain), mode decomposition, wavelet packet transform, etc., to obtain qualitative evaluation of the acoustic emission signal in the time domain, frequency domain, and time-frequency domain.

[0003] Stochastic resonance has received considerable attention in the field of mechanical fault diagnosis in recent years, but its application to acoustic emission signals has been less studied. Stochastic resonance methods are significantly affected by structural parameters, making it difficult to achieve ideal detection results in practical signal processing. Most existing adaptive stochastic resonance methods assign values ​​to each parameter separately, rarely considering the interactions between parameters, making it difficult to achieve synchronous adaptive selection of multiple parameters, resulting in low accuracy in diagnosing early, weak faults.

[0004] Existing signal fault feature extraction methods, such as spectrum analysis, wavelet packet decomposition, local uniform value decomposition, and empirical mode decomposition, are poor at extracting subtle signal features and inevitably weaken the signal to some extent, affecting the diagnostic effect. Summary of the Invention

[0005] To address the problem of poor micro-feature extraction capability in existing signal micro-feature extraction methods, this invention provides a method for extracting micro-defect features in stochastic resonance FSW based on neural networks.

[0006] The present invention provides a method for extracting minute defect features of stochastic resonant stroboscopic swabs (FSWs) based on neural networks, comprising the following steps:

[0007] S1: Establish an acoustic emission acquisition system;

[0008] S2: Extract acoustic emission signal features;

[0009] S3: Establish a neural network model and iteratively optimize the structural parameters of the stochastic resonance system;

[0010] S4: Construct a neural network loss function based on stochastic resonance;

[0011] S5: Update the neural network parameters and determine whether the optimal stochastic resonance effect has been achieved. If not, return to S3; if yes, save the optimal stochastic resonance structure parameters.

[0012] S6: Extract and analyze minute features of acoustic emission signals.

[0013] Preferably, in step S1, an acoustic emission acquisition system is constructed using an R15A acoustic emission sensor, a PCIE-1816H acquisition card, and a sampling rate of 100kHz to acquire acoustic emission signals.

[0014] Preferably, in step S2, the acoustic emission signal is extracted using Mel-frequency spectrum and short-time Fourier transform, respectively, as follows:

[0015] S2-1: Analyze the filtered data using Mel-frequency spectrum, as shown in the following expression:

[0016] mel=f mel (data)

[0017] In the formula: mel is the eigenvector after Mel-frequency transform, and the Mel-frequency curve expression is as follows:

[0018]

[0019] In the formula: f is the original frequency, f mel Mel frequency;

[0020] S2-2: The filtered data 'data' is analyzed using the short-time Fourier transform, as shown in the following expression:

[0021] spec = F(data)

[0022] In the formula: spec is the eigenvector after the short-time Fourier transform, and the short-time Fourier transform is defined as follows:

[0023]

[0024] In the formula: t is the time domain and ω is the frequency domain.

[0025] Preferably, in step S3, the neural network model includes an input layer, a hidden layer, and an output layer. The basic implementation principle of the neural network model is expressed as follows:

[0026] f(x)=G[b i +W i S(b i-1 +W i-1 x)]

[0027] In the formula: x is the input vector, i.e., the feature value extracted in S2; W is the neuron connection weight; b is the bias; i is the number of network layers; functions G and S are ReLU activation functions, and the ReLU activation function expression is as follows:

[0028] ReLU(x) = max(0, x)

[0029] Preferably, in step S3, the solution expression for the stochastic resonance system is as follows:

[0030]

[0031] In the formula: k1, k2, k3, and k4 are process variables of stochastic resonance; a and b are the optimized system structure parameters, output by the neural network model; h = 1 / f s f is the sampling step size; s p is the sampling frequency; n It is the system input signal, x n It is the nth sampling point of the system output signal x(t), x n+1 is the output value of the signal input to the random resonance system, and is the signal feature extracted after random resonance.

[0032] Preferably, in step S4, the expression for the neural network loss function based on stochastic resonance is as follows:

[0033] SI=α*SNR+β*ZCR+γ*SCC+log(α)+log(β)+log(γ)

[0034] In the formula: α, β, γ are balance factors, initialized to 0; SNR is the signal-to-noise ratio; ZCR is the signal zero-crossing rate; SCC is the signal correlation coefficient.

[0035] Preferably, the signal-to-noise ratio expression for the random resonance output is as follows:

[0036]

[0037] In the formula: F0 is the signal frequency; S(F0) is the signal power; P is the total system power, including signal power and noise power; PS(F0) is the noise power;

[0038] The expression for the zero-crossing rate of a signal is as follows:

[0039]

[0040]

[0041] In the formula: sgn(*) is the sign function; N is the signal length; x is the signal value, which is the output value of the stochastic resonance system; x i Let x be the time-domain signal at a certain moment. i+1 The time-domain signal for the next moment; |x i -x i+1 |≥0 is 1 when true, and 0 when false;

[0042] The expression for the signal correlation coefficient is as follows:

[0043]

[0044] In the formula: X is the waveform signal after random resonance; Y is the original waveform signal; Let X and Y be the expected values.

[0045] Preferably, in step S5, the backpropagation method is used to update the neural network parameters. The process stops when the neural network model reaches a maximum of 1000 iterations or when the increment of the loss function value of the neural network based on stochastic resonance is less than a given threshold of 0.01, thus achieving the optimal stochastic resonance effect.

[0046] Preferably, in S6, the optimized stochastic resonance system has already amplified and extracted the acoustic emission signal at the welding defect in S3. It is necessary to combine the internal feature extraction method of the acoustic emission signal for feature extraction and perform feature analysis.

[0047] The filtered data is analyzed using acoustic emission internal feature extraction, as shown in the following expression:

[0048] internal = AE_Event(data)

[0049] In the formula: internal represents the internal feature vector of the acoustic emission signal, and the number of events of the acoustic emission signal AE_Event(*) is the acoustic emission signal generated by a local change in the material, which is called an acoustic emission event.

[0050] This invention discloses a method for extracting minute features of FSW welding defects based on a neural network. Taking FSW welding defects as the analysis target, the method combines acoustic emission signals and extracted features, considers various defect scenarios, and designs a neural network loss function based on random resonance by incorporating signal-to-noise ratio, zero-crossing rate, and signal correlation coefficient. Furthermore, the neural network model is used to more effectively find the structural parameters of the random resonance system, enabling the random resonance system to achieve maximum resonance with the acoustic emission signal, thus more effectively extracting minute features of the welding defect location. Attached Figure Description

[0051] Figure 1 is a flowchart of a method for extracting minute feature defects in stochastic resonant FSW based on neural networks according to the present invention;

[0052] Figure 2 is a structural diagram of the neural network model of the present invention. Detailed Implementation

[0053] The present invention provides a method for extracting minute feature defects in stochastic resonant stroboscopic ...

[0054] S1: Establish an acoustic emission acquisition system using an R15A acoustic emission sensor and a PCIE-1816H acquisition card with a sampling rate of 100kHz to acquire acoustic emission signals.

[0055] S2: Extract acoustic emission signal features. Specifically, Mel spectrum and short-time Fourier transform are applied to extract features from the acoustic emission signal, resulting in a total of 32 feature values.

[0056] S2-1: Analyze the filtered data using Mel spectrum, as shown in the following expression:

[0057] mel=f mel (data)

[0058] In the formula: mel is the eigenvector after Mel-frequency transform, and the Mel-frequency curve expression is as follows:

[0059]

[0060] In the formula: f is the original frequency, f mel Mel frequency;

[0061] S2-2: Analyze the filtered data using the short-time Fourier transform, as shown in the following expression:

[0062] spec = F(data)

[0063] In the formula: spec is the eigenvector after the short-time Fourier transform, and the short-time Fourier transform is defined as follows:

[0064]

[0065] In the formula: t is the time domain and ω is the frequency domain.

[0066] S3: Establish a neural network model and iteratively optimize the structural parameters of the stochastic resonance system. As shown in Figure 2, the neural network model consists of a multi-layer neural network with one input layer of 32 neurons, three hidden layers of 16, 32, and 16 neurons respectively, and one output layer of 2 neurons. It can also be called a complex deep neural network. Therefore, adding a hidden layer (a residual network branch) after the input layer and directing the residual results before the output layer increases the depth of the neural network, improves its fitting ability, and facilitates the optimization of the stochastic resonance system parameters. This multi-layer neural network is an artificial neural network model whose purpose is to map multiple input datasets to a single output dataset, exhibiting better fault tolerance compared to traditional classification methods. The basic implementation principle of the multi-layer neural network is as follows.

[0067] f(x)=G[b i +W i S(b i-1 +W i-1 x)]

[0068] In the formula: x is the input vector, i.e., the feature value extracted in S2; W is the neuron connection weight; b is the bias; i is the number of network layers; functions G and S are ReLU activation functions. Introducing activation functions introduces non-linearity into the neural network, enhancing its learning ability. The expression for the ReLU activation function is as follows:

[0069] ReLU(x) = max(0, x)

[0070] In the formula: x is the input vector.

[0071] Since there is no exact expression for the stochastic resonance system, the fourth-order Runge-Kutta algorithm is used to solve it. The specific expression is as follows:

[0072]

[0073] In the formula: k1, k2, k3, and k4 are process variables of random resonance; a and b are the optimized system structural parameters; h = 1 / f s f is the sampling step size. s p is the sampling frequency; n It is the system input signal, x n It is the nth sampling point of the system output signal x(t), x n+1 is the output value of the signal input to the random resonance system, and is the signal feature extracted after random resonance.

[0074] S4: Construct a neural network loss function based on stochastic resonance by combining signal-to-noise ratio, signal zero-crossing rate, and signal correlation coefficient. The specific steps are as follows:

[0075] The signal-to-noise ratio expression for the output of random resonance is as follows:

[0076]

[0077] In the formula: F0 is the signal frequency; S(F0) is the signal power; P is the total system power, including signal power and noise power; PS(F0) is the noise power;

[0078] The expression for the zero-crossing rate of a signal is as follows:

[0079]

[0080]

[0081] In the formula: sgn(*) is the sign function; N is the signal length; x is the signal value, which is the output x in the expression for solving the stochastic resonance system in S3. n+1 Value; x i Let x be the time-domain signal at a certain moment. i+1 The time-domain signal for the next moment; |x i -x i+1 |≥0 is 1 when true, and 0 when false;

[0082] The expression for the signal correlation coefficient is as follows:

[0083]

[0084] In the formula: X is the waveform signal after random resonance; Y is the original waveform signal; Let X and Y be the expected values.

[0085] The expression for the loss function of a neural network based on stochastic resonance is as follows:

[0086] SI=α*SNR+β*ZCR+γ*SCC+log(α)+log(β)+log(γ)

[0087] In the formula: α, β, γ are balance factors, initialized to 0; SNR is the signal-to-noise ratio; ZCR is the signal zero-crossing rate; SCC is the signal correlation coefficient. The balance factors (α, β, γ) are added to avoid the occurrence of an excessively large value when calculating the SI loss function, and are continuously updated during network model training. At the same time, logarithmic operations are added to avoid the situation where SI is always 0.

[0088] S5: Update the neural network parameters using backpropagation. Stop when the neural network model reaches the maximum number of iterations (1000) or the increment of the loss function value of the stochastic resonance neural network is less than a given threshold of 0.01, thus achieving the optimal stochastic resonance effect. If not, return to S3.

[0089] The specific steps of the backpropagation method are as follows:

[0090] For a neural network with a dataset x vector, we have:

[0091] x = (x1, x2, x3, ..., x k )

[0092] After the activation function, the vector r is calculated:

[0093] r = ReLU(x)

[0094] After the vector r is processed by the neural network, the deviation e is obtained:

[0095] e = forward(r)

[0096] Then calculate each x for the deviation e. i The partial derivative of the deviation e in x is the deviation. i gradient at point:

[0097]

[0098] Update network weights:

[0099]

[0100] w represents the weight of each layer of neurons in the neural network, and LR represents the learning rate.

[0101] S6: Extracting and analyzing minute features of acoustic emission signals: After achieving the optimal stochastic resonance effect, the optimized stochastic resonance system amplifies and extracts the weak acoustic emission signals at the welding defects through the solution expression in S3. Then, it combines the internal feature method of acoustic emission signals for feature extraction and analysis. The steps are as follows:

[0102] The filtered data is analyzed using acoustic emission internal feature extraction, as shown in the following expression:

[0103] internal = AE_Event(data)

[0104] In the formula: internal represents the internal feature vector of the acoustic emission signal, and the number of events of the acoustic emission signal AE_Event(*) is the acoustic emission signal generated by a local change in the material, which is called an acoustic emission event.

[0105] This invention extracts features from acoustic emission signals, builds an iterative neural network model, applies the parameters output by the neural system to stochastic resonance, and evaluates the neural network using a stochastic resonance-based neural network loss function. If the evaluation result reaches the optimal condition or the maximum number of iterations is reached, training stops and parameters a and b are saved. Then, by combining stochastic resonance with the internal parameter features of acoustic emission, minute feature extraction is performed to complete the extraction of minute defect features from the acoustic emission signal of FSW welding defects. This allows the stochastic resonance system and the acoustic emission signal to reach maximum resonance, thus more effectively extracting and analyzing minute features of the welding defect location.

[0106] This invention has been described through embodiments. Those skilled in the art will understand that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of this invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, this invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of this invention.

Claims

1. A method for extracting small features from stochastic resonant free-space swirling waves based on neural networks, characterized in that, Includes the following steps: S1: Establish an acoustic emission acquisition system; S2: Extract acoustic emission signal features; S3: Establish a neural network model and iteratively optimize the structural parameters of the stochastic resonance system; the solution expression for the stochastic resonance system is as follows: In the formula: For random resonance process variables; The optimized system structure parameters are output by the neural network model; This is the sampling step size; The sampling frequency; It is the system input signal. It is the system output signal The One sampling point, S1 represents the output value of the signal input to the stochastic resonance system, and S2 represents the signal features extracted after stochastic resonance. S4: Construct the neural network loss function based on stochastic resonance. The expression for the neural network loss function based on stochastic resonance is as follows: In the formula: This is the balance factor, initialized to 0. Signal-to-noise ratio; The zero-crossing rate of the signal; Let be the signal correlation coefficient; the signal-to-noise ratio expression for the random resonance output is as follows: In the formula: The signal frequency; Signal power; This represents the total system power, including signal power and noise power. The noise power is given; the zero-crossing rate expression is as follows: In the formula: It is a symbolic function; The signal length; Here, represents the signal value, and represents the output value in the stochastic resonance system. For a time-domain signal at a certain moment, This is the time-domain signal for the next moment; The value is 1 when the condition is true and 0 when the condition is false; the expression for the signal correlation coefficient is as follows: In the formula: The waveform signal after random resonance; This is the original waveform signal; for S5: Update the neural network parameters and determine whether the optimal stochastic resonance effect has been achieved. If not, return to S3. If yes, save the optimal stochastic resonance structure parameters. S6: Extract and analyze the subtle features of the acoustic emission signal. The optimized stochastic resonance system has amplified and extracted the acoustic emission signal at the welding defect in S3. It needs to be combined with the internal feature extraction method of the acoustic emission signal for feature extraction and analysis. The filtered data is analyzed using the internal feature extraction method, expressed as follows: internal= (data); where: internal represents the internal feature vector of the acoustic emission signal, and the number of events of the acoustic emission signal AE_Event(*) is the acoustic emission signal generated by a local change in the material, which is called an acoustic emission event.

2. The method for extracting small features of stochastic resonant FSW based on neural networks according to claim 1, characterized in that, In step S1, an acoustic emission acquisition system is constructed using an R15A acoustic emission sensor, a PCIE-1816H acquisition card, and a sampling rate of 100kHz to acquire acoustic emission signals.

3. The method for extracting small features of stochastic resonant FSW based on neural networks according to claim 1, characterized in that, In step S2, Mel-frequency spectrum and short-time Fourier transform are applied to extract the acoustic emission signal, as follows: S2-1: The filtered data is analyzed using the Mel-frequency spectrum, as shown in the following expression: In the formula: mel is the eigenvector after Mel-frequency transform, and the Mel-frequency curve expression is as follows: In the formula: For the original frequency, Where is the Mel frequency; S2-2: Analyze the filtered data using short-time Fourier transform, as shown in the following expression: In the formula: spec is the eigenvector after the short-time Fourier transform, where the short-time Fourier transform is defined as follows: In the formula: For time domain, This is in the frequency domain.

4. The method for extracting small features of stochastic resonant FSW based on neural networks according to claim 1, characterized in that, In S3, the neural network model includes an input layer, a hidden layer, and an output layer. The basic implementation principle of the neural network model is expressed as follows: In the formula: The input vector is the feature value extracted from S2; These are the neuron connection weights; For bias; The number of network layers; function ,function The ReLU activation function is defined as follows: 。 5. The method for extracting small features of stochastic resonant FSW based on neural networks according to claim 1, characterized in that, In step S5, the backpropagation method is used to update the neural network parameters. The neural network model stops when it reaches the maximum number of iterations of 1000 or when the increment of the loss function value of the neural network based on stochastic resonance is less than a given threshold of 0.01, thus achieving the optimal stochastic resonance effect.

Images