Titanium plate performance analysis method and system

Abstract

The invention provides a titanium plate performance analysis method and system, and relates to the technical field of metal material detection.The method comprises the steps that an optical scanner is adopted to establish a three-dimensional structure on the surface of a titanium plate, and test areas are divided; synchronously acquiring stress, hardness and conductivity data of a test point by using a multi-probe collaborative detection device; constructing a dual-branch feature extraction network, wherein the dual-branch feature extraction network comprises a surface topography feature branch and a physical feature branch; performing statistical integration on the features in the test area, calculating distribution features and determining a weight coefficient; and generating a performance analysis result through feature fusion and a multi-task output layer, and finally determining the performance grade of the titanium plate. According to the method, deep fusion of morphology information and physical characteristics is realized, and the accuracy and reliability of titanium plate performance evaluation are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to metal material detection technology, and in particular to a titanium plate performance analysis method and system. BACKGROUND

[0002] The existing titanium plate performance analysis method mainly relies on single physical quantity testing or simple surface morphology detection, which is difficult to comprehensively reflect the comprehensive performance characteristics of the titanium plate. For example, stress testing cannot reflect the microstructure changes of the material, hardness testing has limited prediction of material fatigue performance, and conductivity testing is easily disturbed by environmental factors. These single parameter analysis methods lack effective correlation mechanisms between them, resulting in one-sided analysis results and insufficient prediction accuracy of titanium plate performance, which is difficult to meet the stringent requirements of high-end fields such as aerospace. Although some methods in the prior art attempt to combine multiple testing methods, they are mostly simple superposition or average processing, and fail to establish a physical correlation model between parameters. At the same time, due to the obvious spatial non-uniformity of the titanium plate performance, the traditional method of uniform sampling strategy often ignores the key influence of the local abnormal area on the overall performance, resulting in a large deviation in the evaluation result. In addition, the lack of effective multi-scale feature fusion mechanism also limits the accuracy of the existing method for evaluating and predicting the performance of the titanium plate. SUMMARY

[0003] In view of the deficiencies of the prior art, the present application provides a titanium plate performance analysis method and system, which can solve the problems in the prior art.

[0004] In a first aspect of the embodiments of the present application, a titanium plate performance analysis method is provided, comprising:

[0005] A three-dimensional structure of the titanium plate to be tested is established by scanning the surface of the titanium plate to be tested using an optical scanner, and a plurality of test regions are divided, and a plurality of test points in a matrix distribution are set in each test region; the performance of each test point is detected by using a multi-probe cooperative detection device, the multi-probe cooperative detection device includes stress probes, hardness probes and conductivity probes which are uniformly arranged along the circumferential direction, and stress data, hardness data and conductivity data of the test point are synchronously collected; a double-branch feature extraction network is constructed, including a plane feature branch for extracting surface morphology features and a physical feature branch constructed based on material science knowledge; the physical feature branch includes a physical constraint layer and a physical correlation layer, for establishing the correlation between the stress data, the hardness data and the conductivity data; the distribution characteristics of the surface morphology features and the physical features in each test region are calculated, the corresponding feature weight coefficients are determined according to the distribution characteristics, and the features are weighted and fused; based on the fused features, a titanium plate performance analysis result is generated through a multi-task output layer, and the performance grade of the titanium plate to be tested is determined according to the performance analysis result.

[0006] Optionally, the step of scanning the surface of the titanium plate to be measured to establish a three-dimensional structure and dividing a plurality of test regions, and setting a plurality of test points in a matrix distribution in each test region comprises:

[0007] The image data of the surface of the titanium plate to be measured is collected using a dual-camera cooperative acquisition system; initial three-dimensional point cloud data is obtained based on phase measurement of the image data; vibration errors in the initial three-dimensional point cloud data are eliminated using a birefringence optical path difference method, and temperature drift errors in the initial three-dimensional point cloud data are compensated based on a pre-calibrated temperature-deformation curve to obtain compensated three-dimensional point cloud data; local curvature distribution data of the surface of the titanium plate to be measured is calculated based on the compensated three-dimensional point cloud data, and the surface of the titanium plate to be measured is divided into a plurality of test regions according to the local curvature distribution data, wherein the area of each test region is inversely proportional to the local curvature value in the test region; and a test point matrix is set in each test region based on the local curvature distribution data, wherein the spacing between adjacent test points in each test region is inversely proportional to the local curvature value at the position of the test point.

[0008] Optionally, the step of detecting the performance of each test point using a multi-probe cooperative detection device comprises:

[0009] The stress release characteristic time of the titanium plate to be measured is obtained, and the measurement timing of the stress probe, the hardness probe and the conductivity probe is determined according to the stress release characteristic time; during the measurement process of the stress probe, the load application rate is controlled to be proportional to the yield strength of the titanium plate to be measured, and the stress data of the test point is obtained; during the measurement process of the conductivity probe, the initial conductivity data of the test point is obtained, the disturbance correction coefficient of the stress field on the conductivity is calculated according to the stress data, the initial conductivity data is corrected based on the disturbance correction coefficient to obtain the conductivity data; during the measurement process of the hardness probe, the initial hardness data of the test point is obtained, the deformation compensation factor of the hardness measurement on the stress field is calculated according to the indentation depth, and the initial hardness data is corrected based on the deformation compensation factor to obtain the hardness data; and the stress data and the conductivity data are corrected according to the deformation compensation factor to obtain the corrected stress data and the conductivity data of the test point.

[0010] Optionally, a dual-branch feature extraction network is constructed, including a plane feature branch for extracting surface topography features and a physical feature branch constructed based on material science knowledge; the physical feature branch includes a physical constraint layer and a physical correlation layer, and the step of establishing the correlation between the stress data, the hardness data and the conductivity data comprises:

[0011] According to the height data of the surface of the titanium plate to be measured, local curvature values and height distribution probabilities are calculated, and a topography entropy feature is generated according to the local curvature values and the height distribution probabilities; height deviation data of the surface of the titanium plate to be measured are collected, and surface roughness parameters are calculated; the topography entropy feature and the surface roughness parameters are constructed as surface topography features; a hardness change rule corresponding to the stress data is analyzed, and a corresponding relationship between stress and hardness is established; a change rule of the electrical conductivity data with respect to dislocation density is analyzed, and a corresponding relationship between electrical conductivity and dislocation density is established; the corresponding relationship between stress and hardness and the corresponding relationship between electrical conductivity and dislocation density are taken as physical constraint features; an interactive influence matrix of stress, hardness and electrical conductivity is constructed, the interaction strength of the stress data, the hardness data and the electrical conductivity data is calculated according to the interactive influence matrix, the stress data, the hardness data and the electrical conductivity data are corrected based on the interaction strength, and the corrected data are constructed as physical correlation features.

[0012] Optionally, distribution characteristics of the surface topography features and the physical features in each test area are calculated, corresponding feature weight coefficients are determined according to the distribution characteristics, and the step of feature weighted fusion includes:

[0013] The topography entropy space distribution of the surface topography features in each test area is calculated, the topography entropy space distribution is calculated by the negative logarithm of the surface topography feature probability in the test area; the multi-physical quantity correlation strength of the physical features in each test area is calculated, the multi-physical quantity correlation strength is calculated based on the correlation coefficient matrix between physical quantities; the dispersion coefficient and the anomaly detection index in the test area are calculated based on the topography entropy space distribution and the multi-physical quantity correlation strength; the region importance is weighted by the dispersion coefficient and the multi-physical quantity correlation strength, and the feature reliability is calculated based on the anomaly detection index; the local weight coefficient of each test area is calculated according to the region importance and the feature reliability, and the global weight coefficient is calculated according to the local weight coefficient; the surface topography features and the physical features in each test area are weighted based on the local weight coefficient to obtain local fusion features, and the local fusion features are weighted based on the global weight coefficient to obtain global fusion features.

[0014] Optionally, the step of calculating the dispersion coefficient and the anomaly detection index in the test area based on the topography entropy space distribution and the multi-physical quantity correlation strength includes:

[0015] According to the topographic entropy space distribution, a scale dispersion coefficient of the test region is calculated, and a direction dispersion coefficient is calculated according to the multi-physical quantity correlation strength, wherein the scale dispersion coefficient is calculated by a weighted sum of ratios of standard deviations to mean values at multiple characteristic scales, and the direction dispersion coefficient is obtained by normalization of a deviation of a feature gradient from a mean value; the scale dispersion coefficient and the direction dispersion coefficient are combined by weighting to obtain a comprehensive dispersion coefficient; a feature anomaly score is calculated according to the topographic entropy space distribution and the multi-physical quantity correlation strength, a time sequence weight is updated according to a time sequence change of the feature anomaly score, and an anomaly detection index is obtained by adjusting the feature anomaly score based on the time sequence weight; a region importance of each test region is obtained by combining the comprehensive dispersion coefficient and the anomaly detection index, a feature weight of the test region is updated based on the region importance, and a local weight coefficient of the test region is obtained; a global weight coefficient is obtained by combining the local weight coefficient and a feature similarity of an adjacent test region by weighting, and the global weight coefficient is proportional to the feature similarity of the adjacent test region.

[0016] Optionally, based on the fused features, a titanium plate performance analysis result is generated through a multi-task output layer, and according to the performance analysis result, a performance grade of the titanium plate to be tested is determined.

[0017] The fused features are input into a shared feature extraction layer and a task-specific layer, the shared feature extraction layer performs hierarchical extraction of the fused features at multiple scales to generate multi-scale features, and the task-specific layer includes a performance distribution prediction branch, a defect type identification branch, and a reliability evaluation branch; in the performance distribution prediction branch, spatial distribution data of the multi-scale features is calculated, a distribution interval of a performance index is determined based on the spatial distribution data, the distribution interval is dynamically updated, and a performance distribution prediction result is generated; in the defect type identification branch, the multi-scale features are matched with preset defect features layer by layer to determine optimal matching features, and a defect type and a corresponding confidence thereof are output according to the optimal matching features; in the reliability evaluation branch, the performance distribution prediction result and the defect type and the confidence thereof are taken as evaluation bases, a reliability of each evaluation base is calculated, each evaluation base is combined by weighting according to the reliability, and a reliability evaluation result is obtained; according to a pre-set performance grade determination standard, the performance distribution prediction result, the defect type and the confidence thereof, and the reliability evaluation result are combined to determine the performance grade of the titanium plate to be tested through interval mapping.

[0018] Optionally, the performance distribution prediction result and the defect type and the confidence thereof are taken as evaluation bases, a reliability of each evaluation base is calculated, each evaluation base is combined by weighting according to the reliability, and a reliability evaluation result is obtained.

[0019] calculate a prediction error of a probability distribution deviation of the performance distribution prediction result and a defect type confidence, construct an uncertainty matrix based on the probability distribution deviation and the prediction error; calculate a time sequence fluctuation feature of each evaluation basis according to the uncertainty matrix to obtain a stability index of the evaluation basis; perform tracking analysis on a historical evaluation result of the evaluation basis based on the stability index, calculate an accuracy rate change trend of the evaluation result, and generate a dynamic evaluation weight according to the accuracy rate change trend; combine the stability index and the dynamic evaluation weight to obtain a credibility of each evaluation basis, and perform weighted fusion on the performance distribution prediction result and the defect type and its confidence according to the credibility, and output a reliability evaluation result. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 A flowchart of a titanium plate performance analysis method according to an embodiment of the present application is shown in

[0021] Figure 2 A flowchart of a titanium plate performance analysis double-branch feature extraction method is shown in DETAILED DESCRIPTION

[0022] Figure 1 A flowchart of a titanium plate performance analysis method according to an embodiment of the present application is shown in Figure 1 As shown in the figure, the method comprises:

[0023] A three-dimensional structure is established by scanning the surface of the titanium plate to be tested using an optical scanner, and a plurality of test regions are divided, and a plurality of test points in a matrix distribution are set in each test region;

[0024] A multi-probe cooperative detection device is used to detect the performance of each test point, the multi-probe cooperative detection device comprises stress probes, hardness probes and conductivity probes which are uniformly arranged along the circumferential direction, and stress data, hardness data and conductivity data of the test point are synchronously collected;

[0025] A double-branch feature extraction network is constructed, which comprises a plane feature branch for extracting surface topography features and a physical feature branch constructed based on material science knowledge; the physical feature branch comprises a physical constraint layer and a physical correlation layer, which are used to establish the correlation between stress data, hardness data and conductivity data; the features extracted from the plurality of test points in each test region are statistically integrated to obtain a feature representation of the test region; the distribution features of the surface topography features and the physical features in each test region are calculated, and the corresponding feature weight coefficients are determined according to the distribution features, and the features are weighted and fused;

[0026] Based on the fused features, a titanium plate performance analysis result is generated through a multi-task output layer, and the performance grade of the titanium plate to be tested is determined according to the performance analysis result.

[0027] For example, the surface of the titanium plate to be tested is scanned by an optical scanner to establish a three-dimensional structure. The established three-dimensional structure is regionally divided, and the whole titanium plate is divided into multiple test regions. The region division adopts a grid method, and the 1 m x 1 m titanium plate is uniformly divided into a 10 x 10 grid, each grid being 100 mm x 100 mm, to form 100 test regions. Each test region is identified by a corresponding number, such as A1 to J10. A plurality of test points are arranged in a matrix distribution in each test region. For each 100 mm x 100 mm test region, a 5 x 5 test point matrix is arranged, and the test point spacing is 20 mm. Each test region has a total of 25 test points. The test points are identified by the region number plus the point number, such as A1-1 to A1-25.

[0028] A multi-probe cooperative detection device is used to detect the performance of each test point. The detection device is positioned above each test point by a precision positioning system, the support is lowered to make the probe contact the titanium plate surface, and the stress data, hardness data, and conductivity data of the test point are synchronously collected. The collected data is transmitted to a data processing unit by a high-speed data acquisition system to form a multi-physical quantity data set of each test point. For the 2500 test points of the whole titanium plate, a total of 7500 original data values are generated.

[0029] A double-branch feature extraction network is constructed, including a planar feature branch for extracting surface topography features and a physical feature branch based on material science knowledge. The planar feature branch extracts surface topography features from three-dimensional structure data. The extraction process first calculates the statistical properties of height data within each test region, including mean, variance, skewness, and kurtosis; then calculates the local curvature values by fitting a quadratic surface within a 3×3 region around each test point, and calculates the principal curvature and Gaussian curvature; finally, generates the height distribution probability, divides the height values into 10 equal-interval intervals, and calculates the probability of each interval as the proportion of the number of points in the interval to the total number of points. The topography entropy features include curvature entropy and height entropy. The curvature entropy is calculated by the probability-weighted sum of the principal curvature distribution, reflecting the complexity of the surface curvature distribution; the height entropy is calculated by the weighted sum of the height distribution probability, reflecting the uniformity of the surface height distribution. For well-processed TC4 titanium plates, the curvature entropy is usually between 0.2-0.5, and the height entropy is between 0.8-1.2; for regions with processing marks or defects, the curvature entropy exceeds 0.8, and the height entropy exceeds 1.5. At the same time, the height deviation data of the surface of the titanium plate to be tested is collected, and the surface roughness parameters are calculated. The roughness parameters include the arithmetic average roughness Ra, the maximum peak-to-valley height Rz, and the profile peak density RPc. The arithmetic average roughness Ra is the mean of the absolute values of the height deviation, with a typical value of 0.8-2.5μm; the maximum peak-to-valley height Rz is the height difference between the highest point and the lowest point, with a typical value of 5-15μm; the profile peak density RPc is the number of peaks per unit length, with a typical value of 30-60 / cm. The topography entropy features and surface roughness parameters are constructed as surface topography features, forming a 7-dimensional feature vector that comprehensively represents the topography characteristics of the titanium plate surface. The physical feature branch includes a physical constraint layer and a physical correlation layer, which are used to establish the correlation between stress data, hardness data, and electrical conductivity data. The physical constraint layer analyzes the hardness variation law corresponding to the stress data, and establishes the correspondence between stress and hardness. For TC4 titanium alloy, the correspondence between stress and hardness is based on the strain hardening theory of material science. When the material undergoes plastic deformation under external force, the hardness value increases with the increase of stress. Through statistical analysis of a large amount of test data, it is found that for every 100MPa increase in stress, the hardness value of TC4 titanium alloy increases by about 20HV. The specific correspondence can be expressed as: when the stress is in the range of 800-850MPa, the hardness value is usually in the range of 330-340HV; when the stress is in the range of 850-900MPa, the hardness value is usually in the range of 340-350HV; when the stress is in the range of 900-950MPa, the hardness value is usually in the range of 350-370HV. This correspondence reflects the correlation between the dislocation density inside the material and the macroscopic mechanical properties. The physical correlation layer constructs the interaction matrix of stress, hardness, and electrical conductivity.The interaction matrix is a 3x3 matrix, the diagonal elements are 1, and the non-diagonal elements represent the degree of mutual influence between different physical quantities. Through the analysis of a large number of test data of TC4 titanium alloy, the typical interaction matrix values are determined: the influence of stress on hardness is 0.7, and the influence of stress on electrical conductivity is-0.4; the influence of hardness on stress is 0.5, and the influence of hardness on electrical conductivity is-0.3; the influence of electrical conductivity on stress is-0.2, and the influence of electrical conductivity on hardness is-0.1. Positive values represent positive correlation, negative values represent negative correlation, and the absolute value size represents the degree of influence.

[0030] According to the interaction matrix, the interaction intensity of stress data, hardness data and electrical conductivity data is calculated. The interaction intensity is calculated by the product of the interaction coefficient between physical quantities and the normalized value of the physical quantity. The normalization of the physical quantity is to subtract the average value and divide by the standard deviation. For TC4 titanium alloy, the typical normalization range is-2 to 2. Based on the interaction intensity, the stress data, hardness data and electrical conductivity data are corrected. The correction process adopts an iterative method. In each iteration, according to the influence of the other two physical quantities and the current value, the correction amount of the current physical quantity is calculated, and the iteration is performed for 5-10 times until the correction amount changes little. The corrected stress, hardness and electrical conductivity data are constructed into physical correlation features to form a 9-dimensional feature vector.

[0031] The features extracted from multiple test points in each test area are statistically integrated to obtain the feature representation of the test area. For 25 test points in each test area, the mean, variance, maximum, minimum and median of the 7-dimensional surface morphology features and the 15-dimensional physical features (including physical constraint features and physical correlation features) are calculated to form the comprehensive feature representation of the area. After statistical integration, the feature dimension of each test area remains unchanged, but represents the average level and variation of the entire area.

[0032] The distribution characteristics include spatial distribution characteristics and numerical distribution characteristics. The spatial distribution characteristics are determined by calculating the change gradient of the features between adjacent areas, and the numerical distribution characteristics are determined by calculating the distribution form of the feature values. For features with uniform spatial distribution and stable numerical distribution, a lower weight is given; for features with uneven spatial distribution and large numerical distribution fluctuations, a higher weight is given. Feature weighted fusion adopts a weighted average method, and the surface morphology features and physical features are weighted and averaged according to the weight coefficients to obtain the fused feature vector. Based on the fused features, the titanium plate performance analysis results are generated through a multi-task output layer.

[0033] Optionally, the step of scanning the surface of the titanium plate to be tested by the optical scanner to establish a three-dimensional structure and dividing a plurality of test areas and setting a plurality of test points in a matrix distribution in each test area comprises:

[0034] acquire image data of a surface of a titanium plate to be measured using a dual-camera cooperative acquisition system; and acquire initial three-dimensional point cloud data based on the image data by phase measurement;

[0035] eliminate vibration errors in the initial three-dimensional point cloud data by using a birefringent optical path difference method, and compensate for temperature drift errors in the initial three-dimensional point cloud data based on a pre-calibrated temperature-deformation curve to obtain compensated three-dimensional point cloud data;

[0036] calculate local curvature distribution data of the surface of the titanium plate to be measured based on the compensated three-dimensional point cloud data, and divide the surface of the titanium plate to be measured into a plurality of test regions according to the local curvature distribution data, wherein the area of each test region is inversely proportional to the local curvature value in the test region;

[0037] set a test point matrix in each test region based on the local curvature distribution data, wherein the spacing between adjacent test points in each test region is inversely proportional to the local curvature value at the position of the test point.

[0038] For example, a dual-camera cooperative acquisition system is used to acquire images of the surface of the titanium plate. The dual-camera system includes two industrial-grade cameras, the baseline distance between the cameras is 180 mm, the included angle between the optical axes of the two cameras is 15 degrees, the focal length is 35 mm, and the field of view range is 300 mm x 400 mm. The resolution of the cameras is 4096 x 3072 pixels, and the frame rate is 60 frames / s.

[0039] During image acquisition, the surface of the titanium plate is treated with a diffuse reflection coating to enhance the surface texture features. The dual-camera system uses structured light illumination, projects a grid pattern onto the surface of the titanium plate, and continuously acquires 30 images at different angles with a grid period of 1 mm. The acquired image data is processed by a phase unwrapping algorithm to calculate the correspondence between the phase value and the actual height. The phase value changes by 360 degrees, corresponding to an actual height change of 0.5 mm. By the phase measurement method, initial three-dimensional point cloud data with a resolution of 0.05 mm is generated.

[0040] There are noise and drift errors caused by temperature changes in the initial point cloud data. The birefringent optical path difference method is used to eliminate the vibration error, and the specific implementation is as follows: a birefringent crystal is added to the optical path, so that the incident light is divided into two mutually perpendicular polarized lights, and the optical path difference is proportional to the vibration displacement. By detecting the change of the optical path difference, the displacement error caused by vibration is calculated.

[0041] The compensation of temperature drift error is based on the pre-calibrated temperature-deformation curve. During the calibration process, a high-precision temperature sensor is used to record the environmental temperature changes while monitoring the deformation of the titanium plate surface. Within the temperature range of 15-35 degrees Celsius, a data point is taken every 2 degrees to establish the corresponding relationship between temperature and deformation. For example, when the titanium plate surface height increases by 1 degree Celsius, the average expansion amount is 0.003 millimeters. Based on this relationship, the environmental temperature is measured in real time, and the point cloud data is compensated for temperature drift, improving the accuracy of the compensated point cloud data to 0.003 millimeters.

[0042] Based on the compensated three-dimensional point cloud data, the local curvature distribution of the titanium plate surface is calculated. The curvature calculation uses the neighborhood fitting method, taking the points within a radius of 5 millimeters as the center of each point, fitting a quadratic surface, and calculating the Gaussian curvature and average curvature at that point. The Gaussian curvature value range is usually between 0.001-0.1, and the average curvature value range is between 0.005-0.05. According to the calculated curvature distribution data, the titanium plate surface is divided into multiple test areas, and the area division uses the curvature threshold segmentation method. First, set the reference curvature threshold to 0.01, mark the area with a curvature value greater than the threshold as a high curvature area, and mark the area with a curvature value less than the threshold as a low curvature area. For positions with a curvature change gradient greater than 0.005 / mm, set the area boundary. The area of the test area is inversely proportional to the average local curvature value in the area, and the specific proportion coefficient is 10000 square millimeters. For example, when the average curvature value of a certain area is 0.02, the area is set to 500 square millimeters; when the average curvature value is 0.01, the area is set to 1000 square millimeters. This division method ensures more intensive test coverage in areas with large curvature changes.

[0043] Test points are set in each divided test area. The test point layout is also related to the local curvature value, and the spacing between test points is inversely proportional to the local curvature value at that position. Set the reference spacing to 10 millimeters and the reference curvature to 0.01. When the local curvature value is 0.02, the spacing between adjacent test points is set to 5 millimeters; when the local curvature value is 0.005, the spacing between adjacent test points is set to 20 millimeters. For test points at the boundary, increase the density by 25% to better capture the boundary features. For example, for a titanium plate sample of 300mm x 300mm, it is generally divided into 15-25 test areas, with 16-64 test points in each area, forming a matrix-like distribution.

[0044] This method uses a curvature-based adaptive area division and test point layout strategy, significantly improving test efficiency and data quality; it has stronger adaptability to the geometric features of the titanium plate surface, and can achieve more intensive sampling in complex structure areas while ensuring measurement accuracy.

[0045] Optionally, the step of detecting the performance of each test point by using the multi-probe cooperative detection device comprises:

[0046] obtaining the stress release characteristic time of the titanium plate to be tested, and determining the measurement time sequence of the stress probe, the hardness probe and the conductivity probe according to the stress release characteristic time;

[0047] in the stress probe measurement process, the load application rate is controlled to be proportional to the yield strength of the titanium plate to be tested, and the stress data of the test point are obtained;

[0048] in the conductivity probe measurement process, the initial conductivity data of the test point are obtained, the disturbance correction coefficient of the stress field on the conductivity is calculated according to the stress data, the initial conductivity data are modified based on the disturbance correction coefficient, and the conductivity data are obtained;

[0049] in the hardness probe measurement process, the initial hardness data of the test point are obtained, the deformation compensation factor of the hardness measurement on the stress field is calculated according to the indentation depth, the initial hardness data are modified based on the deformation compensation factor to obtain the hardness data; and the stress data and the conductivity data are modified according to the deformation compensation factor to obtain the modified stress data and the conductivity data of the test point.

[0050] For example, the multi-probe cooperative detection device is composed of a stress probe, a hardness probe and a conductivity probe. The three probes are uniformly distributed along the circumferential direction and are 120 degrees apart from each other to form a three-probe detection unit. Each probe is connected to a central controller to realize synchronous data acquisition and analysis. The stress probe uses a piezoelectric strain sensor with a sensitivity of 0.5 microvolts / microstrain; the hardness probe uses a micro-indentation test principle with a maximum load of 5 Newtons and a diamond Vickers indenter; and the conductivity probe uses a four-point measurement method with a measurement range of 0.5-60 megasiemens / meter and a resolution of 0.01 megasiemens / meter.

[0051] Before the detection process starts, the stress release characteristic time of the titanium plate to be tested needs to be obtained. The stress release characteristic time refers to the time required for the stress value to stabilize after the titanium plate is subjected to a load. This parameter is related to the material composition and heat treatment state of the titanium plate. The determination of the stress release characteristic time is determined by pre-test: a standard load (usually 50% of the yield strength of the titanium plate) is applied to the titanium plate sample, and the change curve of the stress value with time is recorded. The time interval from the peak value to the stable value of the stress value is the stress release characteristic time. For common TC4 titanium alloy plates, the stress release characteristic time is generally between 3-8 seconds; for high-strength titanium alloys subjected to special heat treatment, the time is extended to 10-15 seconds.

[0052] According to the measured stress release characteristic time, the measurement timing of the three probes is determined. The principle of the measurement timing design is that the stress probe is measured first, and then the conductivity probe and the hardness probe are measured in turn after the stress field is stable. The specific timing arrangement is that the stress probe measurement duration is 1.5 times the stress release characteristic time to ensure that the stress field is completely stable; the conductivity probe starts immediately after the stress probe measurement ends, and the measurement duration is 2 seconds; the hardness probe starts after the conductivity probe measurement ends, and the measurement duration is 3 seconds. This timing arrangement can minimize the mutual interference between the measurements of various physical quantities.

[0053] During the measurement process of the stress probe, the load application rate (Newton / second) = yield strength (MPa) x 0.05. For example, for TC4 titanium alloy with a yield strength of 900 MPa, the load application rate is set to 45 Newton / second. The load application adopts a gradient increase method, that is, the load is increased by 70% of the set rate in the first 30% of the time, by 100% of the set rate in the middle 40% of the time, and by 50% of the set rate in the last 30% of the time, forming a smooth load curve. The raw data obtained by the stress probe measurement is collected by a data collection system with a sampling frequency of 1000 Hz, and the high-frequency noise is removed by a filtering algorithm to obtain the stress data of the test points.

[0054] During the measurement process of the conductivity probe, the initial conductivity data of the test point is first obtained. The initial conductivity data is obtained by a four-point measurement method, four equally spaced electrodes are arranged in the measurement area, the outer two electrodes are connected to a constant current (usually 5 milliamps), and the inner two electrodes measure the potential difference. According to Ohm's law, the resistance value is calculated, and then combined with the geometric parameters of the test area (electrode spacing is 2 mm, measurement depth is 0.5 mm) to convert to conductivity value. The original conductivity value is temperature corrected (reference temperature 20 degrees Celsius, correction coefficient is 0.4% increase in conductivity per 1 degree Celsius increase) to obtain the initial conductivity data. Since the stress field will disturb the conductivity measurement, the disturbance correction coefficient of the stress field on the conductivity needs to be calculated. The disturbance correction coefficient is related to the stress value and stress gradient, and the calculation method is: when the stress value of the measurement point is in the range of 0-300 MPa, the disturbance correction coefficient is 1+stress value x 0.0002; when the stress value is in the range of 300-800 MPa, the disturbance correction coefficient is 1+stress value x 0.0003; when the stress value is greater than 800 MPa, the disturbance correction coefficient is 1+stress value x 0.0004. For example, when the stress value of the measurement point is 500 MPa, the disturbance correction coefficient is 1.15. Divide the initial conductivity data by the disturbance correction coefficient to obtain the corrected conductivity data. For TC4 titanium alloy, the conductivity data before correction is generally between 0.58-0.62 megasiemens / meter, and the corrected conductivity data is generally between 0.52-0.56 megasiemens / meter.

[0055] The hardness probe measurement process adopts a multi-stage loading method, and the load is 2 Newton, 3.5 Newton and 5 Newton respectively. Each load level maintains for 0.5 seconds. Record the indentation depth under each load level, and calculate the initial hardness data. The calculation of the initial hardness data is based on the principle of Vickers hardness measurement. The hardness value is calculated by measuring the diagonal length of the indentation (using the built-in optical microscope system, magnification is 400 times, resolution is 0.5 microns). The specific calculation process is to divide the applied load (unit: kilogram force) by the indentation surface area (unit: square millimeter), and then multiply by the constant 1.8544 to get the Vickers hardness value. For each test point, take the average of the hardness values calculated under three loads as the initial hardness data. Since the hardness measurement process will deform the local stress field, it is necessary to calculate the deformation compensation factor. The deformation compensation factor is closely related to the indentation depth. When the indentation depth is in the range of 0-20 microns, the deformation compensation factor is 1+indentation depth x 0.015; when the indentation depth is in the range of 20-40 microns, the deformation compensation factor is 1+indentation depth x 0.02; when the indentation depth is greater than 40 microns, the deformation compensation factor is 1+indentation depth x 0.025. Taking the indentation depth of 25 microns as an example, the deformation compensation factor is 1.5. Multiply the initial hardness data by the deformation compensation factor to get the modified hardness data. For TC4 titanium alloy, the hardness value before correction is usually between 320-350HV, and the hardness value after correction is between 480-525HV.

[0056] The local deformation caused by hardness measurement not only affects the hardness value, but also affects the measured stress data and conductivity data. Therefore, it is necessary to make a secondary correction to the stress data and the conductivity data. The correction method of the stress data is: corrected stress value=original stress value x(1-deformation compensation factor x 0.1). The correction method of the conductivity data is: corrected conductivity value=original conductivity value x(1-deformation compensation factor x 0.05). For the case where the deformation compensation factor is 1.5, the stress value correction coefficient is 0.85, and the conductivity value correction coefficient is 0.925. Through this multi-physical quantity interactive correction mechanism, the corrected stress data, hardness data and conductivity data of the test point are finally obtained.

[0057] The multi-probe collaborative detection method adopted by the method realizes high-precision collaborative measurement of stress, hardness and conductivity by reasonably designing the measurement timing; the measurement timing design based on the stress release characteristic time reduces the measurement interference; through the introduction of the disturbance correction coefficient and the deformation compensation factor, an interactive correction mechanism between multiple physical quantities is established, which significantly improves the accuracy and consistency of the measurement data, and provides a reliable data basis for the comprehensive evaluation of the performance of titanium plate.

[0058] Optionally, a double-branch feature extraction network is constructed, including a plane feature branch for extracting surface topography features and a physical feature branch constructed based on material knowledge; the physical feature branch includes a physical constraint layer and a physical correlation layer, and the step of establishing the correlation between the stress data, the hardness data and the conductivity data includes:

[0059] According to the height data of the surface of the titanium plate to be tested, a local curvature value and a height distribution probability are calculated, and a topography entropy feature is generated according to the local curvature value and the height distribution probability; height deviation data of the surface of the titanium plate to be tested is collected, and a surface roughness parameter is calculated; and the topography entropy feature and the surface roughness parameter are constructed as surface topography features.

[0060] The hardness variation law corresponding to the stress data is analyzed, and the corresponding relationship between stress and hardness is established; the variation law of the conductivity data with dislocation density is analyzed, and the corresponding relationship between conductivity and dislocation density is established; the corresponding relationship between stress and hardness and the corresponding relationship between conductivity and dislocation density are taken as physical constraint features.

[0061] An interaction influence matrix of stress, hardness and conductivity is constructed, the interaction strength of the stress data, the hardness data and the conductivity data is calculated according to the interaction influence matrix, the stress data, the hardness data and the conductivity data are corrected based on the interaction strength, and the corrected data is constructed as physical correlation features.

[0062] In combination with Figure 2 The titanium plate performance analysis double-branch feature extraction flowchart is described as follows: the double-branch network adopts a parallel structure design, and the features extracted by the two branches are merged in the fusion layer after being extracted independently. The plane feature branch adopts a multi-scale convolution structure, including 3 scale layers, and the convolution kernel sizes are 3x3, 5x5 and 7x7 respectively, and each layer includes 16 convolution filters. The physical feature branch includes a physical constraint layer and a physical correlation layer, the physical constraint layer uses a full connection structure, the input node number is the number of physical quantity data of the test points, the hidden layer node number is 128, and the output layer node number is 64; the physical correlation layer adopts a graph convolution structure, the node number is 32, and the edge feature dimension is 16.

[0063] The feature extraction process of the planar feature branch starts from calculating the local curvature value and the height distribution probability according to the height data of the titanium plate surface to be tested. The height data is derived from the aforementioned three-dimensional point cloud data, and the sampling interval is 0.1 mm. The local curvature value is calculated by using the surface fitting method. Taking each point as the center, a circular area with a radius of 0.5 mm is taken, a quadratic surface is fitted, the principal curvatures k1 and k2 at the point are solved, the Gaussian curvature (k1 x k2) and the average curvature ((k1+k2) / 2) are calculated. For a typical TC4 titanium plate surface, the Gaussian curvature value ranges from 0.0001 to 0.01, and the average curvature value ranges from 0.001 to 0.05. The height distribution probability is obtained by statistical analysis of the distribution of height values in the region. The height value range is divided into 10 equal intervals, the number of points in each interval is calculated as a proportion of the total number of points, and a height distribution histogram is obtained.

[0064] The topographic entropy feature includes curvature entropy and height entropy. The curvature entropy calculation method is to divide the local curvature value range into 8 intervals, and calculate the proportion of the number of points in each interval. The negative logarithm of each proportion value is multiplied by the proportion value and then summed. For a region with uniform surface topography, the curvature entropy value is low, usually between 0.5-1.5; for a region with complex topography, the curvature entropy value is high, up to 2.0-3.0. The height entropy is calculated based on the aforementioned height distribution histogram, and the calculation method is similar to that of the curvature entropy. The negative logarithm of each interval proportion value is multiplied by the proportion value and then summed. The height entropy value of a typical titanium plate surface ranges from 1.0 to 2.5.

[0065] The titanium plate surface height deviation data refers to the difference between the height value of each point and the average height of the region. The surface roughness parameters include the arithmetic average roughness Ra, the root mean square roughness Rq and the maximum height difference Rz. The arithmetic average roughness Ra is the average of the absolute value of the height deviation, the root mean square roughness Rq is the square root of the average of the square of the height deviation, and the maximum height difference Rz is the height difference between the highest point and the lowest point in the region. For an aviation-grade TC4 titanium plate, Ra is usually between 0.8-1.5 microns, Rq is between 1.0-1.8 microns, and Rz is between 4.0-8.0 microns. The topographic entropy feature (curvature entropy and height entropy) and the surface roughness parameter (Ra, Rq, Rz) are combined into a 7-dimensional feature vector to form a surface topography feature. The feature vector is mapped through a feature mapping layer, the mapping layer uses a two-layer fully connected network structure, the number of hidden layer nodes is 32, the number of output layer nodes is 16, and the activation function uses ReLU. The 7-dimensional feature vector is mapped to a 16-dimensional feature representation.

[0066] The physical feature branch first analyzes the hardness variation law corresponding to the stress data, and establishes the corresponding relationship between stress and hardness. The stress data and hardness data of each test point are paired and analyzed, and the stress-hardness data pair is extracted. For TC4 titanium alloy, when the stress is in the range of 0-300 MPa, the hardness change rate is about 0.2 HV / Mpa; when the stress is in the range of 300-600 MPa, the hardness change rate is about 0.3 HV / Mpa; when the stress is in the range of 600-900 MPa, the hardness change rate is about 0.4 HV / Mpa. Based on these change rates, a piecewise linear mapping relationship is established to map the stress value to the expected hardness value, and the deviation between the actual hardness value and the expected hardness value is calculated. For stable titanium plates, the deviation value is usually less than 5%; for areas with defects or unstable performance, the deviation value can reach 15%-25%.

[0067] In the process of establishing the corresponding relationship between electrical conductivity and dislocation density, the relationship between dislocation density and material hardness is based on the dislocation strengthening theory in metal materials science. For TC4 titanium alloy, the relationship between dislocation density and hardness is determined by Taylor relationship and Hall-Petch effect: when the material is plastically deformed, the increment of hardness value is proportional to the square root of dislocation density. In specific calculation, the reference hardness value (about 330 HV) and the corresponding reference dislocation density (about 5×10 10 / meter square) of TC4 titanium alloy in standard annealing state are measured first, and then the dislocation density increment is calculated according to the hardness increment. For example, when the hardness value increases from the reference value by 50 HV to 380 HV, the corresponding dislocation density increases by about 6×10 12 / meter square. The relationship between electrical conductivity and dislocation density follows Matthiessen's rule, that is, the resistance increment is proportional to the dislocation density. By measuring the electrical conductivity and corresponding dislocation density of TC4 titanium alloy in different deformation states, an empirical relationship is established: for every 1×10 13 / meter square increase in dislocation density, the electrical conductivity decreases by about 4.8%. Based on this relationship, the expected electrical conductivity value is calculated, and compared with the measured electrical conductivity to obtain the electrical conductivity deviation. The electrical conductivity deviation of normal titanium plate is usually within ±2%; the deviation of areas with micro defects can reach ±5% or more. The deviation of the corresponding relationship between stress and hardness and the deviation of the corresponding relationship between electrical conductivity and dislocation density are combined into a physical constraint feature, forming a 4-dimensional feature vector.

[0068] The interaction matrix is a 3x3 matrix, the diagonal elements are 1, and the non-diagonal elements represent the influence coefficient between two physical quantities. For TC4 titanium alloy, the influence coefficient of stress on hardness is about 0.4, the influence coefficient of stress on electrical conductivity is about 0.3, the influence coefficient of hardness on stress is about 0.2, the influence coefficient of hardness on electrical conductivity is about 0.25, the influence coefficient of electrical conductivity on stress is about 0.1, and the influence coefficient of electrical conductivity on hardness is about 0.15. The construction of the interaction matrix is based on the fitting of experimental data and the theory of materials science. The specific construction method is as follows: first, collect the measured data of hardness and electrical conductivity under different stress states, at least 100 standard sample data are required; then use multivariate regression analysis to determine the coefficient relationship between each physical quantity; finally, adjust the coefficients to minimize the prediction error through iterative optimization method. The matrix is determined in advance before network training, and is used as a fixed parameter of the physical constraint layer.

[0069] The interaction strength of the three physical quantities is calculated according to the interaction matrix. The calculation method of the interaction strength is as follows: for physical quantities i and j, the interaction strength is equal to the corresponding element value in the interaction matrix multiplied by the normalized value of physical quantity j. The normalization process maps each physical quantity to the range of 0-1, and the normalization formula uses the min-max scaling method, i.e. (value-min value) / (max value-min value). For TC4 titanium alloy, the normalized range of stress value is 0-1000 MPa, the normalized range of hardness value is 300-600 HV, and the normalized range of electrical conductivity value is 0.5-0.7 MS / m.

[0070] Based on the interaction strength, the stress data, hardness data and electrical conductivity data are corrected. The correction process uses an iterative method. In each iteration, the correction amount of the current physical quantity is calculated according to the influence coefficient and the current value of the other two physical quantities. The iteration is performed 5-10 times until the correction amount is less than the threshold value (usually set to 0.1% of the original value). When calculating the correction value of the current physical quantity, the original value is first retained, and then the influence of each other physical quantity is accumulated. The influence of each physical quantity is equal to the normalized value of the physical quantity multiplied by the corresponding interaction coefficient and then multiplied by a 0.1 adjustment factor. After accumulating these influences, the original value is multiplied to obtain the final correction value. For example, when correcting the stress value, the influences of hardness and electrical conductivity need to be considered; when correcting the hardness value, the influences of stress and electrical conductivity need to be considered; and when correcting the electrical conductivity value, the influences of stress and hardness need to be considered. The difference between the corrected stress data and the original data is usually between 3%-8%, the difference between the hardness data is between 4%-10%, and the difference between the electrical conductivity data is between 2%-6%. The corrected stress data, hardness data and electrical conductivity data are constructed into physical correlation features to form a 9-dimensional feature vector. The feature vector is encoded by the physical constraint layer to output a 64-dimensional feature representation.

[0071] The physical constraint layer is the key link of how to actually encode the material science knowledge. The weight initialization of the physical constraint layer is not random, but preset based on the physical law of materials. Specifically, the initial value of the weight matrix of the fully connected layer is determined according to the theoretical model of the stress-hardness relationship, the hardness-dislocation density relationship and the dislocation density-conductivity relationship. For example, the first three rows of the weight matrix correspond to the stress-hardness relationship, and the initial value is set according to the aforementioned piecewise linear mapping relationship; the 4th-6th rows correspond to the hardness-dislocation density relationship; the 7th-9th rows correspond to the dislocation density-conductivity relationship. This initialization method enables the network to contain material science knowledge before the training starts, greatly accelerating the training process and improving the material interpretation.

[0072] The physical correlation layer adopts a graph convolution network structure to represent multiple physical quantities and their relationships as a graph. In this graph, nodes represent different physical quantities (stress, hardness, conductivity) and their derived features (such as stress gradient, hardness rate of change, etc.), and edges represent the correlation between physical quantities. The graph convolution operation updates the features of the current node by aggregating the information of adjacent nodes, so that the correlation between physical quantities can be fully utilized in the feature extraction process. The convolution kernel parameters of the graph convolution layer are also initialized according to the interaction influence matrix, ensuring that the network can learn a feature representation that conforms to the physical law. The physical correlation layer outputs a 32-dimensional feature vector, representing the complex interaction between multiple physical quantities.

[0073] The fusion of surface topography features and physical features uses an attention mechanism. The specific implementation is as follows: First, input the 16-dimensional feature vector output by the plane feature branch and the 96-dimensional feature vector (concatenation of the 64-dimensional physical constraint layer and the 32-dimensional physical correlation layer) output by the physical feature branch into the attention layer; the attention layer calculates the correlation score between the two sets of features, and the correlation score calculation method is the dot product of the two sets of features after softmax normalization; then, the two sets of features are weighted and summed according to the correlation score, to obtain the fused features. The dimension of the fused features is 112 (16+96), and the fused features are passed through a dimension reduction layer (fully connected layer with an output dimension of 64) to obtain the final feature representation.

[0074] The training process of the dual-branch feature extraction network is a key step. The training dataset consists of two parts: one is the labeled dataset, which contains titanium plate sample data with known performance levels, a total of 200 groups, each containing surface topography data and physical quantity measurement data; the other is the unlabeled dataset, which contains a large number of unlabeled titanium plate test data, a total of 2000 groups. The training adopts a semi-supervised learning method, combining supervised learning and self-supervised learning strategies. The supervised learning part uses labeled data to train the network, and the loss function includes three parts: feature reconstruction loss, classification loss, and physical consistency loss. Feature reconstruction loss measures the network's ability to reconstruct input features, using mean square error calculation; classification loss measures the accuracy of the network's prediction of titanium plate performance levels, using cross-entropy loss calculation; physical consistency loss measures the degree of conformity between the network's prediction results and physical laws, and the calculation method is the sum of the squared deviations of the predicted physical quantity relationships and the theoretical physical model. The self-supervised learning part uses unlabeled data to train the network, and designs three self-supervised tasks: feature mask reconstruction, physical quantity relationship prediction, and region consistency maintenance. The feature mask reconstruction task is to randomly mask part of the input features and let the network predict the masked feature values; the physical quantity relationship prediction task is to predict other physical quantities based on part of the physical quantities; the region consistency maintenance task is to ensure that the feature representations of adjacent regions have smooth transition properties. These self-supervised tasks help the network learn the internal structure of the data and the physical laws, improving the generalization ability.

[0075] The optimization algorithm uses the Adam optimizer, with an initial learning rate of 0.001 and a cosine annealing strategy to dynamically adjust the learning rate. To prevent overfitting, weight decay (coefficient 0.0001) and Dropout (proportion 0.3) regularization techniques are used. The training round is 200 rounds, using 32 sample batches per round, and the training data is augmented by random cropping, rotation, and noise addition. During the training process, the model performance is evaluated on the validation set every 10 rounds, and the learning rate is dynamically adjusted according to the validation performance, and the model parameters with the best performance are saved.

[0076] Model validation uses a 5-fold cross-validation strategy, randomly dividing the 200 labeled data into 5 parts, using 4 parts as the training set and 1 part as the validation set each time, and repeating 5 times to take the average performance indicators. Evaluation indicators include classification accuracy, root mean square error, correlation coefficient, and physical consistency score. Classification accuracy measures the accuracy of the model's prediction of titanium plate performance levels; root mean square error measures the deviation between predicted physical quantities and actual physical quantities; correlation coefficient measures the correlation between predicted results and actual results; physical consistency score measures the degree of conformity between predicted results and physical laws.

[0077] Feature distribution characteristic analysis is an important step of feature extraction. For each test region, the statistical distribution characteristics of surface topography features and physical features are calculated, including mean, standard deviation, skewness and kurtosis. These statistical characteristics reflect the central tendency and dispersion of the features, which helps to identify abnormal regions. For example, a skewness value greater than 1.5 or a kurtosis value greater than 5 usually indicates that there is a defect in the region. Feature distribution characteristics are also used to calculate feature weights. The more uniform the distribution, the higher the feature weight, and the more abnormal the distribution, the lower the feature weight, ensuring that the final fused features are not dominated by outliers.

[0078] The double-branch feature extraction network constructed by the method realizes the deep fusion of surface topography features and physical features through the design of topography entropy feature extraction and physical constraint layer and physical correlation layer. The physical feature branch establishes the correlation between multiple physical quantities, making the feature extraction process consistent with the laws of material science, significantly improving the physical interpretability and accuracy of the features. The topography entropy analysis and interaction matrix mechanism introduced in the feature extraction process enhance the sensitivity to local abnormal characteristics of titanium plates, providing a reliable feature basis for accurate evaluation of titanium plate performance.

[0079] Optionally, the distribution characteristics of the surface topography features and the physical features in each test region are calculated, and the feature weight coefficients are determined according to the distribution characteristics. The step of weighted fusion of features includes:

[0080] The topography entropy spatial distribution of the surface topography features in each test region is calculated, and the topography entropy spatial distribution is calculated by the negative logarithm of the surface topography feature probability in the test region. The multi-physical quantity correlation strength of the physical features in each test region is calculated, and the multi-physical quantity correlation strength is calculated based on the correlation coefficient matrix between physical quantities;

[0081] Based on the topography entropy spatial distribution and the multi-physical quantity correlation strength, the dispersion coefficient and the abnormal detection index in the test region are calculated. The region importance is weighted by the dispersion coefficient and the multi-physical quantity correlation strength, and the feature reliability is calculated based on the abnormal detection index. The local weight coefficient of each test region is calculated according to the region importance and the feature reliability, and the global weight coefficient is calculated according to the local weight coefficient. The surface topography features and the physical features in each test region are weighted based on the local weight coefficient to obtain the local fusion features, and the local fusion features are weighted based on the global weight coefficient to obtain the global fusion features.

[0082] For example, the calculation of the spatial distribution of the surface topography entropy of the surface topography features in each test area is based on the information entropy theory. The spatial distribution of the surface topography entropy reflects the complexity and uncertainty distribution of the surface topography features. The specific calculation method is to convert the surface topography features (such as the curvature entropy and the height entropy) in the test area into a probability distribution, and then calculate the weighted sum of the negative logarithm of the probabilities. The method of converting into a probability distribution is to divide the feature values into several intervals (usually 10 equal intervals) according to the numerical value, and count the proportion of the number of data points in each interval in the total as the probability of the interval. For the feature probability p of a certain interval, the contribution of the topography entropy is p multiplied by the negative logarithm p. The spatial distribution of the topography entropy of the test area is the accumulation of the topography entropy contributions of all intervals. For the region of the TC4 titanium plate with uniform surface topography, the spatial distribution value of the topography entropy is usually between 1.5-2.0; for the region with slight machining marks, the value is between 2.0-2.5; for the region with obvious defects, the value is more than 3.0.

[0083] The correlation strength of multiple physical quantities reflects the degree of correlation between stress, hardness and conductivity and other physical quantities. The calculation of the correlation strength is based on the correlation coefficient matrix between physical quantities, which is a 3x3 square matrix, the diagonal elements are 1, and the non-diagonal elements are the correlation coefficients between the corresponding physical quantities. The calculation method of the correlation coefficient is the inner product of the normalized data of two physical quantities divided by the sample number. For TC4 titanium alloy, the correlation coefficient between stress and hardness is usually between 0.65-0.80, the correlation coefficient between stress and conductivity is between 0.50-0.65, and the correlation coefficient between hardness and conductivity is between 0.45-0.60. The correlation strength of multiple physical quantities is the weighted average of all non-diagonal elements of the correlation coefficient matrix, and the weight is related to the importance of the physical quantity. Generally, the weight of stress is 0.4, the weight of hardness is 0.35, and the weight of conductivity is 0.25. For the region of the TC4 titanium plate with stable performance, the correlation strength of multiple physical quantities is usually between 0.60-0.75; for the region with unstable performance, the value is less than 0.50 or more than 0.85.

[0084] The dispersion coefficient measures the uniformity of the feature distribution in the region, and the calculation method is the ratio of the standard deviation to the average value of the spatial distribution of the topography entropy. For the titanium plate region with good uniformity, the dispersion coefficient is usually less than 0.2; for the region with local anomalies, the dispersion coefficient reaches more than 0.4. The anomaly detection index is used to identify the abnormal points in the region, and the calculation method is to count the proportion of the points whose feature values deviate from the average value of the region by more than twice the standard deviation. The anomaly detection index of the normal titanium plate region is usually less than 0.05; for the region with obvious defects, the index is more than 0.15.

[0085] The region importance reflects the importance of the test region in the overall performance evaluation, and the calculation method is the dispersion coefficient multiplied by 0.35 plus the multi-physical quantity correlation strength multiplied by 0.65. For key performance regions (such as high stress regions), the weight can be appropriately increased. The region importance distribution of a typical TC4 titanium plate is between 0.4-0.8, and the high importance region is usually the part with high mechanical performance requirement. The feature reliability is calculated based on the anomaly detection index, and the feature reliability is equal to 1 minus the anomaly detection index, reflecting the reliability of the feature data. The feature reliability of the normal region is usually higher than 0.95, and the feature reliability of the defective region is as low as 0.8.

[0086] The local weight coefficient of each test region is calculated according to the region importance and the feature reliability, and the calculation method is to multiply the region importance by the feature reliability, and then perform normalization processing (so that the weight coefficients of all regions sum to 1). The normalization processing is to divide the initial weight of each region by the sum of the initial weights of all regions. The local weight coefficient reflects the importance of a single region in its neighborhood. For a TC4 titanium plate with 100 test regions, the local weight coefficient is usually between 0.005-0.025, and the weight coefficient of the important region is higher. The global weight coefficient is calculated according to the local weight coefficient, and the calculation method is to consider the spatial correlation between regions, and to weight average the local weight of each region with the local weight of its neighborhood region. The neighborhood region weight decays with distance, and a Gaussian decay function is usually used, with a decay radius of 3 grid units. The global weight coefficient reflects the overall importance of the region in the whole titanium plate, and the distribution range is similar to the local weight coefficient, but the spatial distribution is more smooth.

[0087] Based on the local weight coefficient, the surface topography features and physical features in each test region are weighted. For each test region, multiply each component of the 7-dimensional surface topography feature vector by the local weight coefficient of the region to obtain the weighted surface topography feature; similarly, multiply each component of the 9-dimensional physical feature vector by the local weight coefficient to obtain the weighted physical feature. The weighted surface topography feature and physical feature are spliced into a 16-dimensional vector as the local fusion feature of the region. The local fusion feature not only retains the information of the original feature, but also reflects the importance of the region. For important regions, their features have a greater proportion in the fusion process; for regions with low reliability, their feature contribution is appropriately reduced.

[0088] The calculation of the global fusion feature is to weight and sum the local fusion features of all test regions according to the corresponding global weight coefficients to obtain a 16-dimensional feature vector representing the comprehensive performance feature of the whole titanium plate. The first 7 components of the global fusion feature are derived from the surface morphology feature, reflecting the surface quality of the titanium plate; the last 9 components are derived from the physical feature, reflecting the mechanical and electrical properties of the titanium plate. The global fusion feature can be directly used for classification and rating of the titanium plate. For an aviation-grade TC4 titanium plate, the components of the global fusion feature are usually within a certain range: the surface morphology-related components are between 0.5-1.5, and the physical feature-related components are between 0.7-1.3.

[0089] In practical applications, the feature weighting fusion process needs to process a large amount of test data. For a 1m x 1m TC4 titanium plate, it is divided into 100 test regions, and about 1000 test points of data are collected in each region. The data processing flow is: first, calculate the surface morphology feature and physical feature of each test region, then analyze the feature distribution to determine the weight coefficient, and finally perform feature fusion. The whole processing process takes about 3-5 minutes on a computing device with 16GB of memory and 8-core CPU, meeting the real-time requirements of industrial detection.

[0090] The method realizes intelligent weighting of the titanium plate surface morphology feature and physical feature by introducing two core indicators: morphology entropy spatial distribution and multi-physical quantity correlation strength; the introduction of dispersion coefficient and abnormal detection index enhances the sensitivity of the algorithm to abnormal regions; the double-layer weight system of local weight and global weight not only ensures the retention of local detailed features, but also realizes the comprehensive evaluation of global performance, providing an efficient and reliable technical means for accurate analysis and quality control of titanium plate performance.

[0091] Optionally, the step of calculating the dispersion coefficient and the abnormal detection index in the test region based on the morphology entropy spatial distribution and the multi-physical quantity correlation strength comprises:

[0092] According to the morphology entropy spatial distribution, the scale dispersion coefficient in the test region is calculated, and according to the multi-physical quantity correlation strength, the direction dispersion coefficient is calculated, wherein the scale dispersion coefficient is calculated by the weighted sum of the ratio of standard deviation to mean value under multiple feature scales, and the direction dispersion coefficient is obtained by normalizing the deviation of feature gradient from the mean value; the scale dispersion coefficient and the direction dispersion coefficient are combined by weighting to obtain a comprehensive dispersion coefficient;

[0093] According to the morphology entropy spatial distribution and the multi-physical quantity correlation strength, a feature abnormal score is calculated, a time sequence weight is updated according to the time sequence change of the feature abnormal score, and the feature abnormal score is adjusted based on the time sequence weight to obtain an abnormal detection index;

[0094] Combining the comprehensive discrete coefficient with the anomaly detection index to obtain a region importance of each test region, updating a feature weight of the test region based on the region importance to obtain a local weight coefficient of the test region;

[0095] Combining the local weight coefficient with a feature similarity of a neighboring test region to obtain a global weight coefficient, a weight coefficient of the global weight coefficient being proportional to the feature similarity of the neighboring test region.

[0096] For example, the process of calculating the scale discrete coefficient in the test region according to the spatial distribution of the topography entropy can be realized by statistical property analysis under multiple feature scales. For example, for a typical TC4 titanium alloy plate, the surface topography features are analyzed under three different scales: micro scale (5 microns x 5 microns), meso scale (50 microns x 50 microns) and macro scale (500 microns x 500 microns). At each scale, the ratio of the standard deviation to the mean of the topography entropy is calculated, which is denoted as the micro ratio, the meso ratio and the macro ratio, respectively. For a stable performance titanium alloy region, the three ratios are usually 0.15, 0.12 and 0.08, respectively; for a region with defects, the ratios are 0.30, 0.25 and 0.20, respectively. The scale discrete coefficient is calculated by the weighted sum of the three ratios, with the weight distribution being 0.5 for the micro scale, 0.3 for the meso scale and 0.2 for the macro scale, in order to highlight the importance of the microstructure. Therefore, the scale discrete coefficient of the stable performance region is about 0.13, while that of the defect region is about 0.27.

[0097] Calculating the direction discrete coefficient according to the correlation strength of multiple physical quantities involves analysis of feature gradients. For each test region, the gradient values of stress, hardness and conductivity in different directions (horizontal, vertical and diagonal) are calculated. The gradient value distribution of a normal titanium alloy region is relatively uniform, and the degree of deviation of the gradient in each direction from the mean value is small; while the performance of the non-uniform region is significantly directional, and the gradient in each direction deviates significantly from the mean value. The direction discrete coefficient is obtained by calculating the difference between the gradient in each direction and the average gradient, and then normalized. The normalization process is to divide the deviation in each direction by the maximum deviation value, so that the result is controlled between 0 and 1. For a region with uniform performance, the direction discrete coefficient is usually between 0.05 and 0.15; for a region with obvious directional defects, the coefficient reaches 0.3-0.4.

[0098] The scale dispersion coefficient is multiplied by 0.6, and the direction dispersion coefficient is multiplied by 0.4 to obtain the comprehensive dispersion coefficient, so as to balance the influence of scale factor and direction factor. For a test TC4 titanium alloy plate, the scale dispersion coefficient of a certain typical area is 0.18, the direction dispersion coefficient is 0.12, and the calculated comprehensive dispersion coefficient is 0.156. The comprehensive dispersion coefficient reflects the unevenness of the feature distribution in the area, and is an important indicator for evaluating the importance of the area. The higher the value, the more uneven the feature distribution in the area, and more attention should be paid.

[0099] The feature anomaly score measures the deviation of the features in the test area from the normal mode. The calculation method is to compare the spatial distribution of topography entropy with the standard distribution, while considering the change of multi-physical quantity correlation strength. The comparison of spatial distribution of topography entropy and standard distribution is realized by calculating Mahalanobis distance, which considers the covariance structure of the features. For TC4 titanium alloy, the standard distribution is determined by the statistical characteristics of multiple defect-free samples. The change of multi-physical quantity correlation strength is calculated by the difference from the reference value, which is determined by the average correlation strength of a large number of normal samples, usually 0.68. For a certain test area, the Mahalanobis distance of the spatial distribution of topography entropy is 2.3, and the deviation of the multi-physical quantity correlation strength is 0.15. The calculation method of feature anomaly score is to divide the Mahalanobis distance by 10 to obtain the basic score 0.23, and then add the correlation strength deviation multiplied by the weight coefficient 0.5, i.e. 0.15x0.5=0.075, so as to obtain the final feature anomaly score, which is the basic score minus the weighted deviation, calculated as 0.23-0.075=0.225. Mahalanobis distance reflects the degree of topography anomaly, and correlation strength deviation reflects the degree of physical property anomaly. The combination of the two can comprehensively evaluate the abnormal state of the area. The feature anomaly score of normal area is usually less than 0.2, while the score of area with obvious defects is more than 0.4.

[0100] The time series weight reflects the stability of the features over time. In the continuous production or detection scene, the same area is measured multiple times, and the time series of feature anomaly scores is recorded. If the feature anomaly score fluctuates little, it means that the performance of the area is stable, and the time series weight is low; if the fluctuation is large, it means that there are unstable factors in the area, and the time series weight should be increased. The calculation method of time series weight is to divide the standard deviation of the time series of feature anomaly scores by the average value, and then multiply by the adjustment coefficient 0.8. For stable areas, the time series weight is usually between 0.1 and 0.3; for unstable areas, it reaches more than 0.5.

[0101] The feature anomaly score is adjusted based on the time sequence weight. The adjustment method is that the feature anomaly score is multiplied by (1+time sequence weight). For a region with a feature anomaly score of 0.225 and a time sequence weight of 0.2, the calculated anomaly detection index is 0.27. The anomaly detection index comprehensively considers the abnormality degree and time stability of the feature, and provides an important basis for the calculation of the region importance. The anomaly detection index of the normal region is usually lower than 0.25, and the index of the problem region exceeds 0.4.

[0102] The region importance of each test region is obtained by combining the comprehensive dispersion coefficient and the anomaly detection index. The combination method is that the comprehensive dispersion coefficient is multiplied by 0.45, and the anomaly detection index is multiplied by 0.55. For a region with a comprehensive dispersion coefficient of 0.156 and an anomaly detection index of 0.27, the calculated region importance is 0.219. The region importance reflects the criticality of the test region, and the higher the value, the greater the influence of the region on the overall performance, which should be given a higher weight in subsequent analysis. The importance of the normal region is usually between 0.15-0.25, and the critical region reaches 0.3-0.4.

[0103] The feature weight of the test region is updated based on the region importance, and the local weight coefficient of the test region is obtained. The update method is to convert the region importance into an adjustment factor of the feature weight, and to calculate the feature weight basic value plus the adjustment factor. The feature weight basic value is 1 / N, where N is the total number of test regions; the adjustment factor is the region importance minus the average of all region importance, multiplied by the scaling coefficient 0.5. For 100 test regions, the feature weight basic value is 0.01; for a region with a region importance of 0.219 and an average of all region importance of 0.2, the calculated adjustment factor is 0.0095, and thus the local weight coefficient is 0.0195. The local weight coefficient determines the contribution size of the region feature in the subsequent fusion process, and the important region obtains a higher weight, thereby playing a greater role in feature fusion.

[0104] The local weight coefficient is weighted and combined with the feature similarity of the adjacent test region to obtain the global weight coefficient. The feature similarity is obtained by calculating the cosine distance of the feature vectors of the current region and the adjacent region. The smaller the cosine distance, the higher the similarity. For each region, the average similarity is calculated by considering the feature similarity of the four adjacent regions above, below, left and right. The calculation method of the global weight coefficient is that the local weight coefficient is multiplied by (1+average similarity multiplied by 0.3). For a region with a local weight coefficient of 0.0195 and an average similarity of 0.85, the global weight coefficient is 0.022. The global weight coefficient considers the importance of the region itself and the correlation between regions, making the weight distribution more reasonable. The region with higher global weight is obtained by the similar features of adjacent regions, which is conducive to discovering defects or characteristics that are continuously distributed.

[0105] In the actual detection of TC4 titanium alloy plate, after calculating the global weight coefficients of 100 test regions, normalization processing is required to ensure that the sum of all weight coefficients is 1. The normalization method is to divide the global weight coefficient of each region by the sum of the global weight coefficients of all regions. The normalized global weight coefficient is directly used in the subsequent feature weighting fusion process. The higher the global weight coefficient of a region, the greater the influence of its features on the fusion result, thereby achieving the goal of differentiated weighting according to the importance of the region.

[0106] The method realizes accurate evaluation and weight distribution of the importance of the test region by introducing a double-dispersion evaluation mechanism of scale dispersion and direction dispersion, combining time sequence stability analysis and region similarity calculation. The method can effectively identify and highlight the key regions and abnormal features on the surface of titanium alloy, significantly improve the pertinence and accuracy of feature fusion, provide reliable technical support for quality evaluation and defect warning of titanium alloy plate, and solve the technical problems of unreasonable weight distribution and insufficient consideration of region relevance in traditional methods.

[0107] Optionally, based on the fused features, a titanium plate performance analysis result is generated through a multi-task output layer, and the step of determining the performance level of the titanium plate to be tested according to the performance analysis result comprises:

[0108] The fused features are input into a shared feature extraction layer and a task-specific layer. The shared feature extraction layer performs multi-scale hierarchical extraction on the fused features to generate multi-scale features. The task-specific layer includes a performance distribution prediction branch, a defect type identification branch, and a reliability evaluation branch.

[0109] In the performance distribution prediction branch, the spatial distribution data of the multi-scale features is calculated, the distribution interval of the performance index is determined based on the spatial distribution data, the distribution interval is dynamically updated, and the performance distribution prediction result is generated. In the defect type identification branch, the multi-scale features are matched with the preset defect features layer by layer to determine the optimal matching features, and the defect type and its corresponding confidence are output according to the optimal matching features. In the reliability evaluation branch, the performance distribution prediction result and the defect type and its confidence are used as evaluation basis, the credibility of each evaluation basis is calculated, each evaluation basis is weighted and combined according to the credibility, and the reliability evaluation result is obtained.

[0110] According to the pre-set performance level judgment standard, the performance distribution prediction result, the defect type and its confidence, and the reliability evaluation result are combined to determine the performance level of the titanium plate to be tested through interval mapping.

[0111] For example, when the fused features are input into the shared feature extraction layer and the task-specific layer, the shared feature extraction layer adopts a cascaded structure to perform multi-scale hierarchical extraction on the fused features. For example, for a typical TC4 titanium alloy plate, the dimension of the fused features is 16, which first enters the shared feature extraction layer. The shared feature extraction layer includes three cascaded feature extraction modules, each of which is composed of a nonlinear activation unit and a normalization unit. The first level module maps the 16-dimensional features to 32-dimensional features, the second level module maps the 32-dimensional features to 64-dimensional features, and the third level module maps the 64-dimensional features to 128-dimensional features. This cascaded structure can gradually extract the abstract representation of the features and generate multi-scale features. The multi-scale features include micro-scale features (corresponding to the output of the first level module), meso-scale features (corresponding to the output of the second level module), and macro-scale features (corresponding to the output of the third level module), which can capture the performance characteristics of the titanium plate at different spatial scales.

[0112] The task-specific layer includes a performance distribution prediction branch, a defect type identification branch, and a reliability evaluation branch, each of which is designed and optimized for a specific task. The performance distribution prediction branch consists of a feature space transformation module and a distribution parameter estimation module. The feature space transformation module uses an attention mechanism to weight the multi-scale features according to their importance, highlighting key features. The distribution parameter estimation module includes two parallel processing units that estimate the central tendency parameter and the dispersion parameter of the distribution, respectively. For TC4 titanium alloy plates, the performance indicators mainly include tensile strength, yield strength, and elongation, and the distribution parameters of each indicator are estimated by the corresponding processing unit.

[0113] In the performance distribution prediction branch, the spatial distribution data of the multi-scale features is calculated. The spatial distribution data is obtained by statistical characteristics of the features in the spatial domain, including mean distribution, variance distribution, and skewness distribution. For a 1m x 1m TC4 titanium alloy plate, it is divided into a 10 x 10 grid, and each grid represents a spatial unit. In each spatial unit, the statistics of the multi-scale features are calculated to form a spatial distribution matrix. The mean distribution matrix reflects the average performance level of each spatial unit, the variance distribution matrix reflects the volatility of the performance, and the skewness distribution matrix reflects the asymmetry of the performance distribution. Based on the spatial distribution data, the distribution interval of the performance indicator is determined, and the distribution interval is represented by the confidence interval, usually set to 95% confidence. For the tensile strength of TC4 titanium alloy, the typical distribution interval is 920-980MPa; the distribution interval of the yield strength is 830-890MPa; and the distribution interval of the elongation is 10%-14%.

[0114] The distribution interval is dynamically updated, and the dynamic updating process considers historical data and current measurement results. The updating method adopts exponential weighted average, the lower limit of the new distribution interval is equal to the lower limit of the original interval multiplied by 0.9 plus the minimum value of the current measurement multiplied by 0.1, and the upper limit of the new distribution interval is equal to the upper limit of the original interval multiplied by 0.9 plus the maximum value of the current measurement multiplied by 0.1. The dynamic updating mechanism can adapt to the changing trend of the performance of the titanium plate and improve the accuracy of the prediction. The process of generating the performance distribution prediction result is to combine the updated distribution interval with the statistical characteristic data to output complete performance prediction data. Specifically, the statistical characteristics of the spatial distribution matrix are used to calculate the mean, standard deviation, skewness and kurtosis of the distribution and other parameters, and then the most suitable distribution type (such as normal distribution, Weibull distribution or mixed distribution) is determined. For TC4 titanium alloy, the mean tensile strength is 950 MPa, the standard deviation is 15 MPa, and it is determined as normal distribution; the mean yield strength is 860 MPa, the standard deviation is 12 MPa, and it is determined as normal distribution; the mean elongation is 12%, the shape parameter is 3.2, and the scale parameter is 13.5, and it is determined as Weibull distribution. The final performance distribution prediction result includes these distribution parameters, distribution type and predicted over-limit probability, which together constitute the comprehensive prediction result of the performance of the titanium plate. For most TC4 titanium alloy plates, the tensile strength and yield strength usually show normal distribution, and the elongation is closer to Weibull distribution.

[0115] In the defect type identification branch, the multi-scale features are matched with the preset defect features layer by layer. The preset defect feature library contains common defects of typical titanium alloy plates, such as surface cracks, interlayer delamination, pores, inclusions, etc., and each defect has a corresponding feature description. The layer-by-layer matching process adopts a hierarchical comparison strategy, first performs rough matching at the macro scale to filter out candidate defect types, then performs fine matching at the meso scale to narrow down the candidate range, and finally performs accurate matching at the micro scale to determine the final defect type. The calculation of matching degree adopts cosine similarity, and the higher the cosine similarity, the higher the matching degree. The optimal matching feature is the preset defect feature with the highest cosine similarity with the sample feature.

[0116] The defect type and its corresponding confidence are output according to the optimal matching feature. The confidence calculation method is to subtract the cosine similarity of the second optimal matching from the cosine similarity of the optimal matching, and then multiply by the adjustment coefficient 1.5. When the optimal matching is significantly different from other matchings, the confidence is high; when multiple matching results are close, the confidence is low. For typical defects of TC4 titanium alloy plates, such as surface cracks, when the feature matching degree is high, the confidence can reach more than 0.85; for fuzzy defect features, the confidence is as low as 0.6 or less.

[0117] In the reliability evaluation branch, the performance distribution prediction results and the defect types and their confidence are taken as evaluation basis. The reliability evaluation branch is composed of a credibility calculation module and a weighted combination module. The credibility calculation module evaluates the reliability degree of each evaluation basis, and the calculation method varies with the type of basis. For the performance distribution prediction results, the credibility is inversely proportional to the width of the distribution interval and proportional to the sample quantity; for the defect types and the confidence, the credibility is related to the typicality and stability of the defect features. The credibility calculation adopts a method based on evidence theory, which comprehensively considers multiple factors.

[0118] The reliability evaluation results are obtained by weighted combination of each evaluation basis according to the credibility. The weighted combination adopts a method of multiplying the performance distribution credibility by 0.6 and adding the defect type credibility multiplied by 0.4, which balances the influences of performance and defects. For the TC4 titanium alloy plate with stable performance and few defects, the reliability evaluation results are usually between 0.85 and 0.95; for the plate with large performance fluctuation or obvious defects, the reliability evaluation results are as low as below 0.7. The reliability evaluation results are an important index for judging the overall quality of the titanium plate, and directly affect the final performance grade.

[0119] The interval mapping process first establishes a correspondence table of each evaluation index and performance grade. For the performance distribution prediction results, the distribution characteristics of the three indexes of tensile strength, yield strength and elongation are mainly considered, which are divided into three intervals of high, medium and low, each of which corresponds to different performance scores. For TC4 titanium alloy, the high interval of tensile strength is above 960 MPa with a score of 5, the medium interval is 930-960 MPa with a score of 3, and the low interval is 900-930 MPa with a score of 1; the yield strength and elongation are divided in a similar way. For the defect types and the confidence, the common defects are divided into three categories according to the severity: serious defects (such as deep cracks), moderate defects (such as surface scratches) and slight defects (such as slight oxidation), and the defect score is calculated in combination with the confidence. The defect confidence is above 0.8 with a weight of 1, between 0.5 and 0.8 with a weight of 0.7, and below 0.5 with a weight of 0.4. The reliability evaluation results are directly taken as a weight factor, multiplied by the performance score and the defect score to obtain the weighted total score. The final performance grade is determined by the total score range: 8-10 points for A grade, 5-8 points for B grade, 3-5 points for C grade, and below 3 points for unqualified.

[0120] The multi-task output layer of the method integrates performance distribution prediction, defect type identification and reliability evaluation three key functions, realizes the comprehensive analysis and accurate rating of titanium plate performance through the combination of shared feature extraction and task-specific processing. The method can not only accurately predict the performance distribution of titanium plate, but also effectively identify various defects and evaluate the reliability of the analysis results, providing a scientific basis for quality control and performance guarantee of titanium alloy plate, and solving the technical problems of insufficient multi-task cooperation and single rating standard in traditional methods.

[0121] Optionally, the performance distribution prediction result and the defect type and its confidence are used as evaluation basis to calculate the credibility of each evaluation basis, and the reliability evaluation result is obtained by weighting and combining each evaluation basis according to the credibility.

[0122] The probability distribution deviation of the performance distribution prediction result and the prediction error of the defect type confidence are calculated, and an uncertainty matrix is constructed based on the probability distribution deviation and the prediction error. The time sequence fluctuation characteristics of each evaluation basis are calculated according to the uncertainty matrix, and a stability index of the evaluation basis is obtained. Based on the stability index, the historical evaluation results of the evaluation basis are tracked and analyzed to calculate the accuracy rate trend of the evaluation results, and a dynamic evaluation weight is generated according to the accuracy rate trend. The stability index and the dynamic evaluation weight are combined to obtain the credibility of each evaluation basis, and the performance distribution prediction result and the defect type and its confidence are weighted and fused according to the credibility, and the reliability evaluation result is output.

[0123] For example, the probability distribution deviation of the performance distribution prediction result needs to be compared with the reference standard distribution. For TC4 titanium alloy plate, the reference standard distribution is usually established based on a large amount of historical test data. The probability distribution deviation is obtained by calculating the difference between the predicted distribution and the reference distribution. The specific method is to calculate the sum of the absolute values of the frequency difference of the two distributions in each probability interval. In practical application, the distribution interval of tensile strength 920-980MPa is divided into 12 equal intervals, each interval width is 5MPa, and the frequency difference of the predicted distribution and the reference distribution in the 12 intervals is calculated. For TC4 titanium alloy with stable performance, the probability distribution deviation is usually less than 0.15; for samples with large performance fluctuation, the deviation is more than 0.3.

[0124] The prediction error of defect type confidence is calculated by comparing the difference between the predicted defect type and confidence and the actual verification result. The prediction error calculation method is the absolute value of the difference between the predicted confidence and the actual confidence (determined by expert evaluation or destructive testing). For example, for a defect predicted as a surface crack with a confidence of 0.85, if the actual verification confirms that it is a surface crack and the severity is consistent with the prediction, the actual confidence is 0.9, and the prediction error is 0.05; if the actual verification confirms that it is a surface scratch rather than a crack, the actual confidence is only 0.3, and the prediction error is 0.55. For accurate defect prediction, the prediction error is usually less than 0.1; for inaccurate prediction, the error is more than 0.4.

[0125] An uncertainty matrix is constructed based on the probability distribution deviation and the prediction error. The uncertainty matrix is a two-dimensional matrix, with rows representing different evaluation bases (such as distribution prediction of tensile strength, yield strength, elongation, and prediction of various defects), and columns representing different sources of uncertainty (such as measurement error, model error, environmental interference, etc.). For performance evaluation of TC4 titanium alloy plates, the uncertainty matrix is usually a 5x4 matrix, containing 5 evaluation bases and 4 sources of uncertainty. The value of the matrix element is the error contribution of the corresponding evaluation basis under a specific source of uncertainty, and the larger the value, the higher the uncertainty. The construction method of the uncertainty matrix is to decompose the probability distribution deviation and the prediction error into different sources of uncertainty to form the matrix element value.

[0126] The time sequence fluctuation characteristics of each evaluation basis are calculated according to the uncertainty matrix, which describes the changes of the evaluation basis in continuous multiple evaluations. The calculation method is to record the change of the uncertainty value of each evaluation basis in continuous 10 evaluations, and calculate the ratio of the standard deviation to the mean value. For stable evaluation basis, the ratio is usually less than 0.2; for unstable basis, the ratio exceeds 0.5. The stability index is a normalized representation of the time sequence fluctuation characteristics, and the calculation method is 1 minus the fluctuation ratio. For the tensile strength prediction of TC4 titanium alloy, if the mean value of the uncertainty value of continuous 10 evaluations is 0.12 and the standard deviation is 0.02, the fluctuation ratio is 0.167 and the stability index is 0.833; for surface defect prediction, if the mean value of the uncertainty value is 0.25 and the standard deviation is 0.1, the fluctuation ratio is 0.4 and the stability index is 0.6.

[0127] The process of tracking analysis of historical evaluation results based on stability index needs to establish a historical evaluation database to record the results of each evaluation and its actual verification results. For each evaluation basis, the accuracy rate in different stability index intervals is counted. The accuracy rate is the number of correct evaluations divided by the total number of evaluations. For example, for tensile strength prediction, when the stability index is between 0.8-0.9, the historical accuracy rate is 0.92; when the stability index is less than 0.7, the historical accuracy rate drops to 0.75. The accuracy rate trend is the derivative estimate of the accuracy rate with respect to the stability index, indicating the degree of influence of stability change on the accuracy rate. For the performance evaluation of TC4 titanium alloy, the accuracy rate trend is usually between 0.5-1.5, and a higher value indicates that stability has a greater impact on the accuracy rate.

[0128] The dynamic evaluation weight is a weight adjustment factor calculated based on the current stability index and the accuracy rate trend. The calculation method is the stability index multiplied by the adjustment function of the accuracy rate trend, and the adjustment function uses a piecewise linear function to provide greater weight adjustment when the accuracy rate trend is larger. For the tensile strength prediction of TC4 titanium alloy, if the stability index is 0.85 and the accuracy rate trend is 1.2, the dynamic evaluation weight is 0.92; for defect prediction, if the stability index is 0.65 and the accuracy rate trend is 0.8, the dynamic evaluation weight is 0.75.

[0129] The stability index multiplied by 0.4 plus the dynamic evaluation weight multiplied by 0.6 gives the credibility of each evaluation basis. For tensile strength prediction, if the stability index is 0.85 and the dynamic evaluation weight is 0.92, the credibility is 0.892; for defect prediction, if the stability index is 0.65 and the dynamic evaluation weight is 0.75, the credibility is 0.71. The credibility reflects the reliability of the evaluation basis and directly affects its weight in the final evaluation.

[0130] The performance distribution prediction result and the defect type and its confidence are weighted and fused according to the credibility to output the reliability evaluation result. The weighting and fusing process considers the importance and credibility of different evaluation bases. For TC4 titanium alloy plates, the basic weight of the performance distribution prediction result is usually set to 0.6, and the basic weight of the defect type and confidence is set to 0.4. These basic weights are multiplied by the corresponding credibility to obtain the actual weight. For example, if the credibility of the performance distribution prediction result is 0.892 and the credibility of the defect type and confidence is 0.71, the actual weight of the performance distribution prediction result is 0.6x0.892=0.5352, and the actual weight of the defect type and confidence is 0.4x0.71=0.284, and after normalization, they are 0.653 and 0.347 respectively. The reliability evaluation result is the weighted sum of the scores of each evaluation basis multiplied by the corresponding actual weight. For high-quality TC4 titanium alloy plates, the performance distribution prediction score is 0.95, and the defect type prediction score is 0.9, and the calculated reliability evaluation result is 0.95x0.653+0.9x0.347=0.933; for plates with some problems, the scores are 0.8 and 0.6 respectively, and the reliability evaluation result is 0.8x0.653+0.6x0.347=0.731.

[0131] The output of the reliability evaluation result includes the overall reliability score and the credibility and contribution analysis of each evaluation basis. The overall reliability score is a value between 0 and 1, and the closer to 1 indicates that the performance of the titanium plate is more reliable. For aviation-grade TC4 titanium alloy, the overall reliability score is required to be not less than 0.9; for general industrial use, it is required to be not less than 0.75. The credibility analysis shows the reliability of each evaluation basis, and the contribution analysis shows the influence of each evaluation basis on the overall evaluation. These information together constitute the reliability evaluation report of the performance of titanium alloy plates, providing scientific basis for quality control and application decision.

[0132] The method realizes accurate calculation and dynamic adjustment of the credibility of evaluation basis by introducing probability distribution deviation and prediction error analysis, combining time sequence fluctuation characteristics and accuracy change trend. The method not only considers the static evaluation result, but also pays attention to the time stability and historical accuracy of the evaluation process, effectively solving the technical problems of single, static and strong subjectivity in traditional methods, and providing more scientific and comprehensive reliability guarantee for the quality evaluation of titanium alloy plates.

[0133] In a second aspect, a titanium plate performance analysis system is provided, comprising:

[0134] The first unit is configured to scan the surface of the titanium plate to be tested by using an optical scanner to establish a three-dimensional structure, and to divide a plurality of test regions, and to set a plurality of test points in a matrix distribution in each test region; the second unit is configured to detect the performance of each test point by using a multi-probe cooperative detection device, the multi-probe cooperative detection device comprising stress probes, hardness probes and conductivity probes which are uniformly arranged along the circumferential direction, and to synchronously collect stress data, hardness data and conductivity data of the test point; the third unit is configured to construct a double-branch feature extraction network, comprising a plane feature branch for extracting surface topography features and a physical feature branch constructed based on material science knowledge; the physical feature branch comprises a physical constraint layer and a physical correlation layer, and is configured to establish a correlation between the stress data, the hardness data and the conductivity data; to calculate the distribution characteristics of the surface topography features and the physical features in each test region, to determine the corresponding feature weight coefficients according to the distribution characteristics, and to perform weighted fusion on the features; and the fourth unit is configured to generate a titanium plate performance analysis result by using a multi-task output layer based on the fused features, and to determine the performance grade of the titanium plate to be tested according to the performance analysis result.

[0135] In a third aspect, a computer-readable storage medium is provided, and the computer-readable storage medium has stored thereon computer program instructions, which, when executed by a processor, implement the method described above.

Claims

1. A method of analyzing the performance of a titanium sheet, characterized by, The method comprises the steps of: scanning the surface of the titanium plate to be tested by an optical scanner to establish a three-dimensional structure, and dividing a plurality of test areas, and setting a plurality of test points in a matrix distribution in each test area; using a multi-probe cooperative detection device to detect the performance of each test point, the multi-probe cooperative detection device comprising stress probes, hardness probes and conductivity probes arranged uniformly in the circumferential direction, and synchronously collecting stress data, hardness data and conductivity data of the test point; constructing a double-branch feature extraction network, including a plane feature branch for extracting surface topography features and a physical feature branch constructed based on material science knowledge; the physical feature branch comprises a physical constraint layer and a physical correlation layer, for establishing a correlation between stress data, hardness data and conductivity data; calculating the distribution characteristics of the surface topography features and the physical features in each test area, and determining the corresponding feature weight coefficients according to the distribution characteristics, and performing weighted fusion on the features; based on the fused features, generating a titanium plate performance analysis result through a multi-task output layer, and determining the performance grade of the titanium plate to be tested according to the performance analysis result.

2. The method of claim 1, wherein, The step of scanning the surface of the titanium plate to be tested by an optical scanner to establish a three-dimensional structure, and dividing a plurality of test areas, and setting a plurality of test points in a matrix distribution in each test area comprises: using a dual-camera cooperative acquisition system to acquire image data of the surface of the titanium plate to be tested; and performing phase measurement based on the image data to obtain initial three-dimensional point cloud data; eliminating vibration errors in the initial three-dimensional point cloud data by using a birefringence optical path difference method, and compensating temperature drift errors in the initial three-dimensional point cloud data based on a pre-calibrated temperature-deformation curve to obtain compensated three-dimensional point cloud data; calculating local curvature distribution data of the surface of the titanium plate to be tested based on the compensated three-dimensional point cloud data, and dividing the surface of the titanium plate to be tested into a plurality of test areas according to the local curvature distribution data, wherein the area of each test area is inversely proportional to the local curvature value in the test area; based on the local curvature distribution data, setting a test point matrix in each test area, wherein the distance between adjacent test points in each test area is inversely proportional to the local curvature value at the position of the test point.

3. The method of claim 1, wherein, The step of using a multi-probe cooperative detection device to detect the performance of each test point comprises: obtaining the stress release characteristic time of the titanium plate to be tested, and determining the measurement timing of the stress probes, the hardness probes and the conductivity probes according to the stress release characteristic time; controlling the load application rate to be proportional to the yield strength of the titanium plate to be tested during the measurement process of the stress probes, and obtaining the stress data of the test point; during the measurement process of the conductivity probes, obtaining the initial conductivity data of the test point, calculating the disturbance correction coefficient of the stress field on the conductivity according to the stress data, modifying the initial conductivity data based on the disturbance correction coefficient, and obtaining the conductivity data; In the hardness probe measurement process, the initial hardness data of the test point is obtained, the deformation compensation factor of the hardness measurement corresponding to the stress field is calculated according to the indentation depth, and the initial hardness data is corrected based on the deformation compensation factor to obtain the hardness data; the stress data and the conductivity data are corrected according to the deformation compensation factor to obtain the corrected stress data and the conductivity data of the test point.

4. The method of claim 1, wherein, A double-branch feature extraction network is constructed, including a plane feature branch for extracting surface topography features and a physical feature branch constructed based on material science knowledge; The physical feature branch includes a physical constraint layer and a physical correlation layer, and the step of establishing the correlation between the stress data, the hardness data and the conductivity data includes: According to the height data of the surface of the titanium plate to be tested, the local curvature value and the height distribution probability are calculated, and the topography entropy feature is generated according to the local curvature value and the height distribution probability; the height deviation data of the surface of the titanium plate to be tested is collected, and the surface roughness parameter is calculated; the topography entropy feature and the surface roughness parameter are constructed as surface topography features; The hardness variation law corresponding to the stress data is analyzed, and the corresponding relationship between stress and hardness is established; the variation law of the conductivity data with dislocation density is analyzed, and the corresponding relationship between conductivity and dislocation density is established; the corresponding relationship between stress and hardness, and the corresponding relationship between conductivity and dislocation density are taken as physical constraint features; An interaction influence matrix of stress, hardness and conductivity is constructed, the interaction strength of the stress data, the hardness data and the conductivity data is calculated according to the interaction influence matrix, the stress data, the hardness data and the conductivity data are corrected based on the interaction strength, and the corrected data is constructed as physical correlation features.

5. The method of claim 1, wherein, The distribution characteristics of the surface topography features and the physical features in each test area are calculated, and the corresponding feature weight coefficients are determined according to the distribution characteristics, and the steps of feature weighted fusion include: The topography entropy space distribution of the surface topography features in each test area is calculated, and the topography entropy space distribution is calculated by the negative logarithm of the surface topography feature probability in the test area; the multi-physical quantity correlation strength of the physical features in each test area is calculated, and the multi-physical quantity correlation strength is calculated based on the correlation coefficient matrix between physical quantities; Based on the topography entropy space distribution and the multi-physical quantity correlation strength, the dispersion coefficient and the abnormal detection index in the test area are calculated; the regional importance is weighted by the dispersion coefficient and the multi-physical quantity correlation strength, and the feature reliability is calculated based on the abnormal detection index; the local weight coefficient of each test area is calculated according to the regional importance and the feature reliability, and the global weight coefficient is calculated according to the local weight coefficient; the surface topography features and the physical features in each test area are weighted based on the local weight coefficient to obtain the local fusion features, and the local fusion features are weighted based on the global weight coefficient to obtain the global fusion features.

6. The method of claim 5, wherein, The steps of calculating the dispersion coefficient and the abnormal detection index in the test area based on the topography entropy space distribution and the multi-physical quantity correlation strength include: According to the topographic entropy space distribution and the multi-physical quantity correlation strength, a feature anomaly score is calculated, a time sequence weight is updated according to a time sequence change of the feature anomaly score, and an anomaly detection index is obtained by adjusting the feature anomaly score based on the time sequence weight; According to the topographic entropy space distribution and the multi-physical quantity correlation strength, a feature anomaly score is calculated, a time sequence weight is updated according to a time sequence change of the feature anomaly score, and an anomaly detection index is obtained by adjusting the feature anomaly score based on the time sequence weight; The local weight coefficient is combined with the feature similarity of the adjacent test region to obtain a global weight coefficient, and the weight coefficient of the global weight coefficient is proportional to the feature similarity of the adjacent test region. Based on the fused features, a titanium plate performance analysis result is generated through a multi-task output layer, and the performance level of the titanium plate to be tested is determined according to the performance analysis result.

7. The method of claim 1, wherein, The fused features are input into a shared feature extraction layer and a task-specific layer, the shared feature extraction layer performs multi-scale hierarchical extraction on the fused features to generate multi-scale features, and the task-specific layer includes a performance distribution prediction branch, a defect type identification branch, and a reliability evaluation branch. In the performance distribution prediction branch, spatial distribution data of the multi-scale features is calculated, a distribution interval of a performance index is determined based on the spatial distribution data, the distribution interval is dynamically updated to generate a performance distribution prediction result, in the defect type identification branch, the multi-scale features are matched with preset defect features layer by layer to determine optimal matching features, and a defect type and its corresponding confidence are output according to the optimal matching features, and in the reliability evaluation branch, the performance distribution prediction result and the defect type and its confidence are used as evaluation basis to calculate the reliability of each evaluation basis, the evaluation basis is weighted and combined according to the reliability to obtain a reliability evaluation result. According to a pre-set performance level judgment standard, the performance distribution prediction result, the defect type and its confidence, and the reliability evaluation result are combined to determine the performance level of the titanium plate to be tested through interval mapping. The performance distribution prediction result and the defect type and its confidence are used as evaluation basis to calculate the reliability of each evaluation basis, and the evaluation basis is weighted and combined according to the reliability to obtain a reliability evaluation result.

8. The method of claim 7, wherein, ​ The prediction error of the probability distribution deviation of the performance distribution prediction result and the defect type confidence is calculated, an uncertainty matrix is constructed based on the probability distribution deviation and the prediction error, the timing fluctuation characteristics of each evaluation basis are calculated according to the uncertainty matrix, and a stability index of the evaluation basis is obtained; the historical evaluation results of the evaluation basis are tracked and analyzed based on the stability index, the accuracy rate change trend of the evaluation result is calculated, the dynamic evaluation weight is generated according to the accuracy rate change trend, the stability index and the dynamic evaluation weight are combined to obtain the credibility of each evaluation basis, the performance distribution prediction result and the defect type and its confidence are weighted and fused according to the credibility, and a reliability evaluation result is output.

9. A titanium plate performance analysis system for implementing the method according to any one of the preceding claims 1-8, characterized by, Comprise: The first unit is used for scanning the surface of the titanium plate to be tested by an optical scanner to establish a three-dimensional structure, and dividing a plurality of test areas, and a plurality of test points in a matrix distribution are arranged in each test area; The second unit is used for performance detection of each test point by a multi-probe cooperative detection device, the multi-probe cooperative detection device comprises stress probes, hardness probes and conductivity probes which are uniformly arranged along the circumferential direction, and stress data, hardness data and conductivity data of the test point are synchronously collected; The third unit is used for constructing a double-branch feature extraction network, comprising a plane feature branch for extracting surface topography features and a physical feature branch constructed based on material science knowledge; the physical feature branch comprises a physical constraint layer and a physical correlation layer, and is used for establishing the correlation between stress data, hardness data and conductivity data; the distribution characteristics of the surface topography features and the physical features in each test area are calculated, the corresponding feature weight coefficients are determined according to the distribution characteristics, and the features are weighted and fused; The fourth unit is used for generating a titanium plate performance analysis result through a multi-task output layer based on the fused features, and determining the performance grade of the titanium plate to be tested according to the performance analysis result.

10. A computer-readable storage medium having stored thereon computer program instructions, wherein, The computer program instructions are executed by the processor to implement the method of any one of claims 1 to 8. The computer program instructions are executed by the processor to implement the method of any one of claims 1 to 8.

Images