Ultrasonic sensor array scanning beam rapid optimization method based on artificial intelligence

Through the combination of neural networks and genetic algorithms, self-learning and automatic optimization of ultrasonic sensor array focus rules are achieved, solving the problems of low detection efficiency and poor accuracy caused by manual settings in the prior art, and improving detection efficiency and accuracy .

CN118964938APending Publication Date: 2024-11-15JIANGSU UNIV
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202411014366.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-26
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

The existing ultrasonic sensor array focus rules require manual setup, resulting in low detection efficiency, poor accuracy and inability to achieve optimization.

Method used

The self-learning and optimization method of deflection beam focusing law based on neural networks and genetic algorithms is adopted. The mapping relationship between the focus law parameters and evaluation functions is realized through BP neural networks, and the genetic algorithm is used to automatically optimize the focus law parameters to improve detection efficiency and accuracy.

Benefits of technology

The independent learning and optimization of the focus rule of ultrasonic sensor array is realized, which reduces the dependence on the experience of detectors, improves the efficiency of defect detection, and can achieve optimal focus rule setting at different detection targets and array positions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118964938A_ABST
    Figure CN118964938A_ABST
Patent Text Reader

Abstract

The invention discloses an ultrasonic sensor array scanning beam rapid optimization method based on artificial intelligence, and belongs to the technical field of intelligent ultrasonic sensor detection. According to the scanning beam focusing rule optimization method, during ultrasonic sensor array detection, an intelligent algorithm is used for replacing manual work to select deflection beam optimal focusing rule parameters in real time. Through establishing a neural network prediction model between scanning beam focusing rule parameters and an evaluation function F (x), learning of a mapping relation between the focusing rule parameters and evaluation indexes is realized, and the focusing rule parameters after self-learning and the evaluation function F (x) are combined and stored in a neural network as an initial population of a genetic algorithm. And calling a neural network model in a genetic algorithm, and searching an optimal combination of focusing rule parameters through selection, crossover and variation. The current situation of manual traversal of an ultrasonic detection focusing rule is broken through, manual participation in the focusing rule parameter selection process is reduced, and the defect detection sensitivity and timeliness are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the field of ultrasonic nondestructive testing and relates to a rapid optimization method for ultrasonic sensor array scanning beams based on artificial intelligence. Background Art

[0002] Industrial ultrasonic nondestructive testing can detect internal defects without damaging the object being tested. Ultrasonic sensor arrays have the characteristics of variable emission wavefront, flexible beam scanning angle, wide scanning range, etc. At the same time, they can obtain rich information about the target body, which is convenient for accurate quantification. Therefore, they are increasingly valued in the field of industrial ultrasonic nondestructive testing.

[0003] The focusing law refers to controlling the number and timing of element excitation in the ultrasonic sensor array to achieve beam deflection and focusing, forming different wavefronts, and obtaining multi-directional or focused sound field scanning information of the sensor array, thereby achieving the purpose of improving defect detection sensitivity and defect detection rate. The parameters of the focusing law of the ultrasonic array scanning beam include array parameters and delay parameters. The array parameters include element size, element spacing, and number of elements. Different element delay parameters will produce different sound speed deflection angles. There are infinite focusing laws for scanning beams. In order to obtain accurate detection results, it is often necessary to manually set the focusing law parameters multiple times and in multiple directions before detection, which requires high knowledge and experience of the detection personnel and has poor timeliness.

[0004] In view of the problem that the existing focusing law requires human intervention and cannot be optimized, the present invention proposes a self-learning and optimization method for the deflected beam focusing law based on neural network and genetic algorithm, which breaks through the status quo of manual traversal of the focusing law of traditional ultrasonic sensor array and improves the detection efficiency and accuracy. Summary of the invention

[0005] The purpose of the present invention is to provide a method for rapid optimization of an ultrasonic sensor array scanning beam based on artificial intelligence, wherein the method uses array element size, array element spacing, array element number and deflection angle as focus law parameters, and beam main lobe width, side lobe level, grating lobe and defect echo signal amplitude peak as evaluation indicators. The BP neural network is used to realize the self-learning of the mapping relationship between the planar scanning beam focus law parameters and the evaluation function, and the genetic algorithm is used to optimize the focus law parameter combination. This method improves the detection efficiency. In order to achieve the above-mentioned invention purpose, the technical solution provided by the present invention is as follows:

[0006] The present invention provides a method for quickly optimizing an ultrasonic sensor array scanning beam based on artificial intelligence, which is characterized by comprising the following steps:

[0007] Step 1: Construct the basic parameter set of the ultrasonic sensor array, determine the focal law evaluation index system, and design the focal law evaluation function F(x). The basic parameter set includes the array element size a, array element spacing d a , the number of array elements per row n, and the sound velocity deflection angle θ; the evaluation index system includes the directivity D and the peak value A of the defect echo signal amplitude Nor , where directivity D includes beam main lobe width, side lobe level, and grating lobe, and directivity D is a linear weighted function evaluation function of its three indicators.

[0008] Step 2: Self-learning of focal law parameters. A self-learning model of focal law parameters based on back-propagation neural network is constructed. The model input is the basic parameter set. Each parameter in the evaluation index system can be calculated from the basic data set, and finally the evaluation function F(x) is output through the neural network model.

[0009] Step 3: Optimizing the focal law. Use a genetic algorithm to automatically optimize the focal law parameters, taking the basic parameter set as a chromosome, taking each set of focal law parameters as a chromosome individual, taking the evaluation function F(x) as the fitness function of the chromosome individual, and finding the chromosome individual with the best fitness. Through iterative calculation, output the basic data set that makes the evaluation function F(x) reach the minimum value. For a sensor array with a known basic data set, the focal law index system for a specific detection target can be calculated, thereby realizing the optimization of the focal law.

[0010] Furthermore, the step 1 specifically includes the following process:

[0011] (1) Set the evaluation function F(x) to:

[0012]

[0013] Among them, ε j and W j is the main lobe width θ in the evaluation index -3dB , sidelobe level M dB and grating lobe M -40dB The value and its corresponding weight, A Nor It is the ratio of the peak amplitude of the defect echo signal to the peak amplitude of the initial wave.

[0014] (2) A sound field simulation model is established based on the basic data set, and the index system data matrix X composed of the parameters of the directivity evaluation index is calculated through the sound field model:

[0015]

[0016] Among them, x ijRepresents the value of the jth indicator of the i-th basic data set. i is the sample number of the basic data set with different parameters, and j represents the number of the beam main lobe width, side lobe level and grating lobe. The total number of samples m can be selected according to the actual situation.

[0017] (3) Data normalization and standardization processing: sidelobe level M dB and grating lobe M -40dB is a normalized index, the beam main lobe width θ -3dB It is a non-normalized index. In order to eliminate the problem of inconsistent dimensions between indicators, the beam main lobe width θ -3dB The main lobe width θ when the deflection angle is 90° -3dB 90 The ratio of is taken as the normalized main lobe width θ -3dB Nor At the same time, the evaluation indicators are divided into extremely large indicators and extremely small indicators. The larger the value of the extremely large indicator, the better the effect, and the smaller the value of the extremely small indicator, the better the effect. In order to eliminate the problem of different polarity between indicators and enable the data to be measured with a unified standard, it is necessary to standardize each indicator.

[0018] For very large metrics:

[0019]

[0020] For very small indicators:

[0021]

[0022] Where: i and j represent the sample ordinal and evaluation index ordinal respectively.

[0023] (4) Calculate the amount of information C carried by the data j Get the weight W of each indicator j Volatility S j Indicates the difference in the values ​​of the same indicator in different basic data sets, conflict H j Characterize the correlation between indicators, information C j Characterizes the importance of each indicator. Volatility S j , conflict j and information volume C j The calculation formula is as follows:

[0024]

[0025] C j =S j ×H j

[0026] Among them, m is the total number of samples, n is the number of evaluation indicators, and x ijrepresents the value of the jth indicator of the i-th group of data, is the mean value of the jth indicator of the m-group basic data set, r ij represents the Pearson correlation coefficient between the i-th indicator and the j-th indicator. Therefore, the weight of the j-th indicator W j It can be expressed as:

[0027]

[0028] (5) Linear weighting of each evaluation index parameter is used to obtain the directivity D:

[0029] D=W1θ -3dB Nor +W2M dB +W3M -40dB

[0030] Where W1, W2 and W3 are the normalized beam widths θ -3dB Nor , sidelobe level M dB and grating lobe M -40dB The final evaluation function can be expressed as:

[0031] F(x)=W1θ -3dB Nor +W2M dB +W3M -40dB -A Nor

[0032] Furthermore, the normalized defect signal echo amplitude A Nor and the main lobe width θ -3dB Nor It can be expressed as:

[0033]

[0034] Among them, A F and A T are the peak amplitudes of the defect wave and the initial wave of the defect echo signal, θ -3dB 90 is the main lobe width of the beam when the deflection angle is 90°.

[0035] Furthermore, directivity is an extremely small indicator, and the amplitude of the defect echo signal is an extremely large indicator. Therefore, the smaller the value of the evaluation function F(x), the better the focal law detection effect.

[0036] Furthermore, the evaluation index beam main lobe width, side lobe level, grating lobe and defect echo signal amplitude are used to calculate the evaluation function of different basic parameter sets. When the array parameters are constant, different defects and different beam deflection angles affect the peak value of the defect echo signal.

[0037] Furthermore, the step 2 specifically includes the following process:

[0038] (1) Set the network parameters and initialize the weights and biases of each layer. Set the number of layers of the neural network structure to 3, namely the input layer, hidden layer and output layer; the input layer is the basic data set X in ={a,d a ,n,θ}, the number of neurons N in =4; the output layer is the evaluation index function Y out = {F(x)}, number of neurons N out = 1. Initialize the weight w1 and bias b1 between the input layer and the hidden layer, and the weight w2 and bias b2 between the hidden layer and the output layer.

[0039] (2) Solving the optimal number of hidden layer nodes N hidden :

[0040]

[0041] Among them, N input is the number of input layer nodes, N output is the number of nodes in the output layer, and a is an integer in the range of [1,10].

[0042] (3) Divide the data set and train the network. Set the number of training times, learning rate, and minimum error of the training target. Extract a certain number of samples from the basic data set, some of which are used as training samples for supervised learning and training of the neural network; some are used as validation samples to supervise the effectiveness of the training; and the remaining samples are used as test samples to test the predictive ability of the training samples. The number of samples can be selected according to actual conditions. Set the network training function to the traingdx function and the activation function to the sigmoid function, and train the network training samples.

[0043] (4) Update the weights and biases between each layer. Use the neural network to calculate the predicted value of the input training sample, compare the predicted value with the actual value, and calculate the expected value E(i) of the loss function:

[0044]

[0045] Among them, y i is the actual value of the i-th group of samples, is the predicted value of the i-th group of samples.

[0046] Calculate the expected value E(i) of the loss function as the error value MSE. If the MSE is less than the set minimum error or reaches the maximum number of iterations, set the weights and biases of the current inter-layer network to the weights and biases of the neural network. If the MSE is greater than the set minimum error, repeat processes (3) to (4).

[0047] (5) Output and save the results. Put the test sample into the successfully trained neural network and run it to obtain the combined results of the basic data set and evaluation function of the network calculation. The results are saved in the form of a data matrix in the network.

[0048] Furthermore, it is characterized in that the neural network objective function can be expressed as:

[0049] f(x)=f2[w 2r ×f1(w 1r ×x+b 1r )+b 2r ]

[0050] Among them, f1 and f2 are the transfer functions of the hidden layer and the output layer respectively, and w 1r and b 1r are the weights and biases between the input layer and the hidden layer, w 2r and b 2r are the weights and biases between the hidden layer and the output layer, respectively.

[0051] Furthermore, the predictive ability of the neural network model can be expressed as the correlation coefficient R between the predicted value and the actual value. 2 And the average percentage error MAPE can be verified, which can be expressed as:

[0052]

[0053] Among them, n is the number of samples, i is the sample number, and y i To output the actual value of the sample, To output the sample prediction value, is the output sample average. R 2 The closer the value is to 1, the stronger the prediction ability of the network is.

[0054] Furthermore, the step 3 specifically includes the following process:

[0055] (1) Set the genetic algorithm parameters to generate the initial chromosome population. Set the number of chromosome nodes, that is, the number of independent variables, the number of iterations, that is, the number of population generations generated, and the number of individuals in the population, that is, the number of individuals in each generation. Convert the genetic algorithm optimization problem into the problem of finding the minimum value of the evaluation function F(X), import the neural network result matrix as the initial population of the genetic algorithm, and the algorithm objective function and parameter constraints are as follows:

[0056]

[0057] Among them, x i is the optimal variable, n is the number of variables, Fit(x i ) is the fitness function value of different chromosomes, The variables x i The upper and lower limits of .

[0058] (2) Calculate the fitness function value of the population and screen the individuals in the population. Represent each chromosome with binary code, use the evaluation function F(x) as the fitness function Fit(x) of the genetic algorithm, calculate the fitness function value of each individual in the current population, and find the current optimal fitness value Fit best , and its corresponding chromosome is taken as the optimal chromosome.

[0059] (3) Select and merge chromosomes. Sort individuals according to the function value, calculate the probability P(k) of each individual being selected, select the individuals with the probability P(k) in the top 80% of the entire population into the new population, and eliminate the remaining individuals and no longer appear in the new population.

[0060] (4) Crossover and chromosome mutation. Set the crossover probability and crossover the chromosomes of adjacent individuals in the population. Set the mutation probability and perform mutation operations on the individuals in the population to form a new population. Check whether the values ​​of the chromosomes of the individuals in the population are within the constraints at this time, and eliminate the individuals that are not within the constraints so that they no longer appear in the population.

[0061] (5) Check whether the population meets the conditions. If the number of iterations is greater than the set value, output the optimal individual and its fitness function value. If the number of iterations is less than the set value, repeat the process (2) to (5).

[0062] Furthermore, the probability P(k) of the individual being selected is calculated as follows:

[0063]

[0064] Where N is the population size and Fit(k) is the fitness function value of the kth individual.

[0065] Furthermore, the defect echo signal-to-noise ratio (SNR) is used to verify the optimization result of the genetic algorithm. The SNR signal-to-noise ratio can be expressed as:

[0066]

[0067] Where L is the signal length, P n and P s are the power of noise and signal respectively, A n and A sare the amplitudes of noise and signal respectively.

[0068] Furthermore, for an array element size of a and an array element spacing of d a , a two-dimensional array with n array elements in each row. When the wavefront is deflected, the delay of each array element is equal, and the delay of each row of array elements is an arithmetic progression. The acoustic path difference between two adjacent array elements is:

[0069] ΔS=d a ×sinθ

[0070] Among them, d a is the center distance between adjacent array elements, and θ is the deflection angle of the plane wave front. Assuming that the first row of array elements emits sound waves first, the excitation time difference between all array elements in the nth row and the array elements in the first row is the delay Δt n for:

[0071]

[0072] Among them, c is the sound speed of ultrasound in the propagation medium, and t0 is a constant that controls the delay time.

[0073] Furthermore, the optimization algorithm can be applied to the cases where the sensor array parameters are known or unknown by adjusting the setting of the constraint conditions of the sensor array basic data set. When the sensor array basic data set is known, the upper and lower limits of the sensor array parameters are set to the actual array parameter values, and the deflection angle is optimized separately to obtain the optimal sound velocity deflection angle when the sensor array detects defects.

[0074] Furthermore, when the sensor array is in a new position, the focal law needs to be relearned to satisfy the optimality of the focal law when the sensor array is in a different position.

[0075] The present invention has the following beneficial effects:

[0076] As ultrasonic sensors develop towards intelligence, this invention conducts research on ultrasonic sensor arrays with self-learning capabilities of focal laws, realizes autonomous learning and optimization of focal law parameters, can effectively reduce dependence on the experience of inspectors while achieving the expected detection effect, and improves defect detection efficiency. When the basic data set of the sensor array is known and only the deflection angle needs to be optimized separately, the focal law indicator system for a specific detection target can be calculated, thereby realizing the optimization of the focal law. When the basic data set of the sensor array is unknown, the upper and lower limits of each parameter of the genetic algorithm are set to different values ​​according to actual conditions, and the precision of each parameter is set. More gene combinations are provided through crossover and mutation operations, the number of parameter sets is enriched, and the optimal combination of array parameters and deflection angles can be automatically found for the parameter set, providing reference value for the selection and design of sensor array parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0078] Figure 1 A flowchart of rapid optimization of an artificial intelligence ultrasonic sensor array scanning beam proposed by the present invention;

[0079] Figure 2 A schematic diagram of a row-column addressing sensor according to an embodiment of the present invention;

[0080] Figure 3 This is a simulation model diagram of the acoustic field of a through-hole defect when the scanning beam deflection angle is 45° in an embodiment of the present invention;

[0081] Figure 4 This is a simulation model diagram of the acoustic field of a transverse crack defect when the scanning beam deflection angle is 45° in an embodiment of the present invention;

[0082] Figure 5 1 is a diagram of the neural network iteration process of an embodiment of the present invention, (a) is a diagram of the neural network iteration process of a through hole defect, (b) is a diagram of the neural network iteration process of a transverse crack defect;

[0083] Figure 6 Graphs showing optimization results of a genetic algorithm for an embodiment of the present invention, (a) showing optimization results of a genetic algorithm for a through-hole defect, and (b) showing optimization results of a genetic algorithm for a transverse crack defect;

[0084] Figure 7 The focus law evaluation function and signal-to-noise ratio diagram of each defect at different deflection angles with a step length of 5° in the embodiment of the present invention; (a) is the evaluation function value and signal-to-noise ratio of the transverse hole defect at different wavefront deflection angles; (b) is the evaluation function value and signal-to-noise ratio of the transverse crack defect at different wavefront deflection angles;

[0085] Figure 8 They are schematic diagrams of wavefront detection corresponding to the deflection angles of the genetic algorithm optimization results of each defect in the embodiments, (a) is a schematic diagram of wavefront detection of a transverse hole defect, and (b) is a schematic diagram of wavefront detection of a transverse crack defect. DETAILED DESCRIPTION

[0086] The technical features, objectives and effects of the present invention are further described in detail below in conjunction with specific examples and drawings. Figure 1 Obviously, the described embodiments are only some embodiments of the present invention, not all 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.

[0087] Figure 1A flowchart of a method for rapid optimization of ultrasonic sensor array scanning beam based on artificial intelligence is provided in an embodiment of the present invention. The technical solution of the present invention is further described below in conjunction with an embodiment.

[0088] Example:

[0089] In the embodiment, two types of defects are selected for illustration, namely through-hole type and crack type standard defects.

[0090] [1] The standard defect test block size for through holes is 40×15×40mm 3 The defect center is located on the vertical center line of the test block xOz plane (coordinate system as Figure 3 As shown), the radius is 2mm, the depth is 20mm, the detection material is steel, and the ultrasonic sound velocity is 5900m / s. The number of sensor array elements is set to 2×2, 3×3, 4×4, 6×6 and 8×8 respectively; the array element size range is 0.1~1.1mm square array element, the step length is 0.2mm; the array element spacing range is 0.1~1.0mm, the step length is 0.2mm; the beam is set to deflect in the xOz plane, the deflection angle range is 0°~75°, with an interval of 5°, and all the above parameter settings are arranged and combined to obtain the corresponding beam main lobe width, side lobe level and grating lobe. Figure 2 This is a schematic diagram of a row-column addressed sensor array with an element spacing of 1 mm, an element size of 0.9 mm, and an element number of 8×8. The sensor array adopts an 8+8 addressing array, the top electrodes in the same column are connected together and connected to the excitation signal path through an electronic switch, and the bottom electrodes in the same row are connected together and connected to the ground through an electronic switch. When the upper and lower electrodes of the array element are simultaneously selected, the array element is excited. Figure 3 This is the acoustic field distribution diagram of the through-hole defect when the scanning beam deflection angle is 45°.

[0091] [2] The size of the crack type standard defect test block is the same as the through hole type standard defect test block. The length and width of the crack defect are set to 12 mm and 4 mm respectively. The defect center is located on the vertical center line of the xOz plane of the test block, and the depth is 20 mm. Figure 4 This is the acoustic field distribution diagram of the through hole defect in the embodiment of the present invention when the scanning beam deflection angle is 45°. The finite element modeling settings of the sensor array and the defect test block are the same as those in Example 1, and the values ​​of the corresponding beam main lobe width, side lobe level, grating lobe and defect echo signal amplitude are obtained.

[0092] The operation steps of self-learning and optimization of the parameters of the two types of defect focusing laws are as follows:

[0093] Step 1:

[0094] (1) A directional sound field simulation model is established based on the basic data set. The index is set to 1 when there is a grating lobe and to 0 when there is no grating lobe. 100 groups of samples are randomly selected, and the directional index data matrix X can be expressed as:

[0095]

[0096] (2) Calculate the fluctuation coefficient and conflict coefficient of each evaluation index of the directivity function, and obtain the weight of each index as W = [0.12, 0.29, 0.59], and complete the directivity D design. Therefore, directivity D can be expressed as:

[0097] D=0.12θ -3dB Nor +0.29M dB +0.59M -40dB

[0098] Then the evaluation function F(x) of the focal law parameters can be expressed as:

[0099] F(x)=0.12θ -3dB Nor +0.29M dB +0.59M -40dB -A Nor

[0100] Among them, θ -3dB Nor is the normalized beamwidth, M dB is the sidelobe level, M -40dB is the grating lobe, A Nor is the normalized peak value of the defect echo signal amplitude.

[0101] Step 2:

[0102] (1) Set the network parameters and initialize the weights and biases of each layer. Set the number of layers of the neural network structure to 3, namely the input layer, hidden layer and output layer; the input layer is the basic data set X in ={a,d a ,n,θ}, the number of neurons N in =4; the output layer is the evaluation index function Y out = {f(x)}, number of neurons N out = 1. Initialize the weight w1 and bias b1 between the input layer and the hidden layer, and the weight w2 and bias b2 between the hidden layer and the output layer.

[0103] (2) Find the optimal number of hidden layer nodes. Set the number of training times to 1000, the learning rate to 0.01, and the minimum error of the training target to 0.0001.

[0104] (3) Divide the data set and train the network. Extract 200 samples from the basic data set, of which 100 are used as training samples for supervised learning training of the neural network; 50 are used as validation samples to supervise the effectiveness of training; and 50 are used as test samples to test the prediction ability of the training samples. Set the network training function to the traingdx function and the activation function to the sigmoid function, and train the network training samples.

[0105] (4) Complete the update of weights and biases between layers. Use the neural network to calculate the predicted value of the input training sample, compare the predicted value with the actual value, and calculate the mean square error (MSE). If MSE < 0.0001, set the weights and biases of the current network between layers to the weights and biases of the neural network. If MSE > 0.0001, repeat steps (3) to (4).

[0106] (5) Output and save the results. Put the test sample into the successfully trained neural network and run it to obtain the combined results of the basic data set and evaluation function of the network calculation. The results are saved in the form of a data matrix in the network.

[0107] Step 3:

[0108] (1) Set the genetic algorithm parameters and generate the initial chromosome population. Set the number of chromosome nodes to 4, that is, the number of independent variables to 4, the number of iterations to 150, that is, there are a total of 150 generations of individuals, and the number of individuals in the population is 30, that is, each generation of the population has 30 individuals. Convert the genetic algorithm optimization problem into the problem of finding the minimum value of the evaluation function F(x), and import the neural network result matrix as the initial population of the genetic algorithm. Since the array parameters are known, the array element spacing is 1mm, the array element size is 0.9mm, and the number of array elements is 8×8, the algorithm objective function and the parameter constraints are as follows:

[0109]

[0110] Among them, Fit best (x) is the optimal fitness value, F min (x) is the minimum value of the evaluation function.

[0111] (2) Calculate the fitness function value of the population and screen the individuals in the population. Represent each chromosome with binary code, use the evaluation function F(x) of the focusing law parameter as the fitness function Fit(x) of the genetic algorithm, calculate the fitness function value of each individual in the current population, and find the current optimal fitness value Fit best (x), and its corresponding chromosome is taken as the optimal chromosome.

[0112] (3) Select and merge chromosomes. Sort individuals according to the function value, calculate the probability P(k) of each individual being selected, select the individuals with the probability P(k) in the top 80% of the entire population into the new population, and eliminate the remaining individuals and no longer appear in the new population.

[0113] (4) Crossover and chromosome mutation. Set the crossover probability to 0.6 and adopt the uniform crossover method to exchange the genes at each locus of the two paired individuals with the same crossover probability; set the mutation probability to 0.05 and adopt the basic position mutation method to perform mutation calculation on the value of a randomly specified locus to form a new population. Check whether the values ​​of each chromosome of the individuals in the population are within the constraints at this time, and eliminate the individuals that are not within the constraints so that they no longer appear in the population.

[0114] (5) Check whether the population meets the conditions: If the number of iterations is greater than 150, output the optimal individual and its fitness function value; if the number of iterations is less than 150, repeat operations (2) to (5).

[0115] The designed neural network model is used to self-learn all the transverse hole defects and crack defects in turn to find the best deflection angle of the scanning beam. Figure 5 and Figure 6 They are the iterative process diagrams of the neural network and genetic algorithm for two defects respectively. Figure 5 (a) is the neural network self-learning iterative process diagram of through-hole defects. Figure 5 (b) is the neural network self-learning iterative process diagram of transverse crack defects. Figure 6 (a) is the optimization result of genetic algorithm for through-hole defects. Figure 6 (b) is the result of genetic algorithm optimization for transverse crack defect. The correlation coefficient R of all neural network sample sets for through hole defect and transverse crack defect 2 They are 0.9444 and 0.9387 respectively, and the average percentage error MAPE of the test set is 5.7% and 5.4% respectively. The results show that the neural network has good fitting function and can be used in the optimization of genetic algorithm. The genetic algorithm optimization results of the two defects are 26° and 12° respectively, and the signal-to-noise ratio SNR is 31.98dB and 34.79dB respectively. In order to verify the accuracy of the optimization results, the plane wavefront deflection angle θ is set to range from 0° to 75°, with a step size of 5°, and the relationship between the deflection angle of the transverse hole defect and the transverse crack defect and the evaluation system and signal-to-noise ratio is plotted as shown in the figure. Figure 7 As shown, Figure 7 (a) is the evaluation function value and signal-to-noise ratio of the transverse hole defect at different wavefront deflection angles. Figure 7(b) is the evaluation function value and signal-to-noise ratio of the transverse crack defect at different wavefront deflection angles. The results show that for both the transverse hole defect and the transverse crack defect, the designed optimization algorithm can obtain the plane wave deflection angle when the focal law evaluation function value is the smallest, achieving the purpose of optimizing the focal law parameters, and compared with other deflection angles, the defect echo signal-to-noise ratio is also the highest when the evaluation function value is the smallest.

[0116] After obtaining the optimal result of the deflection angle, the delay parameters of each array element are calculated. The column numbers of the sensor array are arranged from left to right. Assuming that the first array element is transmitted first, the delay parameters of the remaining array elements are shown in Table 1. Add the corresponding delay compensation to each array element, and the transmission beam forms a wavefront with a corresponding deflection angle. The schematic diagram of the formed wavefront is shown in Figure 8 .

[0117] Table 1 Array element delay parameters

[0118]

[0119] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementations, which are merely a display of the present invention and not a limitation. The purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. Equivalent implementations and improved changes that do not depart from the principles and design ideas of the present invention are also within the protection scope of the present invention.

Claims

1. A rapid optimization method for ultrasonic sensor array beam based on artificial intelligence, characterized in that: The following steps are involved: Step 1: Construct the basic parameter set of the ultrasonic sensor array, determine the focal law evaluation index system, and design the focal law evaluation function F(x); the basic parameter set includes the array element size a, array element spacing d a , the number of array elements per row n, and the sound velocity deflection angle θ; the evaluation index system includes the directivity D and the ratio of the peak amplitude of the defect echo signal to the peak amplitude of the initial wave A Nor , where the directivity D includes the beam main lobe width, side lobe level, and grating lobe, and the directivity D is a linear weighted function of these three indicators; Step 2: Self-learning of focal law parameters: Construct a focal law parameter self-learning model based on a back-propagation neural network. The model input is a basic parameter set. Each parameter in the evaluation index system can be calculated from the basic data set. Finally, the evaluation function F(x) is output through the neural network model. Step 3: Optimizing the focusing law; Use a genetic algorithm to automatically optimize the focusing law parameters, take the basic parameter set as a chromosome, take each combination of focusing law parameters as a chromosome individual, take the evaluation function F(x) as the fitness function of the chromosome individual, find the chromosome individual with the best fitness, and through iterative calculation, output the basic data set that makes the evaluation function F(x) reach the minimum value; for a sensor array with a known basic data set, the focusing law index system for a specific detection target can be calculated, thereby realizing the optimization of the focusing law; after obtaining the optimization result, calculate the delay parameters of each array element.

2. The array beam fast optimization method according to claim 1, characterized in that: This method is applicable to a two-dimensional array with the same number of row and column elements, the same element spacing, and a square element geometry.

3. The array beam fast optimization method according to claim 1, characterized in that: The step 1 includes the following specific processes: (1) Set the evaluation function F(x) to: Among them, ε j and W j is the main lobe width θ in the evaluation index -3dB , sidelobe level M dB and grating lobe M -40dB The value and its corresponding weight, A Nor It is the ratio of the peak amplitude of the defect echo signal to the peak amplitude of the initial wave; (2) A sound field simulation model is established based on the basic data set, and the index system data matrix X composed of the parameters of the directivity evaluation index is calculated through the sound field model: Among them, x ij represents the value of the jth indicator of the i-th basic data set; i is the sample ordinal number of the basic data set with different parameters selected, and j represents the ordinal number of the beam main lobe width, side lobe level and grating lobe respectively; the total number of samples m can be selected according to the actual situation; (3) Data normalization and standardization processing: sidelobe level M dB and grating lobe M -40dB is a normalized index, the beam main lobe width θ -3dB It is a non-normalized index. In order to eliminate the problem of inconsistent dimensions between indicators, the beam main lobe width θ -3dB The main lobe width θ when the deflection angle is 90° -3dB 90 The ratio of is taken as the normalized main lobe width θ -3dB Nor ; At the same time, the evaluation indicators are divided into extremely large indicators and extremely small indicators. The larger the value of the extremely large indicator, the better the effect, and the smaller the value of the extremely small indicator, the better the effect. In order to eliminate the problem of different polar directions between indicators and enable the data to be measured with a unified standard, it is necessary to standardize each indicator; For very large metrics: For very small indicators: Among them: i and j represent the sample ordinal and evaluation index ordinal respectively; (4) Calculate the amount of information C carried by the data j Get the weight W of each indicator j ; Volatility S j Indicates the difference in the values ​​of the same indicator in different basic data sets, conflict H j Characterize the correlation between indicators, information C j Characterizes the importance of each indicator; volatility S j , conflict j and information volume C j The calculation formula is as follows: Among them, m is the total number of samples, n is the number of evaluation indicators, and x ij Represents the value of the jth indicator of the i-th group of data, is the mean value of the jth indicator of the m-group basic data set, r ij represents the Pearson correlation coefficient between the i-th indicator and the j-th indicator; therefore, the weight of the j-th indicator W j It can be expressed as: (5) Linear weighting of each evaluation index parameter is used to obtain the directivity D: D=W1θ -3dB Nor +W2M dB +W3M -40dB Where W1, W2 and W3 are the normalized beam widths θ -3dB Nor , sidelobe level M dB and grating lobe M -40dB The final evaluation function can be expressed as: F(x)=W1θ -3dB Nor +W2M dB +W3M -40dB -A Nor Normalized defect signal echo amplitude A Nor and the main lobe width θ -3dB Nor It can be expressed as: Among them, A F and A T are the peak amplitudes of the defect wave and the initial wave of the defect echo signal, θ -3dB 90 is the main lobe width of the beam when the deflection angle is 90°.

4. The array beam fast optimization method according to claim 3, characterized in that: Directivity is a very small indicator, and the amplitude of the defect echo signal is a very large indicator. Therefore, the smaller the value of the evaluation function F(x), the better the focal law detection effect. The evaluation index beam main lobe width, side lobe level, grating lobe and defect echo signal amplitude are used to calculate the evaluation function of different basic parameter sets; when the array parameters are constant, different defects and different beam deflection angles affect the peak value of the defect echo signal.

5. The array beam fast optimization method according to claim 1, characterized in that: The step 2 specifically includes the following specific process: (1) Set the network parameters and initialize the weights and biases of each layer; set the number of layers of the neural network structure to 3, namely the input layer, hidden layer and output layer; the input layer is the basic data set X in ={a,d a ,n,θ}, the number of neurons N in =4; the output layer is the evaluation index function Y out = {F(x)}, number of neurons N out =1; initialize the weight w1 and bias b1 between the input layer and the hidden layer, and the weight w2 and bias b2 between the hidden layer and the output layer; (2) Solving the optimal number of hidden layer nodes N hidden : Among them, N input is the number of input layer nodes, N output is the number of nodes in the output layer, and a is an integer ranging from [1,10]; (3) Divide the data set and train the network; set the number of training times, learning rate and minimum error of the training target; extract a certain number of samples from the basic data set, a part of which is used as training samples for supervised learning and training of the neural network; a part of which is used as verification samples to supervise the effectiveness of training; the remaining samples are used as test samples to test the prediction ability of the training samples; the number of samples can be selected according to the actual situation; set the network training function to traingdx function, the activation function to sigmoid function, and train the network training samples; (4) Update the weights and biases between each layer; use the neural network to calculate the predicted value of the input training sample, compare the predicted value with the actual value, and calculate the expected value E(i) of the loss function: Among them, y i is the actual value of the i-th group of samples, is the predicted value of the i-th group of samples, and N is the number of samples; Calculate the expected value E(i) of the loss function as the error value MSE. If the MSE is less than the set minimum error or reaches the maximum number of iterations, set the weights and biases of the current inter-layer network to the weights and biases of the neural network. If the MSE is greater than the set minimum error, repeat processes (3) to (4). (5) Output the results and save them; put the test sample into the successfully trained neural network to run, obtain the combined results of the basic data set and evaluation function of the network calculation, and save the results in the form of a data matrix in the network.

6. The array beam fast optimization method according to claim 5, characterized in that: The neural network objective function can be expressed as: f(x)=f2[W 2r ×f1(w 1r ×x+b 1r )+b 2r ] Among them, f1 and f2 are the activation functions of the hidden layer and the output layer respectively, and w 1r and b1r are the weights and biases between the input layer and the hidden layer, respectively, w 2r and b 2r are the weights and biases between the hidden layer and the output layer respectively; The predictive ability of the neural network model can be measured by the correlation coefficient R between the predicted value and the actual value. 2 And the average percentage error MAPE can be verified, which can be expressed as: Among them, n is the number of samples, i is the sample number, and y i To output the actual value of the sample, To output the sample prediction value, is the output sample average; R 2 The closer the value is to 1, the stronger the prediction ability of the network is.

7. The array beam fast optimization method according to claim 1, characterized in that: The step 3 specifically includes the following process: 1) Set the genetic algorithm parameters to generate the initial chromosome population; set the number of chromosome nodes, i.e. the number of independent variables, the number of iterations, i.e. the number of population generations generated, and the number of individuals in the population, i.e. the number of individuals in each generation; convert the genetic algorithm optimization problem into the problem of finding the minimum value of the evaluation function F(x), import the neural network result matrix as the initial population of the genetic algorithm, and the algorithm objective function and parameter constraints are as follows: Among them, x i is the optimal variable, n is the number of variables, Fit(x i ) is the fitness function value of different chromosomes, The variables x i The upper and lower limits of 2) Calculate the population fitness function value and screen the individuals in the population; represent each chromosome with binary code and use the evaluation function F(x) as the fitness function Fit(x) of the genetic algorithm. i ), calculate the fitness function value of each individual in the current population, and find the current optimal fitness value Fit best , and take its corresponding chromosome as the optimal chromosome; 3) Select and merge chromosomes; sort individuals according to the function value, calculate the probability P(k) of each individual being selected, select the individuals with the probability P(k) in the top 80% of the entire population into the new population, and eliminate the remaining individuals and no longer appear in the new population; 4) Crossover and mutation of chromosomes: set the crossover probability, crossover and exchange the chromosomes of adjacent individuals in the population, set the mutation probability, perform mutation operations on the individuals in the population to form a new population; check whether the values ​​of each chromosome of the individuals in the population are within the constraints at this time, and eliminate the individuals that are not within the constraints so that they no longer appear in the population; 5) Check whether the population meets the conditions; if the number of iterations is greater than the set value, output the optimal individual and its fitness function value; if the number of iterations is less than the set value, repeat process 2) to 5).

8. The array beam fast optimization method according to claim 7, characterized in that: The probability P(k) of the individual being selected is calculated as follows: Where N is the population size and Fit(k) is the fitness function value of the kth individual.

9. The array beam fast optimization method according to claim 7, characterized in that: The optimization result of the genetic algorithm is verified by using the defect echo signal-to-noise ratio (SNR). The SNR signal-to-noise ratio can be expressed as: Where L is the signal length, P n and P s are the power of noise and signal respectively, A n and A s are the amplitudes of noise and signal respectively; For an array element size of a and an array element spacing of d a , a two-dimensional array with n array elements in each row. When the wavefront is deflected, the delay of each row of array elements is equal, and the delay of each row of array elements is an arithmetic progression. The acoustic path difference between two adjacent rows of array elements is: ΔS=d a ×sinθ Among them, d a is the center distance between adjacent array elements, θ is the deflection angle of the plane wave front; assuming that the first row of array elements emits sound waves first, the excitation time difference between all array elements in the nth row and the array elements in the first row is the delay Δt n for: Among them, c is the sound speed of ultrasound in the propagation medium, and t0 is a constant that controls the delay time.

10. The array beam fast optimization method according to claim 7, characterized in that: The optimization algorithm can be applied to the situations where the sensor array parameters are known or unknown by adjusting the setting of the constraint conditions of the sensor array basic data set; when the sensor array basic data set is known, the upper and lower limits of the sensor array parameters are set to the actual array parameter values, so as to achieve separate optimization of the deflection angle and obtain the optimal sound velocity deflection angle when the sensor array detects defects; When the sensor array is in a new position, the focal law needs to be relearned to satisfy the optimality of the focal law when the sensor array is in a different position.

Citation Information

Cited By

  • Parameter optimization method, model and system of multilayer wave absorber and readable storage medium

    CN120373160A

  • Parameter optimization method, model, system and readable storage medium for multilayer absorber

    CN120373160B