Semiconductor coating nondestructive testing method for heating clothes

Abstract

The invention discloses a semiconductor coating nondestructive testing method for heating clothes, and relates to the technical field of nondestructive testing, and the method comprises the following steps: converting a gray scale dot matrix, a multi-wavelength spectrum dot matrix and a wave intensity dot matrix into a gray scale field, a multi-wavelength spectrum field and a wave intensity field; pre-processing the gray field, and analyzing by adopting a multi-scale spectrum analysis auto-encoder to obtain a thickness fluctuation graph; decomposing and disassembling the multi-wavelength spectral field by adopting a sparse non-negative matrix, and processing the input spectral segmentation model to obtain a skip plating probability graph; the wave intensity field is preprocessed and input into a position self-encoder, and a bubble probability graph is generated by considering position deviation analysis; and adopting a Bayesian confidence estimation algorithm, taking the probability graph as prior probability distribution, and updating and obtaining posterior probability distribution based on a Bayesian principle on the basis of causal association of process parameters, skip plating, bubbles and thickness fluctuation and combined with joint conditional probability distribution in historical data so as to determine a skip plating area and a bubble area. And nondestructive testing which considers causal association and is beneficial to traceability analysis is realized.

Description

Technical Field

[0001] The present invention relates to the field of nondestructive testing technology, and in particular to a nondestructive testing method for semiconductor coatings on heating clothing. Background Art

[0002] Semiconductor heating clothing achieves conductive heating function through coating. The quality of coating directly affects the product's heating stability, circuit safety and service life. Accurate detection of missing plating areas, bubble defects and thickness fluctuations is a core requirement to ensure the reliability of wearable devices and process optimization.

[0003] Existing detection technologies are only based on collected data such as images, and directly output defect detection results through data-driven methods such as neural networks. They do not consider the causal relationship between changes in coating thickness and plating leaks, bubbles and process parameters. The defect detection accuracy is low and it is not conducive to tracing the causes of complex defects. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention proposes a non-destructive testing method for semiconductor coatings on heat-generating clothing, which realizes non-destructive testing that takes causal relationships into consideration and facilitates traceability analysis.

[0005] The technical solutions for achieving the purpose of the present invention are:

[0006] A nondestructive testing method for semiconductor coatings on heat-generating clothing, comprising the following steps:

[0007] The measuring equipment is used to obtain the grayscale dot matrix I, multi-wavelength spectrum dot matrix S and wave intensity dot matrix E of the coating layer, and the measurement posture determined by the optical tracking system is converted into the grayscale field G. I , multi-wavelength spectral field G S and wave strength field G E ;

[0008] Preprocessed gray field G I Generate grayscale feature matrix F I And input the multi-scale spectral analysis autoencoder, use multi-scale convolution to mine grayscale spatial correlation and splice to generate a multi-scale grayscale feature matrix The multi-scale gray feature matrix Convert to the frequency domain for weighted summation and restore to generate the weighted feature matrix The weighted feature matrix The deconvolution result and the accumulation of linear transformation minus the standard coating thickness d st Generate thickness fluctuation graph

[0009] The multi-wavelength spectral field G is decomposed using sparse non-negative matrix decomposition. S Perform spectral unmixing and iteratively obtain the basis matrix W S With the coefficient matrix HS And splice the input spectrum segmentation model, reversely splice the output of L-times progressive convolution and deconvolution and aggregate it into the plating leakage probability map P through point-by-point convolution, where the basis matrix W S Represents the spectral characteristics of different materials, the coefficient matrix H S Represents the spatial distribution characteristics of the material;

[0010] Preprocessed wave strength field G E Generate echo feature matrix F E , according to the wave strength field G E The index number of each point in the position code is generated and organized into a position code matrix C P , and the echo feature matrix F E The accumulated input position autoencoder generates the echo feature vector f by introducing the self-attention mechanism of position deviation and nonlinear transformation E , and generate bubble probability map based on fully connected layer decoding

[0011] The Bayesian confidence estimation algorithm is used to determine the causal relationship between process parameters Q, plating leakage, bubbles and thickness fluctuations, and the plating leakage probability map P and bubble probability map As the prior probability distribution of plating leakage and bubble, the process parameter Q is calculated. cu and thickness fluctuation diagram The joint conditional probability distribution under different leak conditions and bubble conditions in the historical data is used for joint causal reasoning based on the Bayesian joint probability formula. The posterior probability distribution of leak and bubble is updated based on the marginal probability formula to determine the leak area and bubble area.

[0012] Furthermore, obtaining the grayscale matrix I, the multi-wavelength spectrum matrix S, and the wave intensity matrix E includes the following steps:

[0013] The emitted light is divided into incident light and reference light by the interferometer's beam splitter, so that the reflected light formed by the incident light reflected by the coating layer interferes with the reference light to generate a grayscale dot matrix I;

[0014] The spectral imager emits detection light to the coating layer and uses an acousto-optic tunable filter to sequentially filter out R wavelengths of monochromatic reflected light signals from the received reflected light, generating a spectral intensity matrix of R wavelengths, which is stacked along the wavelength dimension to generate a three-dimensional multi-wavelength spectral matrix S;

[0015] An array ultrasonic imager is used to transmit ultrasonic pulses to the coating layer, and the received echo signals are processed by filtering and amplification and back-projected onto the imaging plane to generate a wave intensity matrix E.

[0016] Furthermore, the gray field G I , multi-wavelength spectral field GS and wave strength field G E The generation of includes the following steps:

[0017] determining a device coordinate system of a measuring device based on the right-hand rule, wherein the measuring device includes an interferometer, a spectral imager, and an array ultrasonic imager;

[0018] Capturing a device image of the measuring device using K calibrated tracking cameras and identifying pixel coordinates of markers in the device image using a feature extraction algorithm, wherein the markers are pre-installed on the measuring device;

[0019] The pixel coordinates of the marker points in the device image are converted into the corresponding world coordinates in the world coordinate system based on the camera intrinsic parameter matrix and camera extrinsic parameter matrix of the captured tracking camera;

[0020] Obtain the device coordinates of all markers in the device coordinate system, solve the measurement extrinsic parameter matrix that realizes the conversion between device coordinates and world coordinates, and obtain the measurement intrinsic parameter matrix of the measurement device through pre-calibration;

[0021] According to the measurement internal parameter matrix and measurement external parameter matrix of the measuring equipment, the pixel coordinates (u, v) of each point in the grayscale dot matrix I, the multi-wavelength spectrum dot matrix S and the wave intensity dot matrix E are mapped to the corresponding world coordinates (x, y, z) and the corresponding point data are assigned to generate the grayscale field G. I , multi-wavelength spectral field G S and wave strength field G E , point data include grayscale, spectral intensity of R wavelengths and echo intensity.

[0022] Furthermore, preprocess the gray field G I Generate grayscale feature matrix F I And input the multi-scale spectral analysis autoencoder to generate the thickness fluctuation map The following steps are involved:

[0023] The gray field G I The world coordinates (x, y, z) and grayscale I(x, y, z) of each point in are normalized and concatenated into corresponding grayscale feature vectors by column, and the grayscale feature vectors of all points are concatenated by row to generate the grayscale feature matrix F I ;

[0024] The gray feature matrix F I Input the multi-scale spectral analysis autoencoder and use three convolution kernels of different sizes to adjust the grayscale feature matrix F I Perform multi-scale convolution and concatenate the convolution results along the scale dimension into a multi-scale grayscale feature matrix

[0025] The multi-scale grayscale feature matrix is ​​transformed into Converted into multi-scale grayscale amplitude spectrum Pooling is performed at each scale, and the corresponding scale weights are generated through nonlinear mapping and weighted summation of the grayscale amplitude of each scale. The weighted feature matrix is ​​generated through two-dimensional discrete Fourier inverse transform.

[0026] Use deconvolution and linear transformation to expand the weighted feature matrix respectively To the preset dimension and superimpose to generate coating thickness map The coating thickness map Subtract the standard coating thickness d st Generate thickness fluctuation graph Thickness fluctuation graph The pixel coordinates (u, v) in are the projections of the world coordinates (x, y, z) on the xoy plane.

[0027] Furthermore, the joint loss function Loss selected during the training of the multi-scale spectral analysis autoencoder un is the mean square error loss L mse and structural similarity loss L sim The weighted sum, mean square error loss L mse Thickness fluctuation map output during multi-scale spectral analysis autoencoder training The average square error between the actual thickness fluctuation map ΔD and the structural similarity loss L sim Calculated based on the structural similarity index function to measure the thickness fluctuation map The structural similarity with the real thickness fluctuation map ΔD.

[0028] Furthermore, sparse non-negative matrix decomposition is used to decompose the multi-wavelength spectral field G S Perform spectral unmixing and iteratively obtain the basis matrix W S With the coefficient matrix H S The input spectrum segmentation model is spliced, and the outputs of the L-times progressive convolution and deconvolution are reversely spliced ​​and aggregated into a plating leakage probability map P through point-by-point convolution, including the following steps:

[0029] To minimize the multi-wavelength spectral field G S With the basis matrix W S Multiply the coefficient matrix H S The error is the target, and the basis matrix W is obtained by iterative optimization algorithm. S and coefficient matrix H S ;

[0030] Use linear transformation to unify the basis matrix W S and coefficient matrix H S Dimensions and concatenate into input data X in , input spectral segmentation model;

[0031] Through L convolutional layers, feature information of higher levels is gradually extracted. The input of each convolutional layer is the output of the previous convolutional layer, and feature information of L different levels is obtained. Among them, the input of the first convolutional layer is the input data X in ;

[0032] The feature information is gradually restored through L deconvolution layers. The input of each deconvolution layer is the output of the previous deconvolution layer, and L reconstructed feature information is obtained. Among them, the input of the first deconvolution layer is the feature information output by the Lth convolution layer;

[0033] Arrange the L reconstructed feature information in reverse order and splice them correspondingly with the L feature information. Aggregate the L splicing results through point-by-point convolution to generate a plating leakage probability map P. The pixel coordinates (u, v) in the plating leakage probability map P are the projections of the world coordinates (x, y, z) on the xoy plane.

[0034] Furthermore, to minimize the multi-wavelength spectral field G S With the basis matrix W S Multiply the coefficient matrix H S The error is the target, and the basis matrix W is obtained by iterative optimization algorithm. S and coefficient matrix H S , including the following steps:

[0035] Construct the objective function J(W S ,H S ) is to minimize the multi-wavelength spectral field G S Subtract the basis matrix W S Multiply the coefficient matrix H S The square of the F norm;

[0036] Initialize the basis matrix W for round 0 S (0) and the coefficient matrix H S (0), iterative optimization using multiplicative update rule;

[0037] In the nth round, the basis matrix W of the n-1th round is S Transpose of (n-1) and multi-wavelength spectral field G S The product of divided by the basis matrix W of the n-1th round S The transpose of (n-1), the basis matrix W of the n-1th round S (n-1) and the coefficient matrix H of the n-1th round S The product of (n-1) obtains the coefficient update matrix of the nth round And with the coefficient matrix H of the n-1th round S (n-1) multiplication to obtain the coefficient matrix H of the nth round S (n);

[0038] The multi-wavelength spectral field G S and the coefficient matrix H of the nth round S The product of the transpose of (n) divided by the basis matrix W of the n-1th round S (n-1), the coefficient matrix H of the nth round S (n) and the coefficient matrix H of the nth round S The product of the transpose of (n) obtains the basis update matrix of the nth round And with the basis matrix W of the n-1th round S (n-1) multiplication to obtain the basis matrix W of the nth round S (n);

[0039] The basis matrix W of the nth round S (n) and the coefficient matrix H S (n) Substitute the objective function J(W S ,H S ) and determine whether the objective function value is less than the convergence function value. If it is greater than or equal to, continue to the n+1th round;

[0040] If it is less than, the basis matrix W of the nth round S (n) and the coefficient matrix H S (n) as the basis matrix W S and coefficient matrix H S .

[0041] Furthermore, the preprocessing wave intensity field G E Generate echo feature matrix F E , according to the wave strength field G E The index number of each point in the position code is generated and organized into a position code matrix C P , including the following steps:

[0042] Sorting number wave strength field G E For M points, the world coordinates and echo intensity of each point are concatenated into the corresponding echo feature vector;

[0043] Each symbol in the position code of each point is equal to the sine of the ratio of the index number of the point to the power of the symbol number, where the power of the symbol number is equal to the power of 100 divided by the position code dimension;

[0044] The echo feature vectors and position codes of M points are concatenated row by row according to the index number sequence to generate the echo feature matrix F E and the position encoding matrix C P .

[0045] Furthermore, the cumulative echo feature matrix F E With the position encoding matrix C PAnd input the position autoencoder, and generate the echo feature vector f by introducing the self-attention mechanism of position deviation and nonlinear transformation E , and generate bubble probability map based on fully connected layer decoding The following steps are involved:

[0046] Cumulative echo feature matrix F E With the position encoding matrix C P Generate superposition input matrix X plus And input position autoencoder;

[0047] Introducing position bias into the self-attention mechanism to superimpose the input matrix X plus The superimposed input vector of each point in generates the corresponding attention weight, where the position deviation of each point is a matrix that records the difference between the position encoding of the point and the position encoding of the remaining M-1 points;

[0048] The attention weights of M points are concatenated into an attention weight vector a by column, and the attention weight vector a is nonlinearly transformed to generate the echo feature vector f E ;

[0049] The echo feature vector f is processed through three fully connected layers E Perform dimensionality expansion decoding, the first two fully connected layers and the third fully connected layer use ReLU activation function and Sigmoid activation function respectively, and output bubble probability map Bubble Probability Map The pixel coordinates (u, v) in are the projection of the world coordinates (x, y, z) on the xoy plane.

[0050] Furthermore, the Bayesian confidence estimation algorithm is used to determine the plating leakage area and the bubble area, including the following steps:

[0051] Determine the process parameters Q including target sputtering power p mat , coating temperature T emp and gas flow F gas , set the process parameter Q as the cause node, the plating leakage and bubble as the intermediate node, and the thickness fluctuation as the result node;

[0052] The plating probability map P and the bubble probability map As the prior probability distribution of leakage and bubble, the process parameter Q is obtained. cu And generate segment number group Num according to the preset numerical segmentation, for the plating leakage probability map P, bubble probability map and thickness fluctuation diagram For a single point in the image, obtain the prior probability of non-plating, the prior probability of bubble and the thickness fluctuation of the point;

[0053] Counting a first data amount that satisfies the segment number group and the prior probability of plating leakage at the point, and a second data amount that satisfies the segment number group, the prior probability of plating leakage at the point, the prior probability of bubbles, and thickness fluctuation in the historical data, and taking the ratio of the second data amount to the first data amount as the joint conditional probability of the point;

[0054] Based on the Bayesian joint probability formula, the joint posterior probability of the point is equal to the product of the joint conditional probability of the point, the prior probability of the plating leakage, and the prior probability of the bubble, divided by the normalized probability of the point. The normalized probability of the point is the ratio of the third data volume of the segment number group and thickness fluctuation that meets the point in the historical data to the total amount of historical data;

[0055] Based on the plating leakage probability map P and bubble probability map The joint posterior probability of all points in the algorithm is used to construct the joint posterior probability distribution. The posterior probability distribution of leaks and bubbles is obtained by combining the marginal probability formula. The confidence intervals corresponding to leaks and bubbles are generated according to the confidence level. The areas where the posterior probabilities of leaks and bubbles are greater than the upper limits of the corresponding confidence intervals are taken as the leak area and bubble area, respectively.

[0056] Compared with the prior art, the present invention converts the grayscale dot matrix, multi-wavelength spectral dot matrix and wave intensity dot matrix into grayscale field, multi-wavelength spectral field and wave intensity field through the measurement posture of the measuring equipment, realizes the spatial alignment of different measurement data, and assists the subsequent joint causal reasoning; pre-processes the grayscale field and adopts the multi-scale spectrum analysis autoencoder to analyze and obtain the thickness fluctuation map; adopts sparse non-negative matrix decomposition to disassemble the multi-wavelength spectral field and processes the input spectrum segmentation model to obtain the plating leakage probability map; pre-processes the wave intensity field and inputs the position autoencoder, considers the position deviation analysis to generate the bubble probability map; adopts the Bayesian confidence estimation algorithm, takes the plating leakage probability map and the bubble probability map as the prior probability distribution, based on the causal relationship between process parameters, plating leakage, bubbles and thickness fluctuation, combined with the joint conditional probability distribution in the historical data, based on the Bayesian principle to update the posterior probability distribution to determine the plating leakage area and the bubble area, realizes non-destructive testing that considers the causal relationship and is conducive to traceability analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a flow chart of a nondestructive testing method for semiconductor coatings on heat-generating clothing;

[0058] Figure 2 This is the model diagram of the multi-scale spectral analysis autoencoder;

[0059] Figure 3 Generate a flow chart for the skip probability map;

[0060] Figure 4 Flowchart for solving the basis matrix and coefficient matrix. DETAILED DESCRIPTION

[0061] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0062] Example 1

[0063] like Figure 1 As shown, a specific embodiment of the present invention discloses a non-destructive testing method for semiconductor coatings on heat-generating clothing, comprising the following steps:

[0064] Interferometer, spectral imager and array ultrasonic imager are used to measure the coating layer in sequence to obtain grayscale dot matrix I, multi-wavelength spectrum dot matrix S and wave intensity dot matrix E respectively. The measurement postures of interferometer, spectral imager and array ultrasonic imager are determined by optical tracking system respectively, and the grayscale dot matrix I, multi-wavelength spectrum dot matrix S and wave intensity dot matrix E are converted into grayscale field G in the world coordinate system respectively. I , multi-wavelength spectral field G S and wave strength field G E ;

[0065] Normalize the gray field G I The world coordinates and grayscale of the M points in the image are sorted to generate the grayscale feature matrix F. I And input the multi-scale spectral analysis autoencoder, through multi-scale convolution to understand the spatial correlation of grayscale from different perspectives and splice the convolution results to generate a multi-scale grayscale feature matrix The multi-scale gray feature matrix Convert to the frequency domain and adaptively assign scale weights for weighted summation to suppress irrelevant noise, and then reversely restore it to the spatial domain to generate a weighted feature matrix Simultaneous processing of weighted feature matrix using deconvolution and linear transformation And through the residual structure superposition, combined with the standard coating thickness d st Generate thickness fluctuation graph

[0066] The multi-wavelength spectral field G is decomposed using sparse non-negative matrix decomposition. S Perform spectral unmixing and iteratively update the multi-wavelength spectral field G S Decomposed into basis matrix W S With the coefficient matrix H S The product of the basis matrix W S With the coefficient matrix H S Splice the input spectrum segmentation model, use L times convolution to continuously extract higher-level feature information in encoding, and gradually restore the highest-level feature information through L times deconvolution in decoding. Splice the feature information of the same level in encoding and decoding and aggregate the output of the plating leakage probability map P through point-by-point convolution. Among them, the basis matrix W S Represents the spectral characteristics of different materials, the coefficient matrix H SRepresents the spatial distribution characteristics of the material;

[0067] Splicing wave strength field G E The world coordinates and echo intensities of the M points in the image are sorted and the echo feature matrix F is generated. E , generate position codes based on the index numbers of M points and organize them to generate the position code matrix C P , and the echo feature matrix F E The accumulated input position autoencoder generates the echo feature vector f by introducing the self-attention mechanism of position deviation and nonlinear transformation E , through the continuous fully connected layer to the echo feature vector f E Perform dimension expansion and decoding to generate bubble probability map

[0068] The Bayesian confidence estimation algorithm is used to determine the causal relationship between process parameters Q, plating leakage, bubbles and thickness fluctuations, and the plating leakage probability map P and bubble probability map As the prior probability distribution of plating leakage and bubble, the process parameter Q is calculated. cu and thickness fluctuation diagram The joint conditional probability distribution under different leak and bubble conditions in historical data is used for joint causal reasoning based on the Bayesian joint probability formula. The posterior probability distribution of leak and bubble is updated based on the marginal probability formula. The leak and bubble confidence intervals are combined to mark the leak and bubble areas of the coating layer.

[0069] Furthermore, obtaining the grayscale matrix I, the multi-wavelength spectrum matrix S, and the wave intensity matrix E includes the following steps:

[0070] The emitted light is divided into incident light and reference light by using the beam splitter in the interferometer. The incident light is reflected by the coating layer to form reflected light carrying the coating thickness information, which interferes with the reference light to generate a grayscale dot matrix I. The grayscale dot matrix I is a discrete matrix composed of the grayscale of M points. I(u,v) is the grayscale of the point with pixel coordinates (u,v) in the grayscale dot matrix I. u and v are the horizontal and vertical coordinates of the pixel coordinate system respectively.

[0071] A spectral imager is used to emit detection light to the coating layer and receive reflected light. The acousto-optic effect of the acousto-optic tunable filter is used to sequentially screen out R monochromatic reflected light signals of different wavelengths in the order of wavelength, generating R two-dimensional spectral intensity dot matrices of different wavelengths. Channels are stacked along the wavelength dimension to generate a three-dimensional multi-wavelength spectral dot matrix S, where the spectral intensity dot matrix of each wavelength is a discrete matrix composed of the spectral intensities of M points, S r (u,v) is the spectral intensity matrix S of the rth wavelength rThe spectral intensity of the pixel at coordinate (u, v), r = 1,…,R;

[0072] An array ultrasonic imager is used to transmit ultrasonic pulses to the coating layer and receive the echo signals generated by the reflection. After filtering and amplification, the signals are back-projected onto the imaging plane to generate a discretized wave intensity matrix E. The wave intensity matrix E is a discrete matrix composed of the echo intensities of M points, and E(u,v) is the echo intensity of the point with pixel coordinates (u,v) in the wave intensity matrix E.

[0073] Furthermore, the optical tracking system determines the measurement posture of the interferometer, spectral imager and array ultrasonic imager and converts the grayscale dot matrix I, multi-wavelength spectrum dot matrix S and wave intensity dot matrix E into the grayscale field G. I , multi-wavelength spectral field G S and wave strength field G E , including the following steps:

[0074] The installation base point of the measuring device is taken as the origin, and the emission direction of the measuring device is taken as the X coordinate system of the device. e Axis direction, based on the right-hand rule, hold the Z axis of the device coordinate system with your right hand e Axis, thumb pointing to Z e The positive direction of the axis, the natural bending direction of the four fingers is the X axis of the device coordinate system. e The positive direction of the axis is around the Y axis of the device coordinate system e The positive direction of the axis is used to construct the device coordinate system of the measuring equipment, which includes an interferometer, a spectral imager, and an array ultrasonic imager;

[0075] K tracking cameras calibrated using the Zhang Zhengyou calibration method are used to capture device images of the measuring device. The pixel coordinates of each marker point in the device image are identified using a feature extraction algorithm. The Zhang Zhengyou calibration method is used to determine the camera intrinsic parameter matrix and camera extrinsic parameter matrix of the tracking camera. The camera intrinsic parameter matrix is ​​used to convert the pixel coordinates in the device image into camera coordinates in the camera coordinate system, and the camera extrinsic parameter matrix is ​​used to convert the camera coordinates into world coordinates in the world coordinate system. After the K tracking cameras are calibrated, the poses are fixed. The marker points are pre-installed on the measuring device. Generally, 4 markers are installed, with one as the base point and the other three along the X coordinate system of the device coordinate system. e Axis, Y e Axis and Z e Arrangement in the positive direction of the axis;

[0076] For the pixel coordinates of the marker point in the device image captured by the kth tracking camera It is known that the camera intrinsic parameter matrix and camera extrinsic parameter matrix of the kth tracking camera are and The world coordinates of the marker point in the world coordinate system (x ma ,yma ,z ma ) is equal to the camera intrinsic parameter matrix Camera extrinsic matrix and pixel coordinates The product of x and y is used to obtain the world coordinates of all the marking points of the measuring device, where x ma 、y ma and z ma They are the X coordinates of the marked points in the world coordinate system w Axis, Y w Axis and Z w Axis coordinate values, k = 1,…,K;

[0077] Based on the installation positions of all markers of the measuring device, the device coordinates of all markers in the device coordinate system are determined. Given that there is a linear relationship between the world coordinates and the device coordinates of the known markers, the simultaneous equations can be solved to obtain the measurement extrinsic parameter matrix of the measuring device. Since the measurement intrinsic parameter matrix of the measuring device does not change as the measuring device moves, it is obtained through pre-calibration experiment calculation. The measurement extrinsic parameter matrix includes the interference extrinsic parameter matrix T of the interferometer. I,out , the spectral extrinsic parameter matrix T of the spectral imager S,out and the ultrasonic extrinsic parameter matrix T of the array ultrasonic imager E,out , the measurement internal parameter matrix includes the corresponding interference internal parameter matrix T I,in , spectral internal parameter matrix T S,in and ultrasound internal parameter matrix T E,in ;

[0078] According to the measurement internal parameter matrix and measurement external parameter matrix of the measuring equipment, the pixel coordinates (u, v) of each point in the grayscale dot matrix I, the multi-wavelength spectrum dot matrix S and the wave intensity dot matrix E are mapped to the corresponding world coordinates (x, y, z) and the corresponding point data are assigned to generate the grayscale field G. I , multi-wavelength spectral field G S and wave strength field G E , point data includes grayscale, spectral intensity of R wavelengths and echo intensity. Due to the interference external parameter matrix T I,out and the interference internal parameter matrix T I,in , spectral extrinsic matrix T S,out and the spectral internal parameter matrix T S,in , ultrasonic extrinsic parameter matrix T E,out and ultrasound internal parameter matrix T E,in Different, the world coordinates corresponding to the points with the same pixel coordinates (u, v) in the grayscale dot matrix I, the multi-wavelength spectrum dot matrix S and the wave intensity dot matrix E are also different. By measuring the mapping of the internal parameter matrix and the external parameter matrix, the grayscale field G I , multi-wavelength spectral field G S and wave strength field G E The distribution space is aligned.

[0079] like Figure 2 As shown, further, the gray field G I The world coordinates (x, y, z) and grayscale I (x, y, z) of the M points are normalized and sorted to generate the grayscale feature matrix F I And input the multi-scale spectral analysis autoencoder to generate the thickness fluctuation map The following steps are involved:

[0080] The gray field G I The world coordinates (x, y, z) and grayscale I (x, y, z) of each point in the _____ are normalized to the maximum and minimum values ​​and concatenated into corresponding grayscale feature vectors by column. The grayscale feature vector of each point includes the normalized world coordinates and normalized grayscale The grayscale feature vectors of M points are spliced ​​row by row into the grayscale feature matrix F I ;

[0081] The gray feature matrix F I Input the multi-scale spectral analysis autoencoder and use three convolution kernels of different sizes to adjust the grayscale feature matrix F I Perform parallel multi-scale convolution and concatenate the convolution results along the scale dimension into a multi-scale grayscale feature matrix The multi-scale convolution is as follows:

[0082]

[0083] in, Indicates splicing based on convolution kernel size j, Conv j () represents the convolution operation with a convolution kernel size of j×j, j∈{3,5,7}, and convolution kernels of different sizes can capture the grayscale feature matrix F I The information of different scales in the image is extracted, and the spatial correlation information between gray levels is extracted from both local and global perspectives;

[0084] Multi-scale grayscale feature matrix Perform two-dimensional discrete Fourier transform to obtain multi-scale grayscale amplitude spectrum Multi-scale grayscale amplitude spectrum The grayscale amplitudes of the three scales are globally averaged and pooled respectively, and the corresponding scale weights are generated through the fully connected layer and Sigmoid activation function to reflect the importance of each scale. The scale weights are used to perform weighted summation of the grayscale amplitudes of the three scales and generate a weighted feature matrix through the two-dimensional discrete Fourier inverse transform.

[0085] Use deconvolution to gradually expand the weighted feature matrix to the preset dimension and transform the weighted feature matrix through linear transformation Linear expansion to the preset dimension, introducing residual structure for superposition to avoid the gradient disappearance problem, and generating the coating thickness map The specific formula is as follows:

[0086]

[0087] Among them, W I is the grayscale linear weight matrix, DeConv() represents the deconvolution operation, the deconvolution kernel size is 3×3, and the coating thickness map is Subtract the standard coating thickness d st Generate thickness fluctuation graph Thickness fluctuation graph Reflect the coating thickness at different locations of the coating layer and the standard coating thickness d in two-dimensional form st The error, that is, the thickness fluctuation diagram The pixel coordinates (u, v) in are the projection of the world coordinates (x, y, z) on the xoy plane.

[0088] Furthermore, the joint loss function Loss selected during the training of the multi-scale spectral analysis autoencoder un Combined with the mean square error loss L mse and structural similarity loss L sim , as follows:

[0089] Loss un =α1·L mse +α2·L sim ,

[0090] Among them, α1 and α2 are the first loss balance coefficient and the second loss balance coefficient respectively, and satisfy α1+α2=1, the mean square error loss L mse Thickness fluctuation map output during multi-scale spectral analysis autoencoder training The average square error between the actual thickness fluctuation map ΔD and the structural similarity loss L sim Considering the brightness, contrast and structural information of the image to measure the thickness fluctuation map The structural similarity with the true thickness fluctuation map ΔD is calculated using the structural similarity index function, and the training method uses the Adam optimizer.

[0091] like Figure 3 As shown, further, sparse non-negative matrix decomposition is used to transform the multi-wavelength spectral field G S Disassemble and obtain the basis matrix W S and coefficient matrix H S And splicing the input spectrum segmentation model to generate the output plating leakage probability map P, including the following steps:

[0092] Initialize the basis matrix W for round 0 S (0) and the coefficient matrix H S (0), the optimization objective is to minimize the multi-wavelength spectral field G S With the basis matrix W S Multiply the coefficient matrix H S The error is obtained by continuously updating the basis matrix and coefficient matrix of each round through the iterative optimization algorithm to solve the basis matrix W S and coefficient matrix H S ;

[0093] The basis matrix W S and coefficient matrix H S Generate input data X by adjusting to the same dimension and splicing through linear transformation in And further input into the spectral segmentation model;

[0094] In the encoding stage, higher-level feature information is continuously extracted through L convolutional layers of the same structure, that is, the input of the lth convolutional layer is the output of the l-1th convolutional layer, and the output of the lth convolutional layer is the input of the l+1th convolutional layer, obtaining L feature information, which are recorded as in, is the lth feature information, and the input of the first convolutional layer is the input data X in , the convolution kernel size of L convolutional layers is 3×3, 2≤l≤L-1;

[0095] In the decoding stage, feature information is continuously restored through L deconvolution layers of the same structure, that is, the input of the lth deconvolution layer is the output of the l-1th deconvolution layer, and the output of the lth deconvolution layer is the input of the l+1th deconvolution layer, obtaining L reconstructed feature information, which are respectively recorded as in, is the lth reconstructed feature information, and the input of the first deconvolution layer is the Lth feature information in the encoding stage The convolution kernel size of the L deconvolution layers is also 3×3, 2≤l≤L-1;

[0096] For the lth feature information and reconstruct feature information As l increases, the characteristic information The level is increasing, reconstructing feature information The level is decreasing, so the levels of feature information and reconstructed feature information are in reverse order. The L reconstructed feature information are arranged in reverse order and spliced ​​with the corresponding L feature information. The L splicing results are aggregated by point-by-point convolution with a convolution kernel of 1×1 to generate a plating leakage probability map P. The plating leakage probability map P reflects the plating leakage probability at different positions of the coating layer in a two-dimensional form. The dimension and thickness fluctuation map The same, and the pixel coordinates (u, v) in the plating skip probability map P are the projections of the world coordinates (x, y, z) on the xoy plane.

[0097] like Figure 4 As shown, further, the optimization goal is to minimize the multi-wavelength spectral field G S With the basis matrix W S Multiply the coefficient matrix H S The error is obtained by continuously updating the basis matrix and coefficient matrix of each round through the iterative optimization algorithm to solve the basis matrix W S and coefficient matrix H S , including the following steps:

[0098] Construct the objective function J(W S ,H S ) is the multi-wavelength spectral field G S Subtract the basis matrix W S Multiply the coefficient matrix H S The square of the F norm, specifically

[0099] Initialize the basis matrix W for round 0 S (0) and the coefficient matrix H S (0), iterative optimization using multiplicative update rule;

[0100] In the nth round of optimization, the basis matrix W of the n-1th round is S Transpose of (n-1) and multi-wavelength spectral field G S The product of divided by the basis matrix W of the n-1th round S The transpose of (n-1), the basis matrix W of the n-1th round S (n-1) and the coefficient matrix H of the n-1th round S The product of (n-1) obtains the coefficient update matrix of the nth round Update the coefficient matrix of the nth round and the coefficient matrix H of the n-1th round S (n-1) multiplied to update the coefficient matrix H of the nth round S (n), as follows:

[0101]

[0102] The multi-wavelength spectral field G S and the coefficient matrix H of the nth round S The product of the transpose of (n) divided by the basis matrix W of the n-1th round S (n-1), the coefficient matrix H of the nth round S (n) and the coefficient matrix H of the nth round SThe product of the transpose of (n) obtains the basis update matrix of the nth round Update the basis matrix of the nth round and the basis matrix W of the n-1th round S (n-1) multiplication to update the basis matrix W of the nth round S (n), as follows:

[0103]

[0104] The basis matrix W of the nth round S (n) and the coefficient matrix H S (n) Substitute the objective function J(W S ,H S ), determine whether the objective function value of the nth round is less than the convergence function value. If it is greater than or equal to, continue the optimization for the n+1th round;

[0105] If it is less than, then the basis matrix W of the nth round S (n) and the coefficient matrix H S (n) is the basis matrix W obtained as the final disassembly S and coefficient matrix H S .

[0106] Furthermore, the splicing wave strength field G E The world coordinates and echo intensities of the M points in the image are sorted and the echo feature matrix F is generated. E , generate position codes based on the index numbers of M points and organize them to generate the position code matrix C P , including the following steps:

[0107] The wave strength field G E The M points are sorted and numbered according to a certain rule. The sorting rule can be customized. For example, the wave strength field G E The point in the upper left corner is used as the starting point. First, go down along the world coordinate system X w The positive direction of the axis is numbered sequentially with the same Y w Axis point, and then along the Y w Search for the point below the starting point in the negative direction of the X axis, and then move along the w Numbering is performed in the positive direction of the axis until all M points are numbered;

[0108] Get the world coordinates (x m ,y m ,z m ) and echo intensity E(x m ,y m ,z m ) and spliced ​​into the echo feature vector of the mth point m is the index number, m=1,…,M;

[0109] For the index number m of the mth point, the echo feature vector is known The dimension is 4, and the position code of the mth point is generated based on the sine function Positional encoding The dimension is also 4, position encoding The i-th code element The specific formula is as follows:

[0110]

[0111] Where i is the symbol number, i = 1, 2, 3, 4;

[0112] The echo feature vectors and position codes of M points are concatenated row by row according to the index number sequence to generate the echo feature matrix F E and the position encoding matrix C P .

[0113] Furthermore, the cumulative echo feature matrix F E With the position encoding matrix C P And input the position autoencoder to generate a bubble probability map The following steps are involved:

[0114] Cumulative echo feature matrix F E With the position encoding matrix C P Generate superposition input matrix X plus =F E +C P And input the position autoencoder, which includes a spatial encoder and a spatial decoder;

[0115] The spatial encoder is improved based on the self-attention mechanism, and the position deviation is introduced into the self-attention mechanism to superimpose the input matrix X plus The superimposed input vector of each point in generates the corresponding attention weight, and the attention weight a of the mth point m The specific formula is as follows:

[0116]

[0117] Among them, w qui and w key They are the original query vector and key vector in the self-attention mechanism, with dimensions of (M-1)×1, B qui and B key They are the original query bias matrix and key bias matrix in the self-attention mechanism, with dimensions of (M-1)×4. is the superposition input vector of the mth point, with a dimension of 1×4, is the encoding deviation matrix of the mth point, by removing the position encoding matrix CP The position code of the mth point in And the position code of the remaining M-1 points is subtracted from the position code of the mth point Get, that is, record the position code of the mth point Deviation from the position encoding of the remaining M-1 points;

[0118] The attention weights of M points are concatenated into an attention weight vector a by column, and the attention weight vector a is nonlinearly transformed through a linear layer and a ReLU activation function to generate an echo feature vector f E ;

[0119] The spatial decoder uses three fully connected layers to process the echo feature vector f E Perform dimension expansion and decoding. The first two fully connected layers include linear transformation and ReLU activation function. The third fully connected layer includes linear transformation and Sigmoid activation function that limits the output value to the range of 0 to 1, and outputs a bubble probability map. Bubble Probability Map Reflects the probability of bubbles existing at different locations in the coating layer in two-dimensional form, dimension and thickness fluctuation diagram Same, and bubble probability map The pixel coordinates (u, v) in are the projection of the world coordinates (x, y, z) on the xoy plane.

[0120] Furthermore, a Bayesian confidence estimation algorithm is used to mark the non-plating area and the bubble area of ​​the coating layer, including the following steps:

[0121] Construct node causal relationships and determine the process parameters Q that affect the coating effect, including the target sputtering power p mat , coating temperature T emp and gas flow F gas ,Since the process parameter Q directly leads to the generation of plating leakage and bubbles, and plating leakage and bubbles cause thickness changes, the process parameter Q is set as the cause node, plating leakage and bubbles as the intermediate nodes, and thickness fluctuation as the result node;

[0122] The plating probability map P and the bubble probability map As the prior probability distribution of plating leakage and bubble, respectively, for the point with pixel coordinates (u, v), the prior probability of plating leakage, bubble prior probability and thickness fluctuation are P(u, v), and

[0123] Get the process parameter Q for this time cu , according to the target sputtering power p mat , coating temperature T emp and gas flow F gasThe numerical segments are set separately to determine the process parameter Q cu The segment number group Num composed of the segment number of each process parameter is related to the thickness fluctuation Combine and construct the observation data of the point with pixel coordinates (u, v);

[0124] The first data quantity that satisfies the segment number group and the prior probability of plating leakage in the statistical historical data is Num and P(u,v), the first data quantity that satisfies the segment number group, the prior probability of plating leakage, the prior probability of bubble and the thickness fluctuation is Num, P(u,v), and The second data volume is used, and the ratio of the second data volume to the first data volume is used as the joint conditional probability of the point with pixel coordinates (u, v)

[0125] Based on the Bayesian joint probability formula, the joint posterior probability of the point with pixel coordinates (u, v) is Equal to the joint conditional probability and the prior probability of plating leakage P(u,v) and the prior probability of bubble The product of divided by the normalized probability Normalized probability That is, the historical data that satisfies the segment number group and thickness fluctuation are Num and The ratio of the third data volume to the total historical data volume;

[0126] Based on the plating leakage probability map P and bubble probability map The joint posterior probability of all points is used to construct the joint posterior probability distribution

[0127] Based on the marginal probability formula, the joint posterior probability distribution The posterior probability distribution of leak plating and the posterior probability distribution of bubbles are obtained by disassembling, with a confidence level of 0.95. The confidence interval of leak plating and the confidence interval of bubbles are generated based on the basic theory of probability theory. The area where the posterior probability of leak plating in the posterior probability distribution of leak plating is greater than the upper limit of the confidence interval of leak plating and the area where the posterior probability of bubbles in the posterior probability distribution of bubbles is greater than the upper limit of the confidence interval of bubbles are located are respectively regarded as the leak plating area and the bubble area, and the coating layer is marked according to the projection relationship between the pixel coordinates (u, v) and the world coordinates (x, y, z). Since the causal relationship between process parameters, leak plating, bubbles and thickness fluctuations is established, when problems such as thickness anomalies, leak plating and bubbles occur, the cause of the problem can be inferred based on the relevant knowledge of probability theory and the posterior probability distribution inverse analysis.

[0128] The present invention discloses a non-destructive testing method for semiconductor coatings on heat-generating clothing. The method converts a grayscale dot matrix, a multi-wavelength spectral dot matrix, and a wave intensity dot matrix into a grayscale field, a multi-wavelength spectral field, and a wave intensity field through the measurement posture of a measuring device, thereby realizing spatial alignment of different measurement data and assisting subsequent joint causal reasoning; pre-processes the grayscale field and adopts a multi-scale spectral analysis autoencoder to analyze and obtain a thickness fluctuation map; adopts sparse non-negative matrix decomposition to disassemble the multi-wavelength spectral field and processes the input spectral segmentation model to obtain a plating leakage probability map; pre-processes the wave intensity field and inputs it into a position autoencoder, and generates a bubble probability map considering position deviation analysis; adopts a Bayesian confidence estimation algorithm, takes the plating leakage probability map and the bubble probability map as a priori probability distribution, and based on the causal relationship between process parameters, plating leakage, bubbles, and thickness fluctuations, combined with the joint conditional probability distribution in historical data, updates the posterior probability distribution based on the Bayesian principle to determine the plating leakage area and the bubble area, thereby realizing non-destructive testing that considers causal relationship and is conducive to traceability analysis.

[0129] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiment. All technical solutions based on the concept of the present invention are within the scope of protection of the present invention. It should be noted that for those skilled in the art, various improvements and modifications that do not depart from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A non-destructive testing method for semiconductor coatings on heat-generating clothing, characterized in that: The following steps are involved: Measure and obtain grayscale dot matrix, multi-wavelength spectrum dot matrix and wave intensity dot matrix and convert them into grayscale field, multi-wavelength spectrum field and wave intensity field through measuring posture; Preprocess the grayscale field and input it into the multi-scale spectrum analysis autoencoder. Convert the splicing results of the multi-scale convolution to the frequency domain for weighted summation and restore to generate a weighted feature matrix. Subtract the standard coating thickness from the sum of the deconvolution result of the weighted feature matrix and the linear transformation to generate a thickness fluctuation map. The matrices obtained by sparse non-negative matrix decomposition of the multi-wavelength spectral field are spliced ​​and input into the spectral segmentation model. The outputs of multiple convolutions and deconvolutions are reversely spliced ​​and aggregated into a plating leakage probability map through point-by-point convolution. The echo feature matrix generated by the preprocessed wave intensity field and the position encoding matrix generated based on the index number of each point in the wave intensity field are accumulated and input into the position autoencoder. The self-attention mechanism and nonlinear transformation of the position deviation are used to encode the data and the bubble probability map is generated based on the fully connected layer decoding. The Bayesian confidence estimation algorithm is adopted, and the leakage probability map and bubble probability map are used as the prior probability distribution. The joint conditional probability distribution of the process parameters and thickness fluctuation map under different leakage and bubble conditions in the historical data is statistically analyzed. Based on the Bayesian joint probability formula, joint causal reasoning is performed and combined with the marginal probability formula to update the posterior probability distribution to determine the leakage area and bubble area.

2. The method for nondestructive testing of semiconductor coatings for heat-generating clothing according to claim 1, characterized in that: The method of performing joint causal reasoning based on the Bayesian joint probability formula and updating the posterior probability distribution in combination with the marginal probability formula includes the following specific steps: The plating leakage probability map and the bubble probability map are used as prior probability distributions, respectively. The process parameters are obtained and segment number groups are generated according to the preset numerical segmentation. The plating leakage probability map, the bubble probability map and the thickness fluctuation map are obtained for a single point. Prior probability of plating leakage, prior probability of bubbles and thickness fluctuation are obtained. The ratio of the second data amount that satisfies the segment number group, the prior probability of plating leakage at a single point, the prior probability of bubbles, and the thickness fluctuation to the first data amount that satisfies the segment number group and the prior probability of plating leakage at a single point in the historical data is used as the joint conditional probability of the single point; Based on the Bayesian joint probability formula, the joint posterior probability of a single point is equal to the product of the joint conditional probability of the single point, the prior probability of the plating leakage, and the prior probability of the bubble divided by the normalized probability of the single point. The normalized probability of a single point is the ratio of the third data volume in the historical data that meets the segment number group and thickness fluctuation of the single point to the total amount of historical data; The joint posterior probability distribution is constructed based on the joint posterior probability of all points, and the posterior probability distribution is obtained by combining the marginal probability formula.

3. The method for nondestructive testing of semiconductor coatings for heat-generating clothing according to claim 1, characterized in that: The matrix obtained by sparse non-negative matrix decomposition of the spliced ​​multi-wavelength spectral field is input into the spectral segmentation model, and the outputs of multiple convolutions and deconvolutions are reversely spliced ​​and aggregated into a plating leakage probability map through point-by-point convolution, including the following steps: With the goal of minimizing the error between the multi-wavelength spectral field and the basis matrix and the coefficient matrix, the basis matrix and the coefficient matrix are obtained by an iterative optimization algorithm. Use linear transformation to unify the dimensions of the basis matrix and coefficient matrix and splice them into input data, which is then input into the spectral segmentation model; Through L convolutional layers, feature information of higher levels is gradually extracted. The input of each convolutional layer is the output of the previous convolutional layer, and feature information of L different levels is obtained. Among them, the input of the first convolutional layer is the input data; The feature information is gradually restored through L deconvolution layers. The input of each deconvolution layer is the output of the previous deconvolution layer, and L reconstructed feature information is obtained. Among them, the input of the first deconvolution layer is the feature information output by the Lth convolution layer; Arrange the L reconstructed feature information in reverse order and splice them with the corresponding L feature information. Aggregate the L splicing results through point-by-point convolution to generate a plating skip probability map.

4. The method for nondestructive testing of semiconductor coatings on heat-generating clothing according to claim 3, wherein: The method aims to minimize the error between the multi-wavelength spectrum field and the basis matrix and the coefficient matrix, and obtains the basis matrix and the coefficient matrix by an iterative optimization algorithm, including the following steps: Construct an objective function to minimize the square of the norm of the multi-wavelength spectral field minus the basis matrix multiplied by the coefficient matrix, and initialize the initial values ​​of the basis matrix and the coefficient matrix; In this round of optimization, the coefficient matrix of this round is updated based on the ratio of the product of the transpose of the basis matrix of the previous round and the multi-wavelength spectrum field to the transpose of the basis matrix of the previous round, and the product of the basis matrix of the previous round and the coefficient matrix of the previous round; The basis matrix of this round is updated according to the ratio of the product of the multi-wavelength spectrum field and the transpose of the coefficient matrix of this round to the basis matrix of the previous round, the coefficient matrix of this round and the transpose of the coefficient matrix of this round; Substitute the basis matrix and coefficient matrix of this round into the objective function and decide whether to start the next round of optimization or use the basis matrix and coefficient matrix of this round as the basis matrix and coefficient matrix based on the relationship between the objective function value and the convergence function value.

5. The method for nondestructive testing of semiconductor coatings for heat-generating clothing according to claim 1, characterized in that: Preprocessing the grayscale field and inputting it into the multi-scale spectral analysis autoencoder to generate the thickness fluctuation map includes the following steps: Normalize the world coordinates and grayscale of each point in the grayscale field and splice them into the corresponding grayscale feature vector by column, and splice the grayscale feature vectors of all points by row to generate the grayscale feature matrix; The grayscale feature matrix is ​​input into the multi-scale spectral analysis autoencoder, and multi-scale convolution is performed on the grayscale feature matrix through three convolution kernels of different sizes. The convolution results are spliced ​​along the scale dimension into a multi-scale grayscale feature matrix; The multi-scale grayscale feature matrix is ​​converted into a multi-scale grayscale amplitude spectrum through two-dimensional discrete Fourier transform and pooled at each scale. The corresponding scale weights are generated through nonlinear mapping and the grayscale amplitude of each scale is weighted and summed. The weighted feature matrix is ​​generated through two-dimensional inverse discrete Fourier transform. Deconvolution and linear transformation are used to expand the weighted feature matrix to the preset dimension and superimposed to generate a coating thickness map. The coating thickness map is subtracted from the standard coating thickness to generate a thickness fluctuation map.

6. The method for nondestructive testing of semiconductor coatings for heat-generating clothing according to claim 5, characterized in that: The joint loss function selected during the training of the multi-scale spectral analysis autoencoder is the weighted sum of the mean square error loss and the structural similarity loss. The mean square error loss is the average square error between the thickness fluctuation map output during the training of the multi-scale spectral analysis autoencoder and the actual thickness fluctuation map. The structural similarity loss is calculated based on the structural similarity index function and is used to measure the structural similarity between the thickness fluctuation map and the actual thickness fluctuation map.

7. The method for nondestructive testing of semiconductor coatings on heat-generating clothing according to claim 1, wherein: The construction of the echo feature matrix and the position coding matrix includes the following steps: Sort and number the M points in the wave intensity field, and concatenate the world coordinates and echo intensity of each point into the corresponding echo feature vector; Each symbol in the position code of each point is equal to the sine of the ratio of the index number of the point to the power of the symbol number, where the power of the symbol number is equal to the power of 100 divided by the position code dimension; The echo feature vectors and position codes of the M points are concatenated row by row according to the order of index numbers to generate the echo feature matrix and position code matrix.

8. The method for nondestructive testing of semiconductor coatings on heat-generating clothing according to claim 7, wherein: The echo feature matrix and the position encoding matrix are accumulated and input into the position autoencoder. The echo feature vector is generated by introducing a self-attention mechanism with position deviation and nonlinear transformation. The bubble probability map is then generated based on the fully connected layer decoding. The following steps are included: The accumulated echo feature matrix and the position encoding matrix generate a superimposed input matrix and input it into the position autoencoder; Introducing position bias into the self-attention mechanism, generating corresponding attention weights for the superimposed input vector of each point in the superimposed input matrix, where the position bias of each point is a matrix that records the difference between the position encoding of the point and the position encoding of the remaining M-1 points; The attention weights of M points are concatenated into an attention weight vector by column, and the attention weight vector is nonlinearly transformed to generate an echo feature vector; The echo feature vector is expanded and decoded through three fully connected layers. The first two fully connected layers and the third fully connected layer use the ReLU activation function and the Sigmoid activation function respectively to output the bubble probability map.

9. The method for nondestructive testing of semiconductor coatings on heat-generating clothing according to claim 1, wherein: The acquisition of grayscale dot matrix, multi-wavelength spectrum dot matrix and wave intensity dot matrix includes the following steps: The emitted light is divided into incident light and reference light by the interferometer's beam splitter, so that the reflected light formed by the incident light reflected by the coating layer interferes with the reference light to generate a grayscale dot matrix; A spectral imager is used to emit detection light to the coating layer, and an acousto-optic tunable filter is used to sequentially filter out R wavelengths of monochromatic reflected light signals from the received reflected light, generating a spectral intensity matrix of R wavelengths, which is then stacked along the wavelength dimension to generate a three-dimensional multi-wavelength spectral matrix. An array ultrasonic imager is used to transmit ultrasonic pulses to the coating layer, and the received echo signals are processed by filtering and amplification and back-projected onto the imaging plane to generate a wave intensity matrix.

10. The method for nondestructive testing of semiconductor coatings on heat-generating clothing according to claim 1, wherein: The generation of grayscale field, multi-wavelength spectrum field and wave intensity field includes the following steps: determining a device coordinate system of a measuring device based on the right-hand rule, wherein the measuring device includes an interferometer, a spectral imager, and an array ultrasonic imager; Capturing a device image of the measuring device using K calibrated tracking cameras and identifying pixel coordinates of markers in the device image using a feature extraction algorithm, wherein the markers are pre-installed on the measuring device; The pixel coordinates of the marker points in the device image are converted into the corresponding world coordinates in the world coordinate system based on the camera intrinsic parameter matrix and camera extrinsic parameter matrix of the captured tracking camera; Obtain the device coordinates of all markers in the device coordinate system, solve the measurement extrinsic parameter matrix that realizes the conversion between device coordinates and world coordinates, and obtain the measurement intrinsic parameter matrix of the measurement device through pre-calibration; According to the measurement intrinsic parameter matrix and measurement extrinsic parameter matrix of the measuring equipment, the pixel coordinates of each point in the grayscale dot matrix, multi-wavelength spectrum dot matrix and wave intensity dot matrix are mapped to the corresponding world coordinates and assigned corresponding point data to generate the grayscale field, multi-wavelength spectrum field and wave intensity field. The point data includes grayscale, spectral intensity of R wavelengths and echo intensity.

Images