Prediction method based on amplitude denoising and data amplification

Abstract

The invention discloses a prediction method based on amplitude denoising and data amplification, and relates to the technical field of ultrasonic welding. Performing maximum absolute value scaling on the preprocessed amplitude data to generate sample diversity loss, constructing a maximum absolute value scaling-diffusion model in combination with a diffusion model, and amplifying the preprocessed amplitude data; decomposing the amplified amplitude data based on a self-adaptive variational mode decomposition algorithm of intelligent parameter disturbance simulated annealing, and screening a target mode component for reconstruction to obtain reconstructed amplitude data; feature extraction is carried out on the reconstructed amplitude data to obtain data features, the data features are input into a full-connection neural network, a predicted resistance value is output, and the ultrasonic welding quality of the welding sample is evaluated through the resistance value of the welding sample and the predicted resistance value, so that the problems of multiple influence factors and unbalanced sample data can be solved; and the ultrasonic welding quality evaluation accuracy is improved.

Description

Technical Field

[0001] The invention relates to the technical field of ultrasonic welding, and in particular to a prediction method based on amplitude denoising and data amplification. Background Art

[0002] The principle of UMW (Ultrasonic Metal Welding) is a special method of connecting the same or different materials using mechanical vibration energy of ultrasonic frequency. In actual production, the working environment of UMW is very complex. In the prior art, the vibration signal collected during the welding process is usually subject to a lot of noise interference, resulting in signal redundancy, which is not conducive to the subsequent welding quality monitoring. In addition, the quality of ultrasonic welding is affected by multiple factors such as welding pressure, ultrasonic amplitude, welding energy and random external interference. The amount of welding status data is extremely large, and the abnormal data is far less than the normal data. There is a problem of unbalanced sample data, which makes it difficult to achieve online monitoring of welding quality, resulting in poor accuracy in ultrasonic welding quality assessment.

[0003] It can be seen from the above that how to solve the problem of multiple influencing factors and unbalanced sample data and improve the accuracy of ultrasonic welding quality assessment is a problem to be solved in this field. Summary of the invention

[0004] In view of this, the purpose of the present invention is to provide a prediction method based on amplitude denoising and data amplification, which can solve the problems of multiple influencing factors and unbalanced sample data and improve the accuracy of ultrasonic welding quality assessment. The specific scheme is as follows:

[0005] The present application discloses a prediction method based on amplitude denoising and data amplification, comprising:

[0006] Acquiring amplitude data of the ultrasonic welding head in real time, obtaining the resistance value of the welding sample, and performing data preprocessing on the amplitude data to obtain the preprocessed amplitude data;

[0007] Performing maximum absolute value scaling on the preprocessed amplitude data to generate sample diversity loss, and combining with the diffusion model to construct a maximum absolute value scaling-diffusion model, and using the maximum absolute value scaling-diffusion model to amplify the preprocessed amplitude data to obtain the amplified amplitude data;

[0008] Decomposing the amplified amplitude data based on an adaptive variational modal decomposition algorithm of preset intelligent parameter perturbation simulated annealing to obtain modal components, and selecting target modal components from the modal components for reconstruction to obtain the reconstructed amplitude data;

[0009] Feature extraction is performed on the reconstructed amplitude data to obtain data features, and the data features are input into a fully connected neural network to output a predicted resistance value, and the resistance value of the welding sample and the predicted resistance value are used to evaluate the ultrasonic welding quality of the welding sample.

[0010] Optionally, the real-time acquisition of amplitude data of the ultrasonic welding head and the acquisition of the resistance value of the welding sample include:

[0011] The Doppler laser vibrometer is used to measure and collect the amplitude data of the ultrasonic welding head in real time;

[0012] The resistance value of the welding sample is measured and obtained using a loop current generator.

[0013] Optionally, the performing data preprocessing on the amplitude data to obtain the preprocessed amplitude data includes:

[0014] Normalizing the amplitude data to obtain the normalized amplitude data;

[0015] Downsampling is performed on the normalized amplitude data, and padding is performed on the downsampling amplitude data to obtain the pre-processed amplitude data.

[0016] Optionally, performing maximum absolute value scaling on the preprocessed amplitude data to generate sample diversity loss, and combining with a diffusion model to construct a maximum absolute value scaling-diffusion model, includes:

[0017] Using the maximum mean difference method to perform maximum absolute value scaling on the preprocessed amplitude data to generate a sample diversity loss;

[0018] The sample diversity loss is used as the loss function of the diffusion model to construct a maximum absolute value scaling-diffusion model.

[0019] Optionally, the amplifying the preprocessed amplitude data using the maximum absolute value scaling-diffusion model to obtain the amplified amplitude data includes:

[0020] Dividing the preprocessed amplitude data into an augmentation set and a test set according to a preset data division ratio;

[0021] The preprocessed amplitude data in the amplification set is amplified using the maximum absolute value scaling-diffusion model to obtain the amplified amplitude data.

[0022] Optionally, the adaptive variational modal decomposition algorithm based on preset intelligent parameter perturbation simulated annealing decomposes the amplified amplitude data to obtain modal components, and selects target modal components from the modal components for reconstruction to obtain the reconstructed amplitude data, including:

[0023] Construct an adaptive variational mode decomposition algorithm with initial intelligent parameter perturbation simulated annealing, set intelligent parameter perturbation and variational mode decomposition parameter data, use minimum envelope entropy as fitness function, adaptively find the optimal decomposition layer number and penalty factor of different amplitude sequences, so as to generate an adaptive variational mode decomposition algorithm with intelligent parameter perturbation simulated annealing;

[0024] Decomposing the amplified amplitude data using an adaptive variational modal decomposition algorithm based on the intelligent parameter perturbation simulated annealing to obtain modal components;

[0025] The modal components are subjected to fast Fourier transformation, and corresponding target modal components are selected from the modal components according to a preset working frequency range of ultrasonic metal welding, and reconstructed in an accumulative manner to obtain the reconstructed amplitude data.

[0026] Optionally, the step of extracting features from the reconstructed amplitude data to obtain data features, and inputting the data features into a fully connected neural network to output a predicted resistance value comprises:

[0027] performing image data conversion on the reconstructed amplitude data to obtain three-dimensional image data;

[0028] Inputting the three-dimensional image data into a lightweight convolutional neural network for feature extraction to obtain data features;

[0029] The data features are input into a fully connected neural network for resistance value prediction, so that the last layer of neurons in the fully connected neural network outputs the predicted resistance value.

[0030] It can be seen that the present application provides a prediction method based on amplitude denoising and data amplification, including real-time acquisition of amplitude data of an ultrasonic welding head, and obtaining the resistance value of a welding sample, performing data preprocessing on the amplitude data to obtain the preprocessed amplitude data; performing maximum absolute value scaling on the preprocessed amplitude data to generate sample diversity loss, and combining a diffusion model to construct a maximum absolute value scaling-diffusion model, and using the maximum absolute value scaling-diffusion model to amplify the preprocessed amplitude data to obtain the amplified amplitude data; decomposing the amplified amplitude data based on an adaptive variational modal decomposition algorithm of preset intelligent parameter perturbation simulated annealing to obtain modal components, screening target modal components from the modal components for reconstruction to obtain the reconstructed amplitude data; performing feature extraction on the reconstructed amplitude data to obtain data features, inputting the data features into a fully connected neural network to output a predicted resistance value, and using the resistance value of the welding sample and the predicted resistance value to evaluate the ultrasonic welding quality of the welding sample. The present application preprocesses the amplitude data and performs maximum absolute value scaling to construct a maximum absolute value scaling-diffusion model, and uses the model to amplify the amplitude data, thereby solving the problem of unbalanced sample data. An adaptive variational modal decomposition algorithm of intelligent parameter perturbation simulated annealing is introduced to decompose and reconstruct the amplitude data, which can better remove the multi-modal aliasing phenomenon and noise interference in complex amplitude data, so as to solve the problem of many influencing factors. Feature extraction and resistance value prediction are performed on the reconstructed amplitude data, and the resistance value is used as an evaluation indicator of ultrasonic welding quality to realize the evaluation of ultrasonic welding quality, which can improve the accuracy of ultrasonic welding quality evaluation. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0032] Figure 1 A flow chart of a prediction method based on amplitude denoising and data amplification disclosed in the present application;

[0033] Figure 2 A flow chart of a prediction method based on amplitude denoising and data amplification disclosed in the present application;

[0034] Figure 3 A cluster diagram of amplitude data and real data generated by MMD-DDPM disclosed in this application;

[0035] Figure 4 A cluster diagram of amplitude data and real data generated by DCGAN disclosed in this application;

[0036] Figure 5 Amplitude data generated by WGAN disclosed in this application and clustering diagram of real data;

[0037] Figure 6 The time domain diagram of decomposition and reconstruction of the original signal disclosed in this application;

[0038] Figure 7 The spectrum diagram of the decomposition and reconstruction of the original signal disclosed in this application;

[0039] Figure 8 It is a time domain curve diagram of the modal component IMF after PPSA-VMD decomposition disclosed in this application;

[0040] Fig. 9 It is a frequency domain curve diagram of the modal component IMF after PPSA-VMD decomposition disclosed in this application;

[0041] Fig.10 This is a diagram of the overall prediction structure based on amplitude denoising and data amplification disclosed in this application. DETAILED DESCRIPTION

[0042] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] The principle of UMW is a special method of connecting the same or different materials using mechanical vibration energy at ultrasonic frequencies. In actual production, the working environment of UMW is very complex. In the prior art, the vibration signals collected during the welding process are usually subject to a large amount of noise interference, resulting in signal redundancy, which is not conducive to the subsequent monitoring of welding quality. In addition, the quality of ultrasonic welding is affected by multiple factors such as welding pressure, ultrasonic amplitude, welding energy and random external interference. The amount of welding state data is extremely large, and the abnormal data is far less than the normal data. There is a problem of unbalanced sample data, making it difficult to achieve online monitoring of welding quality, resulting in poor accuracy in ultrasonic welding quality assessment. As can be seen from the above, how to solve the problems of multiple influencing factors and unbalanced sample data and improve the accuracy of ultrasonic welding quality assessment are problems to be solved in this field.

[0044] See also Figure 1 As shown, the embodiment of the present invention discloses a prediction method based on amplitude denoising and data amplification, which may specifically include:

[0045] Step S11: collecting amplitude data of the ultrasonic welding head in real time, obtaining the resistance value of the welding sample, and performing data preprocessing on the amplitude data to obtain the preprocessed amplitude data.

[0046] In this embodiment, a Doppler laser vibrometer is used to measure and collect amplitude data of the ultrasonic welding head in real time; a loop current generator is used to measure and obtain the resistance value of the welding sample, and then the amplitude data is preprocessed to obtain the preprocessed amplitude data.

[0047] The present application uses a Doppler laser vibrometer to measure the amplitude data of the ultrasonic welding head. In order to avoid the inconvenience and disadvantages of evaluating the welding sample by shear strength, a 100A current is passed through both ends of the welding sample to measure the resistance value of the welding sample, and the resistance value is used as an evaluation indicator of the welding quality.

[0048] In this embodiment, the process of data preprocessing is as follows: normalizing the amplitude data to obtain the normalized amplitude data; downsampling the normalized amplitude data, and padding the downsampled amplitude data to obtain the preprocessed amplitude data.

[0049] Specifically, the collected amplitude data is normalized so that the data is distributed between [0, 1]:

[0050]

[0051] Among them, any one of the three types of data X={x1, x2, x3, ..., x n}, x i represents the i-th sample data in a certain type of data, n represents the length of the data in this type, max(X) represents the maximum value in this type of data, and min(X) represents the minimum value in this type of data;

[0052] Downsampling and filling processing: The sampling frequency of the laser sensor is 312.5KHz. Two million points will be collected in a welding process, but the effective data length is only about 300,000. Therefore, it is necessary to manually intercept this effective data length and retain 1 point every 100 points for downsampling, reduce the data dimension, fill the amplitude data with 0 value to equal length, and convert it into two-dimensional image data.

[0053] Step S12: Perform maximum absolute value scaling on the preprocessed amplitude data to generate sample diversity loss, and combine with the diffusion model to construct a maximum absolute value scaling-diffusion model, and use the maximum absolute value scaling-diffusion model to amplify the preprocessed amplitude data to obtain the amplified amplitude data.

[0054] In this embodiment, the maximum mean difference method is used to perform maximum absolute value scaling on the preprocessed amplitude data to generate sample diversity loss; the sample diversity loss is used as the loss function of the diffusion model to construct a maximum absolute value scaling-diffusion model, and the preprocessed amplitude data is divided into an augmentation set and a test set according to a preset data division ratio; the preprocessed amplitude data in the augmentation set is amplified using the maximum absolute value scaling-diffusion model to obtain the amplified amplitude data.

[0055] This application amplifies the amplitude data after preprocessing to alleviate the problem of fewer defective samples. The specific steps are as follows: (1) Use the MMD (Maximum Mean Discrepancy) method to scale the maximum absolute value of the preprocessed amplitude data to generate sample diversity loss, and use the sample diversity loss as the loss function of DDPM (Denoising Diffusion Probabilistic Models) to construct MMD-DDPM (maximum absolute value scaling-diffusion model); (2) Divide the 480 groups of sample data sets collected and preprocessed in the experiment (i.e., the amplitude data after preprocessing) into an expansion set and a test set. The expansion set is used for data expansion in MMD-DDPM, and the test set is only used in the final prediction stage to ensure that the data is not leaked. For example, there are 200 test sets and 280 augmentation sets; (3) The amplitude data is processed into a dimension of 4624×1 after (2). In order to make the sample resistance value data have a certain weight, the resistance value is repeatedly spliced ​​to the end of the data, and the data dimension is changed to 5184×1; at this time, the preprocessed amplitude data can be regarded as two types of features. One type is the acquisition data feature, whose value range is between [-30,30]; the other type is the sample label feature, whose value range is between [300,400]. Using feature scaling, the maximum absolute value scaling operation is performed on each type of data. Since the subsequent algorithm for data augmentation is based on convolutional neural network, the data is converted into 72×72×1 three-dimensional features for subsequent data augmentation; (4) The augmentation set data is put into MMD-DDPM for training to obtain 400 sets of amplified amplitude data.

[0056] The purpose of maximum absolute value scaling is to scale the data features according to their maximum absolute value so that the value range of each feature falls between [-1, 1]. By performing such scaling, the influence caused by the difference in numerical values ​​between different features can be eliminated, which is conducive to the subsequent data augmentation algorithm to better process the data and improve the model performance. In addition, maximum absolute value scaling retains the zero value in the data and does not change the distribution shape of the data, so it is suitable for the data distribution in this article. The formula for maximum value scaling is as follows:

[0057]

[0058] Among them, X i,j is the data position, i is the data length [0, 5184], j is the number of data groups [0, n], |X 1：n，j | is the absolute value of the current column data.

[0059] Step S13: Decomposing the amplified amplitude data based on an adaptive variational modal decomposition algorithm of preset intelligent parameter perturbation simulated annealing to obtain modal components, selecting target modal components from the modal components for reconstruction to obtain the reconstructed amplitude data.

[0060] Step S14: extracting features from the reconstructed amplitude data to obtain data features, inputting the data features into a fully connected neural network to output a predicted resistance value, and evaluating the ultrasonic welding quality of the welding sample using the resistance value of the welding sample and the predicted resistance value.

[0061] In this embodiment, the amplitude data of the ultrasonic welding head is collected in real time, and the resistance value of the welding sample is obtained, and the amplitude data is preprocessed to obtain the preprocessed amplitude data; the preprocessed amplitude data is scaled by the maximum absolute value to generate sample diversity loss, and combined with the diffusion model, a maximum absolute value scaling-diffusion model is constructed, and the preprocessed amplitude data is amplified by using the maximum absolute value scaling-diffusion model to obtain the amplified amplitude data; the amplified amplitude data is decomposed by an adaptive variational modal decomposition algorithm based on a preset intelligent parameter perturbation simulated annealing to obtain modal components, and the target modal components are selected from the modal components for reconstruction to obtain the reconstructed amplitude data; feature extraction is performed on the reconstructed amplitude data to obtain data features, and the data features are input into a fully connected neural network to output a predicted resistance value, and the resistance value of the welding sample and the predicted resistance value are used to evaluate the ultrasonic welding quality of the welding sample. The present application preprocesses the amplitude data and performs maximum absolute value scaling to construct a maximum absolute value scaling-diffusion model, and uses the model to amplify the amplitude data, thereby solving the problem of unbalanced sample data. An adaptive variational modal decomposition algorithm of intelligent parameter perturbation simulated annealing is introduced to decompose and reconstruct the amplitude data, which can better remove the multi-modal aliasing phenomenon and noise interference in complex amplitude data, so as to solve the problem of many influencing factors. Feature extraction and resistance value prediction are performed on the reconstructed amplitude data, and the resistance value is used as an evaluation indicator of ultrasonic welding quality to realize the evaluation of ultrasonic welding quality, which can improve the accuracy of ultrasonic welding quality evaluation.

[0062] See also Figure 2 As shown, the embodiment of the present invention discloses a prediction method based on amplitude denoising and data amplification, which may specifically include:

[0063] Step S21: collecting amplitude data of the ultrasonic welding head in real time, obtaining the resistance value of the welding sample, and performing data preprocessing on the amplitude data to obtain the preprocessed amplitude data.

[0064] Step S22: Perform maximum absolute value scaling on the preprocessed amplitude data to generate sample diversity loss, and combine with the diffusion model to construct a maximum absolute value scaling-diffusion model, and use the maximum absolute value scaling-diffusion model to amplify the preprocessed amplitude data to obtain the amplified amplitude data.

[0065] The diffusion model is a generative model based on probability distribution, which is used for image modeling and generating high-quality image data. The core idea of ​​DDPM is to use conditional probability flow to represent the interaction between pixels, which enables the model to effectively consider long-distance pixel correlations when generating images. In addition, DDPM adopts a reversible diffusion process and uses deep learning to transform the potential pixel values ​​at each step, thereby enhancing the expressive power of the model. In the forward denoising process, Gaussian noise is gradually added to the original image data. After completing T steps of addition, the original image data completely becomes random noise and loses the original information. The formula for the denoising process is as follows:

[0066]

[0067] in, is the variance of the noise added at each step, which is 0-1 and is a predefined gradual decay.

[0068] In the forward noise addition process, the image X at the current moment t Only with the image X at the previous moment t-1 This forms a Markov chain, that is, the result at any moment is only related to the value at the previous moment. It satisfies:

[0069]

[0070] The reverse denoising process uses U-net (neural network) to fit the mean of the reverse denoising process, and gradually restores the real image data through the initial noise.

[0071] The forward denoising process and reverse denoising process of DDPM are shown in Table 1:

[0072] Table 1

[0073]

[0074] The training process of forward noise addition: First, randomly select an image X0 from the sample data set ， And randomly select time t from 1, ..., T, randomly sample noise noise from N(0, I), add noise to X0 to get X t ; Secondly, calculate its sinusoidal position code according to time t, and X t Together they serve as the input of the U-net network, and the output of the U-net network is the predicted time t added to the Gaussian noise noise t Finally, according to noise and noise t Calculate the loss and update the parameters of the U-net network until the training is completed.

[0075] The training process of reverse denoising: First determine the image X at the current moment t , if t = T, that is, randomly sample noise with the same dimension as the input from the standard normal distribution as X t , and randomly select time t from 1, ..., T; secondly, encode the sinusoidal position of t and X t As the input of the U-net network, the noise added from the moment t-1 before the forward noise addition process to the current moment t is predicted; finally, based on the noise and X t , using the reparameterization technique, reversely infer X t-1 ; Finally, when the loop ends, that is, at time t=0, the output is the generated image.

[0076] The diffusion model predicts noise using the U-net network, which is used to predict the noise ε added at the previous time t-1 before the current time t θ (X t , t). U-net is a fully convolutional neural network that is widely used in image segmentation tasks. The network architecture consists of two parts: compression path and expansion path, which can effectively capture local features in the image and perform accurate segmentation.

[0077] In order to effectively encourage DDPM to generate more diverse samples during the training process, thereby improving the diversity and fidelity of the generated results, the present invention, based on DDPM and with reference to MMD, designs a generated sample diversity loss based on MMD, that is, designs a loss function to encourage the model to show diversity when generating samples. By measuring the differences between samples and taking the differences as part of the loss, the model is encouraged to generate more diverse samples.

[0078] MMD is a divergence measure that can be used to measure the difference between two distributions. Here, MMD is used to design a diversity loss so that the difference between generated samples becomes the optimization goal. The specific steps are as follows:

[0079] First, use MMD to measure the similarity between DDPM generated samples and real samples, that is, to maximize the authenticity of the generated data. The loss of this part is as follows:

[0080] MMD[P_G(x),P(z)]=E x～P_G(x)，z～P(z) [k(z,G(x))];

[0081] Secondly, maximize the difference between different generated samples to obtain more diverse samples. The loss of this part is as follows:

[0082] MMD[x, P_G(x)]=E x，x～P_G(x) [k(x, G(x))];

[0083] Among them, P_G(x) is the distribution of generated samples, P(z) is the distribution of real samples, k() is a mapping function that exists in the regenerated Hilbert space. This paper uses the Gaussian kernel function as the mapping function. The representation of the Gaussian kernel function is as follows:

[0084]

[0085] Where σ is the image bandwidth of the Gaussian kernel function. If we assume that the generated data set is x = [x1, x2, …, x m ], the real data set is z=[z1,z2,…,z n ], when n,m→∞, the maximum mean difference is calculated as follows:

[0086]

[0087] Finally, the above two losses are combined in a weighted manner to form the final loss function, and the proportion of the two losses is adjusted by the weight coefficient. The final diversity loss function is as follows:

[0088] Loss2max[αE x～P_G(x),z～P(z) [k(z,G(x))]-βE x，x～P_G(x) [k(x, G(x))]];

[0089] By maximizing the similarity between generated samples and real samples and the difference between different generated samples, the generated samples can generate samples with different distributions as much as possible without deviating from the distribution of real samples, thus ensuring the diversity and differentiation of generated data.

[0090] The present invention uses PCA (Principal Component Analysis) dimensionality reduction and t-SNE (t-Distributed Stochastic Neighbor Embedding) dimensionality reduction to visualize the distribution of generated data and real data, and compares them with the currently popular derivative networks DCGAN (Deep Convolutional Generative Adversarial Networks) and WGAN (Wasserstein GAN) based on GAN (Generative Adversarial Networks). The amplitude data generated by MMD-DDPM is clustered with the real data. Figure 3 As shown in Figure 2, the amplitude data generated by DCGAN is clustered with the real data. Figure 4 As shown, the amplitude data generated by WGAN is clustered with the real data as shown in Figure 5 shown.

[0091] Step S23: construct an adaptive variational modal decomposition algorithm of initial intelligent parameter perturbation simulated annealing, set intelligent parameter perturbation and variational modal decomposition parameter data, take minimum envelope entropy as fitness function, adaptively find the optimal decomposition layer number and penalty factor of different amplitude sequences, so as to generate an adaptive variational modal decomposition algorithm of intelligent parameter perturbation simulated annealing.

[0092] In this embodiment, an initial PPSA-VMD (adaptive variational modal decomposition with intelligent parameter perturbation simulated annealing) algorithm is constructed, the intelligent parameter perturbation and variational modal decomposition parameters are set, and the minimum envelope entropy is used as the fitness function to find the optimal decomposition layer number K and penalty factor a of VMD with different amplitude sequences.

[0093] Step S24: Decompose the amplified amplitude data based on the adaptive variational modal decomposition algorithm of the intelligent parameter perturbation simulated annealing to obtain modal components, perform fast Fourier transform on the modal components, and select corresponding target modal components from the modal components according to the preset working frequency range of ultrasonic metal welding, and reconstruct them in a cumulative manner to obtain the reconstructed amplitude data.

[0094] This application filters out interference signals based on the center frequency of the modal components after decomposition, and reconstructs effective vibration data by superposition. The operating frequency of ultrasonic metal welding is usually between 19.5-20KHz, and the quality information of the sample is contained in the modal components within this frequency range. Therefore, after PPSA-VMD decomposition, the modal components are subjected to FFT (Fast Fourier Transformation), and the modal components with an operating frequency of 19.5-20KHz are selected as the target modal components, and the amplitude data are reconstructed in an accumulative manner to obtain the reconstructed amplitude data.

[0095] The traditional simulated annealing algorithm applies a perturbation to the solution space that changes linearly with temperature T, but in actual situations, it is often necessary to consider multiple factors, such as the dimension of the solution space and the rate of change of the solution. For the amplitude data collected from ultrasonic metal welding, it is necessary to find the optimal values ​​of the two parameters, the number of decomposition layers K of VMD and the penalty factor a. The value range of K is [2, 15], and the value range of a is [1000, 9000]. The value ranges of the two parameters are very different. For the parameter K in the entire optimization process, it is only necessary to apply a perturbation of ±1 to it within the solution space. On the other hand, the solution space range of parameter a is large, and the perturbation of the traditional simulated annealing algorithm that changes linearly with temperature T is difficult to meet its change requirements in the entire optimization process. In the early stage of optimization, a is often required to change within a large range to ensure that it can quickly traverse every field in the solution space. In the later stage of optimization, a is required to change within a smaller range so that it can search for the global optimal solution in a smaller field to avoid jumping out of the field or failing to locate the optimal solution due to excessive perturbation.

[0096] In response to the above-mentioned problems, and in combination with the current industrial data of ultrasonic metal welding, this application proposes a PP (Parameter Perturbations, intelligent parameter perturbation strategy): design a more intelligent parameter perturbation strategy to dynamically adjust the amplitude of the perturbation based on the current search state. By adjusting the size of the parameter perturbation according to the characteristic information temperature T and the jump probability P of the current solution, the algorithm convergence can be accelerated to avoid falling into the local optimal solution or missing the global optimal solution. The intelligent parameter perturbation strategy is as follows:

[0097] Set a perturbation strategy for the penalty coefficient a so that the change of a is nonlinearly related to the current temperature T and the jump probability P. The formula is as follows:

[0098]

[0099] Among them, R is the penalty factor disturbance term, A, C, and KP are all adjustment parameters, A is the upper limit of the disturbance, C is the midpoint parameter, KP is the slope of the disturbance curve, which determines the speed of increase and decrease of the disturbance term, T t is the current state temperature, T f is the initial temperature of cooling, and P is the jump probability when E(t+1)>E(t).

[0100] By introducing the intelligent parameter perturbation strategy, the perturbation of the penalty coefficient a is related to the jump-in probability and the current temperature in real time. When the current temperature is high, it means that the algorithm is in the initial optimization state, and a is given a larger change to expand its search range; when the current temperature is low, it means that the algorithm is in the final stage of optimization, and a is given a smaller change to narrow its search space. If the energy change at time t+1 is large, that is, ΔE is large, it means that when the optimal solution is about to be found, there is a certain probability of jumping out of the solution space, further increasing the amount of calculation. However, due to the existence of the perturbation slope KP, ΔE is negatively correlated with the jump probability P. Even if the solution accepts the optimal value at time t+1 with a certain probability, the intelligent parameter perturbation strategy will give a very small perturbation of the penalty factor to reduce the impact.

[0101] In order to apply the improved simulated annealing algorithm to VMD for parameter optimization, the envelope entropy is used as the internal energy function of PPSA. The envelope entropy is used to represent the sparse characteristics of the original signal, that is, there are fewer important feature information and more noise in the signal. When there is more noise and less feature information in the IMF, the envelope entropy value will be larger because the uncertainty and confusion of the system are higher. On the contrary, if there is less noise and more feature information in the IMF, the envelope entropy value will be smaller because the complexity and uncertainty of the system are reduced. Therefore, the characteristics of the original signal can be evaluated and analyzed through the value of the envelope entropy. The formula of envelope entropy is as follows:

[0102]

[0103] Where a(i) is the demodulated signal after the input signal is transformed by Hilbert, and p j The normalized form of the demodulated envelope signal, E P is the envelope entropy.

[0104] The steps of the PPSA-VMD algorithm are as follows:

[0105] Step 1: Perform VMD decomposition based on the input signal and L groups of random parameters in the solution space to obtain L groups of modal components.

[0106] Step 2: Calculate and obtain the minimum envelope entropy E of each group of modal components P .

[0107] Step 3: Apply a random perturbation that changes with temperature T to the L group of parameters in the space, and obtain the minimum envelope entropy E′ in this new solution space P .

[0108] Step 4: Calculate E based on the Metropolis algorithm P and E′ p The jump probability.

[0109] Step 5: Update the disturbance value based on the intelligent parameter disturbance strategy. If the random number ε generated in the interval (0, 1) is greater than the jump probability, return to step 3. Otherwise, go to step 6.

[0110] Step 6: After the inner loop is finished, update the current temperature value T = aT and record the optimal parameter value at this time. If the current temperature T < T f , then return to step 3 and execute a new inner loop until T>=T f , the loop ends and the optimal K and a are output.

[0111] The calculation process of PPSA-VMD is shown in Table 2:

[0112] Table 2

[0113]

[0114] A laser vibration sensor with a sampling frequency of 312.5KHz and a sampling accuracy of 0.01μm was placed horizontally at the front end of the ultrasonic welding head to measure the vibration data of the welding head during ultrasonic metal welding. The effective frequency of ultrasonic welding is between 19.5KHz and 20KHz. After signal downsampling, the number of sampling points is 4624. There is no obvious periodic impact component in the time-frequency diagram of the signal, and no obvious spectral peak can be found at the fault characteristic frequency after envelope demodulation.

[0115] In order to make the improved simulated annealing algorithm find the global optimal solution with the highest probability, its key parameters are set. The dimension of the solution space, L = 100, that is, the number of iterations of the inner loop and the number of groups of K and a under the initial conditions. A sufficiently large L can ensure that more solutions participate in the optimization process. Initial temperature T = 50, cooling coefficient a = 0.8. The higher the initial temperature, the greater the probability of obtaining a high-quality solution, but the longer the algorithm takes. The cooling coefficient is generally between 0.8-0.99, and the convergence speed is exponentially decreasing. Final temperature T f =1, when the temperature drops to the final value temperature, the simulated annealing algorithm is completed. The parameters of the intelligent parameter perturbation strategy are A=100, C=40. The parameters of VMD other than K and a are: noise tolerance tau=0, DC component DC=0, initialization center frequency init=1 obeys uniform distribution, and convergence criterion tolerance τ=1e-7.

[0116] A set of amplitude signals with welding parameters of ultrasonic metal welding, welding energy 400J, amplitude 90%, and pressure 0.45Mpa, were selected for PPSA-VMD processing. The result of PPSA optimization of VMD parameters for this set of data was that the penalty factor a=8881 and the number of decomposition layers K=10.

[0117] The original signal is decomposed and reconstructed in the time domain. Figure 6 As shown, the spectrum diagram of the original signal decomposition and reconstruction is as follows Figure 7 As shown, Figure 6 and Figure 7 The figure includes the original vibration signal, the time domain image and the frequency domain image of the reconstructed vibration signal. The red curve is the vibration curve after the modal component conforming to the ultrasonic frequency is reconstructed. The decomposition result, that is, the time domain curve of the modal component IMF after PPSA-VMD decomposition is shown in Figure 8 As shown in the figure, the amplitude data is decomposed by VMD using the parameters optimized by PPSA to obtain 10 groups of modal components IMF1-10. Since the effective working frequency of ultrasonic metal welding is 19.5KHz-20KHz, the 10 groups of modal components are subjected to fast Fourier transform, and the modal components with frequencies within this range are selected, and the decomposed amplitude data are reconstructed by weighting. The frequency domain curve of the modal component IMF after PPSA-VMD decomposition is shown in the figure. Fig. 9 As shown, Fig. 9 The figure shows 10 groups of modal components after VMD decomposition, where the red curve is the amplitude data reconstructed from the modal components with a frequency of 19.5KHz-20KHz after fast Fourier transform. It can be seen from the figure that the parameter adaptive decomposition algorithm based on PPSA-VMD can effectively decompose the vibration signal working at ultrasonic frequency, and this method can effectively remove the interference of noise.

[0118] Step S25: performing image data conversion on the reconstructed amplitude data to obtain three-dimensional image data, inputting the three-dimensional image data into a lightweight convolutional neural network for feature extraction to obtain data features, and inputting the data features into a fully connected neural network for resistance value prediction, so that the last layer of neurons in the fully connected neural network outputs the predicted resistance value.

[0119] In this embodiment, the amplitude data is decomposed and reconstructed by PPSA-VMD, and then placed in the Mobile Net V2 (lightweight convolutional neural network V2) model to extract features, and the resistance value is predicted by the fully connected network. Specifically, since the convolutional network has a good extraction ability and expressiveness for image features, the reconstructed amplitude data is further converted into a 68×68×1 three-dimensional image. The three-dimensional image data is placed in Mobile net V2 to further extract data features, and placed in a fully connected network for resistance value prediction. The last layer of neurons in the fully connected network outputs the predicted resistance value. The optimizer uses Adam, and the loss is RMSE (Root Mean Squared Error).

[0120] Step S26: evaluating the ultrasonic welding quality of the welding sample by using the resistance value of the welding sample and the predicted resistance value.

[0121] The overall structure of ultrasonic welding quality assessment based on multi-sensor data decision fusion in this application is as follows Fig.10 As shown in the figure, the whole can be divided into the following four parts: the first part, ultrasonic metal welding experiment design and data acquisition. Based on the characteristics of ultrasonic metal welding, the relevant experimental scheme is designed, and the sensor experimental data is collected. The contact resistance value of the welded sample is collected by a special instrument; the second part, the sample data set is expanded. Considering that the collected welding data may be unbalanced, a data augmentation method based on the maximum mean difference (MMD) to construct a diffusion model MMD-DDPM with diversity loss is proposed to verify the improvement of the prediction effect brought by the data augmentation model based on MMD-DDPM; the third part, in view of the non-stationarity of high-frequency amplitude data and the problem of multi-modal aliasing, an adaptive variational mode decomposition algorithm (PPSA-VMD) based on intelligent parameter perturbation simulated annealing is proposed to adaptively decompose and reconstruct the amplitude data; the fourth part, after the original data set and the generated data set are processed by PPSA-VMD, they are put into the Mobile NetV2 model for network training for the prediction of the resistance value of the welding sample.

[0122] In this embodiment, the amplitude data of the ultrasonic welding head is collected in real time, and the resistance value of the welding sample is obtained, and the amplitude data is preprocessed to obtain the preprocessed amplitude data; the preprocessed amplitude data is scaled by the maximum absolute value to generate sample diversity loss, and combined with the diffusion model, a maximum absolute value scaling-diffusion model is constructed, and the preprocessed amplitude data is amplified by using the maximum absolute value scaling-diffusion model to obtain the amplified amplitude data; the amplified amplitude data is decomposed by an adaptive variational modal decomposition algorithm based on a preset intelligent parameter perturbation simulated annealing to obtain modal components, and the target modal components are selected from the modal components for reconstruction to obtain the reconstructed amplitude data; feature extraction is performed on the reconstructed amplitude data to obtain data features, and the data features are input into a fully connected neural network to output a predicted resistance value, and the resistance value of the welding sample and the predicted resistance value are used to evaluate the ultrasonic welding quality of the welding sample. The present application preprocesses the amplitude data and performs maximum absolute value scaling to construct a maximum absolute value scaling-diffusion model, and uses the model to amplify the amplitude data, thereby solving the problem of unbalanced sample data. An adaptive variational modal decomposition algorithm of intelligent parameter perturbation simulated annealing is introduced to decompose and reconstruct the amplitude data, which can better remove the multi-modal aliasing phenomenon and noise interference in complex amplitude data, so as to solve the problem of many influencing factors. Feature extraction and resistance value prediction are performed on the reconstructed amplitude data, and the resistance value is used as an evaluation indicator of ultrasonic welding quality to realize the evaluation of ultrasonic welding quality, which can improve the accuracy of ultrasonic welding quality evaluation.

[0123] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0124] The above is a detailed introduction to a prediction method based on amplitude denoising and data amplification provided by the present invention. Specific examples are used in this article to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

Claims

1. A prediction method based on amplitude denoising and data amplification, characterized in that: include: Acquiring amplitude data of the ultrasonic welding head in real time, obtaining the resistance value of the welding sample, and performing data preprocessing on the amplitude data to obtain the preprocessed amplitude data; Performing maximum absolute value scaling on the preprocessed amplitude data to generate sample diversity loss, and combining with the diffusion model to construct a maximum absolute value scaling-diffusion model, and using the maximum absolute value scaling-diffusion model to amplify the preprocessed amplitude data to obtain the amplified amplitude data; Decomposing the amplified amplitude data based on an adaptive variational modal decomposition algorithm of preset intelligent parameter perturbation simulated annealing to obtain modal components, and selecting target modal components from the modal components for reconstruction to obtain the reconstructed amplitude data; Feature extraction is performed on the reconstructed amplitude data to obtain data features, and the data features are input into a fully connected neural network to output a predicted resistance value, and the resistance value of the welding sample and the predicted resistance value are used to evaluate the ultrasonic welding quality of the welding sample.

2. The prediction method based on amplitude denoising and data amplification according to claim 1, characterized in that: The real-time acquisition of the amplitude data of the ultrasonic welding head and the acquisition of the resistance value of the welding sample include: The Doppler laser vibrometer is used to measure and collect the amplitude data of the ultrasonic welding head in real time; The resistance value of the welding sample is measured and obtained using a loop current generator.

3. The prediction method based on amplitude denoising and data amplification according to claim 1, characterized in that: The performing data preprocessing on the amplitude data to obtain the preprocessed amplitude data includes: Normalizing the amplitude data to obtain the normalized amplitude data; Downsampling is performed on the normalized amplitude data, and padding is performed on the downsampling amplitude data to obtain the pre-processed amplitude data.

4. The prediction method based on amplitude denoising and data amplification according to claim 1, characterized in that: The maximum absolute value scaling is performed on the preprocessed amplitude data to generate sample diversity loss, and combined with the diffusion model to construct a maximum absolute value scaling-diffusion model, including: Using the maximum mean difference method to perform maximum absolute value scaling on the preprocessed amplitude data to generate a sample diversity loss; The sample diversity loss is used as the loss function of the diffusion model to construct a maximum absolute value scaling-diffusion model.

5. The prediction method based on amplitude denoising and data amplification according to claim 1, characterized in that: The amplifying the preprocessed amplitude data by using the maximum absolute value scaling-diffusion model to obtain the amplified amplitude data includes: Dividing the preprocessed amplitude data into an augmentation set and a test set according to a preset data division ratio; The preprocessed amplitude data in the amplification set is amplified using the maximum absolute value scaling-diffusion model to obtain the amplified amplitude data.

6. The prediction method based on amplitude denoising and data amplification according to claim 1, characterized in that: The adaptive variational modal decomposition algorithm based on preset intelligent parameter perturbation simulated annealing decomposes the amplified amplitude data to obtain modal components, and selects target modal components from the modal components for reconstruction to obtain the reconstructed amplitude data, including: Construct an adaptive variational mode decomposition algorithm with initial intelligent parameter perturbation simulated annealing, set intelligent parameter perturbation and variational mode decomposition parameter data, use minimum envelope entropy as fitness function, adaptively find the optimal decomposition layer number and penalty factor of different amplitude sequences, so as to generate an adaptive variational mode decomposition algorithm with intelligent parameter perturbation simulated annealing; Decomposing the amplified amplitude data using an adaptive variational modal decomposition algorithm based on the intelligent parameter perturbation simulated annealing to obtain modal components; The modal components are subjected to fast Fourier transformation, and corresponding target modal components are selected from the modal components according to a preset working frequency range of ultrasonic metal welding, and reconstructed in an accumulative manner to obtain the reconstructed amplitude data.

7. A prediction method based on amplitude denoising and data amplification according to any one of claims 1 to 6, characterized in that: The feature extraction is performed on the reconstructed amplitude data to obtain data features, and the data features are input into a fully connected neural network to output a predicted resistance value, including: performing image data conversion on the reconstructed amplitude data to obtain three-dimensional image data; Inputting the three-dimensional image data into a lightweight convolutional neural network for feature extraction to obtain data features; The data features are input into a fully connected neural network for resistance value prediction, so that the last layer of neurons in the fully connected neural network outputs the predicted resistance value.

Images