Intelligent detection and defect analysis method for gas turbine blade

Abstract

The invention relates to the technical field of artificial neural networks, in particular to a gas turbine blade intelligent detection and defect analysis method, which comprises the following steps of acquiring blade material attribute data, blade geometric structure data and operation load data, calculating a stress transfer matrix and analyzing multi-node stress gradient change. According to the method, based on stress analysis and multi-source data fusion, the accuracy of blade defect detection is improved, misjudgment caused by dependence on single visual features is avoided by calculating the stress transfer matrix, screening the abrupt change area and accurately positioning the stress concentration point, the material stress offset is calculated in combination with the stress disturbance mode of the inner wall of the water flow channel, and the accuracy of blade defect detection is improved. According to the method, hidden defects are identified, cooling channel surface stress features are screened, a stress direction deviation angle is calculated, defect classification basis comprises mechanical features, normalized processing stress and image data, neural network training is optimized, classification precision adapts to complex working conditions, and the multi-type defect identification capability is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial neural networks, and particularly to an intelligent detection and defect analysis method for gas turbine blades. Background Art

[0002] The technical field of artificial neural networks involves constructing and training neural network models using artificial intelligence algorithms to simulate the computational and decision-making processes of the human brain. The core of this technology lies in forming a multi-layer structure through neuron and weight connections and performing iterative training based on a large amount of data to extract features and identify patterns. Artificial neural networks are widely applied in multiple fields such as computer vision, natural language processing, automatic control, and intelligent detection. Among them, in the field of intelligent detection, artificial neural networks can use methods such as convolutional neural networks, recurrent neural networks, or deep reinforcement learning to achieve image recognition, anomaly detection, and predictive analysis in complex environments. The research in this technical field mainly focuses on optimizing the network structure, improving the training efficiency, enhancing the generalization ability, and adjusting the model in combination with specific industry requirements to meet the accuracy requirements of different detection tasks.

[0003] Among them, the intelligent detection and defect analysis method for gas turbine blades refers to a technical solution that uses artificial neural network technology to intelligently detect gas turbine blades and analyze defect characteristics. This method covers technical links such as image acquisition, data preprocessing, neural network training and optimization, defect feature extraction, and classification of gas turbine blades. Specifically, first, surface images of the blades are obtained through high-resolution industrial cameras or laser scanning devices, and image data is preprocessed using methods such as filtering enhancement and edge detection to improve data quality. Subsequently, a detection model based on a convolutional neural network is constructed to perform feature learning on the blade images, and a hierarchical classifier is used to discriminate defect categories. Finally, a specific loss function is used to optimize the model to improve the recognition accuracy, and a defect feature extraction algorithm is combined to analyze the damage location, shape, and expansion trend of the blades.

[0004] The existing technology relies on a single image recognition method, is easily affected by light and surface stains, and it is difficult to identify minor defects. For complex curved surface areas of the blades, light and shadow changes lead to misjudgments, and the consideration of internal stress distribution is lacking, making it difficult to detect deep cracks and stress concentration damages. Classification mainly relies on surface morphology and does not combine stress characteristics, making it difficult to distinguish defects with similar surfaces but different stress distributions. Model optimization does not combine the stress evolution of the blades under high-temperature and high-pressure environments, resulting in limited generalization ability and a high misjudgment rate under different working conditions. Summary of the Invention

[0005] The object of the present invention is to solve the drawbacks existing in the prior art and propose an intelligent detection and defect analysis method for gas turbine blades.

[0006] To achieve the above object, the present invention adopts the following technical solutions: An intelligent detection and defect analysis method for gas turbine blades, comprising the following steps:

[0007] S1: Obtain blade material property data, blade geometric structure data and operating load data, calculate the stress transfer matrix, analyze the multi-node stress gradient change, screen the stress mutation area, extract the stress concentration points and mutation points, and generate stress anomaly eigenvalues;

[0008] S2: Call the stress anomaly eigenvalues, obtain the inner wall data of the water flow channel of the heavy gas turbine blade, calculate the inner wall stress distribution, analyze the stress disturbance mode of the defect area, identify the stress concentration points, calculate the stress offset of the material, screen the stress gradient mutation points, analyze the residual stress accumulation value, and establish a defect stress disturbance coefficient;

[0009] S3: Call the defect stress disturbance coefficient, collect the stress distribution on the surface of the cooling channel of the water-cooled blade, screen the stress characteristics of three types of defects: cracks, spalling, and deformation, calculate the stress direction offset angle, stress mean value and gradient mutation points, and generate a defect feature vector;

[0010] S4: Call the defect feature vector, combine the stress anomaly feature data, normalize the classification features, obtain the blade surface image features, construct a stress feature channel, a defect stress disturbance channel and an image feature channel, and establish a multi-channel input data set.

[0011] As a further solution of the present invention, the stress anomaly eigenvalues are specifically stress concentration points, mutation points, and stress mutation areas. The defect stress disturbance coefficient includes stress concentration points, material stress offset, stress gradient mutation points, and residual stress accumulation values. The defect feature vector is specifically crack stress characteristics, spalling stress characteristics, deformation stress characteristics, stress direction offset angle, stress mean value, and gradient mutation points. The multi-channel input data set includes a stress feature channel, a defect stress disturbance channel, and an image feature channel.

[0012] As a further solution of the present invention, the steps for obtaining the stress anomaly eigenvalues are specifically:

[0013] S101: Based on the blade material property data, blade geometric structure data and operating load data, calculate the stress transfer matrix of multiple nodes, call the stress values transferred between nodes, and obtain the stress gradient values in different directions according to the spatial coordinate relationship between nodes, and calculate the amplitude of the gradient change rate to obtain the stress gradient change coefficient;

[0014] S102: Call the stress gradient change coefficient, calculate the stress mutation interval for the multi-node gradient change situation, screen the stress abnormal area according to the gradient change situation within multiple intervals, obtain the extreme points within the area, and calculate the stress change rate in the adjacent areas of the extreme points to generate a stress mutation point set;

[0015] S103: Based on the stress mutation point set, calculate the stress concentration degree corresponding to multiple mutation points, using the formula:

[0016]

[0017] Perform operations to obtain the stress concentration coefficient of multiple mutation points, call the spatial coordinate information in the mutation point set, establish the morphological distribution of the stress abnormal area, and obtain the stress abnormal characteristic value;

[0018] Among them, S1 represents the stress abnormal characteristic value, Δσ 1,i represents the stress change amount at mutation point i, d 1,i represents the Euclidean distance between mutation point i and its adjacent point, n1 represents the total number of mutation points, and C represents a stable parameter used to avoid the influence of the denominator approaching zero.

[0019] As a further solution of the present invention, the steps for obtaining the defect stress disturbance coefficient are specifically as follows:

[0020] S201: Call the stress abnormal characteristic value, calculate the stress distribution value of the inner wall based on the inner wall data of the water flow channel of the heavy-duty gas turbine blade, call the stress abnormal characteristic value, compare and analyze the stress change amplitude at multiple positions of the inner wall, screen the stress change mutation points, determine the stress abnormal distribution area of the blade inner wall, and obtain the inner wall stress distribution data;

[0021] S202: Based on the inner wall stress distribution data, analyze the stress disturbance mode of the defect area, calculate the gradient change of the local stress, screen the stress gradient mutation points, using the formula:

[0022]

[0023] Perform operations to obtain the local stress gradient change amount, and calculate the stress offset value of the stress concentration point to generate the stress offset amount of the stress concentration point;

[0024] Among them, S2 represents the local stress gradient change amount, Δσ 2,i represents the stress change amount of the i-th point, Δx i represents the stress measurement point spacing, E represents the elastic modulus of the material, ρ represents the material density, and n2 represents the total number of stress measurement points;

[0025] S203: Calculate the cumulative value of the residual stress on the inner wall of the blade water flow channel based on the stress offset of the stress concentration point, extract the residual stress distribution in the defect area, calculate the stress disturbance degree, and establish a defect stress disturbance coefficient.

[0026] As a further solution of the present invention, the steps for obtaining the defect feature vector are specifically as follows:

[0027] S301: Call the defect stress disturbance coefficient, collect the stress distribution on the surface of the cooling channel of the water-cooled blade, calculate the corresponding stress gradient change amount based on the stress values at multiple positions of the cooling channel, detect the stress mutation points in multiple regions of the cooling channel, screen the stress intervals where the stress mutation points are located, and obtain the distribution of stress gradient mutation points;

[0028] S302: Based on the distribution of the stress gradient mutation points, screen the stress characteristics of three types of defects: cracks, spalling, and deformation, calculate the stress mean value in the corresponding area, analyze the local stress direction offset situation, and use the formula:

[0029]

[0030] Perform operations to obtain the stress direction offset angles in multiple defect areas, and establish a stress direction offset angle distribution;

[0031] Among them, V represents the variance of the stress mean value, S 3,i represents the stress value at the i-th place, S 3,i,prev represents the stress value at the previous measurement point of the i-th place, and n3 represents the total number of measurement points in the defect area;

[0032] S303: Call the stress direction offset angle distribution, combine it with the stress gradient mutation point distribution, construct the feature vector of each defect area, and extract the mean value, direction offset angle, and normalized value of the gradient mutation point of the feature vector for different types of defect areas to obtain the defect feature vector.

[0033] As a further solution of the present invention, the steps for obtaining the multi-channel input data set are specifically as follows:

[0034] S401: Based on the defect feature vector, combine the stress anomaly feature data, calculate the root mean square error and standard deviation of the eigenvalue, determine the feature data fluctuation range based on the root mean square error, screen the feature data with larger weights based on the standard deviation, perform normalization processing, and generate a feature normalization weight matrix;

[0035] S402: Call the feature normalization weight matrix, obtain the blade surface image features, calculate the principal component contribution rate of multiple image feature dimensions, screen the features with higher contribution rates, calculate the inter-channel correlation coefficient, construct a stress feature channel, a defect stress disturbance channel, and an image feature channel, and use the formula:

[0036]

[0037] Calculate the influence degree of the comprehensive feature channel to obtain the multi-channel feature influence matrix;

[0038] Among them, M represents the multi-channel feature influence matrix, and W 4,k represents the normalized feature weight, F 4,k represents the normalized single eigenvalue, m4 represents the total number of features, C 4,kp represents the correlation coefficient between channels, and n4 represents the total number of normalized weight features;

[0039] S403: Invoke the multi-channel feature influence matrix, calculate the multi-channel feature contribution rate, screen the channel data with high contribution rate, allocate the data input weight according to the normalized distribution weight of the feature influence matrix, and establish a multi-channel input data set after integration.

[0040] As a further solution of the present invention, the method further includes:

[0041] S5: Invoke the multi-channel input data set, construct the input layer of the neural network, set the weight constraints of the stress anomaly feature channel and the defect stress perturbation channel, optimize the loss function of the hidden layer, train the classification model, and establish the defect classification prediction result;

[0042] The defect classification prediction result specifically refers to the classification labels of three types of defects: crack, spalling, and deformation.

[0043] As a further solution of the present invention, the steps for obtaining the defect classification prediction result are specifically as follows:

[0044] S501: Invoke the multi-channel input data set, preprocess the multi-channel stress data and defect stress perturbation data, calculate the stress feature vector of the multi-channel input data, adjust the normalization parameter according to the data range, calculate and store the initial weight distribution of the multi-feature channel, and obtain the weight constraint matrix of the input layer;

[0045] S502: Based on the weight constraint matrix of the input layer, adjust the connection parameters of the hidden layer, perform weighted summation on the multi-channel stress features, and use the formula:

[0046]

[0047] Calculate the stress response value of the hidden layer;

[0048] Among them, R h represents the stress response value of the hidden layer, W 5,p represents the weight of the input channel p, S 5,p represents the stress eigenvalue of the input channel p, P5,q The weight representing the defective stress perturbation q, D 5,q represents the defective stress perturbation value, n5 is the total number of input channels, and m5 is the total number of defective stress perturbation channels;

[0049] S503: Based on the stress response values of the hidden layer, train a classification model, calculate the classification boundary weights, screen the classification thresholds that meet the classification criteria of the stress anomaly feature channels, calculate the classification boundary parameters and store the classification information, and establish the defective classification prediction results.

[0050] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0051] In the present invention, based on stress analysis and multi-source data fusion, the accuracy of blade defect detection is improved. By calculating the stress transfer matrix, screening the mutation regions, accurately locating the stress concentration points, it avoids misjudgment caused by relying on a single visual feature. Combining the stress perturbation pattern on the inner wall of the water flow channel, calculating the stress offset of the material, and realizing the identification of hidden defects. Screening the stress characteristics on the surface of the cooling channel, calculating the stress direction offset angle, making the defect classification basis include mechanical characteristics, and improving the stability. Normalizing the stress and image data, optimizing the neural network training, making the classification accuracy adapt to complex working conditions, and improving the recognition ability of multiple types of defects. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 is a schematic diagram of the main steps of the present invention;

[0053] Figure 2 is a flowchart of the steps for obtaining the stress anomaly characteristic values of the present invention;

[0054] Figure 3 is a flowchart of the steps for obtaining the defective stress perturbation coefficient of the present invention;

[0055] Figure 4 is a flowchart of the steps for obtaining the defective feature vectors of the present invention;

[0056] Figure 5 is a flowchart of the steps for obtaining the multi-channel input data set of the present invention;

[0057] Figure 6 is a flowchart of the steps for obtaining the defective classification prediction results of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0058] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0059] In the description of the present invention, it should be understood that the orientation or positional relationship indicated by terms such as "length", "width", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc. is based on the orientation or positional relationship shown in the drawings. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation to the present invention. In addition, in the description of the present invention, the meaning of "a plurality of" is two or more, unless otherwise specifically defined.

[0060] Embodiment 1

[0061] Please refer to Figure 1 , the present invention provides a technical solution: an intelligent detection and defect analysis method for gas turbine blades, including the following steps:

[0062] S1: Obtain blade material property data, blade geometric structure data and operating load data, calculate the stress transfer matrix, analyze the multi-node stress gradient change, screen the stress mutation area, extract the stress concentration points and mutation points, and generate stress anomaly eigenvalues;

[0063] S2: Call the stress anomaly eigenvalues, obtain the inner wall data of the water flow channel of the heavy-duty gas turbine blade, calculate the inner wall stress distribution, analyze the stress disturbance mode in the defect area, identify the stress concentration points, calculate the stress offset of the material, screen the stress gradient mutation points, analyze the residual stress accumulation value, and establish a defect stress disturbance coefficient;

[0064] S3: Call the defect stress disturbance coefficient, collect the stress distribution on the surface of the cooling channel of the water-cooled blade, screen the stress characteristics of three types of defects: cracks, spalling, and deformation, calculate the stress direction offset angle, stress mean value and gradient mutation points, and generate defect feature vectors;

[0065] S4: Call the defect feature vectors, combine with the stress anomaly feature data, normalize the classification features, obtain the blade surface image features, construct a stress feature channel, a defect stress disturbance channel and an image feature channel, and establish a multi-channel input data set;

[0066] S5: Call the multi-channel input data set, construct the input layer of the neural network, set the weight constraints of the stress anomaly feature channel and the defect stress disturbance channel, optimize the loss function of the hidden layer, train the classification model, and establish the defect classification prediction result.

[0067] The stress anomaly characteristic values ​​are specifically stress concentration points, mutation points, and stress mutation areas. The defect stress disturbance coefficients include stress concentration points, material stress offsets, stress gradient mutation points, and residual stress accumulation values. The defect characteristic vectors are specifically crack stress characteristics, spalling stress characteristics, deformation stress characteristics, stress direction offset angles, stress mean values, and gradient mutation points. The multi-channel input data set includes stress characteristic channels, defect stress disturbance channels, and image feature channels. The defect classification prediction results specifically refer to the classification labels of three types of defects: cracks, spalling, and deformation.

[0068] See also Figure 2 , the specific steps for obtaining the stress anomaly characteristic value are:

[0069] S101: Based on the blade material property data, blade geometry data and operation load data, a stress transfer matrix of multiple nodes is calculated, the transferred stress values ​​between nodes are called, and the stress gradient values ​​in the differentiated directions are obtained according to the spatial coordinate relationship between the nodes, and the amplitude of the gradient change rate is calculated to obtain the stress gradient change coefficient;

[0070] The blade material property data include parameters such as density, elastic modulus, Poisson's ratio, etc. The blade geometry data include information such as blade length, thickness, chord length, blade curvature, etc. The operating load data involves the aerodynamic load, inertial load, centrifugal force and other influencing factors that the blade is subjected to during operation. After these data are input into the calculation program, the blade model is discretized using the grid division method. The stress distribution value at each grid node of the blade is calculated based on the finite element method, the stress values ​​between adjacent grid nodes are associated according to the spatial coordinate relationship, and the stress gradient value of each node is calculated by the difference method. The rate of change of the stress gradient is , the stress gradient change of each node in the three main directions is obtained by point-by-point calculation. The blade models of different scales can be calculated uniformly through normalization. Then the stress gradient change coefficient is calculated and defined as the ratio of the stress gradient change amplitude between adjacent grid nodes. The area with drastic changes is screened out to further improve the calculation accuracy. Combined with the actual working conditions, it is assumed that the elastic modulus of the blade material is 210GPa, the Poisson's ratio is 0.3, the stress value of a node of the blade is 150MPa, the stress value of the adjacent node is 140MPa, and the distance between the two is 5mm. The stress gradient change in this direction is calculated as After calculations in all directions, a complete set of stress gradient variation coefficients can be obtained.

[0071] S102: calling the stress gradient change coefficient, calculating the stress mutation interval according to the gradient change of multiple nodes, screening the stress abnormal area according to the gradient change in the multiple intervals, obtaining the extreme value points in the area, and calculating the stress change rate of the area adjacent to the extreme value points, and generating a set of stress mutation points;

[0072] Through the stress gradient change coefficient, for the multi-node gradient change situation, the sliding window method is used to determine the stress mutation interval, that is, the stress change amplitude is analyzed within a certain range. When the stress change within a certain interval exceeds the set change rate threshold, it is determined as the stress mutation area. To ensure rationality, this threshold can be set as the 90th percentile of the blade stress change. For example, assuming that the stress change rate corresponding to the 90th percentile of the statistical data is 8 MPa / mm, then when the stress change rate within a certain interval exceeds 8 MPa / mm, this area is marked as the stress mutation interval. Within this area, by calculating the extreme points, that is, screening the local stress maximum and minimum points. For example, if the node stresses within a certain area are [120, 130, 125, 140, 135, 145] MPa respectively, then the extreme points are 130 MPa and 145 MPa respectively. Subsequently, calculate the stress change rates of these extreme points and their adjacent areas to judge the local mutation degree. If the stress change rate of the adjacent area is greater than 10 MPa / mm, then this point is considered as a mutation point, thus generating a set of stress mutation points.

[0073] S103: Based on the set of stress mutation points, calculate the stress concentration degree corresponding to multiple mutation points, using the formula:

[0074]

[0075] Through calculation, obtain the stress concentration coefficient of multiple mutation points, call the spatial coordinate information in the set of mutation points, establish the morphological distribution of the stress abnormal area, and obtain the stress abnormal characteristic value;

[0076] Among them, S1 represents the stress abnormal characteristic value, Δσ 1,i represents the stress change amount at mutation point i, d 1,i represents the Euclidean distance between mutation point i and its adjacent point, n1 represents the total number of mutation points, and C represents a stable parameter used to avoid the influence of the denominator approaching zero.

[0077] The formula is as follows:

[0078]

[0079] Among them, S1 represents the stress abnormal characteristic value, Δσ 1,i represents the stress change amount at mutation point i, d 1,i represents the Euclidean distance between mutation point i and its adjacent point, n1 represents the total number of mutation points, and C is a stable parameter. Assuming that in a certain blade area, 5 stress mutation points have been identified, and their stress change amounts are respectively

[0080] [12, 15, 10, 18, 14] MPa, the Euclidean distances between adjacent points are [3, 4, 3.5, 4.2, 3.8] mm respectively. Taking the stability parameter C = 1 to avoid the influence of the denominator approaching zero, the stress concentration factor is calculated as follows:

[0081]

[0082]

[0083]

[0084] This result indicates that there is a relatively concentrated stress mutation region in this blade area. This value can be further compared with the historical stress abnormal distribution data to judge the abnormal degree of this region. At the same time, the spatial coordinate information in the mutation point set is called, and a three-dimensional fitting method is used to establish the morphological distribution of the stress abnormal region to obtain the stress abnormal characteristic value.

[0085] Table 1 Stress mutation points and calculation parameter table

[0086] Mutation point number Stress change amount (MPa) Euclidean distance between adjacent points (mm) 1 12 3 2 15 4 3 10 3.5 4 18 4.2 5 14 3.8

[0087] As shown in Table 1, the stress change amounts of 5 mutation points and the Euclidean distances between adjacent points are listed. These data are used to calculate the stress concentration factor and finally determine the stress abnormal characteristic value.

[0088] Please refer to Figure 3 , and the specific steps for obtaining the defect stress disturbance coefficient are as follows:

[0089] S201: Call the stress abnormal characteristic value, calculate the stress distribution value of the inner wall based on the data of the inner wall of the water flow channel of the heavy-duty gas turbine blade, call the stress abnormal characteristic value, compare and analyze the stress change amplitudes at multiple positions of the inner wall, screen the stress change mutation points, determine the stress abnormal distribution region of the inner wall of the blade, and obtain the inner wall stress distribution data;

[0090] First, a discretization model of the inner wall surface needs to be established. This model needs to include the blade surface mesh division information and ensure that stress values can be obtained at different positions of the fluid flow. In the layout of discrete points, an equally spaced distribution method is adopted, and the interval of each measurement point is set to 2 mm to ensure the continuity of data. Subsequently, a stress measuring instrument is used to measure the stress at each measurement point, and the stress value of each point is obtained and recorded in the database to obtain the complete inner wall stress distribution data.

[0091] Next, to calculate the stress distribution value of the inner wall, it is necessary to traverse the stress values of each measurement point and perform statistical analysis on them. When calculating the stress abnormal characteristic value, first calculate the stress mean σ avg and the standard deviation σ std , and their calculation formulas are as follows:

[0092]

[0093]

[0094] Among them, σ i represents the stress value of the i-th measurement point, and n represents the total number of measurement points. After the calculation is completed, the method for screening stress outliers is as follows: Set the outlier determination threshold σ thr = σ avg + 2σ std . Any measurement point greater than σ thr is determined as a stress outlier point.

[0095] When comparing and analyzing the stress change amplitudes at multiple positions on the inner wall, select the stress change amount Δσ i = |σ i - σ i+1 | between adjacent measurement points. After calculating the stress change amounts between all measurement points, screen out the measurement points that satisfy Δσ i > 1.5σ std as the stress change mutation points, and determine the stress anomaly distribution area in combination with the positions of the mutation points, and finally obtain the inner wall stress distribution data.

[0096] As shown in Table 2, the stress distribution data of some measurement points on the inner wall of the water flow channel of a gas turbine blade are listed, including the measurement point positions, measured stress values, and calculated stress change amounts:

[0097] Table 2 Stress Distribution Data Table of Measurement Points on the Inner Wall of the Water Flow Channel of the Blade

[0098]

[0099]

[0100] As can be seen from Table 2, the stress change amounts of the 4th and 6th measurement points significantly exceed the average level and belong to the stress mutation points. Therefore, it can be determined that there is a stress anomaly distribution in this area on the inner wall of the water flow channel of the blade.

[0101] S202: Based on the inner wall stress distribution data, analyze the stress disturbance mode in the defect area, calculate the gradient change of the local stress, screen the stress gradient mutation points, and use the formula:

[0102]

[0103] Perform operations to obtain the local stress gradient change amount, and calculate the stress offset value of the stress concentration point to generate the stress offset amount of the stress concentration point;

[0104] Among them, S2 represents the local stress gradient change amount, Δσ 2,iRepresents the stress change of the i-th point, Δx i Represents the stress measurement point spacing, E represents the elastic modulus of the material, ρ represents the material density, and n2 represents the total number of stress measurement points;

[0105] Formula:

[0106]

[0107] Among them, Δσ 2,i = σ i+1 - σ i Represents the stress change of the i-th measurement point, Δx i Represents the measurement point spacing, E is the elastic modulus of the material, ρ is the material density, and n2 is the total number of measurement points.

[0108] Select the elastic modulus E = 210 GPa and density ρ = 7800 kg / m3 of the blade material, and calculate the local stress gradient change of each measurement point:

[0109] Calculate the gradient change of the 1st measurement point:

[0110]

[0111]

[0112] Calculate the gradient change of the 4th measurement point:

[0113]

[0114]

[0115] After calculating all the measurement points, screen the points that satisfy S 2,i > 50 as stress concentration points. For example, the 4th measurement point meets the conditions, determine it as a stress concentration point, and calculate its stress offset value σ 偏移 :

[0116] σ 偏移 = σ max - σ avg = 350 - 280 = 70 MPa;

[0117] This result shows that the stress of the 4th measurement point is 70 MPa higher than the average stress, which can be used as the calculation basis for the subsequent defect stress disturbance coefficient.

[0118] S203: Based on the stress offset of the stress concentration point, calculate the cumulative value of the residual stress on the inner wall of the blade water flow channel, extract the residual stress distribution in the defect area, calculate the stress disturbance degree, and establish the defect stress disturbance coefficient.

[0119] Use the point-by-point integration method to calculate the residual stress. The calculation formula is as follows:

[0120]

[0121] Substitute the data for calculation:

[0122] σ 累积 =(250 - 280)×2+(260 - 280)×2+(280 - 280)×2+(350 -

[0123] 280)×2+(370 - 280)×2+(400 - 280)×2;

[0124] =(-60)+(-40)+0 + 140 + 180 + 240 = 460 MPa·mm;

[0125] This value represents the cumulative amount of residual stress generated by the stress concentration phenomenon on the inner wall of the blade water flow channel. Finally, calculate the stress perturbation degree and establish the defect stress perturbation coefficient:

[0126]

[0127] The result shows that the defect stress perturbation degree of the inner wall of the blade is relatively high, which is 2.55. Subsequently, the stability of the blade stress distribution can be evaluated according to this coefficient.

[0128] Please refer to Figure 4 , and the specific steps for obtaining the defect feature vector are as follows:

[0129] S301: Call the defect stress perturbation coefficient, collect the stress distribution on the surface of the cooling channel of the water-cooled blade, calculate the corresponding stress gradient change amount based on the stress values at multiple positions of the cooling channel, detect the stress mutation points in multiple regions of the cooling channel, screen the stress intervals where the stress mutation points are located, and obtain the stress gradient mutation point distribution;

[0130] First, arrange stress sensors in different regions of the cooling channel to collect the instantaneous stress values of each measurement point. Assume that 8 measurement points are arranged in a certain region, and the obtained instantaneous stress values are σ1, σ2, σ3,..., σ8. Then, for the stress data of these measurement points, calculate the stress gradient between each point and its adjacent point. Define the stress gradient change amount as ΔS = |σ i -σ i-1 |. Calculate the stress gradient change sequence {ΔS2, ΔS3,..., ΔS8} for all measurement points. Subsequently, according to the set stress gradient mutation threshold θ, screen out the mutation points, that is, judge ΔS iMeasurement points of >θ. Assume that θ is set to 5 MPa. When a certain measurement point satisfies conditions such as ΔS5 = 6.2 MPa and ΔS7 = 5.6 MPa, the areas corresponding to measurement points 5 and 7 can be determined as the intervals where the stress mutation points are located. Then, based on the determination of the stress mutation points, the stress intervals where the mutation points are located are screened, that is, regional boundaries are set between multiple adjacent stress mutation points. By comparing the stress change rates before and after the mutation points, it is judged whether they belong to the same region. If the stress gradient mutation amounts of adjacent points are both greater than the set mutation threshold θ and the change trends are the same, they are classified into the same region; otherwise, they are divided independently. Finally, the distribution of stress gradient mutation points in the cooling channel is obtained.

[0131] S302: Based on the distribution of stress gradient mutation points, screen the stress characteristics of three types of defects: cracks, spalling, and deformation. Calculate the stress mean value in the corresponding area, and analyze the local stress direction deviation. Use the formula:

[0132]

[0133] Perform operations to obtain the stress direction deviation angle of the multi-defect area, and establish the stress direction deviation angle distribution;

[0134] Among them, V represents the variance of the stress mean value, S 3,i represents the stress value at the i-th place, S 3,i,prev represents the stress value of the previous measurement point at the i-th place, and n3 represents the total number of measurement points in the defect area;

[0135] First, screen the stress data in each defect area and calculate the stress mean value of each area. Assume that there are n3 measurement points in the crack area, and its stress values are {S 3,1 , S 3,2 ,..., S 3,n3}. When calculating its mean variance V, it is necessary to combine the stress value S 3,i,prev of the previous measurement point of the measurement point, and use the formula:

[0136]

[0137] to calculate. Assume that the crack area contains 5 measurement points, and its stress values are {S 3,1 = 120, S 3,2 = 118, S 3,3 = 110, S 3,4 = 105, S 3,5 = 100} MPa, and the stress values of the previous measurement points are {S 3,0 = 122, S 3,1 = 120, S 3,2 = 118, S 3,3 = 110, S 3,4If = 105} MPa, the mean variance of its stress can be calculated:

[0138]

[0139]

[0140]

[0141] At the same time, analyze the local stress direction offset in each defect area, that is, by comparing the stress distributions of each measurement point, calculate the stress change direction between the measurement points, and calculate the stress direction offset angle of each area through the direction deviation of the stress gradient change. Assume that the direction change angles of the stress gradient in a certain area are θ1 = 10°, θ2 = 15°, and θ3 = 5°, then the stress direction offset angle distribution of this area can be constructed.

[0142] S303: Call the stress direction offset angle distribution, combine it with the stress gradient mutation point distribution, construct the feature vector of each defect area, and for different types of defect areas, extract the mean value, direction offset angle, and normalized value of the gradient mutation point of the feature vector to obtain the defect feature vector.

[0143] For the stress values {S1, S2,..., S n} of all measurement points, set the maximum value S max and the minimum value S min , and calculate the normalized value:

[0144]

[0145] Assume that the stress values of the measurement points in the crack area are {120, 118, 110, 105, 100} MPa, where the maximum value S max = 120 MPa and the minimum value S min = 100 MPa, then the normalized value calculation is as follows:

[0146]

[0147]

[0148]

[0149]

[0150]

[0151] Finally, the normalized stress feature vector [1, 0.9, 0.5, 0.25, 0] is obtained. At the same time, combined with the stress direction offset angles {10°, 15°, 5°} and the gradient mutation point data, the defect feature vector is completely expressed.

[0152] Table 3 Stress data and normalized eigenvectors of the defect area

[0153] Measurement point number Stress value (MPa) Normalized stress value Direction offset angle (°) 1 120 1.00 10 2 118 0.90 15 3 110 0.50 5 4 105 0.25 - 5 100 0.00 -

[0154] As shown in Table 3, after the stress values in the crack defect area are normalized, eigenvectors are formed. Combining with the direction offset angle parameter makes the defect eigenvectors in this area more specific, facilitating subsequent analysis and discrimination.

[0155] Please refer to Figure 5 , and the steps for obtaining the multi-channel input data set are specifically as follows:

[0156] S401: Based on the defect eigenvectors, combined with the stress anomaly characteristic data, calculate the root mean square error (RMSE) and standard deviation of the eigenvalues. Determine the fluctuation range of the characteristic data according to the RMSE, screen the characteristic data with larger weights according to the standard deviation, perform normalization processing, and generate a characteristic normalization weight matrix;

[0157] First, defect characteristic data need to be collected, including but not limited to defect morphological characteristics (such as length 10 mm, width 2 mm), depth characteristics (such as 0.5 mm), and material properties (such as elastic modulus 210 GPa). Subsequently, stress anomaly characteristic data are collected, such as maximum principal stress 150 MPa, minimum principal stress 50 MPa, and shear stress 20 MPa. For these data, calculate the root mean square error (RMSE) to measure the stability of the characteristic data. The RMSE calculation formula is as follows:

[0158]

[0159] where, x i represents each sample eigenvalue, represents the eigenvalue mean. Taking the maximum principal stress as an example, assuming the sample data are 148 MPa, 152 MPa, and 149 MPa respectively, and its mean is 149.67 MPa, then the calculated RMSE is:

[0160]

[0161] Set the RMSE threshold to 2 MPa. When the RMSE is lower than 2 MPa, it is considered that the data fluctuation range is stable. Then calculate the standard deviation (σ). The standard deviation formula is as follows:

[0162]

[0163] After calculating the standard deviation of each characteristic, use the weight calculation formula for normalization processing to obtain the following normalized characteristic weight matrix:

[0164] W = [0.35 0.30 0.20 0.10 0.05];

[0165] The result shows that features with higher weights have a greater impact on the overall fluctuation range of the data. Therefore, these features with higher weights need to be given priority in the subsequent steps and a higher influence degree should be given during the calculation process to ensure the representativeness and stability of the data.

[0166] S402: Call the feature normalization weight matrix, obtain the leaf surface image features, calculate the principal component contribution rate of the multi-image feature dimensions, screen the features with higher contribution rates, calculate the correlation coefficient between multiple channels, construct the stress feature channel, the defect stress perturbation channel, and the image feature channel, and use the formula:

[0167]

[0168] Calculate the influence degree of the comprehensive feature channel to obtain the multi-channel feature influence matrix;

[0169] where M represents the multi-channel feature influence matrix, W 4,k represents the normalized feature weight, F 4,k represents the normalized single eigenvalue, m4 represents the total number of features, C 4,kp represents the correlation coefficient between channels, and n4 represents the total number of normalized weight features;

[0170] For example, if the GLCM contrast of an image is 0.45 and the correlation is 0.80, the calculation method of the contribution rate is as follows:

[0171]

[0172] Assume that the maximum contrast is 0.5 and the maximum correlation is 1, then calculate:

[0173]

[0174] If the contribution rate threshold is set to 1.5, this feature is retained. Then calculate the correlation coefficient C 4,kp , for example, set the correlation of a certain feature with the stress channel to 0.85 and the correlation with the defect perturbation channel to 0.75, then construct the multi-channel feature matrix and calculate the influence degree of the comprehensive feature channel according to the following formula:

[0175]

[0176] For example, set W 4,1 = 0.35, F 4,1 = 0.45, C 4,1p = 0.85, then calculate:

[0177]

[0178] The result shows that the correlation between different channels has a significant effect on the calculation of the multi-channel feature influence matrix. Among them, the high-weight features have a greater influence weight on the stress and defect perturbation channels, which will play a decisive role in the subsequent dataset construction process.

[0179] S403: Call the multi-channel feature influence matrix, calculate the contribution rate of multi-channel features, screen the channel data with high contribution rate, allocate the data input weight according to the normalized distribution weight of the feature influence matrix, and establish a multi-channel input dataset after integration.

[0180] The method is as follows:

[0181]

[0182] Assume that the feature influence degree matrix is M = [0.116, 0.203, 0.178], and the calculated sum is 0.497. Then the normalized input weight is:

[0183] W′ = [0.233 0.409 0.358];

[0184] The finally constructed multi-channel input dataset is shown in Table 4 below:

[0185] Table 4 Multi-channel input dataset

[0186]

[0187] The result shows that when constructing a multi-channel input dataset, using the feature influence degree matrix for normalized distribution weight allocation can ensure that the contribution rate of each channel's data is reasonably measured, and at the same time ensure that the input dataset can reflect the key correlations between channels to the greatest extent and be used for subsequent analysis and processing.

[0188] Please refer to Figure 6 , and the specific steps for obtaining the defect classification prediction result are as follows:

[0189] S501: Call the multi-channel input dataset, preprocess the stress data and defect stress perturbation data of multiple channels, calculate the stress feature vectors of the multi-channel input data, adjust the normalization parameters according to the data range, calculate and store the initial weight distribution of multiple feature channels, and obtain the input layer weight constraint matrix;

[0190] First, preprocess the multi-channel stress data and defect stress perturbation data collected. During this process, perform denoising processing on the stress data, and use the moving average method to eliminate the random interference terms in the data. For example, for the stress value S(t) at a certain moment t, the data can be smoothed by weighted averaging the adjacent five data points before and after, that is, calculate To obtain a smoother stress signal, the data is then normalized. First, the maximum and minimum values of the stress data are calculated. Assuming the stress data range is between [-500, 500] MPa, the normalization formula is used for normalization to map the data range to [0, 1], ensuring that the numerical scales of the data in each channel are consistent. Then, the stress eigenvectors of each channel are calculated, and the dimensionality reduction is performed using the principal component analysis (PCA) method to extract the main stress eigenvalues of each channel. For example, for the three-channel data {S1, S2, S3}, the covariance matrix Σ is calculated, and after eigenvalue decomposition, the main eigenvector v1 is taken as the principal component direction to obtain the reduced-dimensional stress eigenvalues Then, the normalization parameters are adjusted according to the data range. Assuming that the distribution of the input data changes greatly, the mean μ S and the standard deviation σ S need to be recalculated. Then, the normalization can be performed using to perform zero-mean normalization. In addition, the initial weight distribution of the multi-feature channels is calculated and stored. Assuming there are five input channels, the weights are initialized using a uniform distribution W0 = [0.2, 0.2, 0.2, 0.2, 0.2], and the weight constraint matrix W in of the input layer is obtained. Its initialization can be set by the variance equalization method to make the initial contributions of all channels balanced. For example, setting to ensure the balance of the influence of different stress channels, thus completing the preprocessing of the input data and the initialization of the weights.

[0192] S502: Based on the weight constraint matrix of the input layer, adjust the connection parameters of the hidden layer, perform weighted summation on the multi-channel stress features, and use the formula:

[0193]

[0194] to calculate the stress response value of the hidden layer;

[0195] Among them, R h represents the stress response value of the hidden layer, W 5,p represents the weight of the input channel p, S 5,p represents the stress eigenvalue of the input channel p, P 5,q represents the weight of the defect stress perturbation q, D 5,q represents the defect stress perturbation value, n5 is the total number of input channels, and m5 is the total number of defect stress perturbation channels;

[0196] First, determine the weight distribution W5 of each input channel and the weight P5 of the defect stress perturbation channels. Assuming the number of input channels is n5 = 4 and the number of defect stress perturbation channels is m5 = 3, then for the stress eigenvalue S of each channel 5,pPerform absolute value operations to avoid the influence of negative stress on the calculation, and then calculate the weighted summation part For example, if W5 = [0.25, 0.25, 0.3, 0.2] and

[0197] S5 = [50, -30, 20, -10] MPa, then calculate R input = 0.25×50 + 0.25×30 + 0.3×20 + 0.2×10 = 27.5 MPa. Then calculate the root mean square value term of the defect stress perturbation part. Assume the defect stress perturbation values D5 = [9, 16, 25] MPa and the corresponding weights P5 = [0.3, 0.5, 0.2], and calculate the square root value Then calculate the weighted mean That is, R defect = (0.3×3 + 0.5×4 + 0.2×5) / 3 = 1.93 MPa. Finally, calculate the stress response value R of the hidden layer h = R input + R defect = 27.5 + 1.93 = 29.43 MPa, and obtain the stress response value of the hidden layer.

[0198] S503: Based on the stress response value of the hidden layer, train the classification model, calculate the classification boundary weights, screen the classification thresholds that meet the classification criteria of the stress anomaly feature channels, calculate the classification boundary parameters and store the classification information, and establish the defect classification prediction results.

[0199] First, define the classification boundary weight W cls , and use the gradient descent method for optimization, updating in each round of iteration to minimize the classification error L, where the learning rate α = 0.01. Then, screen the classification threshold T cls that meets the classification criteria of the stress anomaly feature channels. Assume that based on historical data statistics, the anomaly classification threshold is set to T cls = 28 MPa. Then, judge whether the stress response value R h = 29.43 MPa of the hidden layer exceeds this threshold. The result satisfies R h > T cls , so this data is classified as the defect category. Then, calculate the classification boundary parameters, including the weight W cls and the bias term b cls . Assume the initial parameters are W cls = 0.8, b cls = 2. Then, the classification decision function is f(R h ) = W cls R h + b cls, calculate f(29.43) = 0.8×29.43 + 2 = 25.54, store the classification information, including the classification result label and calculation parameters, and establish the defect classification prediction result.

[0200] Table 5 Example Table for Calculating Stress Response Values of Hidden Layers

[0201] Input channel number Stress eigenvalue (MPa) Input weight Weighted stress (MPa) 1 50 0.25 12.5 2 -30 0.25 7.5 3 20 0.30 6.0 4 -10 0.20 2.0 Total — — 27.5

[0202] Table 5 lists an example of calculating stress characteristic values. The weighted sum value of each input channel is calculated to be 27.5 MPa, and then the defect stress perturbation part is added. Finally, the stress response value of the hidden layer is 29.43 MPa, which is used for classification model training and defect classification prediction.

[0203] The above is only the preferred embodiment of the present invention, and it is not intended to limit the present invention in other forms. Any person skilled in the art may use the disclosed technical content to make changes or modifications into equivalent embodiments with equivalent changes and apply them to other fields. However, as long as it does not depart from the technical solution content of the present invention, any simple modification, equivalent change, and modification made to the above embodiments based on the technical essence of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. An intelligent detection and defect analysis method for gas turbine blades, characterized in that, It includes the following steps: S1: Obtain the blade material property data, blade geometric structure data and operating load data, calculate the stress transfer matrix, analyze the stress gradient change of multiple nodes, screen the stress mutation regions, extract the stress concentration points and mutation points, and generate stress anomaly eigenvalues; S2: Call the stress anomaly eigenvalues, obtain the inner wall data of the water flow channel of the heavy-duty gas turbine blade, calculate the inner wall stress distribution, analyze the stress disturbance mode of the defect region, identify the stress concentration points, calculate the stress offset of the material, screen the stress gradient mutation points, analyze the residual stress accumulation value, and establish the defect stress disturbance coefficient; S3: Call the defect stress disturbance coefficient, collect the stress distribution on the surface of the cooling channel of the water-cooled blade, screen the stress characteristics of three types of defects, namely cracks, spalling and deformation, calculate the stress direction offset angle, stress mean value and gradient mutation points, and generate the defect feature vector; S4: Call the defect feature vector, combine the stress anomaly feature data, normalize the classification features, obtain the blade surface image features, construct the stress feature channel, defect stress disturbance channel and image feature channel, and establish a multi-channel input data set.

2. The intelligent detection and defect analysis method for gas turbine blades according to claim 1, wherein, The stress anomaly eigenvalues are specifically stress concentration points, mutation points and stress mutation regions. The defect stress disturbance coefficient includes stress concentration points, material stress offset, stress gradient mutation points and residual stress accumulation values. The defect feature vector is specifically crack stress characteristics, spalling stress characteristics, deformation stress characteristics, stress direction offset angle, stress mean value and gradient mutation points. The multi-channel input data set includes a stress feature channel, a defect stress disturbance channel and an image feature channel.

3. The intelligent detection and defect analysis method for gas turbine blades according to claim 2, characterized in that, The specific steps for obtaining the stress anomaly eigenvalues are as follows: S101: Based on the blade material property data, blade geometric structure data and operating load data, calculate the stress transfer matrix of multiple nodes, call the stress values transferred between nodes, and obtain the stress gradient values in different directions according to the spatial coordinate relationship between nodes. Calculate the amplitude of the gradient change rate to obtain the stress gradient change coefficient; S102: Call the stress gradient change coefficient, calculate the stress mutation interval for the gradient change of multiple nodes, screen the stress anomaly regions according to the gradient change in multiple intervals, obtain the extreme points in the regions, and calculate the stress change rate in the adjacent regions of the extreme points to generate a set of stress mutation points; S103: Based on the set of stress mutation points, calculate the stress concentration degree corresponding to multiple mutation points, and use the formula: Operate to obtain the stress concentration coefficient of multiple mutation points, call the spatial coordinate information in the set of mutation points, establish the morphological distribution of the stress anomaly region, and obtain the stress anomaly eigenvalues; Among them, S1 represents the stress anomaly eigenvalue, Δσ 1,i represents the stress change at the mutation point i, d 1,i represents the Euclidean distance between the mutation point i and its adjacent point, n1 represents the total number of mutation points, and C represents the stability parameter, which is used to avoid the influence of the denominator approaching zero.

4. The intelligent detection and defect analysis method for gas turbine blades according to claim 3, characterized in that, The specific steps for obtaining the defect stress disturbance coefficient are as follows: S201: Call the stress anomaly eigenvalues, calculate the stress distribution value of the inner wall based on the inner wall data of the water flow channel of the heavy-duty gas turbine blade, call the stress anomaly eigenvalues, compare and analyze the stress change amplitudes at multiple positions on the inner wall, screen the stress change mutation points, determine the stress anomaly distribution region of the blade inner wall, and obtain the inner wall stress distribution data; S202: Based on the inner wall stress distribution data, analyze the stress perturbation pattern in the defect area, calculate the gradient change of the local stress, screen the stress gradient mutation points, and use the formula: Operate to obtain the local stress gradient change amount, calculate the stress offset value of the stress concentration point, and generate the stress offset amount of the stress concentration point; Among them, S2 represents the change in local stress gradient, Δσ 2,i represents the stress change at the i-th point, Δx i represents the stress measurement point spacing, E represents the elastic modulus of the material, ρ represents the material density, and n2 represents the total number of stress measurement points; S203: Based on the stress offset amount of the stress concentration point, calculate the residual stress accumulation value of the inner wall of the blade water flow channel, extract the residual stress distribution in the defect area, calculate the stress perturbation degree, and establish a defect stress perturbation coefficient.

5. The intelligent detection and defect analysis method for gas turbine blades according to claim 4, characterized in that, The specific steps for obtaining the defect feature vector are as follows: S301: Call the defect stress perturbation coefficient, collect the stress distribution on the surface of the cooling channel of the water-cooled blade, calculate the corresponding stress gradient change amount according to the stress values at multiple positions of the cooling channel, detect the stress mutation points in multiple regions of the cooling channel, screen the stress intervals where the stress mutation points are located, and obtain the stress gradient mutation point distribution; S302: Based on the stress gradient mutation point distribution, screen the stress characteristics of three types of defects: crack, spalling, and deformation, calculate the stress mean value in the corresponding area, and analyze the local stress direction offset situation. Use the formula: Operate to obtain the stress direction offset angle of the multi-defect area, and establish the stress direction offset angle distribution; Among them, V represents the variance of the mean stress, and S 3,i represents the stress value at the i-th location, and S 3,i,prev represents the stress value of the previous measurement point at the i-th location, and n3 represents the total number of measurement points in the defect area; S303: Call the stress direction offset angle distribution, combine it with the stress gradient mutation point distribution, construct the feature vector of each defect area, and extract the mean value of the feature vector, the direction offset angle, and the normalized value of the gradient mutation point for different types of defect areas to obtain the defect feature vector.

6. The intelligent detection and defect analysis method for gas turbine blades according to claim 5, characterized in that, The specific steps for obtaining the multi-channel input data set are as follows: S401: Based on the defect feature vector, combine the stress anomaly feature data, calculate the root mean square error and standard deviation of the eigenvalues, determine the feature data fluctuation range according to the root mean square error, screen the feature data with larger weights according to the standard deviation, perform normalization processing, and generate a feature normalization weight matrix; S402: Call the feature normalization weight matrix, obtain the blade surface image features, calculate the principal component contribution rate of multiple image feature dimensions, screen the features with higher contribution rates, calculate the correlation coefficient between multiple channels, and construct a stress feature channel, a defect stress perturbation channel, and an image feature channel. Use the formula: Calculate the influence degree of the comprehensive feature channel to obtain a multi-channel feature influence matrix; Among them, M represents the multi-channel feature influence matrix, W 4,k represents the normalized feature weight, F 4,k represents the normalized single eigenvalue, m4 represents the total number of features, C 4,kp represents the correlation coefficient between channels, n4 represents the total number of normalized weight features; S403: Call the multi-channel feature influence matrix, calculate the multi-channel feature contribution rate, screen the channel data with high contribution rates, allocate the data input weights according to the normalized distribution weight of the feature influence matrix, and integrate to establish a multi-channel input data set.

7. The intelligent detection and defect analysis method for gas turbine blades according to claim 6, wherein The method further includes: S5: Call the multi-channel input data set, construct the input layer of the neural network, set the weight constraints of the stress anomaly feature channel and the defect stress perturbation channel, optimize the loss function of the hidden layer, train the classification model, and establish a defect classification prediction result; The defect classification prediction result specifically refers to the classification labels of three types of defects: crack, spalling, and deformation.

8. The intelligent detection and defect analysis method for gas turbine blades according to claim 7, characterized in that The specific steps for obtaining the defect classification prediction result are as follows: S501: Call the multi-channel input data set, preprocess the stress data and defect stress perturbation data of multiple channels, calculate the stress feature vectors of the multi-channel input data, adjust the normalization parameters according to the data range, calculate and store the initial weight distribution of multiple feature channels, and obtain the input layer weight constraint matrix; S502: Based on the input layer weight constraint matrix, adjust the connection parameters of the hidden layer, perform weighted summation on the multi-channel stress features, and use the formula: Calculate the stress response value of the hidden layer; Among them, R h represents the stress response value of the hidden layer, W 5,p represents the weight of input channel p, S 5,p represents the stress eigenvalue of input channel p, P 5,q represents the weight of defect stress perturbation q, D 5,q represents the defect stress perturbation value, n5 is the total number of input channels, and m5 is the total number of defect stress perturbation channels; S503: Based on the stress response value of the hidden layer, train the classification model, calculate the classification boundary weights, screen the classification thresholds that meet the classification criteria of the stress anomaly feature channels, calculate the classification boundary parameters and store the classification information, and establish the defect classification prediction result.

Images