Online monitoring method and system for crack propagation of silicon-based new material equipment in high-temperature environment

Abstract

The invention provides an on-line monitoring method and system for crack propagation of silicon-based new material equipment in a high-temperature environment, and relates to the technical field of crack detection, and the method comprises the steps: collecting stress distribution data through arranging a stress sensor, and generating a stress field distribution diagram by using an attention mechanism deep neural network model; identifying a stress concentration region based on a region growing algorithm, when a stress value exceeds a preset threshold value, acquiring temperature distribution data, extracting temperature distribution characteristics through a deep mixed probability model, performing multi-modal data fusion in combination with a stress field distribution diagram, and establishing a crack propagation prediction model by using a graph structure neural network. The crack propagation rate and direction are solved through a swarm intelligence optimization algorithm, when the crack propagation rate exceeds a preset threshold value, a control system adjusts equipment operation parameters, the sampling frequency of a sensor is adjusted according to the crack propagation direction, and model parameters are updated through an incremental learning method to achieve real-time monitoring.

Description

Technical Field

[0001] The present invention relates to the field of crack detection technology, and in particular to an online monitoring method and system for crack propagation in silicon-based new material equipment under high-temperature conditions. Background Art

[0002] Silicon-based new materials are widely used in high-temperature electronic devices, aerospace and other fields due to their excellent high-temperature resistance, high power, and high-frequency properties. As application scenarios become increasingly complex and demanding, silicon-based new material equipment is prone to failure problems such as crack propagation when serving in high-temperature environments, seriously affecting the reliability and service life of the equipment. Therefore, online monitoring of crack propagation in silicon-based new material equipment is crucial to ensure safe operation of the equipment. Traditional crack growth monitoring methods rely primarily on manual visual inspection or non-destructive testing techniques, such as ultrasonic testing and X-ray testing. These methods are subject to limitations in sensor parameters and operating environments, difficulty in accurately describing complex crack growth behavior in high-temperature environments, and a lack of effective online monitoring and early warning mechanisms. Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the Invention

[0003] The embodiments of the present invention provide a method and system for online monitoring crack propagation of silicon-based new material equipment under high-temperature environment, which can at least solve some of the problems existing in the prior art.

[0004] A first aspect of an embodiment of the present invention provides an online crack propagation monitoring method for silicon-based new material equipment in a high-temperature environment, comprising: Multiple stress sensors are evenly arranged along the circumference of the surface of the silicon-based new material device to collect stress distribution data on the surface of the silicon-based new material device. The stress distribution data is then subjected to wavelet transform denoising to obtain the denoised stress distribution data and input into a deep neural network model with an attention mechanism. The spatial attention mechanism is used to process the feature weights of different positions and the channel attention mechanism is combined to extract the correlation between feature channels. The residual backbone network is then used to fine-tune the historical crack growth data to generate a stress field distribution map of the silicon-based new material device. The stress field distribution map is segmented using a region growing algorithm to identify stress concentration areas and compare the stress values ​​with a preset stress threshold. If the stress value in the stress concentration area exceeds the preset stress threshold, the temperature distribution data of the current area is obtained. The temperature distribution characteristics are established using a deep hybrid probability model. The probability model parameters are estimated using a variational inference method. The extracted temperature distribution characteristics and stress field distribution map are then subjected to multimodal data fusion. The spatiotemporal characteristic relationship of crack propagation is captured using a graph-structured neural network to establish a crack propagation prediction model. The crack propagation prediction model is solved according to the swarm intelligence optimization algorithm to calculate the crack propagation rate and direction. If the crack propagation rate exceeds the preset rate threshold, the control system adjusts the operating parameters of the silicon-based new material equipment according to the predictive control algorithm. According to the crack propagation direction, the sampling frequency of the sensor is adjusted through an adaptive sampling strategy. The crack propagation rate and propagation direction data are used to update the deep neural network model parameters through the incremental learning method. Real-time monitoring is performed based on the updated deep neural network model to obtain a monitoring data set.

[0005] In an optional embodiment, Multiple stress sensors are evenly arranged along the circumference of the surface of the silicon-based new material device to collect stress distribution data on the surface of the silicon-based new material device. The stress distribution data is denoised by wavelet transform to obtain the denoised stress distribution data and input into the attention mechanism deep neural network model. The spatial attention mechanism is used to process the feature weights of different positions and the channel attention mechanism is combined to extract the correlation between feature channels. The residual backbone network is used to fine-tune the historical crack growth data to generate the stress field distribution map of the silicon-based new material device, including: Stress sensors are placed on the surface of silicon-based new material equipment to collect stress distribution data. The collected stress distribution data is decomposed into multiple layers using an adaptive wavelet transform algorithm. The optimal wavelet basis function is determined based on the energy distribution of the decomposed signal. Dynamic threshold processing is performed on high-frequency coefficients of different scales. The threshold parameters are adaptively adjusted based on the local characteristics of the signal. The processed high-frequency coefficients and low-frequency coefficients are then subjected to adaptive weighted wavelet reconstruction to obtain noise-reduced stress distribution data. The denoised stress distribution data is input into the attention mechanism deep neural network model. The multi-scale features of the stress distribution are extracted layer by layer through a multi-layer convolutional structure. The feature weights at different positions of the feature map are calculated using the spatial attention mechanism. The correlation between feature channels is extracted using the channel attention mechanism. The spatial feature weights and channel correlations are adaptively fused to construct an attention feature map. The attention feature map is then weightedly combined with the original feature map through the feature fusion module. The residual backbone network is divided into a basic feature extraction unit and a deep feature learning unit. The network parameters of the basic feature extraction unit are fixed, and the parameters of the deep feature learning unit are fine-tuned by migrating historical crack growth data. The historical crack growth condition data and the corresponding stress distribution characteristics are divided into a training sample set and a validation sample set according to an adaptive stratified sampling strategy. Based on the training sample set, the network model is trained through a phased dynamic learning strategy. In the first phase, a larger learning rate is used to optimize the deep feature learning unit. In the second phase, the learning rate is reduced to fine-tune the network parameters. In the third phase, the learning rate is further reduced to improve the generalization performance of the model. An early stopping judgment mechanism is set in each training phase, and the verification sample set is used for verification. If there is no improvement in the performance after multiple rounds of verification, the learning rate is reduced, and the training is repeated to obtain a trained attention mechanism deep neural network model. The trained attention mechanism deep neural network model is used to process the denoised stress distribution data to generate a stress field distribution map corresponding to the silicon-based new material device.

[0006] In an optional embodiment, The residual backbone network is divided into a basic feature extraction unit and a deep feature learning unit. The network parameters of the basic feature extraction unit are fixed, and the parameters of the deep feature learning unit are fine-tuned by migrating historical crack growth data. The historical crack growth condition data and the corresponding stress distribution characteristics are divided into a training sample set and a validation sample set according to an adaptive stratified sampling strategy. The residual backbone network is divided into a basic feature extraction unit and a deep feature learning unit. The basic feature extraction unit adopts a multi-level cascade convolution layer structure to extract the basic texture features of stress distribution data layer by layer through convolution kernels of different receptive field sizes. The deep feature learning unit adopts a parallel multi-scale convolution layer structure to extract features from input features through convolution branches of different scales. The features are weighted and fused through the channel attention module to obtain high-level semantic features of crack propagation. A bidirectional residual connection structure is constructed in the deep feature learning unit, shallow features are transmitted through the forward path, deep semantic information is fed back through the reverse path, a gating unit is set to regulate the flow of information, and a feature adaptive fusion module is used to calculate the weights of features at different levels according to the spatial distribution of feature responses, and the parameters of the deep feature learning unit are fine-tuned and trained; A dynamic window sliding method is used to segment the historical crack growth data. The window size is adjusted according to the crack growth rate. A multi-scale feature pyramid is constructed for the stress distribution feature map, local texture features and global structural features are extracted, and the crack growth annotation information is converted into a probability density distribution heat map. Based on the probability density distribution heat map, the training data was stratified in combination with the temporal evolution characteristics and spatial distribution characteristics of crack propagation. The feature similarity between samples was calculated, and the categories were divided using the spectral clustering method. The importance weight was calculated according to the distance between the sample and the category center. The training sample set and the validation sample set were divided based on the importance weight.

[0007] In an optional embodiment, The stress field distribution map is segmented using a region growing algorithm to identify stress concentration areas and compare the stress value with a preset stress threshold. If the stress value in the stress concentration area exceeds the preset stress threshold, the temperature distribution data of the current area is obtained. The temperature distribution characteristics are established using a deep hybrid probability model. The probability model parameters are estimated using a variational inference method. The extracted temperature distribution characteristics and stress field distribution map are then subjected to multimodal data fusion. The spatiotemporal characteristic relationship of crack propagation is captured through a graph-structured neural network to establish a crack propagation prediction model, including: The stress field distribution map is segmented using an adaptive region growing algorithm. Local maximum points are extracted from the stress field distribution map as seed points. A growth threshold is dynamically determined based on the stress value differences between adjacent pixels. Based on the growth threshold, regions with similar stress values ​​are merged through region growing iterations to obtain multiple stress concentration regions. The average stress value of each stress concentration region is calculated and compared with a preset stress threshold to identify high-risk areas. An infrared thermal imager is used to obtain temperature distribution data of the high-risk area. The temperature distribution data is spatially registered with the stress field distribution map. A deep hybrid probability model is constructed containing multiple Gaussian distribution components, each corresponding to a different temperature pattern characteristic. A variational posterior distribution is established to approximate the true posterior distribution. The probability model parameters are iteratively optimized by maximizing the lower bound of evidence to obtain a probabilistic representation of the temperature distribution. A dual-branch feature fusion network is constructed. The temperature feature extraction branch of the dual-branch feature fusion network uses a fully connected layer to reduce the dimensionality of the probability model parameters, and the stress feature extraction branch uses a convolutional layer to extract the spatial features of the stress field. The temperature features and stress features are adaptively weighted and fused through a feature attention mechanism to obtain multimodal features that characterize the characteristics of the crack region. A multi-layer graph convolutional network is constructed, and the multimodal features are used as attributes of the graph network nodes. Connection relationships are established based on the spatial distance and feature similarity between nodes. Features are transferred and updated between connected nodes through a message passing mechanism. Residual connections and layer normalization structures are set after each layer of graph convolution. The multi-layer graph convolutional network is trained in an end-to-end manner. The historical crack propagation path is used as supervision information, and the network parameters are optimized through back propagation to obtain a crack propagation prediction model.

[0008] In an optional embodiment, The connection relationship is established based on the spatial distance and feature similarity between nodes. The features are transferred and updated between connected nodes through the message passing mechanism. After each layer of graph convolution, a residual connection and layer normalization structure are set. The multi-layer graph convolution network is trained in an end-to-end manner. The historical crack propagation path is used as supervision information. The network parameters are optimized through back propagation. The crack propagation prediction model is obtained, which includes: Calculating an inter-node distance matrix based on the spatial distance between nodes, setting an adaptive distance threshold, wherein the adaptive distance threshold takes a smaller value in a node-dense area and a larger value in a node-sparse area, establishing initial connections for node pairs whose distance is less than the adaptive distance threshold to obtain an initial connection matrix, calculating the feature similarity of the node feature vector corresponding to each node, setting a dynamic similarity threshold according to the number of network layers, wherein the dynamic similarity threshold gradually increases as the number of network layers deepens, and retaining the node pairs in the initial connection matrix that meet the dynamic similarity threshold to obtain an optimized node connection relationship; A bidirectional message transmission mechanism is constructed. In the forward transmission phase, a multilayer perceptron is used to perform nonlinear mapping on the source node features to obtain forward transformation features. In the backward transmission phase, an inverse mapping network is constructed to obtain reverse transformation features. Based on the optimized node connection relationship, the forward transformation features and the reverse transformation features are respectively input into a bidirectional multi-head attention module to calculate the bidirectional message weights. An adaptive gating mechanism is used to fuse the bidirectional messages to obtain the aggregated features of the target node. A multi-layer graph convolutional network is constructed based on the aggregated features of the target node. The hidden dimension of each layer decreases layer by layer. Message passing is performed in each layer of graph convolution to obtain node update features. The node update features are fused with the original features through residual connections to obtain fused features. The fused features are layer-normalized to obtain standardized features. An end-to-end training approach is adopted. In the pre-training phase, the historical crack propagation path and the standardized features are used to reconstruct the current crack state to obtain a reconstructed feature. In the fine-tuning phase, the reconstructed feature and target scene data are used to predict the crack propagation at the next moment to obtain a predicted feature. Based on the predicted feature, a local prediction error term and a global morphological similarity term are constructed to obtain a multi-scale loss function. Based on the prediction features, the crack propagation speed and the degree of direction change are calculated, the difficulty of the training samples is evaluated and graded to obtain the difficulty level, and the samples of the same difficulty level are organized into the same training batch; based on the multi-scale loss function, the model is verified at three levels: node position prediction, regional expansion trend, and rationality of stress distribution to obtain verification indicators, and the crack propagation prediction model is obtained by optimizing the network parameters through back propagation.

[0009] In an optional embodiment, The crack growth prediction model is solved using a swarm intelligence optimization algorithm to calculate the crack growth rate and direction. If the crack growth rate exceeds the preset rate threshold, the control system adjusts the operating parameters of the silicon-based new material equipment according to the predictive control algorithm. Based on the crack growth direction, the sensor sampling frequency is adjusted using an adaptive sampling strategy. The crack growth rate and growth direction data are used to update the deep neural network model parameters through an incremental learning method. Real-time monitoring is performed based on the updated deep neural network model, resulting in a monitoring data set including: Initialize a population individual set, randomly generate a vector containing a crack growth rate and a growth direction as an initial population individual, calculate the fitness value of each initial population individual based on stress distribution data and crack morphology characteristics, the fitness value including a prediction error term, a stress field consistency term, and an energy dissipation term, obtain an iterative prediction deviation, calculate a target weight coefficient based on the iterative prediction deviation, and assign the target weight coefficient to the prediction error term, the stress field consistency term, and the energy dissipation term; A swarm intelligence optimization algorithm is used to solve the crack growth prediction model, calculate the crack growth rate and growth direction, and a hybrid mutation operator is designed to perform mutation operations on the individuals in the initial population; Comparing the crack growth rate with a preset rate threshold, and when the crack growth rate exceeds the preset rate threshold, adjusting the operating parameters of the silicon-based new material equipment through a predictive control algorithm; dynamically partitioning the monitoring area according to the growth direction, dividing the monitoring area into a high-risk area, a transition area, and a safe area; Adopting an adaptive sampling strategy to adjust the sampling frequency of the sensor, the working mode of the sensor array is determined according to the degree of danger, and the stress signal mutation index is calculated in real time. When the stress signal mutation index exceeds the preset mutation threshold, the adjacent sensor group is switched to the highest sampling frequency mode; Inputting the crack growth rate and the growth direction into an incremental learning module, constructing a training data set using a sliding window mechanism, extracting features, constructing a hybrid loss function using the original deep neural network model prediction results as soft labels, and incrementally updating the deep neural network model parameters based on the hybrid loss function; The deep neural network model is used for real-time monitoring to obtain stress field data, temperature field data and acoustic emission signal data. The stress field data, temperature field data and acoustic emission signal data are adaptively filtered and denoised. A sliding window method is used for feature extraction, and the extracted features are input into the deep neural network model to obtain a monitoring data set.

[0010] In an optional embodiment, Inputting the crack growth rate and the growth direction into an incremental learning module, constructing a training data set using a sliding window mechanism, extracting features, constructing a hybrid loss function using the original deep neural network model prediction results as soft labels, and incrementally updating the deep neural network model parameters based on the hybrid loss function includes: receiving a crack growth rate and a crack growth direction, normalizing the crack growth rate and the crack growth direction, removing outliers using a median filter, and performing noise reduction using a wavelet transform; A sliding window mechanism is used to construct a training data set. The basic window length is set to one-tenth of the original data length, and the window sliding step is set to one-quarter of the basic window length. Time domain statistical features are extracted within each window, and the signal is subjected to Fourier transform to extract frequency domain features. Instantaneous features are obtained through Hilbert transform. Calculating the mutual information values ​​between the time-domain statistical features, the frequency-domain features, the instantaneous features, and the target variable, and constructing a feature importance score based on the correlation coefficients between the features. The feature importance score simultaneously considers the mutual information values ​​between the features and the target variable and the redundancy between the features. Based on the feature importance score, the features are hierarchically screened, first selecting the features with the highest scores, and then sequentially selecting features with lower correlations with the selected features and higher scores, until the cumulative importance of the feature subset reaches a preset threshold; The prediction results of the original deep neural network model are used as soft labels. The soft labels are combined with the real labels in a mixing ratio of 3 to 7 to construct a hybrid loss function. The underlying feature extraction part of the deep neural network model is frozen, and only the parameters of the fully connected layer are updated. Then all layers are unfrozen, and the entire network is fine-tuned with a randomly selected learning rate less than 0.01, and the learning rate is adjusted by exponential decay. Elastic network regularization is applied to the network weights, and a random dropout strategy is used. The dropout ratio of the random dropout strategy decreases layer by layer from the input layer to the output layer. Batch normalization is used to maintain the intermediate feature distribution. The prediction error is calculated on the validation set. When the prediction error exceeds a preset threshold, the proportion of historical data in the training set is increased. Training is stopped when there is no improvement in performance after multiple rounds of verification, and an updated deep neural network model is obtained.

[0011] A second aspect of an embodiment of the present invention provides an online crack propagation monitoring system for silicon-based new material equipment in a high-temperature environment, comprising: The first unit is used to collect stress distribution data on the surface of the silicon-based new material device through multiple stress sensors evenly arranged along the circumference of the surface of the silicon-based new material device, and perform wavelet transform denoising on the stress distribution data to obtain the denoised stress distribution data and input it into the attention mechanism deep neural network model. The spatial attention mechanism is used to process the feature weights of different positions and the channel attention mechanism is combined to extract the correlation between feature channels. The residual backbone network is used to fine-tune the historical crack growth data to generate a stress field distribution map of the silicon-based new material device; The second unit is used to segment the stress field distribution map using a region growing algorithm, identify stress concentration areas, and compare the stress values ​​with a preset stress threshold. If the stress value in the stress concentration area exceeds the preset stress threshold, the temperature distribution data of the current area is obtained, and the temperature distribution characteristics are established using a deep hybrid probability model. The probability model parameters are estimated using a variational inference method. The extracted temperature distribution characteristics and stress field distribution map are then subjected to multimodal data fusion. The spatiotemporal characteristic relationship of crack propagation is captured through a graph-structured neural network to establish a crack propagation prediction model. The third unit is used to solve the crack propagation prediction model based on the swarm intelligence optimization algorithm and calculate the crack propagation rate and direction. If the crack propagation rate exceeds the preset rate threshold, the control system adjusts the operating parameters of the silicon-based new material equipment according to the predictive control algorithm. According to the crack propagation direction, the sampling frequency of the sensor is adjusted through an adaptive sampling strategy. The crack propagation rate and propagation direction data are used to update the deep neural network model parameters through the incremental learning method. Real-time monitoring is performed based on the updated deep neural network model to obtain a monitoring data set.

[0012] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0013] According to a fourth aspect of the embodiments of the present invention, A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0014] In the present invention, by combining the stress field distribution and temperature distribution characteristics and using a graph structure neural network to capture the spatiotemporal characteristic relationship of crack propagation, a more accurate crack propagation prediction model is established, which can more accurately predict the crack propagation rate and direction. By monitoring the stress value in the stress concentration area in real time and issuing an early warning when the crack propagation rate exceeds the preset threshold, the risk of crack propagation can be discovered in time to avoid equipment failure. By monitoring the stress value in the stress concentration area in real time and issuing an early warning when the crack propagation rate exceeds the preset threshold, the risk of crack propagation can be discovered in time to avoid equipment failure. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a flow chart of a method for online monitoring crack growth in silicon-based new material equipment under high-temperature conditions according to an embodiment of the present invention; Figure 2This is a structural schematic diagram of an online monitoring system for crack propagation in silicon-based new material equipment under a high-temperature environment according to an embodiment of the present invention. DETAILED DESCRIPTION

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0017] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0018] Figure 1 FIG. 1 is a flow chart of an online crack propagation monitoring method for silicon-based new material equipment under high temperature environment according to an embodiment of the present invention. Figure 1 As shown, the method includes: S1. Multiple stress sensors are evenly arranged along the circumference of the silicon-based new material device surface to collect stress distribution data. The stress distribution data is denoised by wavelet transform to obtain the de-noised stress distribution data. This data is then input into a deep neural network model with an attention mechanism. The spatial attention mechanism is used to process feature weights at different locations, and the channel attention mechanism is combined to extract the correlation between feature channels. The residual backbone network is then fine-tuned and trained on historical crack growth data to generate a stress field distribution map of the silicon-based new material device. In an optional embodiment, Multiple stress sensors are evenly arranged along the circumference of the surface of the silicon-based new material device to collect stress distribution data on the surface of the silicon-based new material device. The stress distribution data is denoised by wavelet transform to obtain the denoised stress distribution data and input into the attention mechanism deep neural network model. The spatial attention mechanism is used to process the feature weights of different positions and the channel attention mechanism is combined to extract the correlation between feature channels. The residual backbone network is used to fine-tune the historical crack growth data to generate the stress field distribution map of the silicon-based new material device, including: Stress sensors are placed on the surface of silicon-based new material equipment to collect stress distribution data. The collected stress distribution data is decomposed into multiple layers using an adaptive wavelet transform algorithm. The optimal wavelet basis function is determined based on the energy distribution of the decomposed signal. Dynamic threshold processing is performed on high-frequency coefficients of different scales. The threshold parameters are adaptively adjusted based on the local characteristics of the signal. The processed high-frequency coefficients and low-frequency coefficients are then subjected to adaptive weighted wavelet reconstruction to obtain noise-reduced stress distribution data. The denoised stress distribution data is input into the attention mechanism deep neural network model. The multi-scale features of the stress distribution are extracted layer by layer through a multi-layer convolutional structure. The feature weights at different positions of the feature map are calculated using the spatial attention mechanism. The correlation between feature channels is extracted using the channel attention mechanism. The spatial feature weights and channel correlations are adaptively fused to construct an attention feature map. The attention feature map is then weightedly combined with the original feature map through the feature fusion module. The residual backbone network is divided into a basic feature extraction unit and a deep feature learning unit. The network parameters of the basic feature extraction unit are fixed, and the parameters of the deep feature learning unit are fine-tuned by migrating historical crack growth data. The historical crack growth condition data and the corresponding stress distribution characteristics are divided into a training sample set and a validation sample set according to an adaptive stratified sampling strategy. Based on the training sample set, the network model is trained through a phased dynamic learning strategy. In the first phase, a larger learning rate is used to optimize the deep feature learning unit. In the second phase, the learning rate is reduced to fine-tune the network parameters. In the third phase, the learning rate is further reduced to improve the generalization performance of the model. An early stopping judgment mechanism is set in each training phase, and the verification sample set is used for verification. If there is no improvement in the performance after multiple rounds of verification, the learning rate is reduced, and the training is repeated to obtain a trained attention mechanism deep neural network model. The trained attention mechanism deep neural network model is used to process the denoised stress distribution data to generate a stress field distribution map corresponding to the silicon-based new material device.

[0019] Multiple stress sensors are evenly distributed across the surface of the silicon-based new material device to ensure full coverage of the test area. These sensors can be piezoresistive, piezoelectric, or fiber optic, with the appropriate type selected based on the actual situation. The data collected by the sensors is the stress values ​​at different locations on the device surface, forming the initial stress distribution data. For example, if the device surface is divided into 100 measurement points, and each sensor collects the stress value at one measurement point, an initial dataset containing 100 stress values ​​can be obtained.

[0020] The collected stress distribution data is subjected to wavelet transform denoising. The data is decomposed into multiple layers using an adaptive wavelet transform algorithm. Appropriate wavelet basis functions are selected, such as Daubechies wavelet, Symlets wavelet or Coiflets wavelet, and the optimal wavelet basis function is determined based on the energy distribution of the decomposed signal. Dynamic threshold processing is performed on the decomposed high-frequency coefficients, and the threshold parameters are adaptively adjusted according to the local characteristics of the signal to remove noise signals. For example, the threshold can be dynamically adjusted according to the variance or standard deviation of the signal. The processed high-frequency coefficients and low-frequency coefficients are subjected to adaptive weighted wavelet reconstruction to obtain the denoised stress distribution data. Assume that after wavelet denoising, the stress values ​​of the 100 measurement points are optimized to form a new data set.

[0021] The denoised stress distribution data is input into a deep neural network model with an attention mechanism. This model consists of a multi-layer convolutional architecture, a spatial attention mechanism, a channel attention mechanism, and a residual backbone network. The multi-layer convolutional architecture extracts multi-scale features of the stress distribution layer by layer. The spatial attention mechanism calculates feature weights at different locations in the feature map, highlighting features in important areas. The channel attention mechanism extracts correlations between feature channels and enhances the feature representation of related channels. The spatial feature weights are adaptively fused with the channel correlations to construct an attention feature map. The feature fusion module then performs a weighted combination of the attention feature map and the original feature map to obtain more representative features.

[0022] The residual backbone network consists of a basic feature extraction unit and a deep feature learning unit. The network parameters of the basic feature extraction unit are fixed, and pre-trained model weights are used. The parameters of the deep feature learning unit are fine-tuned using historical crack growth data. The historical crack growth data and corresponding stress distribution features are divided into training and validation sets using an adaptive stratified sampling strategy. For example, 80% of the data can be used for training and 20% for validation.

[0023] Based on the training sample set, the network model is trained using a phased dynamic learning strategy. In the first phase, a high learning rate is used to rapidly optimize the parameters of the deep feature learning unit. In the second phase, the learning rate is reduced to fine-tune the network parameters. In the third phase, the learning rate is further reduced to improve model generalization performance. An early stopping mechanism is implemented in each training phase, and validation is performed using a validation sample set. If performance does not improve after multiple rounds of validation, the learning rate is reduced and training is repeated until the model converges.

[0024] The trained attention mechanism deep neural network model processes the de-noised stress distribution data to generate a stress field distribution map corresponding to the silicon-based new material device. This map visually displays the stress distribution on the device surface, providing a basis for predicting crack propagation and other factors. For example, the stress field distribution map can be used to identify areas of stress concentration and predict the possible crack initiation location and propagation path.

[0025] In this embodiment, an adaptive wavelet transform algorithm is used to denoise the collected stress distribution data, effectively removing noise interference and improving the signal-to-noise ratio of the data, thereby improving the accuracy of the stress field distribution map finally generated. The attention mechanism deep neural network model can effectively extract the multi-scale features of the stress distribution. The combination of the spatial attention mechanism and the channel attention mechanism further enhances the feature expression ability of the model, enabling the model to better capture the key features of the stress distribution. The staged dynamic learning strategy and transfer learning method are adopted to effectively train the deep neural network model, improve the generalization performance of the model, and enable the model to better adapt to different silicon-based new material equipment and working conditions.

[0026] In an optional embodiment, The residual backbone network is divided into a basic feature extraction unit and a deep feature learning unit. The network parameters of the basic feature extraction unit are fixed, and the parameters of the deep feature learning unit are fine-tuned by migrating historical crack growth data. The historical crack growth condition data and the corresponding stress distribution characteristics are divided into a training sample set and a validation sample set according to an adaptive stratified sampling strategy. The residual backbone network is divided into a basic feature extraction unit and a deep feature learning unit. The basic feature extraction unit adopts a multi-level cascade convolution layer structure to extract the basic texture features of stress distribution data layer by layer through convolution kernels of different receptive field sizes. The deep feature learning unit adopts a parallel multi-scale convolution layer structure to extract features from input features through convolution branches of different scales. The features are weighted and fused through the channel attention module to obtain high-level semantic features of crack propagation. A bidirectional residual connection structure is constructed in the deep feature learning unit, shallow features are transmitted through the forward path, deep semantic information is fed back through the reverse path, a gating unit is set to regulate the flow of information, and a feature adaptive fusion module is used to calculate the weights of features at different levels according to the spatial distribution of feature responses, and the parameters of the deep feature learning unit are fine-tuned and trained; A dynamic window sliding method is used to segment the historical crack growth data. The window size is adjusted according to the crack growth rate. A multi-scale feature pyramid is constructed for the stress distribution feature map, local texture features and global structural features are extracted, and the crack growth annotation information is converted into a probability density distribution heat map. Based on the probability density distribution heat map, the training data was stratified in combination with the temporal evolution characteristics and spatial distribution characteristics of crack propagation. The feature similarity between samples was calculated, and the categories were divided using the spectral clustering method. The importance weight was calculated according to the distance between the sample and the category center. The training sample set and the validation sample set were divided based on the importance weight.

[0027] A deep learning model for crack growth prediction is constructed. The model is based on a residual network structure and is divided into a basic feature extraction unit and a deep feature learning unit.

[0028] The basic feature extraction unit is designed to extract basic texture features from stress distribution data. It employs a multi-layer cascaded convolutional layer architecture, with each layer using different convolution kernel sizes to capture texture information at different scales. For example, the first layer might use a 3x3 convolution kernel to extract fine texture features, the second layer might use a 5x5 convolution kernel to extract larger-scale texture features, and so on. The cascaded convolutional layers allow the network to gradually learn from simple edge information to more complex texture patterns.

[0029] The deep feature learning unit is designed to capture high-level semantic features of crack growth. It employs a parallel multi-scale convolutional layer structure, passing input features to multiple convolutional branches of different scales. Each branch uses a different-sized convolution kernel to extract features at a different scale. For example, one branch uses a 3x3 kernel, another a 5x5 kernel, and yet another a 7x7 kernel. These features at different scales are then weighted and fused using a channel attention module. This module dynamically adjusts the weights of each channel based on its importance, highlighting the features most useful for crack growth prediction.

[0030] A bidirectional residual connection structure is also constructed within the deep feature learning unit. The forward path transmits shallow-layer features to deeper layers, while the reverse path feeds deep-layer semantic information back to shallow layers. This bidirectional information flow enhances feature representation capabilities. Furthermore, a gating unit is implemented to regulate information flow and further optimize the feature learning process. The feature adaptive fusion module calculates weights for features at different levels based on the spatial distribution of feature responses, enabling the model to adaptively fuse features from different levels.

[0031] Historical crack growth data is used to transfer and fine-tune the parameters of the deep feature learning unit. A dynamic window sliding method is used to segment the historical crack growth data. The window size is dynamically adjusted based on the crack growth rate. For example, if the crack growth rate is fast, a smaller window is used, and vice versa. A multi-scale feature pyramid is constructed for the stress distribution feature map. For example, the original feature map can be downsampled to obtain feature maps of different resolutions, thereby constructing a feature pyramid. Local texture features and global structural features are extracted from the feature pyramid. Finally, the crack growth annotation information is converted into a probability density distribution heat map. For example, if the probability of crack growth at a certain location is high, the value of that location in the heat map will be larger.

[0032] Based on the probability density distribution heat map, the training data is stratified by combining the temporal evolution and spatial distribution characteristics of crack growth. For example, the data can be divided into different strata based on the speed and direction of crack growth, and the feature similarity between samples can be calculated. For example, cosine similarity can be used to measure the similarity between samples. Spectral clustering is used to divide the data into categories. Importance weights are calculated based on the distance between the sample and the category center. For example, the closer the sample is to the category center, the greater its importance weight. Based on the importance weights, a training set and a validation set are generated. For example, samples with higher importance weights can be selected as the training set, and the remaining samples as the validation set. For example, 1000 sample data are divided into five categories after spectral clustering, with 200 samples in each category. The distance between each sample and its category center is calculated, and each sample is assigned an importance weight based on the distance. The top 150 samples in each category are selected as the training set, and the remaining 50 samples are selected as the validation set. This results in 750 training samples and 250 validation samples, resulting in a crack growth prediction model trained through transfer learning.

[0033] In this embodiment, by using residual networks and transfer learning technology, features can be extracted more effectively and crack propagation patterns can be learned, thereby improving prediction accuracy. Transfer learning can use historical data for pre-training, thereby reducing dependence on new data and lowering the cost of data collection and annotation. Through multi-scale feature extraction and adaptive feature fusion, the model can better adapt to different types of crack propagation data, thereby improving the generalization ability of the model.

[0034] S2. Segment the stress field distribution map using a region growing algorithm, identify stress concentration areas, and compare the stress values ​​with a preset stress threshold. If the stress value in the stress concentration area exceeds the preset stress threshold, obtain the temperature distribution data for the current area. Establish temperature distribution features using a deep hybrid probability model. Estimate the parameters of the probability model using a variational inference method. Perform multimodal data fusion on the extracted temperature distribution features and stress field distribution map. Capture the spatiotemporal characteristics of crack propagation using a graph-structured neural network to establish a crack propagation prediction model. In an optional embodiment, The stress field distribution map is segmented using a region growing algorithm to identify stress concentration areas and compare the stress value with a preset stress threshold. If the stress value in the stress concentration area exceeds the preset stress threshold, the temperature distribution data of the current area is obtained. The temperature distribution characteristics are established using a deep hybrid probability model. The probability model parameters are estimated using a variational inference method. The extracted temperature distribution characteristics and stress field distribution map are then subjected to multimodal data fusion. The spatiotemporal characteristic relationship of crack propagation is captured through a graph-structured neural network to establish a crack propagation prediction model, including: The stress field distribution map is segmented using an adaptive region growing algorithm. Local maximum points are extracted from the stress field distribution map as seed points. A growth threshold is dynamically determined based on the stress value differences between adjacent pixels. Based on the growth threshold, regions with similar stress values ​​are merged through region growing iterations to obtain multiple stress concentration regions. The average stress value of each stress concentration region is calculated and compared with a preset stress threshold to identify high-risk areas. An infrared thermal imager is used to obtain temperature distribution data of the high-risk area. The temperature distribution data is spatially registered with the stress field distribution map. A deep hybrid probability model is constructed containing multiple Gaussian distribution components, each corresponding to a different temperature pattern characteristic. A variational posterior distribution is established to approximate the true posterior distribution. The probability model parameters are iteratively optimized by maximizing the lower bound of evidence to obtain a probabilistic representation of the temperature distribution. A dual-branch feature fusion network is constructed. The temperature feature extraction branch of the dual-branch feature fusion network uses a fully connected layer to reduce the dimensionality of the probability model parameters, and the stress feature extraction branch uses a convolutional layer to extract the spatial features of the stress field. The temperature features and stress features are adaptively weighted and fused through a feature attention mechanism to obtain multimodal features that characterize the characteristics of the crack region. A multi-layer graph convolutional network is constructed, and the multimodal features are used as attributes of the graph network nodes. Connection relationships are established based on the spatial distance and feature similarity between nodes. Features are transferred and updated between connected nodes through a message passing mechanism. Residual connections and layer normalization structures are set after each layer of graph convolution. The multi-layer graph convolutional network is trained in an end-to-end manner. The historical crack propagation path is used as supervision information, and the network parameters are optimized through back propagation to obtain a crack propagation prediction model.

[0035] The acquired stress field distribution map undergoes preprocessing, such as noise reduction and filtering, to improve image quality. The preprocessed stress field distribution map is then segmented using an adaptive region growing algorithm. Local maximum points are extracted from the stress field distribution map as seed points. For example, in a 1000x1000 pixel stress field distribution map, five local maximum points are extracted as seed points. Next, a growth threshold is dynamically determined based on the stress value differences between adjacent pixels. For example, the difference between the stress values ​​of the eight pixels surrounding the seed point and the stress value at the seed point is calculated, and the average of these differences is used as the growth threshold. Assume that the stress value of a seed point is 100 MPa, and the stress values ​​of the surrounding 8 pixels are 95 MPa, 98 MPa, 102 MPa, 105 MPa, 97 MPa, 101 MPa, 99 MPa, and 103 MPa, respectively. The growth threshold is |(100-95)+(100-98)+(100-102)+(100-105)+(100-97)+(100-101)+(100-99)+(100-103)| / 8=2.5 MPa. Based on this growth threshold, regions with similar stress values ​​are iteratively merged through region growing to generate multiple stress concentration regions. This process continues until all pixels meeting the criteria are merged into the corresponding region. The average stress value of each stress concentration region is calculated. For example, if a stress concentration region contains 100 pixels with stress values ​​of 100 MPa, 101 MPa, ..., and 199 MPa, the average stress value of the region is 150 MPa. The calculated average stress value is compared with a preset stress threshold (for example, 200 MPa) to identify high-risk areas.

[0036] Use an infrared thermal imager to obtain temperature distribution data in high-risk areas. For example, use an infrared thermal imager to collect temperature data in high-risk areas at a rate of 30 frames per second. Spatial registration is performed on the obtained temperature distribution data and the stress field distribution map to ensure that the two are aligned in spatial position. For example, an image registration algorithm is used to align the temperature distribution data and the stress field distribution map to the same coordinate system. Then, a deep mixture probability model containing multiple Gaussian distribution components is constructed. Each Gaussian distribution component corresponds to a different temperature pattern feature. For example, assume that the model contains three Gaussian components, representing low temperature, medium temperature, and high temperature patterns, respectively. Next, a variational posterior distribution is established to approximate the true posterior distribution, and the probability model parameters are iteratively optimized by maximizing the evidence lower bound. For example, a variational autoencoder (VAE) is used to model the temperature distribution data, and the model parameters are optimized by maximizing the evidence lower bound (ELBO). Finally, a probabilistic representation of the temperature distribution is obtained.

[0037] A dual-branch feature fusion network is constructed. The temperature feature extraction branch uses a fully connected layer to reduce the dimensionality of the probability model parameters. For example, the probability distribution parameters learned by the VAE are used as input and the dimensionality is reduced through a neural network containing two fully connected layers to obtain a low-dimensional temperature feature vector. The stress feature extraction branch uses a convolutional layer to extract the spatial features of the stress field. For example, multiple convolutional layers and pooling layers are used to extract spatial features in the stress field distribution map. The temperature features and stress features are adaptively weighted and fused through a feature attention mechanism to obtain multimodal features that characterize the characteristics of the crack region. For example, the attention mechanism is used to calculate the weights of the temperature features and stress features, and the weighted features are concatenated to obtain the final multimodal features.

[0038] Construct a multi-layer graph convolutional network. Use multimodal features as attributes of graph network nodes. Construct connection relationships based on spatial distance and feature similarity between nodes. For example, calculate the Euclidean distance and feature cosine similarity between nodes, and combine the two to construct connection relationships between nodes. Transfer and update features between connected nodes through a message passing mechanism. For example, use a graph convolutional network (GCN) to transfer and aggregate information between nodes. Set residual connections and layer normalization structures after each layer of graph convolution. Train the multi-layer graph convolutional network in an end-to-end manner. Use historical crack propagation paths as supervision information, optimize network parameters through backpropagation, and obtain a crack propagation prediction model. For example, use known crack propagation path data as a training set, and optimize network parameters by minimizing the difference between the predicted path and the true path.

[0039] In this embodiment, the multimodal information of stress field and temperature field is integrated, and the spatiotemporal characteristic relationship of crack propagation is captured using graph neural network, which can more accurately predict the crack propagation path. The adaptive region growing algorithm and variational inference method can effectively handle noise and uncertainty, improve the robustness of the model, and adopt an end-to-end training method to avoid the tedious feature engineering and model selection process, thereby improving the prediction efficiency.

[0040] In an optional embodiment, The connection relationship is established based on the spatial distance and feature similarity between nodes. The features are transferred and updated between connected nodes through the message passing mechanism. After each layer of graph convolution, a residual connection and layer normalization structure are set. The multi-layer graph convolution network is trained in an end-to-end manner. The historical crack propagation path is used as supervision information. The network parameters are optimized through back propagation. The crack propagation prediction model is obtained, which includes: Calculating an inter-node distance matrix based on the spatial distance between nodes, setting an adaptive distance threshold, wherein the adaptive distance threshold takes a smaller value in a node-dense area and a larger value in a node-sparse area, establishing initial connections for node pairs whose distance is less than the adaptive distance threshold to obtain an initial connection matrix, calculating the feature similarity of the node feature vector corresponding to each node, setting a dynamic similarity threshold according to the number of network layers, wherein the dynamic similarity threshold gradually increases as the number of network layers deepens, and retaining the node pairs in the initial connection matrix that meet the dynamic similarity threshold to obtain an optimized node connection relationship; A bidirectional message transmission mechanism is constructed. In the forward transmission phase, a multilayer perceptron is used to perform nonlinear mapping on the source node features to obtain forward transformation features. In the backward transmission phase, an inverse mapping network is constructed to obtain reverse transformation features. Based on the optimized node connection relationship, the forward transformation features and the reverse transformation features are respectively input into a bidirectional multi-head attention module to calculate the bidirectional message weights. An adaptive gating mechanism is used to fuse the bidirectional messages to obtain the aggregated features of the target node. A multi-layer graph convolutional network is constructed based on the aggregated features of the target node. The hidden dimension of each layer decreases layer by layer. Message passing is performed in each layer of graph convolution to obtain node update features. The node update features are fused with the original features through residual connections to obtain fused features. The fused features are layer-normalized to obtain standardized features. An end-to-end training approach is adopted. In the pre-training phase, the historical crack propagation path and the standardized features are used to reconstruct the current crack state to obtain a reconstructed feature. In the fine-tuning phase, the reconstructed feature and target scene data are used to predict the crack propagation at the next moment to obtain a predicted feature. Based on the predicted feature, a local prediction error term and a global morphological similarity term are constructed to obtain a multi-scale loss function. Based on the prediction features, the crack propagation speed and the degree of direction change are calculated, the difficulty of the training samples is evaluated and graded to obtain the difficulty level, and the samples of the same difficulty level are organized into the same training batch; based on the multi-scale loss function, the model is verified at three levels: node position prediction, regional expansion trend, and rationality of stress distribution to obtain verification indicators, and the crack propagation prediction model is obtained by optimizing the network parameters through back propagation.

[0041] The crack propagation prediction method analyzes historical crack propagation paths and constructs a prediction model based on graph convolutional networks to accurately predict future crack propagation.

[0042] Discretize the crack propagation region and represent it as a series of nodes. Calculate the spatial distances between nodes and construct a distance matrix. Set an adaptive distance threshold, with a lower threshold (e.g., 0.5 mm) in densely populated areas and a higher threshold (e.g., 2 mm) in sparsely populated areas. Node pairs whose distance is less than the adaptive distance threshold are connected to form an initial connection matrix. For example, if the distance between nodes A and B is 0.3 mm, which is less than the threshold of 0.5 mm in densely populated areas, then A and B will be connected in the initial connection matrix.

[0043] Calculate the similarity between the eigenvectors corresponding to each node. The node eigenvector can contain information such as the node's location coordinates, stress value, material properties, etc. For example, the eigenvector of node A is [x_A, y_A, stress_A, material_A], and the eigenvector of node B is [x_B, y_B, stress_B, material_B]. Use cosine similarity to calculate the feature similarity between nodes A and B. Set a dynamic threshold for similarity based on the number of network layers. Assuming the network has 3 layers, the similarity thresholds for the 1st, 2nd, and 3rd layers are set to 0.6, 0.7, and 0.8, respectively. Retain the node pairs in the initial connection matrix that meet the dynamic threshold for similarity to obtain the optimized node connection relationship. For example, if the similarity between nodes A and B is 0.65, which is greater than the first-layer threshold of 0.6, then the connection between A and B is retained.

[0044] Construct a bidirectional message passing mechanism. In the forward transmission phase, a multilayer perceptron is used to perform nonlinear mapping on the source node features to obtain forward transformation features. For example, the feature vector of node A is input into a multilayer perceptron containing two hidden layers to obtain forward transformation features. In the backward transmission phase, an inverse mapping network is constructed to obtain reverse transformation features. For example, the feature vector of node B is input into an inverse mapping network with the opposite structure of the forward mapping network to obtain reverse transformation features. Based on the optimized node connection relationship, the forward transformation features and reverse transformation features are respectively input into the bidirectional multi-head attention module to calculate the bidirectional message weights. An adaptive gating mechanism is used to fuse bidirectional messages to obtain the aggregated features of the target node. For example, the forward transformation features of node A and the reverse transformation features of node B are input into the attention module to calculate the message weights from A to B and the message weights from B to A. Then, the messages in both directions are fused through the gating mechanism to obtain the aggregated features of node B.

[0045] Based on the aggregated features of the target node, a multi-layer graph convolutional network is constructed. The hidden layer dimension of each layer decreases gradually. For example, the hidden layer dimension of the first layer is 256, the hidden layer dimension of the second layer is 128, and the hidden layer dimension of the third layer is 64. Within each graph convolution layer, message passing is performed sequentially to obtain updated node features. The updated node features are fused with the original features through residual connections to obtain fused features. The fused features are then layer-normalized to obtain standardized features.

[0046] An end-to-end training approach is employed. During the pre-training phase, the current crack state is reconstructed using historical crack propagation paths and standardized features to obtain reconstructed features. During the fine-tuning phase, the reconstructed features and target scene data are used to predict crack propagation at the next moment, obtaining predicted features. Based on the predicted features, local prediction error terms and global morphological similarity terms are constructed to obtain a multi-scale loss function. For example, the local prediction error term can use the mean squared error to calculate the difference between the predicted crack location and the actual crack location, while the global morphological similarity term can use the structural similarity index (SSIM) to evaluate the similarity between the predicted crack morphology and the actual crack morphology.

[0047] Based on the predicted features, the crack growth rate and directional change are calculated, and the difficulty of the training samples is assessed and graded, for example, into three levels: easy, medium, and difficult. Samples of the same difficulty level are grouped into the same training batch. Model validation is performed using a multi-scale loss function at three levels: node location prediction, regional growth trend, and stress distribution rationality. Verification metrics such as accuracy, recall, and F1 value are obtained. Backpropagation is used to optimize network parameters to obtain a crack growth prediction model.

[0048] In this embodiment, by constructing a bidirectional message passing mechanism and a multi-layer graph convolutional network, the complex nonlinear characteristics of crack propagation can be effectively captured, thereby improving the accuracy of crack propagation prediction. The use of adaptive distance thresholds and similarity dynamic thresholds can adapt to the node distribution and feature differences of different crack scenarios, thereby improving the generalization ability of the model. The end-to-end training method is adopted to avoid the tedious feature engineering and model selection process, thereby improving computational efficiency.

[0049] S3. Solve the crack propagation prediction model according to the swarm intelligence optimization algorithm, and calculate the crack propagation rate and direction. If the crack propagation rate exceeds the preset rate threshold, the control system adjusts the operating parameters of the silicon-based new material equipment according to the predictive control algorithm. According to the crack propagation direction, the sampling frequency of the sensor is adjusted through an adaptive sampling strategy. The crack propagation rate and propagation direction data are used to update the deep neural network model parameters through the incremental learning method. Real-time monitoring is performed based on the updated deep neural network model to obtain a monitoring data set.

[0050] In an optional embodiment, The crack growth prediction model is solved using a swarm intelligence optimization algorithm to calculate the crack growth rate and direction. If the crack growth rate exceeds the preset rate threshold, the control system adjusts the operating parameters of the silicon-based new material equipment according to the predictive control algorithm. Based on the crack growth direction, the sensor sampling frequency is adjusted using an adaptive sampling strategy. The crack growth rate and growth direction data are used to update the deep neural network model parameters through an incremental learning method. Real-time monitoring is performed based on the updated deep neural network model, resulting in a monitoring data set including: Initialize a population individual set, randomly generate a vector containing a crack growth rate and a growth direction as an initial population individual, calculate the fitness value of each initial population individual based on stress distribution data and crack morphology characteristics, the fitness value including a prediction error term, a stress field consistency term, and an energy dissipation term, obtain an iterative prediction deviation, calculate a target weight coefficient based on the iterative prediction deviation, and assign the target weight coefficient to the prediction error term, the stress field consistency term, and the energy dissipation term; A swarm intelligence optimization algorithm is used to solve the crack growth prediction model, calculate the crack growth rate and growth direction, and a hybrid mutation operator is designed to perform mutation operations on the individuals in the initial population; Comparing the crack growth rate with a preset rate threshold, and when the crack growth rate exceeds the preset rate threshold, adjusting the operating parameters of the silicon-based new material equipment through a predictive control algorithm; dynamically partitioning the monitoring area according to the growth direction, dividing the monitoring area into a high-risk area, a transition area, and a safe area; Adopting an adaptive sampling strategy to adjust the sampling frequency of the sensor, the working mode of the sensor array is determined according to the degree of danger, and the stress signal mutation index is calculated in real time. When the stress signal mutation index exceeds the preset mutation threshold, the adjacent sensor group is switched to the highest sampling frequency mode; Inputting the crack growth rate and the growth direction into an incremental learning module, constructing a training data set using a sliding window mechanism, extracting features, constructing a hybrid loss function using the original deep neural network model prediction results as soft labels, and incrementally updating the deep neural network model parameters based on the hybrid loss function; The deep neural network model is used for real-time monitoring to obtain stress field data, temperature field data and acoustic emission signal data. The stress field data, temperature field data and acoustic emission signal data are adaptively filtered and denoised. A sliding window method is used for feature extraction, and the extracted features are input into the deep neural network model to obtain a monitoring data set.

[0051] Crack propagation prediction is performed based on swarm intelligence optimization algorithm and deep learning, which is used to monitor and control crack propagation of silicon-based new material equipment in real time.

[0052] Initialize the crack growth prediction model. A set of vectors is randomly generated, each containing two components: crack growth rate and growth direction. These vectors constitute the initial population of individuals. Then, based on stress distribution data and crack morphological characteristics (such as crack length, width, and shape) acquired during equipment operation, the fitness value of each individual in the initial population is calculated. The fitness value consists of three components: a prediction error term, a stress field consistency term, and an energy dissipation term. The prediction error term reflects the deviation between the predicted and actual values; the stress field consistency term reflects the consistency between the predicted crack growth direction and the actual stress field distribution; and the energy dissipation term reflects the rationality of energy release during crack growth. The difference between the fitness value of the individual in the previous generation of population and the fitness value of the individual in the current generation is calculated as the iterative prediction deviation. Based on the magnitude of the iterative prediction deviation, the target weight coefficient is dynamically adjusted and assigned to the prediction error term, stress field consistency term, and energy dissipation term, respectively, to balance the importance of the three indicators at different iteration stages. For example, in the initial stage, the weight coefficient of the prediction error term can be set larger, while in the later stage, the weight coefficients of the stress field consistency term and the energy dissipation term can be set larger.

[0053] A swarm intelligence optimization algorithm is used to solve the crack growth prediction model. The particle swarm optimization algorithm is used as an example. The initial swarm individuals are considered particles. Each particle moves in the solution space, continuously updating its position and velocity. During each iteration, the particle adjusts its movement direction and velocity based on its historical best position and the global best position, ultimately approaching the optimal solution. To improve the algorithm's search efficiency, a hybrid mutation operator is designed to mutate the particles. This hybrid mutation operator can combine multiple mutation strategies, such as Gaussian mutation and Cauchy mutation, to enhance the algorithm's global and local search capabilities. Through iterative calculations using the particle swarm optimization algorithm, predicted values ​​for the crack growth rate and direction are ultimately obtained. For example, assume a swarm contains 100 particles, each representing a possible combination of crack growth rate and direction. After 100 iterations, the algorithm converges to an optimal solution, predicting a crack growth rate of 0.1 mm / s and a crack growth direction of 45 degrees.

[0054] After obtaining the predicted values ​​of the crack propagation rate and propagation direction, the crack propagation rate is compared with the preset rate threshold. If the crack propagation rate exceeds the preset rate threshold, it indicates that the risk of crack propagation is high and the operating parameters of the silicon-based new material equipment need to be adjusted. For example, the operating power of the equipment can be reduced, the operating temperature of the equipment can be adjusted, etc. to slow down the crack propagation rate. According to the predicted crack propagation direction, the monitoring area is divided into a high-risk area, a transition area, and a safe area. For example, if the predicted crack propagation direction is 45 degrees, the area near the 45-degree direction can be divided into a high-risk area, the area farther away from the high-risk area can be divided into a safe area, and the area between the high-risk area and the safe area can be divided into a transition area.

[0055] An adaptive sampling strategy adjusts the sensor sampling frequency based on the crack propagation direction and the resulting monitoring area. In high-risk areas, the sensor sampling frequency is set to a higher value to obtain more detailed monitoring data; in transitional areas, the sensor sampling frequency is set to a medium value; and in safe areas, the sensor sampling frequency is set to a lower value to conserve resources. The stress signal mutation index, which reflects the degree of stress signal change, is calculated in real time. When the stress signal mutation index exceeds a preset mutation threshold, it indicates that a sudden change in crack propagation may occur, requiring improved monitoring accuracy. At this point, the adjacent sensor group is switched to the highest sampling frequency mode to capture the sudden change in crack propagation. For example, if the preset mutation threshold is 10, when the stress signal mutation index of a sensor reaches 12, that sensor and its surrounding sensors are switched to the highest sampling frequency mode.

[0056] The predicted crack growth rate and direction are input into the incremental learning module to update the deep neural network model. A sliding window mechanism is used to construct a training dataset, adding the latest monitoring data to the training dataset and removing the oldest monitoring data. Features such as stress field features, temperature field features, and acoustic emission signal features are extracted from the training dataset. The prediction results of the original deep neural network model are used as soft labels, and a hybrid loss function is constructed together with the actual monitoring data. The hybrid loss function includes hard label loss and soft label loss. The hard label loss reflects the deviation between the model prediction value and the actual value, while the soft label loss reflects the deviation between the model prediction value and the original model prediction value. Based on the hybrid loss function, the parameters of the deep neural network model are incrementally updated to improve the model's prediction accuracy and generalization ability.

[0057] An updated deep neural network model is used for real-time monitoring. Stress field data, temperature field data, and acoustic emission signal data are acquired from sensors. Adaptive filtering and noise reduction are performed on these data to remove noise interference. A sliding window approach is used for feature extraction, which is then fed into the deep neural network model to generate a monitoring dataset, including predicted values ​​for crack growth rate and direction.

[0058] In this embodiment, by combining the swarm intelligence optimization algorithm and the deep learning method, the crack propagation rate and direction can be predicted more accurately, providing protection for the safe operation of the equipment. Through the adaptive sampling strategy and predictive control algorithm, the crack propagation status can be monitored in real time, and the equipment operating parameters can be dynamically adjusted according to the crack propagation situation to effectively control the crack propagation risk. Through the incremental learning method, the parameters of the deep neural network model can be continuously updated so that it can adapt to the ever-changing working conditions and improve the robustness and generalization ability of the model.

[0059] In an optional embodiment, Inputting the crack growth rate and the growth direction into an incremental learning module, constructing a training data set using a sliding window mechanism, extracting features, constructing a hybrid loss function using the original deep neural network model prediction results as soft labels, and incrementally updating the deep neural network model parameters based on the hybrid loss function includes: receiving a crack growth rate and a crack growth direction, normalizing the crack growth rate and the crack growth direction, removing outliers using a median filter, and performing noise reduction using a wavelet transform; A sliding window mechanism is used to construct a training data set. The basic window length is set to one-tenth of the original data length, and the window sliding step is set to one-quarter of the basic window length. Time domain statistical features are extracted within each window, and the signal is subjected to Fourier transform to extract frequency domain features. Instantaneous features are obtained through Hilbert transform. Calculating the mutual information values ​​between the time-domain statistical features, the frequency-domain features, the instantaneous features, and the target variable, and constructing a feature importance score based on the correlation coefficients between the features. The feature importance score simultaneously considers the mutual information values ​​between the features and the target variable and the redundancy between the features. Based on the feature importance score, the features are hierarchically screened, first selecting the features with the highest scores, and then sequentially selecting features with lower correlations with the selected features and higher scores, until the cumulative importance of the feature subset reaches a preset threshold; The prediction results of the original deep neural network model are used as soft labels. The soft labels are combined with the real labels in a mixing ratio of 3 to 7 to construct a hybrid loss function. The underlying feature extraction part of the deep neural network model is frozen, and only the parameters of the fully connected layer are updated. Then all layers are unfrozen, and the entire network is fine-tuned with a randomly selected learning rate less than 0.01, and the learning rate is adjusted by exponential decay. Elastic network regularization is applied to the network weights, and a random dropout strategy is used. The dropout ratio of the random dropout strategy decreases layer by layer from the input layer to the output layer. Batch normalization is used to maintain the intermediate feature distribution. The prediction error is calculated on the validation set. When the prediction error exceeds a preset threshold, the proportion of historical data in the training set is increased. Training is stopped when there is no improvement in performance after multiple rounds of verification, and an updated deep neural network model is obtained.

[0060] The collected crack growth rate and direction data are preprocessed. These data are normalized, for example, by scaling them to between 0 and 1 to eliminate the effects of different dimensions. Then, a median filter is used to remove outliers. For example, a median filter with a window size of 5 is used to replace each data point with the median of its neighborhood. Finally, noise reduction is performed using a wavelet transform. For example, a 5-layer decomposition using the db4 wavelet is used to remove high-frequency noise. Assume that the crack growth rate data is [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0] and the growth direction data is [30, 32, 35, 33, 36, 38, 40, 39, 42, 41].

[0061] A sliding window mechanism is used to construct the training dataset. The base window length is set to one-tenth the original data length. For example, if the original data length is 10, the base window length is 1. The window sliding step is set to one-quarter the base window length. For example, if the base window length is 1, the sliding step is 0.25. Within each window, time-domain statistical features such as mean, standard deviation, maximum, and minimum are extracted. A Fourier transform is performed on the signal to extract frequency-domain features such as amplitude, phase, and frequency. A Hilbert transform is used to obtain instantaneous features such as instantaneous frequency and amplitude.

[0062] Perform feature selection. Calculate the mutual information between time-domain statistical features, frequency-domain features, and instantaneous features and the target variable (e.g., future crack growth rate or direction). Construct a feature importance score based on the correlation coefficients between features. This score considers both the correlation between the feature and the target variable and the redundancy between features. Perform a hierarchical screening of features based on the feature importance scores. First, select the features with the highest scores, e.g., the features with the highest mutual information values ​​with the target variable and low correlation with other features. Then, select features with lower correlations with the selected features and higher scores, until the cumulative importance of the feature subset reaches a preset threshold, e.g., 95%.

[0063] Perform incremental model updates. Use the original deep neural network model's predictions as soft labels, and combine them with the true labels in a 3:7 ratio to construct a hybrid loss function. For example, if the true label is 1 and the soft label is 0.8, the hybrid label is 0.7*1+0.3*0.8=0.94. Freeze the underlying feature extraction layer of the deep neural network model, updating only the fully connected layer parameters. Then, unfreeze all layers and fine-tune the entire network with a randomly selected learning rate less than 0.01, for example, 0.005. Adjust the learning rate using exponential decay, for example, by a factor of 0.9 every 100 epochs.

[0064] Elastic net regularization is applied to the network weights to prevent overfitting. A random dropout strategy is used, with the dropout ratio decreasing layer by layer from the input layer to the output layer. For example, the dropout ratio for the input layer is 0.5, and the dropout ratio for the output layer is 0.1. Batch normalization is used to maintain a stable distribution of intermediate features. The prediction error is calculated on the validation set. When the prediction error exceeds a preset threshold, for example, 0.05, the proportion of historical data in the training set is increased, for example, from 20% to 30%. Training is terminated when there is no improvement in performance after multiple rounds of validation, resulting in an updated deep neural network model.

[0065] In this embodiment, by combining new data and original model knowledge, the incremental learning method can effectively improve the accuracy of crack propagation prediction. Compared with retraining the model, the incremental learning method can significantly reduce training time and computing resource consumption. By continuously learning new data, the incremental learning method can enable the model to better adapt to the ever-changing crack propagation pattern and improve the generalization ability of the model.

[0066] Figure 2 FIG. 1 is a schematic diagram of the structure of an online monitoring system for crack propagation of silicon-based new material equipment under high temperature environment according to an embodiment of the present invention. Figure 2 As shown, the system includes: The first unit is used to collect stress distribution data on the surface of the silicon-based new material device through multiple stress sensors evenly arranged along the circumference of the surface of the silicon-based new material device, and perform wavelet transform denoising on the stress distribution data to obtain the denoised stress distribution data and input it into the attention mechanism deep neural network model. The spatial attention mechanism is used to process the feature weights of different positions and the channel attention mechanism is combined to extract the correlation between feature channels. The residual backbone network is used to fine-tune the historical crack growth data to generate a stress field distribution map of the silicon-based new material device; The second unit is used to segment the stress field distribution map using a region growing algorithm, identify stress concentration areas, and compare the stress values ​​with a preset stress threshold. If the stress value in the stress concentration area exceeds the preset stress threshold, the temperature distribution data of the current area is obtained, and the temperature distribution characteristics are established using a deep hybrid probability model. The probability model parameters are estimated using a variational inference method. The extracted temperature distribution characteristics and stress field distribution map are then subjected to multimodal data fusion. The spatiotemporal characteristic relationship of crack propagation is captured through a graph-structured neural network to establish a crack propagation prediction model. The third unit is used to solve the crack propagation prediction model based on the swarm intelligence optimization algorithm and calculate the crack propagation rate and direction. If the crack propagation rate exceeds the preset rate threshold, the control system adjusts the operating parameters of the silicon-based new material equipment according to the predictive control algorithm. According to the crack propagation direction, the sampling frequency of the sensor is adjusted through an adaptive sampling strategy. The crack propagation rate and propagation direction data are used to update the deep neural network model parameters through the incremental learning method. Real-time monitoring is performed based on the updated deep neural network model to obtain a monitoring data set.

[0067] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0068] According to a fourth aspect of the embodiments of the present invention, A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0069] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An online monitoring method for crack propagation of silicon-based new material equipment under high temperature environment, characterized in that: include: Multiple stress sensors are evenly arranged along the circumference of the surface of the silicon-based new material device to collect stress distribution data on the surface of the silicon-based new material device. The stress distribution data is then subjected to wavelet transform denoising to obtain the denoised stress distribution data and input into a deep neural network model with an attention mechanism. The spatial attention mechanism is used to process the feature weights of different positions and the channel attention mechanism is combined to extract the correlation between feature channels. The residual backbone network is then used to fine-tune the historical crack growth data to generate a stress field distribution map of the silicon-based new material device. The stress field distribution map is segmented using a region growing algorithm to identify stress concentration areas and compare the stress values ​​with a preset stress threshold. If the stress value in the stress concentration area exceeds the preset stress threshold, the temperature distribution data of the current area is obtained. The temperature distribution characteristics are established using a deep hybrid probability model. The probability model parameters are estimated using a variational inference method. The extracted temperature distribution characteristics and stress field distribution map are then subjected to multimodal data fusion. The spatiotemporal characteristic relationship of crack propagation is captured using a graph-structured neural network to establish a crack propagation prediction model. The crack propagation prediction model is solved according to the swarm intelligence optimization algorithm to calculate the crack propagation rate and direction. If the crack propagation rate exceeds the preset rate threshold, the control system adjusts the operating parameters of the silicon-based new material equipment according to the predictive control algorithm. According to the crack propagation direction, the sampling frequency of the sensor is adjusted through an adaptive sampling strategy. The crack propagation rate and propagation direction data are used to update the deep neural network model parameters through the incremental learning method. Real-time monitoring is performed based on the updated deep neural network model to obtain a monitoring data set.

2. The method according to claim 1, characterized in that Multiple stress sensors are evenly arranged along the circumference of the surface of the silicon-based new material device to collect stress distribution data on the surface of the silicon-based new material device. The stress distribution data is denoised by wavelet transform to obtain the denoised stress distribution data and input into the attention mechanism deep neural network model. The spatial attention mechanism is used to process the feature weights of different positions and the channel attention mechanism is combined to extract the correlation between feature channels. The residual backbone network is used to fine-tune the historical crack growth data to generate the stress field distribution map of the silicon-based new material device, including: Stress sensors are placed on the surface of silicon-based new material equipment to collect stress distribution data. The collected stress distribution data is decomposed into multiple layers using an adaptive wavelet transform algorithm. The optimal wavelet basis function is determined based on the energy distribution of the decomposed signal. Dynamic threshold processing is performed on high-frequency coefficients of different scales. The threshold parameters are adaptively adjusted based on the local characteristics of the signal. The processed high-frequency coefficients and low-frequency coefficients are then subjected to adaptive weighted wavelet reconstruction to obtain noise-reduced stress distribution data. The denoised stress distribution data is input into the attention mechanism deep neural network model. The multi-scale features of the stress distribution are extracted layer by layer through a multi-layer convolutional structure. The feature weights at different positions of the feature map are calculated using the spatial attention mechanism. The correlation between feature channels is extracted using the channel attention mechanism. The spatial feature weights and channel correlations are adaptively fused to construct an attention feature map. The attention feature map is then weightedly combined with the original feature map through the feature fusion module. The residual backbone network is divided into a basic feature extraction unit and a deep feature learning unit. The network parameters of the basic feature extraction unit are fixed, and the parameters of the deep feature learning unit are fine-tuned by migrating historical crack growth data. The historical crack growth condition data and the corresponding stress distribution characteristics are divided into a training sample set and a validation sample set according to an adaptive stratified sampling strategy. Based on the training sample set, the network model is trained through a phased dynamic learning strategy. In the first phase, a larger learning rate is used to optimize the deep feature learning unit. In the second phase, the learning rate is reduced to fine-tune the network parameters. In the third phase, the learning rate is further reduced to improve the generalization performance of the model. An early stopping judgment mechanism is set in each training phase, and the verification sample set is used for verification. If there is no improvement in the performance after multiple rounds of verification, the learning rate is reduced, and the training is repeated to obtain a trained attention mechanism deep neural network model. The trained attention mechanism deep neural network model is used to process the denoised stress distribution data to generate a stress field distribution map corresponding to the silicon-based new material device.

3. The method according to claim 2, characterized in that The residual backbone network is divided into a basic feature extraction unit and a deep feature learning unit. The network parameters of the basic feature extraction unit are fixed, and the parameters of the deep feature learning unit are fine-tuned by migrating historical crack growth data. The historical crack growth condition data and the corresponding stress distribution characteristics are divided into a training sample set and a validation sample set according to an adaptive stratified sampling strategy. The residual backbone network is divided into a basic feature extraction unit and a deep feature learning unit. The basic feature extraction unit adopts a multi-level cascade convolution layer structure to extract the basic texture features of stress distribution data layer by layer through convolution kernels of different receptive field sizes. The deep feature learning unit adopts a parallel multi-scale convolution layer structure to extract features from input features through convolution branches of different scales. The features are weighted and fused through the channel attention module to obtain high-level semantic features of crack propagation. A bidirectional residual connection structure is constructed in the deep feature learning unit, shallow features are transmitted through the forward path, deep semantic information is fed back through the reverse path, a gating unit is set to regulate the flow of information, and a feature adaptive fusion module is used to calculate the weights of features at different levels according to the spatial distribution of feature responses, and the parameters of the deep feature learning unit are fine-tuned and trained; A dynamic window sliding method is used to segment the historical crack growth data. The window size is adjusted according to the crack growth rate. A multi-scale feature pyramid is constructed for the stress distribution feature map, local texture features and global structural features are extracted, and the crack growth annotation information is converted into a probability density distribution heat map. Based on the probability density distribution heat map, the training data was stratified in combination with the temporal evolution characteristics and spatial distribution characteristics of crack propagation. The feature similarity between samples was calculated, and the categories were divided using the spectral clustering method. The importance weight was calculated according to the distance between the sample and the category center. The training sample set and the validation sample set were divided based on the importance weight.

4. The method according to claim 1, wherein The stress field distribution map is segmented using a region growing algorithm to identify stress concentration areas and compare the stress value with a preset stress threshold. If the stress value in the stress concentration area exceeds the preset stress threshold, the temperature distribution data of the current area is obtained. The temperature distribution characteristics are established using a deep hybrid probability model. The probability model parameters are estimated using a variational inference method. The extracted temperature distribution characteristics and stress field distribution map are then subjected to multimodal data fusion. The spatiotemporal characteristic relationship of crack propagation is captured through a graph-structured neural network to establish a crack propagation prediction model, including: The stress field distribution map is segmented using an adaptive region growing algorithm. Local maximum points are extracted from the stress field distribution map as seed points. A growth threshold is dynamically determined based on the stress value differences between adjacent pixels. Based on the growth threshold, regions with similar stress values ​​are merged through region growing iterations to obtain multiple stress concentration regions. The average stress value of each stress concentration region is calculated and compared with a preset stress threshold to identify high-risk areas. An infrared thermal imager is used to obtain temperature distribution data of the high-risk area. The temperature distribution data is spatially registered with the stress field distribution map. A deep hybrid probability model is constructed containing multiple Gaussian distribution components, each corresponding to a different temperature pattern characteristic. A variational posterior distribution is established to approximate the true posterior distribution. The probability model parameters are iteratively optimized by maximizing the lower bound of evidence to obtain a probabilistic representation of the temperature distribution. A dual-branch feature fusion network is constructed. The temperature feature extraction branch of the dual-branch feature fusion network uses a fully connected layer to reduce the dimensionality of the probability model parameters, and the stress feature extraction branch uses a convolutional layer to extract the spatial features of the stress field. The temperature features and stress features are adaptively weighted and fused through a feature attention mechanism to obtain multimodal features that characterize the characteristics of the crack region. A multi-layer graph convolutional network is constructed, and the multimodal features are used as attributes of the graph network nodes. Connection relationships are established based on the spatial distance and feature similarity between nodes. Features are transferred and updated between connected nodes through a message passing mechanism. Residual connections and layer normalization structures are set after each layer of graph convolution. The multi-layer graph convolutional network is trained in an end-to-end manner. The historical crack propagation path is used as supervision information, and the network parameters are optimized through back propagation to obtain a crack propagation prediction model.

5. The method according to claim 4, characterized in that The connection relationship is established based on the spatial distance and feature similarity between nodes. The features are transferred and updated between connected nodes through the message passing mechanism. After each layer of graph convolution, a residual connection and layer normalization structure are set. The multi-layer graph convolution network is trained in an end-to-end manner. The historical crack propagation path is used as supervision information. The network parameters are optimized through back propagation. The crack propagation prediction model is obtained, which includes: Calculating an inter-node distance matrix based on the spatial distance between nodes, setting an adaptive distance threshold, wherein the adaptive distance threshold takes a smaller value in a node-dense area and a larger value in a node-sparse area, establishing initial connections for node pairs whose distance is less than the adaptive distance threshold to obtain an initial connection matrix, calculating the feature similarity of the node feature vector corresponding to each node, setting a dynamic similarity threshold according to the number of network layers, wherein the dynamic similarity threshold gradually increases as the number of network layers deepens, and retaining the node pairs in the initial connection matrix that meet the dynamic similarity threshold to obtain an optimized node connection relationship; A bidirectional message transmission mechanism is constructed. In the forward transmission phase, a multilayer perceptron is used to perform nonlinear mapping on the source node features to obtain forward transformation features. In the backward transmission phase, an inverse mapping network is constructed to obtain reverse transformation features. Based on the optimized node connection relationship, the forward transformation features and the reverse transformation features are respectively input into a bidirectional multi-head attention module to calculate the bidirectional message weights. An adaptive gating mechanism is used to fuse the bidirectional messages to obtain the aggregated features of the target node. A multi-layer graph convolutional network is constructed based on the aggregated features of the target node. The hidden dimension of each layer decreases layer by layer. Message passing is performed in each layer of graph convolution to obtain node update features. The node update features are fused with the original features through residual connections to obtain fused features. The fused features are layer-normalized to obtain standardized features. An end-to-end training approach is adopted. In the pre-training phase, the historical crack propagation path and the standardized features are used to reconstruct the current crack state to obtain a reconstructed feature. In the fine-tuning phase, the reconstructed feature and target scene data are used to predict the crack propagation at the next moment to obtain a predicted feature. Based on the predicted feature, a local prediction error term and a global morphological similarity term are constructed to obtain a multi-scale loss function. Based on the prediction features, the crack propagation speed and the degree of direction change are calculated, the difficulty of the training samples is evaluated and graded to obtain the difficulty level, and the samples of the same difficulty level are organized into the same training batch; based on the multi-scale loss function, the model is verified at three levels: node position prediction, regional expansion trend, and rationality of stress distribution to obtain verification indicators, and the crack propagation prediction model is obtained by optimizing the network parameters through back propagation.

6. The method according to claim 1, characterized in that The crack growth prediction model is solved using a swarm intelligence optimization algorithm to calculate the crack growth rate and direction. If the crack growth rate exceeds the preset rate threshold, the control system adjusts the operating parameters of the silicon-based new material equipment according to the predictive control algorithm. Based on the crack growth direction, the sensor sampling frequency is adjusted using an adaptive sampling strategy. The crack growth rate and growth direction data are used to update the deep neural network model parameters through an incremental learning method. Real-time monitoring is performed based on the updated deep neural network model, resulting in a monitoring data set including: Initialize a population individual set, randomly generate a vector containing a crack growth rate and a growth direction as an initial population individual, calculate the fitness value of each initial population individual based on stress distribution data and crack morphology characteristics, the fitness value including a prediction error term, a stress field consistency term, and an energy dissipation term, obtain an iterative prediction deviation, calculate a target weight coefficient based on the iterative prediction deviation, and assign the target weight coefficient to the prediction error term, the stress field consistency term, and the energy dissipation term; A swarm intelligence optimization algorithm is used to solve the crack growth prediction model, calculate the crack growth rate and growth direction, and a hybrid mutation operator is designed to perform mutation operations on the individuals in the initial population; Comparing the crack growth rate with a preset rate threshold, and when the crack growth rate exceeds the preset rate threshold, adjusting the operating parameters of the silicon-based new material equipment through a predictive control algorithm; dynamically partitioning the monitoring area according to the growth direction, dividing the monitoring area into a high-risk area, a transition area, and a safe area; Adopting an adaptive sampling strategy to adjust the sampling frequency of the sensor, the working mode of the sensor array is determined according to the degree of danger, and the stress signal mutation index is calculated in real time. When the stress signal mutation index exceeds the preset mutation threshold, the adjacent sensor group is switched to the highest sampling frequency mode; Inputting the crack growth rate and the growth direction into an incremental learning module, constructing a training data set using a sliding window mechanism, extracting features, constructing a hybrid loss function using the original deep neural network model prediction results as soft labels, and incrementally updating the deep neural network model parameters based on the hybrid loss function; The deep neural network model is used for real-time monitoring to obtain stress field data, temperature field data and acoustic emission signal data. The stress field data, temperature field data and acoustic emission signal data are adaptively filtered and denoised. A sliding window method is used for feature extraction, and the extracted features are input into the deep neural network model to obtain a monitoring data set.

7. The method according to claim 6, characterized in that Inputting the crack growth rate and the growth direction into an incremental learning module, constructing a training data set using a sliding window mechanism, extracting features, constructing a hybrid loss function using the original deep neural network model prediction results as soft labels, and incrementally updating the deep neural network model parameters based on the hybrid loss function includes: receiving a crack growth rate and a crack growth direction, normalizing the crack growth rate and the crack growth direction, removing outliers using a median filter, and performing noise reduction using a wavelet transform; A sliding window mechanism is used to construct a training data set. The basic window length is set to one-tenth of the original data length, and the window sliding step is set to one-quarter of the basic window length. Time domain statistical features are extracted within each window, and the signal is subjected to Fourier transform to extract frequency domain features. Instantaneous features are obtained through Hilbert transform. Calculating the mutual information values ​​between the time-domain statistical features, the frequency-domain features, the instantaneous features, and the target variable, and constructing a feature importance score based on the correlation coefficients between the features. The feature importance score simultaneously considers the mutual information values ​​between the features and the target variable and the redundancy between the features. Based on the feature importance score, the features are hierarchically screened, first selecting the features with the highest scores, and then sequentially selecting features with lower correlations with the selected features and higher scores, until the cumulative importance of the feature subset reaches a preset threshold; The prediction results of the original deep neural network model are used as soft labels. The soft labels are combined with the real labels in a mixing ratio of 3 to 7 to construct a hybrid loss function. The underlying feature extraction part of the deep neural network model is frozen, and only the parameters of the fully connected layer are updated. Then all layers are unfrozen, and the entire network is fine-tuned with a randomly selected learning rate less than 0.01, and the learning rate is adjusted by exponential decay. Elastic network regularization is applied to the network weights, and a random dropout strategy is used. The dropout ratio of the random dropout strategy decreases layer by layer from the input layer to the output layer. Batch normalization is used to maintain the intermediate feature distribution. The prediction error is calculated on the validation set. When the prediction error exceeds a preset threshold, the proportion of historical data in the training set is increased. Training is stopped when there is no improvement in performance after multiple rounds of verification, and an updated deep neural network model is obtained.

8. An online crack growth monitoring system for silicon-based new material equipment under high temperature environment, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to collect stress distribution data on the surface of the silicon-based new material device through multiple stress sensors evenly arranged along the circumference of the surface of the silicon-based new material device, and perform wavelet transform denoising on the stress distribution data to obtain the denoised stress distribution data and input it into the attention mechanism deep neural network model. The spatial attention mechanism is used to process the feature weights of different positions and the channel attention mechanism is combined to extract the correlation between feature channels. The residual backbone network is used to fine-tune the historical crack growth data to generate a stress field distribution map of the silicon-based new material device; The second unit is used to segment the stress field distribution map using a region growing algorithm, identify stress concentration areas, and compare the stress values ​​with a preset stress threshold. If the stress value in the stress concentration area exceeds the preset stress threshold, the temperature distribution data of the current area is obtained, and the temperature distribution characteristics are established using a deep hybrid probability model. The probability model parameters are estimated using a variational inference method. The extracted temperature distribution characteristics and stress field distribution map are then subjected to multimodal data fusion. The spatiotemporal characteristic relationship of crack propagation is captured through a graph-structured neural network to establish a crack propagation prediction model. The third unit is used to solve the crack propagation prediction model based on the swarm intelligence optimization algorithm and calculate the crack propagation rate and direction. If the crack propagation rate exceeds the preset rate threshold, the control system adjusts the operating parameters of the silicon-based new material equipment according to the predictive control algorithm. According to the crack propagation direction, the sampling frequency of the sensor is adjusted through an adaptive sampling strategy. The crack propagation rate and propagation direction data are used to update the deep neural network model parameters through the incremental learning method. Real-time monitoring is performed based on the updated deep neural network model to obtain a monitoring data set.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

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

Images