Wind power high-strength bolt axial stress detection method based on GA-BP neural network

Abstract

The invention discloses a wind power high-strength bolt axial stress detection method based on a GA-BP neural network, and the method comprises the following steps: 1, selecting high-strength bolts with different isopolarities and lengths, and obtaining an ultrasonic transverse and longitudinal wave sound-time ratio and loading stress data; 2, establishing a BP neural network structure between the ultrasonic transverse and longitudinal sound time ratio and the three parameters; 3, randomly distributing the original data to different groups; 4, optimizing the BP neural network model through a GA (Genetic Algorithm); 5, forming a GA-BP neural network model; and 6, substituting GA-BP neural network data into a bolt axial stress formula to obtain the axial stress of the high-strength bolt. According to the detection method, a GA-BP neural network is established to accurately predict the three parameters of the bolt and the ultrasonic transverse and longitudinal wave sound-time ratio, the accurate numerical value of the axial stress of the bolt is obtained through calculation, and then the accurate numerical value is compared with the axial stress obtained by actually measuring the ultrasonic transverse and longitudinal wave sound-time ratio, so that whether the high-strength bolt fails or not is judged. And safety accidents are avoided.

Description

Technical Field

[0001] The present invention relates to the technical field of nondestructive testing, in particular to a method for detecting the axial stress of bolts based on the acoustic time ratio of ultrasonic transverse and longitudinal waves. Background Art

[0002] As a representative of clean energy, China's new energy wind power generation industry is in a stage of high-quality and high-speed development, and the quality and safety issues of wind turbine generators have attracted increasing attention. High-strength bolts, as key connectors, are under high load stress for a long time. Especially under the influence of high pre-tightening force, fatigue cracks will occur, which may lead to serious consequences. Therefore, it is very necessary to regularly conduct safety inspections on high-strength bolts. High-strength bolt connections clamp the connected plates by applying high pre-tightening force inside the bolt rod, and thus generate strong frictional force; if the axial stress of the bolt decreases, it may cause the connection to fail, and in severe cases, it may also cause the instability of the structure, thus triggering safety accidents.

[0003] The ultrasonic stress detection technology is a rapid and non-destructive stress measurement method. Based on the acoustoelastic principle (the ultrasonic wave velocity changes with the stress magnitude), it establishes the mathematical relationship between elastic stress and the velocities of ultrasonic transverse and longitudinal waves through experiments, and then measures the value of elastic stress. When the ultrasonic stress detection is applied to high-strength bolts, the axial stress of high-strength bolts is indirectly measured by measuring the acoustic time when the ultrasonic wave velocity reflects back to the end face inside the bolt. However, the relationship between ultrasonic signals and axial stress is very complex, and different materials and specifications will affect the results, making it difficult to fit with a simple function.

[0004] Patent CN117077441A discloses a method for calculating the axial stress of bolts based on double waves. Its disadvantages are: it does not consider that different bolt grades and lengths will lead to different acoustic time ratios of ultrasonic transverse and longitudinal waves, and different parameter combinations will have different effects on the ultrasonic transverse and longitudinal wave ratios. Therefore, the accuracy of the axial stress of bolts calculated by this method is not high; at the same time, if a large number of experiments are carried out to find the relationship between the acoustic time ratios of ultrasonic transverse and longitudinal waves of different bolts and the axial stress of bolts, it is not only costly but also time-consuming and laborious. Summary of the Invention

[0005] The problem to be solved by the present invention is to provide a method for detecting the axial stress of high-strength bolts for wind power generation based on a GA-BP neural network. First, a GA-BP neural network is established to accurately predict the three parameters of bolt loading stress, grade, and length and the acoustic time ratio of ultrasonic transverse and longitudinal waves, and the accurate value of the bolt axial stress is calculated. Then, it is compared with the axial stress obtained from the measured acoustic time ratio of ultrasonic transverse and longitudinal waves to determine whether the high-strength bolt fails and avoid safety accidents.

[0006] A method for detecting the axial stress of high-strength bolts for wind power based on a GA-BP neural network includes the following steps: Step 1: Select high-strength bolts of different grades and lengths, obtain the ultrasonic shear-longitudinal wave time ratio and loading stress data by stretching with an ultrasonic detector and a tensile testing machine, draw a scatter plot, and perform linear fitting. Use the fitted data set as the sample set for training the BP neural network; Step 2: According to the influence results of different loading stresses, bolt grades and lengths on the ultrasonic shear-longitudinal wave time ratio, establish a BP neural network structure between the ultrasonic shear-longitudinal wave time ratio and the above three parameters; Step 3: When using the BP neural network, randomly assign the original data to different groups and perform data normalization processing; Step 4: Optimize the BP neural network model through the GA genetic algorithm, using floating-point coding, fitness function, roulette wheel selection, two-point crossover and adaptive mutation operator to improve the prediction accuracy of the ultrasonic shear-longitudinal wave time ratio and the three parameters; Step 5: Test and train the BP neural network through the GA algorithm, and at the same time screen and optimize the thresholds and initial weights of the neural network to form a GA-optimized BP neural network model; Step 6: Substitute the obtained GA-BP neural network data into the bolt axial stress calculation formula for calculation to obtain the axial stress of the bolt.

[0007] Among them, Step 1 specifically includes the following steps: Step 1.1: Tensile test preparation: First, polish the upper and lower end faces of the bolt, and then calibrate the bolt; Step 1.2: Tensile test: For the bolt placed on the tensile testing machine, load it at a rate of 5 mm / min. Before reaching the yield strength, hold the load every 100 MPa, and hold the load every 50 MPa after reaching the yield strength, and end before reaching the tensile strength; The holding time is 45 s; Use an ultrasonic detector to record the ultrasonic shear-longitudinal wave time ratio data when the bolt is holding the load; Step 1.3: Use data processing software to draw a scatter plot of the shear-longitudinal wave time ratio and the loading stress, and perform linear fitting on the data using the formula: y = B0 + B1x, where B0 and B1 are the coefficients of each term in the above formula.

[0008] Among them, in Step 1.1: The specific steps of calibration are: Place the bolt on the tensile testing machine and apply tensile force. The stresses applied five times are 0.12, 0.24, 0.36, 0.48 and 0.6 times of σ' (σ' is the theoretical yield strength of the test bolt). Hold the load for 45 seconds after each loading, and take the average of any three values during the holding time as the hydraulic press value.

[0009] Among them, step 2 specifically includes the following steps: Step 2.1: In the tensile test, the single control variable method is adopted to study the influence of bolt grade and length on the ratio of longitudinal and transverse wave acoustic time when applying different tensile forces. The specific ranges of its parameters are: grade 8.8, length 150 mm, grade 8.8, length 200 mm, grade 10.9, length 200 mm, grade 10.9, length 250 mm, grade 12.9, length 150 mm, grade 12.9, length 200 mm; Step 2.2: Build a three-layer BP neural network (input layer, hidden layer, output layer) on the programming software: Take the three variables of applied stress (kN), bolt grade, and bolt length (mm) as the input layer data (corresponding to Figure 2 the three neurons in the input layer of the BP network in Figure 2 ), and the ratio of longitudinal and transverse wave acoustic time as the output layer data (corresponding to one neuron in the output layer of the BP network in ), and establish the BP neural network structure of different bolts and the ratio of longitudinal and transverse wave acoustic time. The formula of the BP prediction model for the ratio of longitudinal and transverse wave acoustic time is: In the formula, the weight refers to the connection weight between the nth node in the jth layer and the mth node in the ith layer, and the threshold represents the activation threshold of the qth node in the pth layer; Step 2.3: In the training process of the BP neural network, use the mean square error function to evaluate the accuracy of the neural network prediction: In the formula, m is the number of mini-batch samples, y i represents the predicted value made by the model for the ith sample, and y i is the actual value corresponding to the ith sample; Step 2.4: In the sample training of the BP neural network, select the Gauss-Newton method as the most suitable training algorithm; Step 2.5: Determine the optimal number of nodes in the hidden layer of the BP neural network. The number of nodes is determined according to the following formula: In the formula, n1 is the number of nodes in the hidden layer, n is the number of neurons in the input end, m is the number of neurons in the output end, and a is a constant between 1 and 10.

[0010] Among them, in step 2.2: The hidden layer uses the non-linear tansig function as its activation function, and the output layer uses the purelin function for identity mapping:

[0011] Among them, step 3 specifically includes the following steps: Step 3.1: Randomly select and divide the data as follows: 70% of the ratio data of longitudinal and transverse wave acoustic time constitutes the training set, 15% of the ratio data of longitudinal and transverse wave acoustic time constitutes the validation set, and the remaining 15% is reserved for testing; Step 3.2: The input data of the BP neural network is standardized through the following formula to ensure the consistency of the data set: In the formula, x is the original data set of the input, xnorm is the dataset after standardization processing, x max and x min represent the maximum and minimum values in the original dataset respectively.

[0012] Among them, step 4 specifically includes the following steps: Step 4.1, the GA algorithm provides an encoding solution for setting the thresholds and initial weights of the BP neural network through floating-point encoding: L = S1×S2 + S2 + S2×S3 + S3, where L is the total length of the individual encoding, S1 is the number of nodes in the input layer, S2 is the number of nodes in the hidden layer, and S3 is the number of nodes in the output layer; Step 4.2, use the reciprocal of the mean square error function as the fitness function to evaluate the probability that an individual in the population is inherited into the next generation population; Step 4.3, the process of individuals in the population being selected into the next generation is realized through the roulette wheel algorithm, and the probability of an individual being selected is: where M is the total number of individuals in the population, f i is the fitness value of individual i, and p i is the probability of the individual being selected; Step 4.4, in the GA algorithm, randomly select two parent individuals, and use the two-point crossover algorithm of randomly selecting two gene positions for exchange to generate two offspring individuals with genetic characteristics; Step 4.5, use the adaptive mutation operator in the creation of new individuals, and the algorithm is: where f is the fitness value of the mutant individual, f max is the maximum fitness value in the population, f avg is the fitness value of the entire population, p m1 and p m2 are set to 0.1 and 0.01 respectively.

[0013] Among them, the model of the GA-BP genetic neural network obtained in step 5 is:

[0014] y1 = -0.65995*tansig(-3.1602*x1 - 1.3748*x2 - 0.28143*x3 + 2.87) -1.5792*tansig(-0.18161*x1 + 3.7103*x2 + 0.19573*x3 - 3.1665) -1.447*tansig(2.2003*x1 - 0.70863*x2 + 0.2737*x3 - 1.7175) +0.4517*tansig(3.2792*x1 + 0.68242*x2 - 0.34785*x3 - 1.4032) -0.00030917 * tansig(0.055479 * x1 - 1.4917 * x2 + 2.6305 * x3 - 0.27168) +3.2778 * tansig(0.22129 * x1 + 0.15112 * x2 - 0.1873 * x3 + 0.20342) -0.19163 * tansig(2.0896 * x1 - 2.2765 * x2 - 0.48421 * x3 + 1.8166) +0.68537 * tansig(-0.2008 * x1 - 3.4 * x2 - 0.2871 * x3 - 0.7213) +0.66295 * tansig(-1.5088 * x1 + 2.6315 * x2 + 0.04775 * x3 - 1.665) +0.00088341 * tansig(0.72797 * x1 - 2.2588 * x2 - 2.3195 * x3 + 0.7726) +0.0036288 * tansig(-1.2528 * x1 - 0.59376 * x2 - 1.7108 * x3 - 1.656) +0.011317 * tansig(1.4017 * x1 + 0.41311 * x2 - 2.1459 * x3 + 2.5602) -1.2677

[0028] In the model, y1 is the prediction model of the longitudinal and transverse wave acoustic time ratio of different bolts, and x1, x2, and x3 correspond to three variables: bolt grade, bolt length (mm), and applied stress (kN), respectively.

[0029] Among them, the calculation formula for the axial stress of the bolt in step 6 is as follows: In the formula, σ is the axial stress, and C R is the acoustoelastic coefficient of the sound velocity ratio, are the acoustic times of the ultrasonic longitudinal wave and transverse wave under stress-free conditions, respectively, is the ratio of the ultrasonic longitudinal and transverse wave acoustic times; among them, C R is related to the material of the bolt, is related to the material and precision of the bolt and is a constant.

[0030] The advantages of the detection method of the present invention are as follows: By establishing a GA-BP neural network model, it is possible to accurately predict the ultrasonic transverse-longitudinal wave acoustic time ratio based on three parameters of the high-strength bolt, namely the loading stress (kN), grade, and length (mm), and then obtain the axial stress of the high-strength bolt. In this way, when the high-strength bolt is installed, according to the three parameters of the loading stress (kN), grade, and length (mm), the transverse-longitudinal wave acoustic time ratio is obtained, and then the axial stress of the high-strength bolt is calculated and used as the reference value during normal operation. After the high-strength bolt has worked for a period of time, the axial stress of the high-strength bolt is calculated by the measured ultrasonic transverse-longitudinal wave acoustic time ratio. The bolt axial stress at this time is compared with the bolt axial stress during the installation stage. When it exceeds the safety threshold, it is confirmed that it may fail, and thus it is replaced. Description of the Drawings

[0031] Figure 1 is a schematic diagram of the ultrasonic double-wave method tensile test.

[0032] Figure 2 is a schematic diagram of a three-layer BP network of the ultrasonic transverse-longitudinal wave acoustic time ratio.

[0033] Figure 3 is a comparison chart of the predicted value and the true value of the transverse-longitudinal wave acoustic time ratio under the GA-BP and BP network models.

[0034] Figure 4 is a correlation chart for evaluating the predicted value and the test value of the transverse-longitudinal wave acoustic time ratio by the GA-BP network model. Detailed Embodiment

[0035] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be described in detail below in conjunction with the relevant drawings and embodiments. However, the application of the present invention is not limited to these specific embodiments.

[0036] Embodiment 1

[0037] A method for detecting the axial stress of high-strength bolts for wind power based on a GA-BP neural network according to the present invention includes the following steps:

[0038] Step 1: Select high-strength bolts of different grades and lengths, obtain the ultrasonic transverse-longitudinal wave acoustic time ratio and loading stress data through an ultrasonic detector and a tensile testing machine (tensile test and ultrasonic double-wave method), draw a scatter plot, and perform linear fitting. The fitted data set is used as the sample set for training the BP neural network.

[0039] Step 2: According to the influence results of different loading stresses, bolt grades, and lengths on the ultrasonic transverse-longitudinal wave acoustic time ratio, establish a BP neural network structure between the ultrasonic transverse-longitudinal acoustic time ratio and the above three parameters.

[0040] Step 3: When using the BP neural network, randomly assign the original data to different groups and perform data normalization processing;

[0041] Step 4: Optimize the BP neural network model through the GA genetic algorithm, and use floating-point coding, fitness function, roulette wheel selection, two-point crossover, and adaptive mutation operator to improve the accuracy of predicting the ultrasonic longitudinal and transverse wave time ratio and three parameters;

[0042] Step 5: Test and train the BP neural network through the GA algorithm, and simultaneously screen and optimize the thresholds and initial weights of the neural network to form a BP neural network model optimized by GA;

[0043] Step 6: Substitute the obtained GA-BP neural network data into the bolt axial stress calculation formula for calculation to obtain the axial stress of the bolt.

[0044] Example 2

[0045] Specifically, Step 1 is as follows:

[0046] Step 1.1: Tensile test preparation: First, polish the upper and lower end faces of the bolt, and then calibrate the bolt;

[0047] Step 1.2: Tensile test: According to Figure 1 For assembly, load the bolt placed on the tensile testing machine at a rate of 5 mm / min. Before reaching the yield strength, hold the load every 100 MPa, and hold the load every 50 MPa after reaching the yield strength, and end before reaching the tensile strength; the holding time is 45 s; when the bolt is holding the load, use an ultrasonic detector to record the ultrasonic longitudinal and transverse wave time ratio data;

[0048] Step 1.3: Use data processing software (Origin) to draw a scatter plot of the longitudinal and transverse wave time ratio and the applied stress, and linearly fit the data using the formula: y = B0 + B1x, where B0 and B1 are the coefficients of each term in the above formula.

[0049] The scatter plot directly reflects the relationship between two variables through the distribution of independent points, and can clearly show the overall density, dispersion degree, and potential trend of the data.

[0050] Example 3

[0051] In Step 1.1: Polish the upper and lower end faces of the bolt to ensure receiving ultrasonic echo signals.

[0052] In Step 1.1, the specific calibration steps are as follows: Place the bolt on a tensile testing machine and apply tensile force. The stresses for the five loadings are 0.12, 0.24, 0.36, 0.48, and 0.6 times of σ’ (where σ’ is the theoretical yield strength of the test bolt). After each loading, hold the load for 45 seconds. During the holding time, randomly select three values and calculate the average, which is used as the value for the hydraulic press.

[0053] Example 4

[0054] Step 2 is specifically as follows:

[0055] In Step 2.1, the single control variable method is used in the tensile test to study the influence of bolt grade and length on the ratio of longitudinal and transverse wave acoustic time when applying different tensile forces. The specific ranges of its parameters are: grade 8.8, length 150 mm, grade 8.8, length 200 mm, grade 10.9, length 200 mm, grade 10.9, length 250 mm, grade 12.9, length 150 mm, grade 12.9, length 200 mm;

[0056] In Step 2.2, on the programming software (MatLab), establish a three-layer BP neural network (input layer, hidden layer, output layer): As Figure 2 shown, take the three variables of applied stress (kN), bolt grade, and bolt length (mm) as the input layer data (corresponding to the three neurons in the input layer of the BP network in Figure 2 ), and take the ratio of longitudinal and transverse wave acoustic time as the output layer data (corresponding to one neuron in the output layer of the BP network in Figure 2 ), and establish a BP neural network structure for different bolts and the ratio of longitudinal and transverse wave acoustic time. The formula for the BP prediction model of the ratio of longitudinal and transverse wave acoustic time is:

[0057]

[0058] In the formula, the weight refers to the connection weight between the nth node in the jth layer and the mth node in the ith layer, and the threshold represents the activation threshold of the qth node in the pth layer;

[0059] In Step 2.3, during the training process of the BP neural network, use the mean square error function to evaluate the accuracy of the neural network prediction: In the formula, m is the number of small batch samples, y i represents the predicted value of the model for the ith sample, and y i is the actual value corresponding to the ith sample;

[0060] In Step 2.4, in the sample training of the BP neural network, select the Gauss-Newton method as the most suitable training algorithm;

[0061] Step 2.5: Determine the optimal number of nodes in the hidden layer of the BP neural network. The number of nodes is determined according to the following formula: In the formula, n1 is the number of nodes in the hidden layer, n is the number of neurons in the input layer, m is the number of neurons in the output layer, and a is a constant between 1 and 10.

[0062] Example 5

[0063] In Step 2.2: The hidden layer uses the non-linear tansig function as its activation function, and the output layer uses the purelin function for identity mapping:

[0064]

[0065] tansig'(x) = tansig(x)·(1 - tansig(x))

[0066] purelin(x) = x

[0067] In Step 2.5: The neural network has the best effect when a is 3. Therefore, the number of nodes n1 in the corresponding hidden layer is preferably 5.

[0068] Example 6

[0069] Step 3: When using the BP neural network, randomly assign the original data to different groups and perform data normalization processing to enhance the stability of the BP neural network training process and prevent unnecessary fluctuations in the gradient.

[0070] Step 3 is specifically as follows:

[0071] Step 3.1: To optimize, improve, and verify the BP neural network model and ensure its good generalization performance on new data, randomly select the data for the following division: 70% of the shear-to-compressional wave travel time ratio data constitutes the training set, 15% of the shear-to-compressional wave travel time ratio data constitutes the validation set, and the remaining 15% is reserved for testing;

[0072] Step 3.2: The input data of the BP neural network is standardized through the following formula to ensure the consistency of the data set: In the formula, x is the original data set of the input, x norm is the data set after standardization processing, x max and x min represent the maximum and minimum values in the original data set respectively. The input data here is the data in the training set.

[0073] Among them: The training set enables the model to learn and master specific features and patterns of the data; the validation set monitors the training process and prevents the model from overfitting, thus maintaining its generalization ability for new data; the test set evaluates the training effect and generalization performance of the BP model to ensure that it can make accurate predictions on new data sets.

[0074] Example 7

[0075] Step 4: Optimize the BP neural network model through the GA genetic algorithm, using floating-point encoding, fitness function, roulette wheel selection, two-point crossover, and adaptive mutation operator to reduce the error between the prediction result and the real data and improve the accuracy of the prediction of the ultrasonic travel time ratio and the three parameters.

[0076] Step 4 is specifically as follows:

[0077] Step 4.1: The GA algorithm provides an encoding solution for setting the thresholds and initial weights of the BP neural network through floating-point encoding: L = S1×S2 + S2 + S2×S3 + S3, where L is the total length of the individual encoding, S1 is the number of nodes in the input layer, S2 is the number of nodes in the hidden layer, and S3 is the number of nodes in the output layer. Since all the BP neural network models for the shear wave and longitudinal wave travel time ratios of the constructed bolts adopt the same structure (3×5×1), the encoding length of each individual is calculated as 3×5 + 5 + 5 + 1 = 26. In addition, the number of individuals in the population is set to 50.

[0078] Step 4.2: Use the reciprocal of the mean square error function as the fitness function to evaluate the probability of an individual in the population being inherited into the next generation population.

[0079] Step 4.3: The process of individuals in the population being selected to enter the next generation is achieved through the roulette wheel algorithm, and the probability of an individual being selected is: where M is the total number of individuals in the population, f i is the fitness value of individual i, and p i is the probability of the individual being selected.

[0080] In the genetic algorithm, the selection process of an individual depends on the evaluation of its fitness. As the algorithm iterates, the probability of a less fit individual being selected and passed on to the next generation decreases. On the contrary, a more fit individual is more likely to be inherited into the next generation. Therefore, whether an individual is eligible to be inherited into the next generation population depends on the proportion of its fitness value in the total fitness of the entire population.

[0081] Step 4.4: In the GA algorithm, randomly select two parent individuals and use the two-point crossover algorithm that exchanges two randomly selected gene positions to generate two offspring individuals with genetic characteristics.

[0082] In the genetic algorithm, the crossover operator is the main way to generate new individuals, which is crucial for the global search performance of the algorithm. This method can maintain the genetic diversity of the population, expand the search domain, and thus improve the overall performance of the algorithm.

[0083] Step 4.5: Use the adaptive mutation operator in the creation of new individuals. This algorithm ensures that individuals with good adaptability can be preferentially retained, while those with poor adaptability will gradually be replaced:

[0084] In the formula, f is the fitness value of the mutated individual, f max is the maximum fitness value in the population, f avg is the fitness value of the entire population, p m1 and p m2 are set to 0.1 and 0.01 respectively.

[0085] By imposing random perturbations on the gene values of the chromosome, the mutation operator can create new individuals. In the framework of the genetic algorithm, this gene mutation helps to avoid the algorithm converging prematurely to the local optimal solution, helps to maintain the diversity of the population, and enhances the local exploration ability of the algorithm.

[0086] In summary, the thresholds and initial weights of the BP network are encoded through the floating-point coding scheme; the fitness function is used to evaluate the probability of an individual being passed to the next generation population; individuals are selected into the next generation through the roulette wheel algorithm; the two-point crossover algorithm is used to generate offspring individuals with genetic characteristics; the adaptive mutation operator is used to generate new individuals; in this way, the GA algorithm not only optimizes the training process of the BP neural network but also improves the accuracy of bolt axial stress prediction.

[0087] Example 8

[0088] Step 5: Test and train the BP neural network through the GA algorithm, and at the same time screen and optimize the thresholds and initial weights of the neural network to form a BP neural network model optimized by GA:

[0089] y1 = -0.65995 * tansig(-3.1602 * x1 - 1.3748 * x2 - 0.28143 * x3 + 2.87) -1.5792 * tansig(-0.18161 * x1 + 3.7103 * x2 + 0.19573 * x3 - 3.1665) -1.447 * tansig(2.2003 * x1 - 0.70863 * x2 + 0.2737 * x3 - 1.7175) +0.4517 * tansig(3.2792 * x1 + 0.68242 * x2 - 0.34785 * x3 - 1.4032) -0.00030917 * tansig(0.055479 * x1 - 1.4917 * x2 + 2.6305 * x3 - 0.27168) +3.2778 * tansig(0.22129 * x1 + 0.15112 * x2 - 0.1873 * x3 + 0.20342) -0.19163 * tansig(2.0896 * x1 - 2.2765 * x2 - 0.48421 * x3 + 1.8166) +0.68537 * tansig(-0.2008 * x1 - 3.4 * x2 - 0.2871 * x3 - 0.7213) +0.66295 * tansig(-1.5088 * x1 + 2.6315 * x2 + 0.04775 * x3 - 1.665) +0.00088341 * tansig(0.72797 * x1 - 2.2588 * x2 - 2.3195 * x3 + 0.7726) +0.0036288 * tansig(-1.2528 * x1 - 0.59376 * x2 - 1.7108 * x3 - 1.656) +0.011317 * tansig(1.4017 * x1 + 0.41311 * x2 - 2.1459 * x3 + 2.5602) -1.2677

[0103] In the model, y1 is the prediction model of the longitudinal and transverse wave acoustic time ratio of different bolts, and x1, x2, and x3 correspond to three variables: bolt grade, bolt length (mm), and applied stress (kN).

[0104] Figure 3 The left figure in [reference] is the fitting result of the BP neural network model optimized by the GA algorithm, and the right figure is the fitting result of the BP neural network model without optimization. Figure 4 The left figure in [reference] is the regression plot of the longitudinal and transverse wave acoustic time ratio under the GA-BP model, and the right figure is the regression plot of the validation set of the longitudinal and transverse wave acoustic time ratio under the GA-BP model. From Figure 3 and Figure 4 it can be seen that after the BP neural network model is optimized by the GA algorithm, its prediction value shows higher accuracy in terms of the consistency with the actual value, making this model more suitable for predicting the longitudinal and transverse wave acoustic time ratio, applied stress, bolt grade, and bolt length.

[0105] Example 9

[0106] Step 6: Substitute the obtained GA - BP neural network data into the bolt axial stress calculation formula for calculation: In the formula, σ is the axial stress, C R is the acoustic velocity ratio acoustic elasticity coefficient, are the acoustic times of ultrasonic longitudinal wave and transverse wave under stress - free conditions respectively, is the acoustic time ratio of ultrasonic transverse wave to longitudinal wave; among them, C R is related to the material of the bolt, is related to the material and precision of the bolt and is a constant.

[0107] In the present invention, 6 groups of experimental data of the acoustic time ratio of transverse wave to longitudinal wave of high - strength bolts with different parameters (bolt grade, bolt length (mm)) under loading stress (kN) are obtained through experiments. The data is linearly fitted to construct a three - layer BP neural network model, that is, a prediction model between the three parameters and the acoustic time ratio of ultrasonic transverse wave to longitudinal wave, which can effectively predict the relationship between the three parameters of the bolt and the acoustic time ratio of transverse wave to longitudinal wave; meanwhile, the GA algorithm is used to optimize the BP neural network model to construct a high - precision axial stress prediction model, showing higher accuracy in the consistency between its predicted output value and the actual measured value, and the relationship between the three parameters of the high - strength bolt and the acoustic time ratio of ultrasonic transverse wave to longitudinal wave can be accurately obtained.

[0108] The advantages of the detection method of the present invention are as follows: By establishing a GA - BP neural network model, it is possible to accurately predict the acoustic time ratio of ultrasonic transverse wave to longitudinal wave based on the three parameters of the loading stress (kN), grade, and length (mm) of the high - strength bolt, and then obtain the axial stress of the high - strength bolt; in this way, when installing the high - strength bolt, according to the three parameters of the loading stress (kN), grade, and length (mm), the acoustic time ratio of transverse wave to longitudinal wave is obtained, and then the axial stress of the high - strength bolt is calculated as the reference value during normal operation; after the high - strength bolt has worked for a period of time, the axial stress of the high - strength bolt is calculated through the actually measured acoustic time ratio of ultrasonic transverse wave to longitudinal wave. When comparing the bolt axial stress at this time with the bolt axial stress at the installation stage, when it exceeds the safety threshold, it is confirmed that it may fail, and thus it is replaced.

[0109] The present invention provides a reference for quickly and accurately measuring the axial stress of bolts in practical engineering, reduces the waste of manpower and material resources, reduces costs, and avoids safety accidents.

Claims

1. An axial stress detection method for high-strength bolts used in wind power based on GA-BP neural network, comprising the following steps: Step 1: Select high-strength bolts of different grades and lengths, obtain the ultrasonic transverse-longitudinal wave time ratio and loading stress data by stretching with an ultrasonic detector and a tensile testing machine, draw a scatter plot, and perform linear fitting. The fitted data set is used as the sample set for training the BP neural network. Step 2: According to the influence results of different loading stresses, bolt grades and lengths on the ultrasonic transverse-longitudinal wave time ratio, establish a BP neural network structure between the ultrasonic transverse-longitudinal wave time ratio and the above three parameters. Step 3: When using the BP neural network, randomly assign the original data to different groups and perform data normalization processing. Step 4: Optimize the BP neural network model through the GA genetic algorithm, and use floating-point coding, fitness function, roulette wheel selection, two-point crossover and adaptive mutation operator to improve the prediction accuracy of the ultrasonic transverse-longitudinal wave time ratio and the three parameters. Step 5: Test and train the BP neural network through the GA algorithm, and at the same time screen and optimize the thresholds and initial weights of the neural network to form a GA-optimized BP neural network model. Step 6: Substitute the obtained GA-BP neural network data into the bolt axial stress calculation formula for calculation, so as to obtain the axial stress of the bolt.

2. The detection method according to claim 1, characterized in that: Step 1 specifically includes the following steps: Step 1.1: Tensile test preparation: First, polish the upper and lower end faces of the bolt, and then calibrate the bolt. Step 1.2: Tensile test: For the bolt placed on the tensile testing machine, load it at a rate of 5 mm / min. Before reaching the yield strength, hold the load once every 100 MPa, and hold the load once every 50 MPa after reaching the yield strength, and end before reaching the tensile strength; the holding time is 45 s; when the bolt is holding the load, use an ultrasonic detector to record the ultrasonic transverse-longitudinal wave time ratio data. Step 1.3: Use data processing software to draw a scatter plot of the transverse-longitudinal wave time ratio and the loading stress, and perform linear fitting on the data using the formula: y = B0 + B1x, where B0 and B1 are the coefficients of each term in the above formula.

3. The detection method according to claim 2, characterized in that: In Step 1.1: The specific steps of calibration are: Place the bolt on the tensile testing machine and apply tensile force. The stresses applied five times are 0.12, 0.24, 0.36, 0.48 and 0.6 times of σ’ (σ’ is the theoretical yield strength of the experimental bolt). After each loading, hold the load for 45 seconds, and take the average of any three values during the holding time as the hydraulic press value.

4. The detection method according to claim 1, characterized in that: Step 2 specifically includes the following steps: Step 2.1: The tensile test adopts the single control variable method to study the influence of bolt grade and length on the transverse-longitudinal wave time ratio when applying different tensile forces. The specific ranges of its parameters are: grade 8.8, length 150 mm, grade 8.8, length 200 mm, grade 10.9, length 200 mm, grade 10.9, length 250 mm, grade 12.9, length 150 mm, grade 12.9, length 200 mm; Step 2.2: Establish a three-layer BP neural network (input layer, hidden layer, output layer) on the programming software. Take the three variables of applied stress (kN), bolt grade, and bolt length (mm) as the input layer data (corresponding to the three neurons in the input layer of the BP network in Figure 2), and the longitudinal-to-transverse wave acoustic time ratio as the output layer data (corresponding to one neuron in the output layer of the BP network in Figure 2). Establish the BP neural network structure for different bolts and the longitudinal-to-transverse wave acoustic time ratio. The formula for the BP prediction model of the longitudinal-to-transverse wave acoustic time ratio is: where the weight refers to the connection weight between the nth node in the jth layer and the mth node in the ith layer, and the threshold represents the activation threshold of the qth node in the pth layer; Step 2.

3. During the training process of the BP neural network, the mean square error function is used to evaluate the accuracy of the neural network prediction: where m is the number of mini-batch samples, and y i represents the predicted value made by the model for the i-th sample, and y i is the actual value corresponding to the i-th sample; Step 2.4: In the sample training of the BP neural network, select the Gauss-Newton method as the most suitable training algorithm. Step 2.

5. Determine the optimal number of nodes in the hidden layer of the BP neural network. The number of nodes is determined according to the following formula: In the formula, n1 is the number of nodes in the hidden layer, n is the number of neurons at the input end, m is the number of neurons at the output end, and a is a constant between 1 and 10.

5. The detection method according to claim 4, characterized in that: In Step 2.2: The hidden layer uses the non-linear tansig function as its activation function, and the output layer uses the purelin function for identity mapping.

6. The detection method according to claim 1, characterized in that: Step 3 specifically includes the following steps: Step 3.1: Randomly select and divide the data as follows: 70% of the longitudinal-to-transverse wave acoustic time ratio data constitutes the training set, 15% of the longitudinal-to-transverse wave acoustic time ratio data constitutes the validation set, and the remaining 15% is reserved for testing. Step 3.2: The input data of the BP neural network is processed through standardization by the following formula to ensure the consistency of the data set: In the formula, x is the original data set of the input, and x norm is the data set after standardization processing, and x max and x min represent the maximum value and the minimum value in the original data set respectively.

7. The detection method according to claim 1, wherein: Step 4 specifically includes the following steps: Step 4.1: The GA algorithm provides an encoding solution for the thresholds and initial weights of the BP neural network through floating-point encoding: L = S1×S2 + S2 + S2×S3 + S3, where L is the total length of the individual encoding, S1 is the number of nodes in the input layer, S2 is the number of nodes in the hidden layer, and S3 is the number of nodes in the output layer. Step 4.2: Use the reciprocal of the mean square error function as the fitness function to evaluate the probability of an individual in the population being inherited into the next generation population. Step 4.3: The process of individuals in the population being selected to enter the next generation is implemented through the roulette wheel algorithm, and the probability of an individual being selected is as follows: In the formula, M is the total number of individuals in the population, f i is the fitness value of individual i, and p i is the probability of the individual being selected; Step 4.4: In the GA algorithm, randomly select two parent individuals and use the two-point crossover algorithm that randomly selects two gene positions for exchange to generate two offspring individuals with genetic characteristics. Step 4.5: Use the adaptive mutation operator in the creation of new individuals. The algorithm is as follows: where f is the fitness value of the mutant individual, f max is the maximum fitness value in the population, f avg is the fitness value of the entire population, p m1 and p m2 are set to 0.1 and 0.01 respectively.

8. The detection method according to claim 1, wherein: The model of the GA-BP genetic neural network obtained in Step 5 is: y1＝ -0.65995*tansig(-3.1602*x1 - 1.3748*x2 - 0.28143*x3 + 2.87) -1.5792*tansig(-0.18161*x1 + 3.7103*x2 + 0.19573*x3 - 3.1665) -1.447*tansig(2.2003*x1 - 0.70863*x2 + 0.2737*x3 - 1.7175) +0.4517*tansig(3.2792*x1 + 0.68242*x2 - 0.34785*x3 - 1.4032) -0.00030917*tansig(0.055479*x1 - 1.4917*x2 + 2.6305*x3 - 0.27168) +3.2778*tansig(0.22129*x1 + 0.15112*x2 - 0.1873*x3 + 0.20342) -0.19163 * tansig(2.0896 * x1 - 2.2765 * x2 - 0.48421 * x3 + 1.8166) +0.68537 * tansig(-0.2008 * x1 - 3.4 * x2 - 0.2871 * x3 - 0.7213) +0.66295 * tansig(-1.5088 * x1 + 2.6315 * x2 + 0.04775 * x3 - 1.665) +0.00088341 * tansig(0.72797 * x1 - 2.2588 * x2 - 2.3195 * x3 + 0.7726) +0.0036288 * tansig(-1.2528 * x1 - 0.59376 * x2 - 1.7108 * x3 - 1.656) +0.011317 * tansig(1.4017 * x1 + 0.41311 * x2 - 2.1459 * x3 + 2.5602) -1.2677 In the model, y1 is the prediction model of the longitudinal and transverse wave acoustic time ratio of different bolts, and x1, x2, and x3 correspond to three variables: bolt grade, bolt length (mm), and applied stress (kN), respectively.

9. The detection method according to claim 1, characterized in that: The calculation formula for the axial stress of the bolt in Step 6 is as follows: In the formula, σ is the axial stress, and C R is the sound velocity ratio acoustic elasticity coefficient, are the acoustic times of the longitudinal wave and the transverse wave of ultrasonic waves under the stress-free condition respectively, is the ratio of the acoustic times of the transverse and longitudinal waves of ultrasonic waves; among them, C R is related to the material of the bolt, is related to the material and precision of the bolt and is a constant.

Images