Probe contact impedance real-time monitoring and compensation method, device and storage medium

Abstract

The present invention relates to a real-time monitoring and compensation method, device and storage medium for probe contact impedance. The method: obtain a first measurement data set including probe contact resistance value and contact force, and input it into a two-layer deep learning neural network for training to obtain a probe contact impedance monitoring model; perform dual-parameter fault detection to obtain a contact state classification result; generate a two-dimensional compensation matrix based on the contact state classification result, and perform nonlinear mapping calculation on the second measurement data set to obtain a compensation coefficient; input the compensation coefficient, the original probe structure parameter and the original probe operation parameter into a graph attention stacked autoencoder for parameter prediction to obtain a target probe structure parameter and a target probe operation parameter; perform cloud model calculation and weighted operation to obtain a target compensation instruction, which is used to adjust the probe position parameter and the probe pressure parameter. The implementation of the present invention improves the accuracy of the probe test of the chip.

Description

Technical Field

[0001] The present invention relates to the technical field of probe contact impedance, and in particular to a method, device and storage medium for real-time monitoring and compensation of probe contact impedance. Background Art

[0002] With the continuous development of integrated circuit manufacturing technology, chip size continues to shrink and integration continues to increase, which puts higher requirements on probe test technology. The stability of probe contact impedance directly affects the accuracy and reliability of test results, and has become a key technical problem in the current probe card test field.

[0003] Traditional probe contact impedance monitoring methods mainly rely on fixed threshold judgments and simple mathematical models, which cannot effectively cope with complex test environments and changeable contact states. These methods have great limitations in dealing with transient faults and progressive faults, and it is difficult to accurately identify contact anomalies caused by factors such as material deformation and surface oxidation. Most existing compensation strategies use empirical parameter settings or a single feedback adjustment mechanism, lacking comprehensive consideration of probe structural characteristics and operating parameters, resulting in unsatisfactory compensation effects and prone to over-compensation or under-compensation problems. At the same time, due to the lack of intelligent parameter optimization methods, existing methods are difficult to adapt to changes in contact characteristics under different materials and different test conditions. Summary of the invention

[0004] The main purpose of the present invention is to provide a method, device and storage medium for real-time monitoring and compensation of probe contact impedance, which improves the accuracy of probe testing of chips.

[0005] To achieve the above object, the present invention provides a probe contact impedance real-time monitoring and compensation method, comprising the following steps:

[0006] A first measurement data set including a probe contact resistance value and a contact force is obtained, and input into a double-layer deep learning neural network for training to obtain a probe contact impedance monitoring model;

[0007] A second measurement data set is collected in real time, and input into the intelligent contact impedance monitoring model to perform dual-parameter fault detection to obtain a contact state classification result;

[0008] generating a two-dimensional compensation matrix based on the contact state classification result, and performing nonlinear mapping calculation on the second measurement data set to obtain a compensation coefficient;

[0009] Inputting the compensation coefficient, the original probe structure parameter and the original probe operation parameter into the graph attention stacked autoencoder for parameter prediction to obtain the target probe structure parameter and the target probe operation parameter;

[0010] Cloud model calculation and weighted operation are performed on the target probe structural parameters and the target probe operating parameters to obtain target compensation instructions, and the target compensation instructions are used to adjust probe position parameters and probe pressure parameters.

[0011] The present invention also provides a probe contact impedance real-time monitoring and compensation device, comprising:

[0012] An acquisition module is used to acquire a first measurement data set including a probe contact resistance value and a contact force, and input the data set into a double-layer deep learning neural network for training to obtain a probe contact impedance monitoring model;

[0013] A detection module, used for collecting a second measurement data set in real time, and inputting the second measurement data set into the intelligent contact impedance monitoring model for dual-parameter fault detection to obtain a contact state classification result;

[0014] A calculation module, used for generating a two-dimensional compensation matrix based on the contact state classification result, and performing nonlinear mapping calculation on the second measurement data set to obtain a compensation coefficient;

[0015] A prediction module, used for inputting the compensation coefficient, the original probe structure parameter and the original probe operation parameter into the graph attention stacked autoencoder for parameter prediction to obtain the target probe structure parameter and the target probe operation parameter;

[0016] The output module is used to perform cloud model calculation and weighted operation on the target probe structural parameters and the target probe operating parameters to obtain target compensation instructions, and the target compensation instructions are used to adjust the probe position parameters and the probe pressure parameters.

[0017] The present invention also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the above methods when executing the computer program.

[0018] The present invention also provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of any of the above-mentioned methods are implemented.

[0019] In summary, the technical solution provided by the present invention monitors the contact impedance of the probe by constructing a two-layer deep learning neural network, combines physical model encoding and material property response, significantly enhances the model's ability to characterize the contact state, adopts graph attention stacked autoencoder for parameter optimization, realizes adaptive adjustment of probe structure parameters and operation parameters, introduces cloud model and hierarchical analysis method for weight calculation, establishes a complete compensation instruction generation mechanism, designs a dual-parameter-based fault detection algorithm, and realizes accurate identification of transient faults and progressive faults through the construction of resistance-force coupling feature space; adopts adaptive grid division and nonlinear mapping technology to solve the problem of parameter space division in the compensation process and improve the degree of refinement of compensation; through multimodal feature fusion and cross-modal feature interaction, enhances the system's processing capability of different types of data, and makes the compensation strategy more comprehensive and stable. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a schematic diagram of the steps of a method for real-time monitoring and compensation of probe contact impedance in one embodiment of the present invention;

[0021] Figure 2 It is a structural block diagram of a probe contact impedance real-time monitoring and compensation device in one embodiment of the present invention;

[0022] Figure 3 It is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention.

[0023] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0025] Reference Figure 1 , this embodiment provides a probe contact impedance real-time monitoring and compensation method, comprising the following steps:

[0026] S1, obtaining a first measurement data set including a probe contact resistance value and a contact force, and inputting the data set into a two-layer deep learning neural network for training to obtain a probe contact impedance monitoring model;

[0027] Among them, a first measurement data set including the probe contact resistance value and the contact force is obtained, and the measurement data set is subjected to data stratification processing to obtain a normal contact data subset, a transient fault data subset and a progressive degradation data subset. The normal contact data subset, the transient fault data subset and the progressive degradation data subset are input into the data preprocessing unit for standardization processing to obtain standardized training data. In the process of standardization, the data is normalized to a unified scale range, eliminating the influence caused by the scale difference between different physical quantities, and improving the convergence speed and stability of the model. Feature extraction is performed on the standardized training data to construct a feature matrix including the contact resistance feature vector and the contact force feature vector, and the dimension of the feature matrix is ​​M×N, where M is the number of samples and N is the number of features. The feature matrix is ​​input into the physical model encoding layer in the two-layer deep learning neural network for encoding calculation. The physical model encoding layer includes five convolution units, each of which includes a convolution layer, a batch normalization layer and a ReLU activation function layer. The convolution kernel size of the convolution layer is 3×3, and the step size is 1. The convolution layer is used to extract local features. The output is normalized through the batch normalization layer to reduce the deviation in the neural network and improve the training speed and stability of the model. At the same time, the ReLU activation function layer is used to increase the nonlinear expression ability of the network and obtain a richer feature description. After this step, the first intermediate feature is obtained. The first intermediate feature is input into the material characteristic response layer in the two-layer deep learning neural network for characteristic analysis. This layer contains three fully connected layers and two LSTM (long short-term memory network) layers. In the fully connected layer, the number of neurons is 512, 256 and 128 respectively. Through these layers, the feature dimension is gradually reduced, while the most important information is retained, reducing the amount of calculation and the possibility of overfitting. In the material characteristic response analysis, the number of hidden units in the two LSTM layers is 64. The LSTM layer is used to capture the timing dependency during the probe contact process, which can effectively extract the dynamic features in the time series data and obtain the second intermediate feature. The second intermediate feature is subjected to feature fusion operation, which includes attention mechanism calculation and residual connection calculation. The attention mechanism is used to weight the parts of the features with important information, improve the model's attention to important features, and improve the overall feature expression ability. The residual connection calculation can alleviate the gradient vanishing problem in neural network training, ensure the effective training of deep networks, and obtain a fused feature vector. The fused feature vector is input into the bidirectional gated recurrent unit for temporal feature learning. The bidirectional gated recurrent unit contains two hidden layers, forward and reverse, with 32 units in each hidden layer. The forward and backward dependencies in the feature sequence are captured by bidirectional learning to obtain a more complete description of the temporal features and obtain a temporal feature sequence. Based on the temporal feature sequence, a loss function is constructed and back-propagation optimization is performed. The loss function contains mean square error loss term, cross entropy loss term, and regularization loss term.The mean square error loss term is used to measure the difference between the model output and the true value, the cross entropy loss term is used for classification problems to measure the classification performance of the model, and the regularization loss term is used to prevent overfitting. Through the combined effect of these loss terms, an effective loss function is constructed, and the Adam optimizer is used to update the parameters. The Adam optimizer combines the advantages of momentum and adaptive learning rate, and has a good convergence effect in the model training process, and obtains the probe contact impedance monitoring model.

[0028] S2, real-time acquisition of a second measurement data set, and input into an intelligent contact impedance monitoring model for dual-parameter fault detection to obtain a contact state classification result;

[0029] Specifically, a second measurement data set including the probe contact resistance value and the contact force is collected in real time, and the second measurement data set is segmented by a sliding window method to obtain a segmented data sequence. The segmented data sequence is subjected to wavelet denoising to obtain a denoised data sequence. Wavelet denoising is an effective signal processing method that can effectively remove high-frequency noise in the measurement process while retaining important signal features, making the data cleaner and smoother, and improving the detection accuracy of the model. The denoised data sequence is input into the probe contact impedance monitoring model, and forward calculation is performed through the physical model encoding layer and the material characteristic response layer to obtain a fault feature vector. The function of the physical model encoding layer is to extract features related to the contact physical characteristics in the data, while the material characteristic response layer performs in-depth analysis of the physical characteristics and material properties of the data, comprehensively extracts important features related to the contact process, and forms a representative fault feature vector. The fault feature vector is calculated by a dual-parameter joint probability distribution to establish a resistance-force coupling feature space. There is a certain coupling relationship between the two parameters of resistance and contact force, and the mutual influence between them is described by a joint probability distribution. After the feature space is established, the distance between the sample point and the standard mode is calculated by the Mahalanobis distance metric to obtain the abnormal measurement value. The Mahalanobis distance is a distance measurement method that can simultaneously consider the correlation between multiple variables. The Mahalanobis distance is used to effectively evaluate the difference between the current measurement data and the standard mode and measure the degree of abnormality of the sample. The larger the abnormal measurement value, the more obvious the difference between the sample and the normal state. Based on the abnormal measurement value, a fault detection function based on a sliding time window is constructed. The fault detection function uses the exponential weighted average method to calculate the detection statistic and obtain the fault detection value. The exponential weighted average method is a method for smoothing time series data. By assigning different weights to historical data, the detection statistic can better reflect the trend and change of the current state, which helps to detect fault signs in time. The fault detection value is compared with the preset classification threshold to determine the category of the current contact state. The preset classification threshold includes the normal state threshold and the fault state threshold, which are used to divide different state categories. In order to improve the accuracy of fault detection, a three-level decision rule is used to divide the state and obtain the contact state classification result. The three-level decision rule is a more robust decision strategy that can ensure the accuracy and reliability of fault detection through multi-level threshold settings and decision conditions. When the fault detection value is lower than the normal state threshold, it is judged as a normal state; when the detection value is higher than the fault state threshold, it is judged as a fault state; and for the detection value between the two, further judgment is made to determine whether it belongs to a potential fault or a normal state, making the state classification more refined and reliable.

[0030] S3, generating a two-dimensional compensation matrix based on the contact state classification result, and performing nonlinear mapping calculation on the second measurement data set to obtain a compensation coefficient;

[0031] It should be noted that the contact state classification results are state-encoded, the normal state is assigned to 1, the minor fault state is assigned to 2, and the serious fault state is assigned to 3, and the state encoding matrix is ​​obtained. The state encoding matrix can effectively characterize the different states in the probe contact process, so that the subsequent compensation process can perform appropriate compensation operations according to different states. Based on the state encoding matrix, a compensation reference value lookup table is constructed, and the compensation reference value lookup table contains a resistance compensation reference value and a force compensation reference value to obtain compensation reference data. The compensation reference data is input into the kernel function mapping unit, and the radial basis function is used to perform nonlinear feature transformation on the data to obtain a transformation feature space. The radial basis function can map the original data to a high-dimensional space, in which the linear separability of the data is improved. The transformation feature space is adaptively meshed, and 10 uniformly distributed mesh nodes are set along the resistance dimension and the force dimension respectively to obtain a two-dimensional mesh structure. Through adaptive meshing, the relationship between resistance and force can be described in a more detailed manner, making the compensation operation in different states more refined and controllable. The setting of the mesh nodes enables the compensation calculation to have a high resolution in both the resistance and force dimensions. According to the two-dimensional grid structure, the probe contact resistance value and contact force in the second measurement data set are bilinearly interpolated to obtain the initial compensation coefficient. Bilinear interpolation interpolates the two-dimensional grid data of resistance and force to obtain the compensation value between grid nodes, ensuring the smoothness and continuity of the compensation coefficient in the entire compensation area. The initial compensation coefficient is time-series smoothed to obtain a smoothed compensation coefficient. The compensation coefficient is smoothed in the time dimension to reduce the drastic changes in the compensation coefficient caused by measurement noise or instantaneous fluctuations of the system and improve the stability of the compensation effect. The smoothed compensation coefficient is multiplied by the state weight factor to obtain a weighted compensation coefficient. The state weight factor is used to weight the compensation coefficient under different states, dynamically adjust the compensation operation according to different fault levels, ensure greater compensation strength under severe fault conditions, and minimize unnecessary adjustments under normal conditions. The weighted compensation coefficient is subjected to boundary constraint processing, and the upper and lower limits of the compensation are set. The saturation function is used to perform a limiting operation to obtain the final compensation coefficient. The boundary constraint processing of the compensation coefficient is to prevent the compensation coefficient from being too large or too small, thereby adversely affecting the system. The compensation coefficient is limited by a saturation function to ensure that the compensation value does not exceed a preset safety range, so that the compensation process does not threaten the stability of the system while maintaining its effectiveness.

[0032] The compensation benchmark data is normalized, and the resistance value and force value are mapped to the interval [-1,1] respectively to obtain the normalized compensation data. Through normalization, the numerical scales between different physical quantities are consistent, reducing the error caused by the difference in numerical range during model training. The normalized compensation data is input into the radial basis kernel function for nonlinear transformation, where the radial basis kernel function adopts the form of Gaussian kernel function, and the kernel width parameter is set to 0.5 to obtain the initial feature map. The Gaussian kernel function is a radial basis function that can map the input data to a higher-dimensional feature space and enhance the separability of the data. The setting of the kernel width parameter directly affects the smoothness and generalization ability of the feature transformation. The initial feature map is subjected to singular value decomposition operation to extract the main feature directions, and select the feature vectors with a cumulative contribution rate of more than 95% to obtain the reduced dimension feature space. Singular value decomposition is an effective dimensionality reduction technology that extracts the most important feature directions, removes redundant information, reduces computational complexity and improves data processing efficiency. The reduced-dimensional feature space is divided into two orthogonal subspaces, resistance dimension and force dimension, and the probability density function of data distribution is calculated in each subspace to obtain the edge distribution characteristics. The edge distribution characteristics are used to describe the independent distribution of data in the resistance and force dimensions, which helps to determine the node positions more reasonably in the feature space. Based on the edge distribution characteristics, the optimal distribution position of the grid nodes is determined, and the K-means clustering algorithm is used to cluster 10 nodes in each dimension to obtain the adaptive node positions. The K-means clustering algorithm is used to divide the data set into several nodes with similar characteristics, so that the distribution of grid nodes in the feature space can better reflect the actual distribution of the data and improve the accuracy of the grid processing. Based on the adaptive node position, Delaunay triangulation is constructed to grid the feature space. Delaunay triangulation is a gridding method that divides the feature space into several grid cells through triangulation. Each grid cell records the corresponding local feature statistics to obtain the initial grid structure. The local feature statistics contain the relationship information between nodes and the feature distribution in the area. In order to optimize the mesh structure, the mesh quality of the initial mesh structure is evaluated, and the shape factor and size factor of each mesh unit are calculated. The purpose of mesh quality evaluation is to ensure that the geometric shape and size of the mesh unit are within a reasonable range to facilitate the accuracy and numerical stability of subsequent calculations. For mesh units that do not meet the quality requirements, local encryption or coarsening is performed to obtain an optimized mesh structure. Local encryption refers to adding mesh nodes to refine the mesh unit, while coarsening refers to reducing mesh nodes to expand the size of the mesh unit. The optimized mesh structure is mapped and transformed with the original coordinate system, and the corresponding relationship between the mesh nodes and the physical quantity is established, and finally a two-dimensional mesh structure is obtained.By linking the optimized grid with the physical coordinate system, the mapping between the grid nodes and the probe contact resistance and contact force is realized, making the physical meaning of each grid node clearer.

[0033] S4, inputting the compensation coefficient, the original probe structure parameter and the original probe operation parameter into the graph attention stacked autoencoder for parameter prediction to obtain the target probe structure parameter and the target probe operation parameter;

[0034] Specifically, the compensation coefficient, the original probe structure parameter and the original probe operation parameter are parameter fused to construct a multimodal feature vector containing compensation information, structure information and operation information. The multimodal feature vector is input into the first attention layer of the graph attention stacked autoencoder for feature association calculation. In the first attention layer, a multi-head self-attention mechanism is adopted, in which the number of heads is set to 8 and the attention dimension is 64. The role of the multi-head self-attention mechanism is to capture the diversity and correlation between features through multiple different attention heads, and to weight the features from different angles to obtain attention weighted features. The attention weighted features are subjected to the first encoding operation. The first encoding operation includes three graph convolution layers, the number of output channels is set to 128, 256 and 512 respectively, and each graph convolution layer is followed by a LeakyReLU activation function. The graph convolution layer can effectively capture the local features and topological relationships in the graph structure data by aggregating the nodes and their neighborhoods, and the LeakyReLU activation function can introduce nonlinearity and improve the expression ability of the model. Through multi-layer graph convolution operations, the first encoding feature is obtained, which contains the spatial and structural information of the multimodal feature vector. The first encoded feature is input into the second attention layer for cross-modal feature interaction. The second attention layer contains a channel attention module and a spatial attention module, which respectively calculate the channel importance and spatial importance weights of the feature. The channel attention module is used to calculate the importance of different feature channels and improve the attention to key feature channels, while the spatial attention module is used to capture the differences and importance of features in spatial positions. Through the joint action of these two modules, multi-dimensional attention features are obtained. The second encoding operation is performed on the multi-dimensional attention features. The second encoding operation includes two fully connected layers with 1024 and 512 neurons respectively. Each fully connected layer is followed by a batch normalization layer and a ReLU activation function. The fully connected layer is used to extract global features. The batch normalization layer normalizes each batch of data to reduce internal covariate shift and improve the training stability and convergence speed of the model. The ReLU activation function provides the model with nonlinear mapping capabilities to ensure the complex expression of features. Through the operation, the encoded features are obtained. The encoded features are input into the decoder for decoding operation. The decoder consists of three transposed convolutional layers with a kernel size of 3×3, a stride of 2, and output channels of 256, 128, and 64, respectively. The transposed convolutional layer is a deconvolution operation that is used to upsample features and restore the spatial resolution of the features. In this process, the original dimension of the features is gradually restored by gradually increasing the number of output channels to ensure that the decoded features can contain enough information for parameter prediction, and finally the decoded features are obtained. The decoded features are subjected to parameter separation operations, and the target probe structure parameters and target probe operation parameters are predicted respectively through two parallel fully connected layers. The information in the decoded features is mapped to two different output spaces of target structure parameters and operation parameters, respectively, to obtain the final prediction results.

[0035] S5, performing cloud model calculation and weighting operation on the target probe structure parameters and the target probe operation parameters to obtain a target compensation instruction, which is used to adjust the probe position parameters and the probe pressure parameters.

[0036] Among them, the target probe structure parameters and target probe operation parameters are input into the cloud model generator respectively, and the expected value, entropy value and super entropy value are generated based on the normal distribution function to obtain the structure parameter cloud and the operation parameter cloud. The normal distribution function is used to characterize the distribution characteristics of the parameters. The expected value represents the central tendency of the parameters, the entropy value represents the measure of uncertainty, and the super entropy value reflects the stability and randomness of the cloud model. Through the generation of these parameters, a cloud model that can effectively reflect the uncertainty of the probe parameters is established. The cloud droplet generation calculation is performed on the structure parameter cloud and the operation parameter cloud. 200 cloud droplets are generated for each parameter, and the cloud droplets are mapped to the parameter space through the inverse cloud transformation to obtain the cloud droplet feature sequence. Cloud droplets are the basic elements in the cloud model. The inverse cloud transformation is used to reflect the distribution characteristics of the structure and operation parameters in the parameter space. Generating multiple cloud droplets and performing inverse cloud transformation can effectively represent the randomness and fuzziness of the probe parameters in the parameter space. The fuzzy membership matrix is ​​constructed based on the cloud droplet feature sequence. In order to calculate the membership value of each parameter, the triangular membership function is used to represent the membership relationship between the cloud droplet and the specific fuzzy set, and the fuzzy evaluation matrix is ​​obtained. The fuzzy evaluation matrix is ​​input into the hierarchical analysis system, and the judgment matrix is ​​constructed by the pairwise comparison method to calculate the weight vector of the parameters. The parameter weight vector is checked for consistency to ensure the rationality of the judgment matrix. During the consistency check process, the consistency ratio (CR) is calculated. If the consistency ratio is less than the preset threshold, the judgment matrix is ​​considered to have good consistency, so that the hierarchical single ranking and hierarchical total ranking are further calculated. Through the hierarchical single ranking and hierarchical total ranking, the comprehensive weight coefficient of each parameter is obtained. The comprehensive weight coefficient reflects the relative importance of different probe parameters in the entire control system. According to the comprehensive weight coefficient, the target probe structure parameters and the target probe operation parameters are converted into an execution control instruction sequence. By assigning corresponding weights to different parameters, it is ensured that the generation of control instructions can be dynamically adjusted according to the importance of different parameters to achieve the optimal probe position and pressure regulation. These instructions are formatted through the control protocol, and the control instructions are converted into a standard format that conforms to the communication protocol of the probe device to obtain the target compensation instructions.

[0037] In one example, a first measurement data set including a probe contact resistance value and a contact force is obtained and input into a two-layer deep learning neural network for training to obtain a probe contact impedance monitoring model, including:

[0038] Acquire a first measurement data set including probe contact resistance value and contact force, and perform data stratification processing on the first measurement data set to obtain a normal contact data subset, a transient fault data subset, and a progressive degradation data subset;

[0039] The normal contact data subset, the instantaneous fault data subset and the progressive degradation data subset are input into the data preprocessing unit for standardization processing to obtain standardized training data, and feature extraction is performed on the standardized training data to construct a feature matrix containing the contact resistance feature vector and the contact force feature vector. The dimension of the feature matrix is ​​M×N, where M is the number of samples and N is the number of features.

[0040] The feature matrix is ​​input into the physical model encoding layer in the two-layer deep learning neural network for encoding calculation. The physical model encoding layer contains five convolution units, each of which includes a convolution layer, a batch normalization layer, and a ReLU activation function layer. The convolution kernel size of the convolution layer is 3×3, and the step size is 1, and the first intermediate feature is obtained;

[0041] The first intermediate feature is input into the material characteristic response layer in the two-layer deep learning neural network for characteristic analysis. The material characteristic response layer includes three fully connected layers and two LSTM layers. The number of neurons in the fully connected layers is 512, 256, and 128, respectively. The number of hidden units in the LSTM layer is 64, and the second intermediate feature is obtained.

[0042] Perform feature fusion operation on the second intermediate feature, the feature fusion operation includes attention mechanism calculation and residual connection calculation, and obtain a fused feature vector;

[0043] The fused feature vector is input into the bidirectional gated recurrent unit for time series feature learning. The bidirectional gated recurrent unit contains two hidden layers, the forward layer and the reverse layer. The number of units in each hidden layer is 32, and the time series feature sequence is obtained.

[0044] A loss function is constructed based on the time series feature sequence and back-propagation optimization is performed. The loss function includes mean square error loss term, cross entropy loss term and regularization loss term. The parameters are updated through the Adam optimizer to obtain the probe contact impedance monitoring model.

[0045] In this example, the contact resistance value and contact force data in the actual measurement are collected to obtain the first measurement data set. The first measurement data set is subjected to data stratification processing and divided into three subsets, namely the normal contact data subset, the instantaneous fault data subset and the progressive degradation data subset. The normal contact data subset represents the situation where everything is running normally during the probe contact process, the instantaneous fault data subset corresponds to the short-term fault phenomenon during the probe contact process, such as poor contact or instantaneous disconnection, and the progressive degradation data subset describes the situation where the probe contact state gradually deteriorates, such as the gradual increase of contact resistance due to wear and other reasons. The normal contact data, instantaneous fault data and progressive degradation data are input into the data preprocessing unit for standardization processing to obtain standardized training data. The data of different physical quantities are unified within the same scale range to reduce the errors caused by different dimensions. The standardized training data is feature extracted, the contact resistance and contact force data are converted into numerical values ​​with characteristic significance, and a feature matrix containing the contact resistance feature vector and the contact force feature vector is constructed. The dimension of the feature matrix is ​​M×N, where M represents the number of samples and N represents the number of features. The feature matrix is ​​input into the physical model encoding layer in the two-layer deep learning neural network for encoding calculation. The physical model encoding layer contains five convolution units, each of which consists of a convolution layer, a batch normalization layer, and a ReLU activation function layer. The convolution kernel size of the convolution layer is 3×3, and the step size is 1. The convolution layer extracts the local spatial features of the data and identifies the patterns and regularities in the data by continuously performing convolution operations on the input data. The batch normalization layer is used to normalize the convolution output to reduce the deviation in the training process and speed up the convergence of the model. The ReLU activation function is used to introduce nonlinearity so that the model can express more complex relationships. Through the operation of the convolution unit, the first intermediate feature is obtained. The first intermediate feature is input into the material property response layer in the two-layer deep learning neural network for feature analysis. The material property response layer contains three fully connected layers and two LSTM layers (long short-term memory network). The number of neurons in the fully connected layers is 512, 256, and 128, respectively. Through these fully connected layers, the features are gradually abstracted at a high level, retaining the most important information and removing redundancy. The number of hidden units in the LSTM layer is 64. LSTM is a special recurrent neural network that can effectively capture the temporal dependencies in data sequences and is suitable for processing time series data or time-related data. Through LSTM processing, the second intermediate feature is obtained, which contains the time dependency information during the probe contact process. The second intermediate feature is subjected to feature fusion operation, which includes attention mechanism calculation and residual connection calculation. The attention mechanism is used to weight the important parts of the feature, increase the attention to key information, and increase the effectiveness of feature representation. For example, assuming that the weight calculation formula of the attention mechanism is:

[0046] ;

[0047] in, Indicates The attention weight of each feature, is the original value of the feature, and the importance weight of each feature is calculated by the softmax function. The residual connection calculation is used to avoid the gradient vanishing problem after the model is deepened, to ensure the effective propagation of information in the network, and to obtain the fused feature vector. The fused feature vector is input into the bidirectional gated recurrent unit for time series feature learning. The bidirectional gated recurrent unit contains two hidden layers, forward and reverse, and the number of units in each hidden layer is 32. The bidirectional gated recurrent unit is an improved recurrent neural network similar to LSTM, which can effectively capture long-term dependencies. Through forward and reverse learning, it can capture the global information in the feature sequence and obtain the time series feature sequence. Based on the time series feature sequence, a loss function is constructed and back propagation optimization is performed. The construction of the loss function includes the mean square error loss term, the cross entropy loss term, and the regularization loss term. The mean square error loss term is used to measure the difference between the model output and the actual value, and its expression is:

[0048] ;

[0049] in, Indicates the actual value, is the predicted value, is the number of samples. The cross entropy loss term is used to measure the performance of the model in the classification task, and the regularization loss term is used to prevent the overfitting of the model. These loss terms are combined to form the total loss function, and the Adam optimizer is used to update the model parameters. The Adam optimizer combines the advantages of momentum and adaptive learning rate to update parameters quickly and stably, so that the model can converge efficiently during the training process. Through the back propagation algorithm, the gradient of the loss function is used to optimize the parameters of the model to obtain the final probe contact impedance monitoring model. This monitoring model can effectively monitor the contact impedance of the probe in real time, and provide a reliable basis for judgment by classifying different states.

[0050] In one example, a second measurement data set is collected in real time and input into an intelligent contact impedance monitoring model for dual-parameter fault detection to obtain a contact state classification result, including:

[0051] collecting a second measurement data set including a probe contact resistance value and a contact force in real time, and performing sliding window segmentation on the second measurement data set to obtain a segmented data sequence;

[0052] Perform wavelet denoising on the segmented data sequence to obtain a denoised data sequence, and input the denoised data sequence into the probe contact impedance monitoring model, perform forward calculation through the physical model encoding layer and the material characteristic response layer to obtain the fault feature vector;

[0053] The fault feature vector is subjected to dual-parameter joint probability distribution calculation, and the resistance-force coupling feature space is established. The distance between the sample point and the standard mode is calculated by the Mahalanobis distance metric to obtain the abnormal measurement value.

[0054] According to the abnormal measurement value, a fault detection function based on a sliding time window is constructed. The fault detection function uses the exponential weighted average method to calculate the detection statistics and obtain the fault detection value.

[0055] The fault detection value is compared with the preset classification threshold, which includes a normal state threshold and a fault state threshold. The state is divided through a three-level decision rule to obtain a contact state classification result.

[0056] In this example, a second measurement data set containing the probe contact resistance value and contact force is collected in real time, and the second measurement data set is segmented by a sliding window to obtain a segmented data sequence. The continuous time series data is divided into multiple small time segments, each segment contains a certain number of samples, so as to capture the short-term dynamic changes of the data. Assume that the length of the sliding window is , the sliding step length is , then the measured data Divide into several time segments , each time segment . Perform wavelet denoising on the segmented data sequence. Wavelet denoising is a signal processing method that decomposes the signal into sub-bands of different frequencies, effectively removing high-frequency noise while retaining useful information in the signal. Assume that the original data sequence is , after wavelet transform, we get several coefficients of different frequencies Then, by removing the high-frequency part and performing inverse wavelet transform, the denoised data sequence is obtained. The denoised data sequence is input into the probe contact impedance monitoring model, which contains a physical model encoding layer and a material characteristic response layer. The spatial characteristics of contact resistance and contact force are extracted through the convolution operation of the physical model encoding layer, and then the characteristics are analyzed through the material characteristic response layer to obtain the fault feature vector. The fault feature vector is subjected to a two-parameter joint probability distribution calculation to establish a resistance-force coupling feature space. In this process, the relationship between resistance value and contact force is described by a joint probability distribution, assuming that the resistance value is , the contact force is , then its joint probability density function is expressed as The coupling relationship between resistance and force is characterized by joint distribution. In order to evaluate the difference between the current sample and the standard mode, the Mahalanobis distance is used to measure the distance between the sample point and the standard mode. The definition of Mahalanobis distance is as follows:

[0057] ;

[0058] in, represents the Mahalanobis distance, is the feature vector of the current sample, is the mean vector of the standard mode, is the covariance matrix. The Mahalanobis distance takes into account the mean and covariance of the sample when calculating, which can effectively measure the differences between different features. By calculating the Mahalanobis distance, the abnormality measurement value of each sample relative to the standard mode is obtained. , the larger the anomaly metric value, the greater the deviation of the sample from the normal state. Based on the anomaly metric value, a fault detection function based on a sliding time window is constructed. The fault detection function uses the exponential weighted average method to calculate the detection statistic to capture the trend of abnormal changes. Assume that at time The detection value is , the exponentially weighted average is expressed as:

[0059] ;

[0060] in, Indicates at time The test statistic, is the smoothing coefficient, and its value range is is the current abnormality metric value, is the detection statistic of the previous moment. By setting an appropriate smoothing coefficient, the detection statistic is made more sensitive to new abnormal changes, while avoiding over-response to short-term noise. The calculated fault detection value effectively reflects the trend change in the probe contact process. The fault detection value is compared with the preset classification threshold. The preset classification threshold includes the normal state threshold and the fault state threshold, and the contact state of the probe is divided into different categories by these thresholds. Assume that the threshold of the normal state is , the threshold of the fault state is , then when the fault detection value When , it is judged as normal state; when When , it is judged as a minor fault state; and when When the probe contact state is , it is judged as a serious fault state. The method of state classification by threshold comparison is a three-level judgment rule to ensure accurate classification of the probe contact state.

[0061] In one example, a two-dimensional compensation matrix is ​​generated based on the contact state classification result, and a nonlinear mapping calculation is performed on the second measurement data set to obtain a compensation coefficient, including:

[0062] The contact state classification results are coded, the normal state is assigned to 1, the minor fault state is assigned to 2, and the serious fault state is assigned to 3, and the state coding matrix is ​​obtained;

[0063] Constructing a compensation reference value lookup table according to the state encoding matrix, wherein the compensation reference value lookup table includes a resistance compensation reference value and a force compensation reference value, and obtaining compensation reference data;

[0064] The compensation reference data is input into the kernel function mapping unit, and the radial basis function is used to perform nonlinear feature transformation to obtain the transformation feature space, and the transformation feature space is adaptively meshed, and 10 evenly distributed mesh nodes are set along the resistance dimension and the force dimension respectively to obtain a two-dimensional mesh structure;

[0065] Performing bilinear interpolation calculation on the probe contact resistance value and the contact force in the second measurement data set according to the two-dimensional grid structure to obtain an initial compensation coefficient;

[0066] Performing time series smoothing processing on the initial compensation coefficient to obtain a smoothed compensation coefficient, and performing a product operation on the smoothed compensation coefficient and the state weight factor to obtain a weighted compensation coefficient;

[0067] The weighted compensation coefficient is subjected to boundary constraint processing, the upper and lower limits of compensation are set, and the compensation coefficient is obtained by limiting the amplitude through a saturation function.

[0068] In this example, the contact state classification results are state-encoded. The contact state of the probe is divided into normal state, slight fault state and severe fault state, and they are assigned values ​​of 1, 2 and 3 respectively to obtain a state encoding matrix. Different states are numerically processed so that the subsequent calculation and compensation process is more systematic and automated. A compensation reference value lookup table is constructed according to the state encoding matrix, and the lookup table contains resistance compensation reference values ​​and force compensation reference values ​​to obtain compensation reference data. The compensation reference value lookup table determines the corresponding compensation value according to different state encodings. The generation of compensation reference data is to effectively compensate and adjust the resistance and force of the probe under different states. For example, when the probe is in a normal state, the compensation reference value is a smaller fine-tuning value, while when the probe is in a severe fault state, a larger compensation value is required for effective adjustment. The compensation reference data is input into the kernel function mapping unit, and the radial basis function is used for nonlinear feature transformation to obtain the transformed feature space. The role of the radial basis function is to map the original data to a higher-dimensional feature space to enhance the nonlinear separability of the data, so that the compensation reference data can be more effectively fitted in the new space. Assume that the input compensation reference data is , the radial basis function takes the form of a Gaussian kernel, which is defined as:

[0069] ;

[0070] in, represents the features after kernel function transformation, The core center, is the kernel width parameter. Through the nonlinear transformation of the Gaussian kernel function, the originally indistinguishable data becomes more linearly separable in the feature space. The transformed feature space is adaptively gridded, and 10 evenly distributed grid nodes are set along the resistance dimension and the force dimension respectively to obtain a two-dimensional grid structure. These grid nodes are evenly distributed in the resistance and force dimensions, so that the grid division can effectively cover the entire feature space. By refining the different intervals of resistance and force, more accurate compensation can be performed in different intervals. Based on the two-dimensional grid structure, the probe contact resistance value and contact force in the second measurement data set are bilinearly interpolated to obtain the initial compensation coefficient. Bilinear interpolation is a method of interpolating two-dimensional data of resistance and force, and its formula is as follows:

[0071] ;

[0072] in, is the interpolation result, are the function values ​​of the grid nodes, is the weight of the interpolation position in two dimensions. Through bilinear interpolation, the values ​​between different grid nodes are smoothly transitioned to obtain the initial compensation coefficient in the entire feature space. The initial compensation coefficient is time-series smoothed to obtain a smooth compensation coefficient, reduce the fluctuation of the compensation coefficient caused by noise or instantaneous changes, and improve the stability of compensation. Time-series smoothing is achieved through exponentially weighted moving average, and its formula is:

[0073] ;

[0074] in, For time The smoothing compensation coefficient is For time The initial compensation coefficient, is a smoothing coefficient, and its value is between (0,1). Through smoothing, the drastic changes of the compensation coefficient in the time dimension are effectively reduced, ensuring the stability of the probe during operation. The smoothed compensation coefficient is multiplied by the state weight factor to obtain a weighted compensation coefficient. The compensation strength is adjusted according to the contact state of the probe. For example, in a severe fault state, a larger weight factor is required to increase the compensation strength, while in a normal state, the weight factor is relatively small to maintain subtle adjustments to the probe contact. The weighted compensation coefficient can more flexibly respond to needs under different states. The weighted compensation coefficient is subjected to boundary constraints, and the upper and lower limits of the compensation are set. The saturation function is used to limit the compensation coefficient to ensure that the compensation coefficient is within a reasonable range. The saturation function is defined as:

[0075] ;

[0076] in, is the final compensation coefficient, and are the upper and lower limits of compensation respectively, is the weighted compensation coefficient. Through the limiting operation, the compensation coefficient is prevented from being too large or too small, ensuring that the compensation is controlled within a reasonable range, thereby avoiding adverse effects on the probe system.

[0077] In one example, the compensation reference data is input into the kernel function mapping unit, and the radial basis function is used to perform nonlinear feature transformation to obtain the transformation feature space, and the transformation feature space is adaptively meshed, and 10 uniformly distributed mesh nodes are set along the resistance dimension and the force dimension respectively to obtain a two-dimensional mesh structure, including:

[0078] Normalize the compensation benchmark data, map the resistance value and force value to the interval [-1,1] respectively, and obtain the normalized compensation data;

[0079] The normalized compensation data is input into the radial basis kernel function for nonlinear transformation. The radial basis kernel function adopts the Gaussian kernel function form, and the kernel width parameter is set to 0.5 to obtain the initial feature map;

[0080] Perform singular value decomposition on the initial feature map, extract the main feature directions, select feature vectors with cumulative contribution rates exceeding 95%, and obtain the reduced-dimensional feature space;

[0081] The reduced-dimensional feature space is divided into two orthogonal subspaces: resistance dimension and force dimension. The probability density function of data distribution is calculated in each subspace to obtain the marginal distribution characteristics.

[0082] The optimal distribution position of the grid nodes is determined according to the edge distribution characteristics, and the K-means clustering algorithm is used to cluster 10 nodes in each dimension to obtain the adaptive node position;

[0083] Delaunay triangulation is constructed based on the adaptive node positions, and the feature space is gridded. Each grid unit records the corresponding local feature statistics to obtain the initial grid structure.

[0084] Perform mesh quality assessment on the initial mesh structure, calculate the shape factor and size factor of each mesh unit, and perform local encryption or coarsening on the mesh that does not meet the quality requirements to obtain an optimized mesh structure;

[0085] The optimized grid structure is mapped and transformed with the original coordinate system, the corresponding relationship between the grid nodes and the physical quantities is established, and a two-dimensional grid structure is obtained.

[0086] In this example, the compensation benchmark data is normalized. The resistance value and force value are mapped to the interval [-1,1] to obtain the normalized compensation data. Assume that the original resistance value and force value are and , the normalized value is obtained by the following formula:

[0087] ;

[0088] ;

[0089] in, and are the normalized resistance and force values, are the minimum and maximum values ​​of resistance and force, respectively. Through linear transformation, the original data is mapped to the interval [-1,1] to ensure that physical quantities of different dimensions can be subsequently calculated on the same scale. The normalized compensation data is input into the radial basis kernel function for nonlinear feature transformation. The radial basis kernel function adopts the form of Gaussian kernel function, and the kernel width parameter is set to 0.5. Gaussian kernel function is a radial basis function, which is defined as follows:

[0090] ;

[0091] in, is the feature map obtained by the Gaussian kernel function, To normalize the compensation data, The core center, is the kernel width parameter, . Through the nonlinear transformation of the radial basis kernel function, the original data is mapped to a high-dimensional feature space, which enhances the linear separability of the data and makes the fitting of classification and compensation in the feature space more accurate. The initial feature map is subjected to singular value decomposition to extract the main feature directions. Singular value decomposition is a matrix decomposition method that decomposes the feature matrix into several eigenvectors and singular values. By selecting eigenvectors with a cumulative contribution rate of more than 95%, the dimension of the feature space is effectively reduced to obtain a reduced-dimensional feature space. Assume that the initial feature mapping matrix is , its singular value decomposition is expressed as:

[0092] ;

[0093] in, and is an orthogonal matrix, is a diagonal matrix whose diagonal elements are singular values. By selecting the eigenvectors corresponding to the singular values ​​with a cumulative contribution rate of more than 95% in the diagonal matrix, the reduced-dimensional feature space is obtained, thereby reducing the computational complexity and retaining the most important information in the original data. The reduced-dimensional feature space is divided into two orthogonal subspaces, the resistance dimension and the force dimension, and the probability density function of the data distribution is calculated in each subspace to obtain the edge distribution characteristics. The edge distribution characteristics are used to describe the independent distribution of resistance and force in their respective dimensions. Based on the edge distribution characteristics, the optimal distribution position of the grid nodes is determined. The K-means clustering algorithm is used to cluster 10 nodes in each dimension to obtain the adaptive node position. K-means clustering is a commonly used clustering algorithm. By minimizing the square error within each cluster, the clustering center of the data in the feature space is effectively determined, so that the position of the grid node is more reasonable and effective. Based on the adaptive node position, the Delaunay triangulation is constructed to grid the feature space. Delaunay triangulation is a geometric method for constructing meshes. It divides the feature space into several triangular units so that the circumscribed circle of any triangle does not contain other nodes, so that the mesh division has good geometric properties. Each mesh unit records the corresponding local feature statistics to obtain the initial mesh structure. The local statistics are used to calculate the subsequent compensation coefficients to ensure that the compensation process can be fine-tuned based on the local features. The mesh quality of the initial mesh structure is evaluated, and the shape factor and size factor of each mesh unit are calculated. The shape factor is used to measure whether the shape of the triangular mesh unit is close to the ideal equilateral triangle, and the size factor is used to evaluate whether the area of ​​the mesh unit meets the expected size range. If the shape factor and size factor of some mesh units do not meet the quality requirements, these mesh units are locally encrypted or coarsened to obtain an optimized mesh structure. Local encryption refers to refining the mesh by adding mesh nodes, while coarsening refers to reducing mesh nodes to expand the area of ​​the mesh, so that the entire mesh structure is more uniform and reasonable, meeting the quality standards. The optimized mesh structure is mapped and transformed with the original coordinate system, and the corresponding relationship between the mesh nodes and the physical quantity is established to obtain the final two-dimensional mesh structure. By mapping the grid structure back to the original physical coordinate system, the physical meaning of each grid node is ensured to be clear, providing a basis for the execution of compensation. For example, during the compensation process, the resistance and force of the probe contact are adjusted according to the position of each grid node and its corresponding compensation value to achieve accurate compensation.

[0094] In one example, the compensation coefficient, the original probe structure parameter, and the original probe operation parameter are input into the graph attention stacked autoencoder for parameter prediction to obtain the target probe structure parameter and the target probe operation parameter, including:

[0095] Perform parameter fusion on the compensation coefficient, the original probe structure parameter and the original probe operation parameter to construct a multimodal feature vector containing compensation information, structure information and operation information;

[0096] The multimodal feature vector is input into the first attention layer of the graph attention stacking autoencoder for feature association calculation. The first attention layer adopts a multi-head self-attention mechanism with the number of heads set to 8 and the attention dimension set to 64 to obtain the attention weighted feature.

[0097] Perform the first encoding operation on the attention weighted feature. The first encoding operation includes three graph convolution layers. The number of output channels of each graph convolution layer is 128, 256, and 512 respectively. The activation function is LeakyReLU to obtain the first encoding feature.

[0098] The first encoded feature is input into the second attention layer for cross-modal feature interaction. The second attention layer includes a channel attention module and a spatial attention module, which respectively calculate the channel importance and spatial importance weights of the feature to obtain a multi-dimensional attention feature.

[0099] Perform the second encoding operation on the multi-dimensional attention features. The second encoding operation includes two fully connected layers with 1024 and 512 neurons respectively. Each layer is followed by a batch normalization layer and a ReLU activation function to obtain the encoded features.

[0100] The encoded features are input into the decoder for decoding operation. The decoder includes three transposed convolutional layers, the convolution kernel size is 3×3, the step size is 2, and the number of output channels is 256, 128, and 64 respectively, to obtain the decoded features;

[0101] A parameter separation operation is performed on the decoded features, and the target probe structure parameters and target probe operation parameters are predicted respectively through two parallel fully connected layers.

[0102] In this example, the compensation coefficient, the original probe structure parameters, and the original probe operation parameters are parameter fused. The compensation coefficient, the probe structure parameters, and the operation parameters are combined to construct a multimodal feature vector containing compensation information, structure information, and operation information. The multimodal feature vector is input to the first attention layer of the graph attention stacked autoencoder for feature association calculation. The first attention layer adopts a multi-head self-attention mechanism, in which the number of attention heads is set to 8 and the attention dimension is 64. The multi-head self-attention mechanism calculates the association of features from different angles through multiple different attention heads to capture the complex correlation between features. The calculation of the attention mechanism is expressed by the following formula:

[0103] ;

[0104] in, , and denote the query matrix, key matrix and value matrix respectively, is the dimension of the key matrix. In the multi-head attention mechanism, the input feature vector is divided into multiple parts, each part is calculated by an independent attention mechanism, and then these parts are spliced ​​to obtain the attention-weighted features. The first encoding operation is performed on the attention-weighted features. The first encoding operation includes three graph convolutional layers, and the number of output channels of each graph convolutional layer is 128, 256, and 512, respectively. The function of the graph convolutional layer is to extract and aggregate features through the relationship between nodes and their neighborhoods. The purpose of the convolution operation is to mine local feature patterns in the data, and by stacking multiple layers of convolution, gradually learn more complex and abstract features. Each graph convolutional layer is followed by a LeakyReLU activation function, whose expression is:

[0105] ;

[0106] in, is the activation function output, is the output of the convolutional layer, It is usually a small positive value (such as 0.01). The introduction of LeakyReLU effectively solves the problem of neuron "death" caused by the traditional ReLU activation function. Through the above-mentioned graph convolution operation and the application of the activation function, the first encoded feature is obtained, which contains the spatial correlation characteristics in the fused multimodal information. The first encoded feature is input into the second attention layer for cross-modal feature interaction. The second attention layer contains a channel attention module and a spatial attention module, which are used to calculate the channel importance and spatial importance weights of the features respectively. The channel attention module determines which feature channels are more important to the model by calculating the weight coefficient of each channel, thereby enhancing these channels; while the spatial attention module is used to capture the importance of the feature in the spatial position, and enhances the features of the key positions by calculating the weight of each position. Through the combination of channel and spatial attention, a multidimensional attention feature is obtained. The second encoding operation is performed on the multidimensional attention feature. The second encoding operation includes two fully connected layers, with the number of neurons being 1024 and 512 respectively, and each fully connected layer is followed by a batch normalization layer and a ReLU activation function. The function of the fully connected layer is to extract global features from the input features. Through the fully connected operation, the global information representation is captured from the features. The batch normalization layer is used to normalize the data of each batch to reduce the internal covariate shift and improve the training speed and stability of the model. The ReLU activation function is used to introduce nonlinearity so that the model has sufficient expressive power. After being processed by two layers of fully connected layers, the encoded features are obtained. The encoded features are input into the decoder for decoding operation. The decoder includes three transposed convolution layers with a convolution kernel size of 3×3, a stride of 2, and output channels of 256, 128, and 64, respectively. Transposed convolution is a deconvolution operation used to upsample the features to restore them to a higher resolution. During the decoding process, by gradually increasing the number of output channels, the decoder can map the encoded features back to a spatial representation close to the original input to generate decoded features. Parameter separation operation is performed on the decoded features. Through two parallel fully connected layers, the structural parameters and operational parameters of the target probe are predicted respectively. Each fully connected layer focuses on a specific output task, one for predicting the structural parameters of the target probe and the other for predicting the operational parameters of the target probe. In the process of fusing multimodal information, different tasks can effectively utilize the encoded features and finally obtain accurate output.

[0107] In one example, a cloud model calculation and weighted operation are performed on the target probe structure parameters and the target probe operation parameters to obtain a target compensation instruction, which is used to adjust the probe position parameters and the probe pressure parameters, including:

[0108] Input the target probe structure parameters and the target probe operation parameters into the cloud model generator respectively, generate the expected value, entropy value and super entropy value based on the normal distribution function, and obtain the structure parameter cloud and the operation parameter cloud;

[0109] The cloud droplet generation calculation is performed on the structural parameter cloud and the operational parameter cloud. 200 cloud droplets are generated for each parameter. The cloud droplets are mapped to the parameter space through inverse cloud transformation to obtain the cloud droplet characteristic sequence.

[0110] Based on the cloud droplet feature sequence, a fuzzy membership matrix is ​​constructed, and the membership value of each parameter is calculated using a triangular membership function to obtain a fuzzy evaluation matrix. The fuzzy evaluation matrix is ​​input into the hierarchical analysis system, and a judgment matrix is ​​constructed through a pairwise comparison method to calculate the parameter weight vector.

[0111] The parameter weight vector is checked for consistency, the consistency ratio is calculated, and the comprehensive weight coefficient is obtained through hierarchical single sorting and hierarchical total sorting. According to the comprehensive weight coefficient, the target probe structure parameters and target probe operation parameters are converted into an execution control instruction sequence, which is formatted through the control protocol to obtain the target compensation instruction.

[0112] In this example, the target probe structural parameters and target probe operation parameters are input into the cloud model generator respectively. Based on the normal distribution function, the cloud model generator generates the expected value, entropy value and hyperentropy value for each input parameter, and obtains the structural parameter cloud and the operation parameter cloud. In the cloud model, the expected value is used to describe the central tendency of the parameter, the entropy value is used to reflect the uncertainty of the parameter, and the hyperentropy value describes the volatility of the uncertainty itself. For example, assuming that the mean of the target probe structural parameter is , the standard deviation is , then the expected value is defined as , the entropy value is calculated by the standard deviation, specifically:

[0113] ;

[0114] in, represents the entropy value of the structure parameter cloud, Represents a small disturbance term, which is used to ensure the robustness of the entropy value and the stability of the calculation. The super entropy value is further calculated by the volatility of the parameters. Cloud droplet generation calculation is performed on the structural parameter cloud and the operating parameter cloud. 200 cloud droplets are generated for each parameter. Cloud droplets are the basic elements in the cloud model and represent a possible value of the parameter within its fuzzy range. Through the inverse cloud transformation, the cloud droplets are mapped to the parameter space to obtain the cloud droplet feature sequence. The generation and inverse cloud transformation of cloud droplets are to describe the uncertainty of the parameters so that the model can handle the impact caused by parameter fluctuations. In the inverse cloud transformation, cloud droplets are generated by random sampling so that each cloud droplet is evenly distributed in the uncertainty interval of the parameter, reflecting the possible value range of the parameter. The fuzzy membership matrix is ​​constructed based on the cloud droplet feature sequence. Fuzzy membership is used to describe the degree to which an element belongs to a fuzzy set. The triangular membership function is used to calculate the membership value of each parameter. The triangular membership function is simple and efficient, and its expression is:

[0115] ;

[0116] in, represents the membership value, is the cloud droplet characteristic value, is the center position of the triangle membership function, is the width of the membership function. By calculating the membership of each cloud droplet, a fuzzy evaluation matrix is ​​constructed. The fuzzy evaluation matrix is ​​used to describe the degree of membership of different parameters under different fuzzy sets. The fuzzy evaluation matrix is ​​input into the hierarchical analysis system, and a judgment matrix is ​​constructed by pairwise comparison to calculate the weight vector of the parameters. In the pairwise comparison, the judgment matrix is ​​constructed by scoring the importance of each pair of parameters. For example, the judgment matrix is ​​expressed as , where the elements Indicates Parameters and The importance ratio of each parameter. Through normalization, the weight vector of each parameter is obtained. Assume that the judgment matrix is:

[0117] ;

[0118] By normalizing the judgment matrix, the relative weight of each parameter is obtained, thereby obtaining the parameter weight vector. In order to ensure the rationality of the judgment matrix, the parameter weight vector is subjected to consistency check. The consistency ratio (CR) is calculated to verify whether the judgment matrix has good consistency. The formula of the consistency ratio is:

[0119] ;

[0120] in, is the consistency index, is the random consistency index. When , the judgment matrix is ​​considered to have good consistency. If the consistency check passes, the hierarchical single sorting and hierarchical total sorting are performed to obtain the comprehensive weight coefficient of each parameter. The comprehensive weight coefficient is used to describe the relative importance of each parameter in the entire system. According to the calculated comprehensive weight coefficient, the target probe structure parameters and target probe operation parameters are converted into execution control instruction sequences. For example, assuming that the comprehensive weight of the structure parameter is , the comprehensive weight of the operating parameters is , then the control instruction is expressed as:

[0121] '

[0122] in, Indicates control instructions. is the target probe structure parameter, The target probe operation parameters are calculated by weighted calculations to dynamically adjust the control instructions according to the importance of different parameters to ensure that the probe can maintain stable performance and high precision during operation. The generated control instructions are formatted through the control protocol to obtain the target compensation instructions. The control instructions are converted into a standard format that conforms to the device communication protocol to ensure that the instructions can be correctly identified and executed by the probe control system.

[0123] Reference Figure 2 , this embodiment provides a probe contact impedance real-time monitoring and compensation device, including:

[0124] Acquisition module 1, used to acquire a first measurement data set including probe contact resistance value and contact force, and input it into a two-layer deep learning neural network for training to obtain a probe contact impedance monitoring model;

[0125] Detection module 2, used for real-time acquisition of a second measurement data set, and inputting the data into an intelligent contact impedance monitoring model for dual-parameter fault detection to obtain a contact state classification result;

[0126] A calculation module 3 is used to generate a two-dimensional compensation matrix based on the contact state classification result, and perform nonlinear mapping calculation on the second measurement data set to obtain a compensation coefficient;

[0127] Prediction module 4, used for inputting the compensation coefficient, the original probe structure parameter and the original probe operation parameter into the graph attention stacked autoencoder for parameter prediction, and obtaining the target probe structure parameter and the target probe operation parameter;

[0128] The output module 5 is used to perform cloud model calculation and weighted operation on the target probe structural parameters and the target probe operating parameters to obtain target compensation instructions, and the target compensation instructions are used to adjust the probe position parameters and the probe pressure parameters.

[0129] In this embodiment, for the specific implementation of each unit in the above device embodiment, please refer to the above method embodiment, which will not be repeated here.

[0130] Reference Figure 3 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3 As shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0131] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0132] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0133] In summary, a probe contact impedance real-time monitoring and compensation method is proposed in an embodiment of the present invention.

[0134] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided by the present invention and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM.

[0135] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0136] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A real-time monitoring and compensation method for probe contact impedance, characterized in that: The following steps are involved: A first measurement data set including a probe contact resistance value and a contact force is obtained, and input into a double-layer deep learning neural network for training to obtain a probe contact impedance monitoring model; collecting a second measurement data set in real time, and inputting the probe contact impedance monitoring model to perform dual-parameter fault detection to obtain a contact state classification result; generating a two-dimensional compensation matrix based on the contact state classification result, and performing nonlinear mapping calculation on the second measurement data set to obtain a compensation coefficient; Inputting the compensation coefficient, the original probe structure parameter and the original probe operation parameter into the graph attention stacked autoencoder for parameter prediction to obtain the target probe structure parameter and the target probe operation parameter; Cloud model calculation and weighted operation are performed on the target probe structural parameters and the target probe operating parameters to obtain target compensation instructions, and the target compensation instructions are used to adjust probe position parameters and probe pressure parameters.

2. The probe contact impedance real-time monitoring and compensation method according to claim 1, characterized in that: The first measurement data set including the probe contact resistance value and the contact force is obtained, and input into a double-layer deep learning neural network for training to obtain a probe contact impedance monitoring model, including: Acquire a first measurement data set including probe contact resistance value and contact force, and perform data stratification processing on the first measurement data set to obtain a normal contact data subset, a transient fault data subset, and a progressive degradation data subset; Inputting the normal contact data subset, the instantaneous fault data subset and the progressive degradation data subset into a data preprocessing unit for standardization processing to obtain standardized training data, and performing feature extraction on the standardized training data to construct a feature matrix including a contact resistance feature vector and a contact force feature vector, wherein the dimension of the feature matrix is ​​M×N, where M is the number of samples and N is the number of features; Inputting the feature matrix into the physical model encoding layer in the double-layer deep learning neural network for encoding calculation, the physical model encoding layer includes five convolution units, each convolution unit includes a convolution layer, a batch normalization layer and a ReLU activation function layer, the convolution kernel size of the convolution layer is 3×3, and the step size is 1, to obtain the first intermediate feature; Inputting the first intermediate feature into the material characteristic response layer in the two-layer deep learning neural network for characteristic analysis, the material characteristic response layer includes three fully connected layers and two LSTM layers, the number of neurons in the fully connected layers is 512, 256, and 128 respectively, and the number of hidden units in the LSTM layer is 64, to obtain a second intermediate feature; Performing a feature fusion operation on the second intermediate feature, wherein the feature fusion operation includes an attention mechanism calculation and a residual connection calculation to obtain a fused feature vector; Inputting the fused feature vector into a bidirectional gated recurrent unit for time series feature learning, wherein the bidirectional gated recurrent unit comprises two hidden layers, a forward layer and a reverse layer, and the number of units in each hidden layer is 32, to obtain a time series feature sequence; A loss function is constructed based on the time series feature sequence and back propagation optimization is performed. The loss function includes a mean square error loss term, a cross entropy loss term and a regularization loss term. The parameters are updated through the Adam optimizer to obtain the probe contact impedance monitoring model.

3. The probe contact impedance real-time monitoring and compensation method according to claim 2, characterized in that: The real-time acquisition of the second measurement data set and inputting the probe contact impedance monitoring model to perform dual-parameter fault detection to obtain a contact state classification result include: collecting a second measurement data set including a probe contact resistance value and a contact force in real time, and performing sliding window segmentation on the second measurement data set to obtain a segmented data sequence; Performing wavelet denoising on the segmented data sequence to obtain a denoised data sequence, and inputting the denoised data sequence into the probe contact impedance monitoring model, performing forward calculation through the physical model encoding layer and the material characteristic response layer to obtain a fault feature vector; Performing a dual-parameter joint probability distribution calculation on the fault feature vector, establishing a resistance-force coupling feature space, calculating the distance between the sample point and the standard mode by Mahalanobis distance measurement, and obtaining an abnormal measurement value; Constructing a fault detection function based on a sliding time window according to the abnormal measurement value, wherein the fault detection function calculates the detection statistic using an exponential weighted average method to obtain a fault detection value; The fault detection value is compared with a preset classification threshold, wherein the preset classification threshold includes a normal state threshold and a fault state threshold, and the state is divided by a three-level decision rule to obtain the contact state classification result.

4. The probe contact impedance real-time monitoring and compensation method according to claim 3, characterized in that: The generating of a two-dimensional compensation matrix based on the contact state classification result and performing nonlinear mapping calculation on the second measurement data set to obtain a compensation coefficient includes: Performing state coding on the contact state classification result, assigning a normal state value of 1, a minor fault state value of 2, and a serious fault state value of 3, to obtain a state coding matrix; Constructing a compensation reference value lookup table according to the state encoding matrix, wherein the compensation reference value lookup table includes a resistance compensation reference value and a force compensation reference value, and obtaining compensation reference data; Input the compensation reference data into a kernel function mapping unit, use radial basis function to perform nonlinear feature transformation to obtain a transformation feature space, and perform adaptive grid division on the transformation feature space, respectively set 10 evenly distributed grid nodes along the resistance dimension and the force dimension to obtain a two-dimensional grid structure; Performing bilinear interpolation calculation on the probe contact resistance value and the contact force in the second measurement data set according to the two-dimensional grid structure to obtain an initial compensation coefficient; Performing time series smoothing processing on the initial compensation coefficient to obtain a smoothed compensation coefficient, and performing a product operation on the smoothed compensation coefficient and a state weight factor to obtain a weighted compensation coefficient; The weighted compensation coefficient is subjected to boundary constraint processing, an upper limit value and a lower limit value of compensation are set, and the compensation coefficient is obtained by performing amplitude limiting through a saturation function.

5. The probe contact impedance real-time monitoring and compensation method according to claim 4, characterized in that: The compensation reference data is input into the kernel function mapping unit, and a radial basis function is used to perform nonlinear feature transformation to obtain a transformation feature space, and the transformation feature space is adaptively gridded, and 10 evenly distributed grid nodes are respectively set along the resistance dimension and the force dimension to obtain a two-dimensional grid structure, including: Normalizing the compensation reference data, mapping the resistance value and the force value to the interval [-1, 1] respectively, to obtain normalized compensation data; Inputting the normalized compensation data into a radial basis kernel function for nonlinear transformation, wherein the radial basis kernel function adopts a Gaussian kernel function form and a kernel width parameter is set to 0.5 to obtain an initial feature map; Performing a singular value decomposition operation on the initial feature map, extracting the main feature directions, selecting feature vectors with a cumulative contribution rate exceeding 95%, and obtaining a reduced-dimensional feature space; The dimension reduction feature space is divided into two orthogonal subspaces, namely, resistance dimension and force dimension, and a probability density function of data distribution is calculated in each subspace to obtain marginal distribution features; Determine the optimal distribution position of the grid nodes according to the edge distribution characteristics, use the K-means clustering algorithm to cluster 10 nodes in each dimension, and obtain adaptive node positions; Based on the adaptive node positions, a Delaunay triangulation is constructed, and the feature space is gridded, and each grid unit records the corresponding local feature statistics to obtain an initial grid structure; Performing a mesh quality assessment on the initial mesh structure, calculating the shape factor and size factor of each mesh unit, and locally encrypting or coarsening the meshes that do not meet the quality requirements to obtain an optimized mesh structure; The optimized grid structure is mapped and transformed with the original coordinate system, a corresponding relationship between grid nodes and physical quantities is established, and the two-dimensional grid structure is obtained.

6. The probe contact impedance real-time monitoring and compensation method according to claim 4, characterized in that: The compensation coefficient, the original probe structure parameter and the original probe operation parameter are input into the graph attention stacked autoencoder for parameter prediction to obtain the target probe structure parameter and the target probe operation parameter, including: Performing parameter fusion on the compensation coefficient, the original probe structure parameter and the original probe operation parameter to construct a multimodal feature vector including compensation information, structure information and operation information; Input the multimodal feature vector into the first attention layer of the graph attention stacking autoencoder for feature association calculation, the first attention layer adopts a multi-head self-attention mechanism, the number of heads is set to 8, the attention dimension is 64, and the attention weighted feature is obtained; Performing a first encoding operation on the attention weighted feature, the first encoding operation includes three graph convolution layers, the number of output channels of each graph convolution layer is 128, 256, and 512, respectively, and the activation function is LeakyReLU, to obtain a first encoding feature; The first encoded feature is input into the second attention layer for cross-modal feature interaction. The second attention layer includes a channel attention module and a spatial attention module, which respectively calculate the channel importance and spatial importance weights of the feature to obtain a multi-dimensional attention feature; Performing a second encoding operation on the multidimensional attention feature, the second encoding operation includes two fully connected layers, the number of neurons is 1024 and 512 respectively, each layer is followed by a batch normalization layer and a ReLU activation function, to obtain an encoded feature; Input the encoded features into a decoder for decoding operation, wherein the decoder includes three transposed convolutional layers, the convolution kernel size is 3×3, the step size is 2, and the number of output channels is 256, 128, and 64 respectively, to obtain the decoded features; A parameter separation operation is performed on the decoded features, and the target probe structure parameters and the target probe operation parameters are predicted respectively through two parallel fully connected layers.

7. The probe contact impedance real-time monitoring and compensation method according to claim 4, characterized in that: The cloud model calculation and weighted operation are performed on the target probe structural parameters and the target probe operating parameters to obtain a target compensation instruction, wherein the target compensation instruction is used to adjust the probe position parameters and the probe pressure parameters, including: Inputting the target probe structural parameters and the target probe operating parameters into a cloud model generator respectively, generating an expected value, an entropy value and a super entropy value based on a normal distribution function, and obtaining a structural parameter cloud and an operating parameter cloud; Performing cloud droplet generation calculation on the structural parameter cloud and the operating parameter cloud, generating 200 cloud droplets for each parameter, mapping the cloud droplets to the parameter space through inverse cloud transformation, and obtaining a cloud droplet feature sequence; A fuzzy membership matrix is ​​constructed based on the cloud droplet feature sequence, and the membership value of each parameter is calculated using a triangular membership function to obtain a fuzzy evaluation matrix. The fuzzy evaluation matrix is ​​input into a hierarchical analysis system, and a judgment matrix is ​​constructed by a pairwise comparison method to calculate a parameter weight vector. The parameter weight vector is checked for consistency, the consistency ratio is calculated, and a comprehensive weight coefficient is obtained through hierarchical single sorting and hierarchical total sorting. The target probe structure parameters and the target probe operation parameters are converted into an execution control instruction sequence according to the comprehensive weight coefficient, and the target compensation instruction is obtained through control protocol formatting.

8. A probe contact impedance real-time monitoring and compensation device, characterized in that: For implementing the steps of the method according to any one of claims 1 to 7, the device comprises: An acquisition module is used to acquire a first measurement data set including a probe contact resistance value and a contact force, and input the data set into a double-layer deep learning neural network for training to obtain a probe contact impedance monitoring model; A detection module, used for collecting a second measurement data set in real time, and inputting the probe contact impedance monitoring model to perform dual-parameter fault detection to obtain a contact state classification result; A calculation module, used for generating a two-dimensional compensation matrix based on the contact state classification result, and performing nonlinear mapping calculation on the second measurement data set to obtain a compensation coefficient; A prediction module, used for inputting the compensation coefficient, the original probe structure parameter and the original probe operation parameter into the graph attention stacked autoencoder for parameter prediction to obtain the target probe structure parameter and the target probe operation parameter; The output module is used to perform cloud model calculation and weighted operation on the target probe structural parameters and the target probe operating parameters to obtain target compensation instructions, and the target compensation instructions are used to adjust the probe position parameters and the probe pressure parameters.

9. A computer device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images