A Stress Detection Method for Aluminum Alloy Thick Plates Based on the PSO-GSA-GRNN Model

Abstract

The present invention discloses a stress detection method for thick aluminum alloy plates based on a PSO-GSA-GRNN model, including: S1: Select materials, process the sample blanks into tensile specimens, and after treating the sample blanks, place them in a material testing machine for tensile experiments. Use ultrasonic waves to measure the acoustic time difference at different temperatures in a constant temperature bath and calculate the stress; S2: Establish a stress detection model, incorporate the dendritic neural network (DD) and the acoustic time difference parameter of traditional ultrasonic detection into the deep learning model, and use temperature, tensile rate, and acoustic time difference as inputs and the true stress as the output to establish the dendritic neural network expression. This stress detection method for thick aluminum alloy plates based on the PSO-GSA-GRNN model: The ultrasonic measurement results of 7065 aluminum alloy at different temperatures vary greatly and this variation is non-linear. Conducting error compensation based on the conventional linear fitting method will cause large errors, but the acoustic time difference changes of aluminum alloy templates with different tensile rates show similar change trends.

Description

Technical Field

[0001] The present invention relates to the technical field of aluminum alloy thick plate detection, and specifically to a stress detection method for aluminum alloy thick plates based on a PSO-GSA-GRNN model. Background Art

[0002] In the stress non-destructive testing method, the ultrasonic testing method has the characteristics of widespread use, high speed and low cost. And because ultrasonic waves have strong penetrability [1], at the same time, the ultrasonic method can continuously detect the macroscopic stress of the entire plate, and is very suitable for stress detection of thick plates.

[0003] The traditional ultrasonic testing method is based on the acoustoelastic theory, and calculates the stress by using the time difference between ultrasonic wave emission and reception. And it discloses the influence of different ultrasonic powers on the maximum contact stress and springback, and proposes a mathematical model for predicting the maximum contact stress and springback, and predicts the maximum contact stress and depth of ultrasonic surface rolling. The linear ultrasonic testing method is used to detect and analyze the transverse stress of 7050 thick plates in different processing states and different thicknesses. By using the time of longitudinal wave propagation along the stress direction, and the propagation times of transverse waves with polarization directions perpendicular and parallel to the stress direction, the stress value is characterized and a residual stress test system is established. In order to improve the detection accuracy, a calibration method for stress coefficients is proposed and applied to detection in multiple fields, proving that the ultrasonic detection technology has high practicability and wide application.

[0004] Regarding the influence of factors such as temperature on ultrasonic stress measurement, Zhang Hongbo [7] et al. studied the influence of different temperatures on ultrasonic testing, and then detected the defects of Q235A test blocks. As the temperature increases, the overall measurement error gradually increases. Chen Gang [8] et al. adopted a power exponential equation containing parameters to describe the influence of temperature and strain rate on rheological behavior. By using the method of linear regression analysis, the variation law of material constants with strain under different temperatures and strain rates was studied. Zhao Zhiran [9] et al. established an ultrasonic detection model of carbon steel under uniform temperature field and non-uniform temperature field, indicating that when the temperature of metal materials increases in the uniform temperature field and the temperature difference increases in the non-uniform temperature field, the error of defect location will increase. Finally, a regression fitting is used to establish a correction formula for ultrasonic defect detection under a specific temperature field.

[0005] The above-mentioned literature mainly tests and improves the ultrasonic stress measurement method, and uses the method of linear fitting for stress calculation. The complex mathematical expressions are difficult to calculate, and the simple ideal physical model is difficult to improve the model accuracy. In recent years, deep learning methods represented by neural networks have made rapid progress and have been applied in various fields, solving many problems that are difficult to solve by traditional calculations. Summary of the Invention

[0006] The purpose of the present invention is to provide a stress detection method for thick aluminum alloy plates based on the PSO-GSA-GRNN model to solve the problems presented in the above-mentioned background technology.

[0007] To achieve the above object, the present invention provides the following technical solution: A stress detection method for thick aluminum alloy plates based on the PSO-GSA-GRNN model, including;

[0008] S1: Select materials, process the test blanks into tensile specimens, and after treating the test blanks, place them in a material testing machine for tensile tests. Measure the acoustic time difference using ultrasonic waves at different temperatures in a constant temperature bath and calculate the stress.

[0009] S2: Establish a stress detection model. Incorporate the dendritic neural network (DD) and the acoustic time difference parameter of traditional ultrasonic detection into the deep learning model. Use temperature, tensile rate, and acoustic time difference as inputs and the true stress as the output to establish the expression of the dendritic neural network. Then normalize the data. Compare the established model with the deep learning model that does not incorporate the acoustic time difference parameter of traditional ultrasonic detection, and at the same time compare and analyze it with the BP neural network that does not introduce the dendritic information processing ability. The model error performance uses the root mean square error RMSE.

[0010] S3: Establish a temperature compensation model. Incorporate the particle swarm algorithm with the ability to exchange information between groups into the gravitational search algorithm, and construct the PSO-GSA-GRNN model. Perform temperature compensation on the values detected for aluminum alloy at different temperatures through GSA-GRNN.

[0011] Preferably, in S1, the test blanks are treated by quenching at 475°C / 1.5h followed by water cooling.

[0012] Preferably, the expression of the dendritic neural network model in S2 is:

[0013]

[0014] Preferably, the architecture of the multi-layer DD in S2 is:

[0015]

[0016] In the multi-layer DD, the temperature matrix X 0 and the tensile rate X 1 are used as inputs to establish the expression of the dendritic neural network:

[0017]

[0018] Preferably, the specific construction steps of the PSO-GSA-GRNN model:

[0019] Step 1: Divide the data into two groups as the training set and the test set respectively; initialize the particle swarm, and use the mean error of the particles as the fitness.

[0020] Step 2: Use the improved GSA algorithm for iterative optimization to find the particle with the best fitness, and retrain the GRNN to calculate the fitness using the optimal particle in the next generation. Repeat the cycle until reaching the stopping requirement.

[0021] Step 3: Verify the PSO-GSA-GRNN model with the test set, and calculate the correlation coefficient, relative error, etc.

[0022] Compared with the prior art, the beneficial effects of the present invention are that the stress detection method for aluminum alloy thick plates based on the PSO-GSA-GRNN model:

[0023] 1. The ultrasonic measurement results of 7065 aluminum alloy vary greatly at different temperatures and this variation is non-linear. Error compensation based on the conventional linear fitting method will cause large errors, but the change trends of the acoustic time difference of aluminum alloy templates with different tensile rates are similar.

[0024] 2. A stress detection model for 7065 aluminum alloy is established through the method of deep learning. The neural network introducing the dendritic information processing ability has better accuracy compared with the BP neural network. The root mean square error and correlation coefficient of the detection method integrating the deep learning method and the traditional ultrasonic detection method are 0.84636 and 0.99743 respectively, and the effect is better than directly using the deep learning method for stress detection. At the same time, the method integrating the traditional physical model has better interpretability.

[0025] 3. Temperature compensation is performed on the detected values of 7065 aluminum alloy at different temperatures through the improved GSA-GRNN. The root mean square error and correlation coefficient of the compensated stress can reach 0.78848 and 0.99844 respectively, and the accuracy is 0.4 and 0.02 higher than that of the model before improvement. There is an obvious improvement in accuracy at a tensile rate of 1.0% - 1.5%. The model also has high accuracy at other tensile rates. Description of the Drawings

[0026] Figure 1 It is a schematic diagram of the neural network structure of the present invention;

[0027] Figure 2 It is a flow chart of PSO-GSA-GRNN of the present invention;

[0028] Figure 3 It is a diagram of the fitting process of the dendritic neural network of the present invention;

[0029] Figure 4 It is a diagram of the fitting process of the dendritic neural network of the present invention;

[0030] Figure 5 This is the temperature compensation comparison chart of the present invention. Specific Embodiments

[0031] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0032] Embodiment 1: Select a 7065 plate specimen with a thickness of 19 mm, cut it into 8 tensile specimen blanks and process them into tensile specimens. All specimens are treated by quenching at 475 °C / 1.5 h and water cooling. Tensile tests are carried out on an E45 material testing machine (300 kN). The 8 specimens are respectively pre-tensioned from 0 to 3.5% (at 0.5% intervals), and then the time difference of sound is measured by ultrasonic waves at different temperatures in a constant temperature bath and the stress is calculated. Among them, the temperature is measured once every 1 °C from 10 to 40 °C, and the reverse cooling is from 40 - 25 °C, and the measurement is taken once every 1 °C. The calibration coefficient of the time difference of sound is carried out at room temperature of 24 °C.

[0033] Finally, a total of 46×8 groups of data are measured for the 8 specimens, and the eight groups of data are put into a stacked line chart. It can be seen that due to the different propagation speeds of the ultrasonic time difference of sound at different temperatures, the change of the measured stress size at different temperatures is not stable, but the change trends of the stress measured by the eight specimens at different temperatures are similar. Among them, at the calibration temperature of 24 °C, when the pre-tension amount is 2.5%, the measured stress is 0. Therefore, in the data preprocessing stage, the stress changes measured at each temperature based on the 2.5% pre-tension amount are compensated for each pre-tension amount to reduce the error caused by ultrasonic measurement.

[0034] Embodiment 2:

[0035] Add the dendritic information processing mechanism to the neural network model to construct a dendritic neural network model, which is expressed as:

[0036]

[0037] In the formula, o represents the Hadamard product, X is the vector representation form of the input value, DD is similar to the BP neural network structure, and DD performs the Hadamard product with the original data before output.

[0038] Among them, the architecture of the multi-layer DD is as follows:

[0039]

[0040] Wherein, L and L-1 represent from the (L-1)th weight module to the Lth weight module. The input data is the weight matrix from the first layer to the nth layer of the neural network.

[0041] In this formula, the temperature matrix X 0 and the elongation rate X 1 are used as inputs to establish the dendritic neural network expression:

[0042]

[0043] Wherein, the initialization of the weight matrix follows a uniform distribution from 0 to 1. Each layer of the dendritic neural network will be continuously trained based on the white-box test method using Y through forward propagation and error backpropagation until the stop requirement is met.

[0044] Using the dendritic neural network, the acoustic time difference parameter of traditional ultrasonic detection is incorporated into the deep learning model. Temperature, elongation rate, and acoustic time difference are used as inputs, and the true stress is used as the output to establish the dendritic neural network expression. There are a total of 322 pieces of data in this experiment. 70% of them are extracted as the training set, and 30% are used as the test set. Then, the data is normalized. The established model is compared with the deep learning model that does not incorporate the acoustic time difference parameter of traditional ultrasonic detection, and at the same time, it is compared and analyzed with the BP neural network that does not introduce the dendritic information processing ability. The model error performance is processed using the root mean square error.

[0045] Example 3:

[0046] Introduce the particle swarm optimization algorithm (PSO) with the ability of information exchange between groups to improve the velocity and displacement formulas of the gravitational search algorithm.

[0047] Since the particle position update of the GSA algorithm only considers the current position, after introducing the PSO algorithm, the particle motion formula becomes:

[0048]

[0049] Wherein: is the current global optimal solution; ω is the inertia weight; c1 and c2 are learning factors; r1 and r2 are random numbers between 0 and 1, and xdi(t) is the position of particle i at the t-th generation. Among them, the inertia parameter ω adopts an adaptive transformation strategy as follows:

[0050]

[0051] Among them, the max and min subscripts respectively represent the maximum and minimum values of the inertia factor, f i t i$(t)$ is the error of particle $i$ at the $t$-th iteration. $best(t)$ and $worst(t)$ are the minimum error and maximum error at the $t$-th iteration respectively. In PSO-GSA, there is a current optimal solution to guide the particles with large inertial mass to gather towards the global optimal direction. Therefore, the improved GSA can accelerate the population movement and has stronger optimization ability.

[0052] Using PSO-GSA to optimize GRNN is to find the smoothing factor $\sigma$ suitable for GRNN. Select the particle with the best effect in each generation of particles. Based on this, the optimal particle finally iterated contains this important parameter of the smoothing factor. The specific steps to construct the PSO-GSA-GRNN model are as follows:

[0053] Step 1: Divide the data into two groups as the training set and the test set respectively; initialize the particle swarm, and use the mean error of the particles as the fitness.

[0054] Step 2: Use the improved GSA algorithm for iterative optimization to find the particle with the best fitness. In the next generation, use the optimal particle to retrain GRNN to calculate the fitness, and repeat the cycle until it stops when reaching the required limit.

[0055] Step 3: Use the test set to verify the PSO-GSA-GRNN model, and calculate the correlation coefficient, relative error, etc.

[0056] Take the elongation rate, temperature, and stress predicted by the dendritic neural network as the inputs, and the true stress as the output, with a total of 322 data. Extract 70% as the training set and 30% as the test set, and then normalize the data. The number of GSA particles is 50, and the iteration is 100 times. The model error performance uses the root mean square error RMSE.

[0057] The overall experimental model of this study is as Figure 5 shown. First, introduce the important parameter of ultrasonic detection in the traditional physical model, the acoustic time difference, and conduct a preliminary prediction of the stress based on the dendritic neural network. Then, input the predicted stress value into the improved GSA-GRNN temperature compensation model. Finally, output the resulting stress and conduct model performance evaluation and error analysis.

[0058] To sum up: The cell structure of the dendritic neural network is still based on the BP neural network. Therefore, in this study, the errors of the BP neural network and the dendritic neural network are compared in the experiment. Figure 3 For the fitting convergence process of the dendritic neural network, it can be found that the dendritic neural network can converge to a very small error and can converge to the global optimum with a high probability.

[0059] In the experiment, a 3-layer neural network model was adopted for the BP neural network, and the number of neurons in the hidden layer was 2n + 1 (n is the number of input parameters). The dendritic neural network also adopted a 3-layer network structure. The root mean square error (RMSE) of the BP neural network without introducing the ultrasonic time difference detection method was 2.8809, and the correlation coefficient was 0.96785. The root mean square error of the dendritic neural network was 1.1549, and the correlation coefficient was 0.99494. The root mean square error of the BP neural network with the ultrasonic time difference detection method introduced was 2.2679, and the correlation coefficient was 0.97959. The root mean square error of the dendritic neural network was 0.84636, and the correlation coefficient was 0.99743.

[0060] The results show that the fusion detection method of deep learning and traditional ultrasonic detection has better interpretability and accuracy than the deep learning method based entirely on data. The dendritic neural network that introduces the dendritic information processing ability has better prediction ability than the BP neural network. The detection method of the fusion of the dendritic neural network and traditional ultrasonic detection has the best accuracy, and the results obtained by this method are temperature compensated.

[0061] The stress obtained from the fusion model was used as the measured stress and input into three temperature compensation models respectively. The individual learning factor was set to 0.3, and the social learning factor was set to 0.5. It can be seen from the iterative process of the three temperature compensation models that the PSO-GRNN no longer decreased when the error dropped to 1.196. The main reason is that the PSO algorithm is prone to falling into local optimum and the PSO as a local optimum algorithm has a slow convergence speed. GSA can converge faster, but its ability to find local optimum is relatively weak. Therefore, the final error of the GRNN optimized by GSA is slightly higher than that of the GRNN optimized by PSO-GSA. The error of the improved PSO-GRNN temperature compensation model is smaller after iterative convergence.

[0062] The root mean square error of PSO-GRNN was 1.196, and the correlation coefficient was 0.99687. The root mean square error of GSA-GRNN was 0.80423, and the correlation coefficient was 0.99838. The root mean square error of PSO-GSA-GRNN was 0.78848, and the correlation coefficient was 0.99844. The low accuracy of the temperature compensation model based on traditional deep learning under partial tensile rates is the main reason for the high error, and the improved model effectively reduces this part of the error.

[0063] For the improved GSA-GRNN temperature compensation model, most of the data errors are less than 0.02 MPa, and the stress compensation error at the tensile rate of 1.0% - 1.5% is significantly reduced. Therefore, this model has good temperature compensation performance. The comparison of the evaluation indexes of each model is shown in Table 1. It can be seen that the temperature compensation model of GSA-GRNN improved by PSO has certain advantages in all evaluation indexes.

[0064] Although embodiments of the present invention have been shown and described, those of ordinary skill in the art will appreciate that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A stress detection method for aluminum alloy thick plate based on PSO-GSA-GRNN model, characterized in that: The following steps are involved: S1: Select materials, select samples to be processed into tensile specimens, and put them into the material testing machine for tensile test after processing. Use ultrasonic wave to measure the acoustic time difference at different temperatures in the constant temperature bath and calculate the stress; S2: Establish a stress detection model, integrate the dendritic neural network (DD) and the acoustic time difference parameters of traditional ultrasonic testing into the deep learning model, and use temperature, stretch rate, and acoustic time difference as inputs, and the true stress as output. Establish the dendritic neural network expression, and then normalize the data. Compare the established model with the deep learning model that does not integrate the acoustic time difference parameters of traditional ultrasonic testing, and compare and analyze it with the BP neural network that does not introduce dendritic information processing capabilities. The model error performance uses the root mean square error RMSE; S3: Establish a temperature compensation model, add the particle swarm algorithm with the ability of information exchange between groups to the gravitational search algorithm, and construct a PSO-GSA-GRNN model, and use GSA-GRNN to perform temperature compensation on the values ​​detected by aluminum alloy at different temperatures; The structure of the multi-layer DD in step S2 is: Where X is the vector representation of the input value, W is the weight matrix, o represents the Haddam product, and L, L-1 represents from the L-1th weight module to the Lth weight module; In the multi-layer DD, the temperature matrix X0 and the stretching rate X1 are used as input to establish the dendritic neural network expression: Where, the temperature matrix is ​​X0 and the stretching rate is X1; The specific construction steps of the PSO-GSA-GRNN model are: Step 1: Divide the data into two groups as training set and test set respectively; initialize the particle swarm and take the mean error of the particles as the fitness; Step 2: Use the improved GSA algorithm to iteratively search for the best particle, and use the best particle to retrain the GRNN to calculate the fitness in the next generation, and repeat the cycle until the end requirement is met; Step 3: Use the test set to verify the PSO-GSA-GRNN model and calculate the correlation coefficient and relative error.

2. The stress detection method for aluminum alloy thick plate based on PSO-GSA-GRNN model according to claim 1 is characterized in that: The sample processing method in step S1 is to quench at 475° C. / 1.5 h and cool with water.

Images