Plywood gluing defect intelligent identification method and system based on deep learning

Abstract

The present application relates to the technical field of intelligent manufacturing, and more particularly to a plywood gluing defect intelligent identification method and system based on deep learning. The multi-source data on the gluing surface of the plywood is subjected to parameter inversion and separation of geometric, thermal and reflection characteristic components, the variational autoencoder is mapped to a characteristic embedding space and independent constraints and mutual information loss are applied to obtain a physical decoupling embedding vector, three encoders share a decoupling loss for joint training; a causal relationship graph is constructed to calculate the causal effect strength of the process parameters on defect generation, the physical decoupling embedding vector is associated and matched with the causal effect strength to generate a defect candidate area; a time series enhancement characteristic embedding vector is obtained by jointly optimizing the time series prediction and time series reconstruction tasks, and then distance measurement and fusion matching are performed with a multi-granularity prototype feature library to output a defect identification result. The present application realizes high-precision physical decoupling and time series evolution modeling of defect characteristics, and significantly improves the recognition accuracy and robustness.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology, and in particular to a method and system for intelligent identification of adhesive defects in plywood based on deep learning. Background Technology

[0002] In the plywood gluing process, the identification of gluing defects is a crucial step in ensuring the bonding quality and mechanical properties of the plywood. Current conventional practices primarily rely on detection schemes combining machine vision technology and deep learning models. This typically involves using sensors such as industrial cameras or thermal imagers to acquire images or signal data of the gluing surface. After preprocessing, high-dimensional features are extracted using deep networks such as convolutional neural networks, and then defect localization and type identification are performed based on classifiers or object detection networks. Some technical solutions employ multimodal data fusion, stitching together or cascading features from different sources such as visible light images, infrared thermal images, or laser point clouds to obtain richer defect characterization information. Furthermore, some systems introduce temporal sampling mechanisms into the production line, analyzing the dynamic changes of the gluing surface through inter-frame difference or optical flow methods to assist in identifying defects that occur in real time.

[0003] Data collected by multi-source sensors often contains feature information of different physical properties, such as geometric contours, thermal conduction anomalies, and reflectivity changes. These features differ fundamentally in physical meaning and dimensions. However, current solutions typically input multimodal data directly into a unified feature extraction network or perform simple concatenation in the fusion layer, lacking separation and constraint on the physical source of the features. This leads to the entanglement of features of different physical properties in the embedding space, causing feature redundancy and making it difficult for the model to focus on the essential attributes directly related to the defect type, thereby reducing recognition accuracy and noise immunity.

[0004] Existing deep learning models primarily rely on statistical correlations in data samples for defect identification, neglecting the deep causal relationships between coating process parameters and defect generation mechanisms. For example, variations in process parameters such as coating pressure, adhesive viscosity, and coating speed directly lead to specific types of defects, but the model only learns apparent correlations from image features, failing to distinguish between causal dependencies and irrelevant spurious correlations. When process parameters shift, the model's generalization ability significantly decreases because it cannot understand the driving effect of parameter changes on defects. Furthermore, most existing methods are based on static analysis of single frames or short-term images, lacking modeling of the temporal evolution of defects during the coating process. This results in an inability to predict defect development trends or distinguish between transient anomalies and permanent defects, limiting the proactiveness and timeliness of online quality control. Summary of the Invention

[0005] This invention provides a method and system for intelligent identification of adhesive defects in plywood based on deep learning, which can solve the problems in the prior art.

[0006] A first aspect of the present invention provides a deep learning-based intelligent identification method for adhesive defects in plywood, comprising: Frame-by-frame parameter inversion is performed on the continuous time-series multi-source data of the glued surface of plywood to separate geometric feature components, thermal feature components and reflection feature components. These are then mapped to the feature embedding space by a variational autoencoder. In the feature embedding space, independence constraints are applied to each feature component to make the feature components of different physical properties orthogonally distributed. Redundant dependencies between feature components are suppressed by minimizing mutual information loss, resulting in a physically decoupled feature embedding vector. Construct a causal relationship map characterizing the causal relationship between process parameters and defect generation mechanisms, and calculate the causal effect intensity of each process parameter on defect generation based on the causal relationship map; The physical decoupled feature embedding vector is associated and matched with the causal effect intensity to identify the feature subspace affected by the causal influence of specific process parameters and generate a defect-sensitive feature map characterizing the spatial distribution of defects. Connected regions are extracted from the defect-sensitive feature map as defect candidate regions. A temporal prediction task and a temporal reconstruction task are constructed. The temporal prediction task predicts the feature evolution state of the defect candidate region at the next time step, and the temporal reconstruction task reconstructs the feature of the defect region of the occluded part at the current time step. By jointly optimizing the reconstruction error of the time-series prediction task and the time-series reconstruction task, and learning the temporal consistency feature embedding vector of the evolution law of defects in the time dimension, a time-enhanced defect feature embedding vector is obtained. The distance metric between the temporally enhanced defect feature embedding vector and the fine-grained, medium-grained, and coarse-grained prototype feature vectors in the multi-granularity prototype feature library is calculated. The matching results of different granularities are fused according to the confidence weight and a voting decision is made to obtain the defect identification result. The defect identification result and the corresponding abnormal process parameters are sent to the glue coating machine control system to realize the online control of the process parameters.

[0007] Frame-by-frame parameter inversion is performed on continuous temporal multi-source data of the glued surface of plywood to separate geometric feature components, thermal feature components, and reflection feature components. These components are then mapped to the feature embedding space via a variational autoencoder, including: Based on the thermal conduction model and optical reflection model of the adhesive material, the physical parameters of the data vector at each spatial location in the multi-source data are solved, and an inversion mapping relationship from the observed data vector to the physical parameter vector is established to obtain the spatially distributed physical parameter field. The geometric parameter tensor characterizing the spatial distribution of the coating thickness is extracted from the physical parameter field as the geometric feature component, the thermal parameter tensor characterizing the spatial distribution of the temperature field of the coating material is extracted as the thermal feature component, and the reflection parameter tensor characterizing the spatial distribution of the surface optical response is extracted as the reflection feature component. For the geometric feature components, the thermal feature components, and the reflection feature components, corresponding variational autoencoders are constructed respectively, and each feature component is input into the corresponding encoder. Each encoder outputs probability distribution parameters containing a mean vector and a variance vector. The probability distribution parameters output by each encoder are reparameterized and sampled to obtain geometric feature embedding vector, thermal feature embedding vector and reflection feature embedding vector. The geometric feature embedding vector, the thermal feature embedding vector and the reflection feature embedding vector are respectively located in the geometric subspace, thermal subspace and reflection subspace of the feature embedding space.

[0008] In the feature embedding space, independence constraints are applied to each feature component to ensure that feature components of different physical properties are orthogonally distributed. Redundant dependencies between feature components are suppressed by minimizing mutual information loss, resulting in a physically decoupled feature embedding vector comprising: Calculate the covariance matrix between each pair of the geometric feature embedding vector, the thermal feature embedding vector, and the reflection feature embedding vector, and construct the orthogonality constraint loss. The orthogonality constraint loss quantifies the degree of deviation of the off-diagonal elements in the covariance matrix. For all pairwise combinations of the geometric feature embedding vector, the thermal feature embedding vector, and the reflection feature embedding vector, a corresponding joint probability distribution estimation network and a marginal probability distribution estimation network are constructed respectively. Based on the product of the joint probability distribution output by each joint probability distribution estimation network and the marginal probability distribution output by the corresponding marginal probability distribution estimation network, the mutual information between each feature embedding vector pair is calculated, and the mutual information is summed to obtain the total mutual information loss. A joint optimization objective is constructed, which includes a weighted combination of the orthogonality constraint loss and the total mutual information loss. The network parameters of the encoders of the geometric variational autoencoder, the thermal variational autoencoder, and the reflection variational autoencoder are updated by minimizing the joint optimization objective, and the decoupled geometric feature embedding vector, thermal feature embedding vector, and reflection feature embedding vector are obtained as the feature embedding vectors of the physical decoupling.

[0009] Construct a causal relationship map characterizing the causal relationship between process parameters and defect generation mechanisms, and calculate the causal effect strength of each process parameter on defect generation based on the causal relationship map, including: Identify the independent and dependent variable nodes from the process parameters; Based on the change sequence of the physically decoupled feature embedding vector in the time series, where the time series is a physical decoupled feature sequence of more than 3 consecutive frames, the temporal dependency relationship between the independent variable node and the dependent variable node is calculated, and based on the temporal dependency relationship, it is determined whether the change of the independent variable node precedes the change of the dependent variable node in time. For independent variable nodes and dependent variable nodes that satisfy the time-series prior conditions, a conditional independence test mechanism is constructed. By introducing a set of control variables and calculating the difference in conditional probability distributions, the mechanism identifies whether there is a direct causal path between the independent variable nodes and the dependent variable nodes. Using the independent variable nodes and the dependent variable nodes as vertices in the graph structure, and the identified direct causal paths as directed edges in the graph structure, a causal relationship graph is constructed from the process parameter nodes to the defect type nodes. For each directed edge in the causal relationship graph, a numerical perturbation is simulated on the process parameters corresponding to the starting node, and the response change generated when the numerical perturbation propagates to the defect type corresponding to the ending node is calculated; the ratio of the response change to the numerical perturbation is taken as the causal effect strength corresponding to the directed edge.

[0010] Associating and matching the physically decoupled feature embedding vector with the causal effect strength, identifying the feature subspace causally affected by specific process parameters, and generating a defect-sensitive feature map characterizing the spatial distribution of defects, extracting connected regions as defect candidate regions from the defect-sensitive feature map includes: For the causal paths of process parameters and defect types in the causal relationship graph where the intensity of the causal effect exceeds a preset causal effect threshold, the parameter identifiers corresponding to the process parameters and the defect identifiers corresponding to the defect types are extracted to form strong causal association pairs. For each strong causal relationship pair, extract the feature embedding vector corresponding to the process parameter in terms of physical properties from the feature embedding vector of the physical decoupling as the feature vector to be analyzed, and calculate the response intensity distribution of the feature vector to be analyzed in each dimension of the feature embedding space; Based on the response intensity distribution, dimensions with response intensity exceeding a preset response threshold are selected to form a sensitive dimension set, and the feature subspace corresponding to the sensitive dimension set is marked as a feature subspace causally affected by the process parameters. Spatial back projection is performed on the feature embedding vectors in the feature subspace, and the feature embedding vectors are mapped back to the spatial coordinate system of the original coated surface to obtain the correspondence between spatial position and feature response value; Based on the correspondence, spatial locations where the feature response value exceeds the response threshold are marked in the two-dimensional space of the coated surface. The marked areas are then subjected to connectivity analysis and adjacent marked areas are merged to generate a defect-sensitive feature map characterizing the spatial distribution of the defect type. The regions marked in the defect-sensitive feature map are extracted as the defect candidate regions.

[0011] A temporal prediction task and a temporal reconstruction task are constructed. The temporal prediction task predicts the feature evolution state of the defect candidate region at the next time step, and the temporal reconstruction task reconstructs the feature of the occluded defect region at the current time step, including: Defect candidate region samples are extracted from historically collected adhesive surface data and arranged in chronological order to form a time-series training set. For each time-series sample in the time-series training set, the physical decoupling feature embedding vectors corresponding to the defect candidate regions at the current time and the next adjacent time are extracted as the current state features and target evolution features, respectively. A temporal predictive encoder and a temporal predictive decoder are constructed. The temporal predictive encoder receives the current state features and outputs a temporal encoded representation. The temporal predictive decoder receives the temporal encoded representation and outputs the predicted feature evolution state. The network parameters of the temporal predictive encoder and the temporal predictive decoder are updated by calculating the reconstruction error between the predicted feature evolution state and the target evolution feature; For each time-series sample in the time-series training set, a spatial occlusion operation is applied to the feature embedding vector of the physically decoupled defect candidate region at the current time to generate partially visible features that include the occluded region and the visible region. A temporal reconstruction encoder and a temporal reconstruction decoder are constructed. The temporal reconstruction encoder receives the partially visible features and outputs an occlusion-aware encoded representation. The temporal reconstruction decoder receives the occlusion-aware encoded representation and outputs the reconstructed complete features. The network parameters of the temporal reconstruction encoder and the temporal reconstruction decoder are updated by calculating the reconstruction error between the reconstructed complete features and the original complete features.

[0012] By jointly optimizing the reconstruction errors of the temporal prediction task and the temporal reconstruction task, and learning the temporal consistency feature embedding vector of the evolution law of defects in the time dimension, the temporal enhanced defect feature embedding vector is obtained, including: The reconstruction error between the predicted feature evolution state and the target evolution feature is extracted from the time series prediction task and used as the time series prediction loss; The reconstruction error between the reconstructed complete features and the original complete features is extracted from the temporal reconstruction task and used as the temporal reconstruction loss. A joint optimization objective function is constructed, which sums the time series prediction loss and the time series reconstruction loss by weighting them. The contribution of the time series prediction task and the time series reconstruction task to the joint optimization is controlled by adjusting the weight coefficients. Based on the joint optimization objective function, the network parameters of the temporal prediction encoder, the temporal prediction decoder, the temporal reconstruction encoder, and the temporal reconstruction decoder are updated synchronously, so that the temporal prediction task and the temporal reconstruction task share the underlying feature representation space; After the joint optimization objective function converges, the temporal coding representation output by the temporal prediction encoder and the occlusion-aware coding representation output by the temporal reconstruction encoder are extracted and merged into a unified temporal consistency feature embedding vector through feature fusion operation; The temporal consistency feature embedding vector and the physically decoupled feature embedding vector are concatenated to generate a temporally enhanced defect feature embedding vector.

[0013] A second aspect of the present invention provides a deep learning-based intelligent identification system for adhesive defects in plywood, comprising: The physical decoupling unit is used to perform frame-by-frame parameter inversion on continuous time-series multi-source data of the glued surface of plywood, separate geometric feature components, thermal feature components and reflection feature components, and map them to the feature embedding space through variational autoencoder. In the feature embedding space, independence constraints are applied to each feature component to make the feature components of different physical properties orthogonally distributed, and the redundant dependence between feature components is suppressed by minimizing mutual information loss, so as to obtain the physical decoupling feature embedding vector. The causal effect unit is used to construct a causal relationship map characterizing the causal relationship between process parameters and defect generation mechanisms, and to calculate the causal effect intensity of each process parameter on defect generation based on the causal relationship map. Candidate region units are used to associate and match the physical decoupled feature embedding vectors with the causal effect intensity, identify the feature subspace causally affected by specific process parameters and generate a defect-sensitive feature map characterizing the spatial distribution of defects, and extract connected regions from the defect-sensitive feature map as defect candidate regions. The temporal task unit is used to construct a temporal prediction task and a temporal reconstruction task. The temporal prediction task predicts the feature evolution state of the defect candidate region at the next time step, and the temporal reconstruction task reconstructs the feature of the defect region of the occluded part at the current time step. The temporal enhancement unit is used to jointly optimize the reconstruction error of the temporal prediction task and the temporal reconstruction task, learn the temporal consistency feature embedding vector of the evolution law of defects in the time dimension, and obtain the defect feature embedding vector of temporal enhancement. The defect identification unit is used to calculate the distance metric between the time-enhanced defect feature embedding vector and the fine-grained, medium-grained, and coarse-grained prototype feature vectors in the multi-grained prototype feature library. It fuses the matching results of different granularities according to the confidence weight and performs voting decision to obtain the defect identification result. The defect identification result and the corresponding abnormal process parameters are sent to the glue coating machine control system to realize the online control of the process parameters.

[0014] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0015] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0016] The beneficial effects of this application are as follows: The physically decoupled feature embedding vectors, through independence constraints and mutual information loss, ensure orthogonal distribution of feature components of different physical properties, completely eliminating redundant dependencies between geometric, thermal, and reflection features. This design significantly improves the efficiency of multi-source data fusion, makes defect feature representation physically interpretable, and effectively overcomes the misjudgment problem caused by feature coupling in traditional methods for identifying defects on glued plywood surfaces. It is particularly sensitive to minute defects against complex texture backgrounds.

[0017] Causal relationship mapping quantifies the causal effect of process parameters on defect generation. This quantification is then correlated and matched with physical decoupling features to generate a defect-sensitive feature map. This mechanism accurately locates the feature subspace causally influenced by specific process parameters, revealing not only the physical root causes of defects but also enhancing model interpretability. This helps process engineers quickly locate abnormal parameters while reducing noise interference introduced by irrelevant parameters, significantly improving the accuracy and robustness of defect candidate regions.

[0018] Joint optimization of temporal prediction and temporal reconstruction tasks ensures accurate capture of the evolutionary patterns of defects over time. The learned temporal consistency feature embedding vectors enable the model to predict defect expansion trends and reconstruct occluded areas, significantly enhancing recognition stability during dynamic glue application. Combining distance metrics and fusion strategies with a multi-granularity prototype feature library, the complementary and synergistic matching results of prototypes at different granularities comprehensively improve the fine-grained classification accuracy and recall rate of plywood glue application defects, achieving comprehensive intelligent defect recognition from macro to micro levels.

[0019] This invention can achieve an online detection speed of over 3m / min, with a minimum detectable defect size of 0.3mm, a missed detection rate of ≤0.5%, a false detection rate of ≤1%, and is compatible with the production cycle of mainstream plywood gluing machines. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating a deep learning-based intelligent identification method for adhesive defects in plywood. Figure 2 A schematic diagram of the process for generating embedding vectors for decoupled physical features. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0023] Figure 1 This is a flowchart illustrating the intelligent identification method for plywood adhesive defects based on deep learning, as described in an embodiment of the present invention.

[0024] Deep learning-based intelligent identification methods for plywood adhesive defects include: Frame-by-frame parameter inversion is performed on the continuous time-series multi-source data of the glued surface of plywood to separate geometric feature components, thermal feature components and reflection feature components. These are then mapped to the feature embedding space by a variational autoencoder. In the feature embedding space, independence constraints are applied to each feature component to make the feature components of different physical properties orthogonally distributed. Redundant dependencies between feature components are suppressed by minimizing mutual information loss, resulting in a physically decoupled feature embedding vector. Construct a causal relationship map characterizing the causal relationship between process parameters and defect generation mechanisms, and calculate the causal effect intensity of each process parameter on defect generation based on the causal relationship map; The physical decoupled feature embedding vector is associated and matched with the causal effect intensity to identify the feature subspace affected by the causal influence of specific process parameters and generate a defect-sensitive feature map characterizing the spatial distribution of defects. Connected regions are extracted from the defect-sensitive feature map as defect candidate regions. A temporal prediction task and a temporal reconstruction task are constructed. The temporal prediction task predicts the feature evolution state of the defect candidate region at the next time step, and the temporal reconstruction task reconstructs the feature of the defect region of the occluded part at the current time step. By jointly optimizing the reconstruction error of the time-series prediction task and the time-series reconstruction task, and learning the temporal consistency feature embedding vector of the evolution law of defects in the time dimension, a time-enhanced defect feature embedding vector is obtained. The distance metric between the temporally enhanced defect feature embedding vector and the fine-grained, medium-grained, and coarse-grained prototype feature vectors in the multi-granularity prototype feature library is calculated. The matching results of different granularities are fused according to the confidence weight and a voting decision is made to obtain the defect identification result. The defect identification result and the corresponding abnormal process parameters are sent to the glue coating machine control system to realize the online control of the process parameters.

[0025] In one optional implementation, frame-by-frame parameter inversion is performed on the continuous temporal multi-source data of the plywood adhesive surface to separate geometric feature components, thermal feature components, and reflection feature components, which are then mapped to the feature embedding space via a variational autoencoder, including: Based on the thermal conduction model and optical reflection model of the adhesive material, the physical parameters of the data vector at each spatial location in the multi-source data are solved, and an inversion mapping relationship from the observed data vector to the physical parameter vector is established to obtain the spatially distributed physical parameter field. The geometric parameter tensor characterizing the spatial distribution of the coating thickness is extracted from the physical parameter field as the geometric feature component, the thermal parameter tensor characterizing the spatial distribution of the temperature field of the coating material is extracted as the thermal feature component, and the reflection parameter tensor characterizing the spatial distribution of the surface optical response is extracted as the reflection feature component. For the geometric feature components, the thermal feature components, and the reflection feature components, corresponding variational autoencoders are constructed respectively, and each feature component is input into the corresponding encoder. Each encoder outputs probability distribution parameters containing a mean vector and a variance vector. The probability distribution parameters output by each encoder are reparameterized and sampled to obtain geometric feature embedding vector, thermal feature embedding vector and reflection feature embedding vector. The geometric feature embedding vector, the thermal feature embedding vector and the reflection feature embedding vector are respectively located in the geometric subspace, thermal subspace and reflection subspace of the feature embedding space.

[0026] Multi-source data collected from the glued surface of plywood typically includes structured light scanning data, infrared thermal imaging data, and visible / near-infrared reflectance spectral data. These three types of data describe the state of the glued surface from three physical dimensions: geometric morphology, thermal field distribution, and optical response, respectively. The core of parameter inversion lies in mapping these observation data from the "measurement space" back to the "physical parameter space," that is, for each spatial location of the glued surface. Solve for the corresponding physical parameter vector This ensures that the observed values ​​predicted positively by the physical model match the vector of the actually collected observation data. Minimize the residuals between them.

[0027] Specifically, based on the thermal conduction model of the coating material, the temperature-time response curve of each pixel in the infrared thermographic sequence is fitted, and thermal parameters such as thermal conductivity and heat capacity of the coating layer are extracted by solving the inverse problem of the thermal diffusion equation. Based on an optical reflection model (such as the Kubelka-Munk dual-flux model or a multilayer thin-film interference model), the spectral response curve of each spatial location in the reflection spectrum data is fitted, and optical parameters such as absorption coefficient, scattering coefficient, and refractive index of the coating layer are extracted. Based on a structured light phase unwrapping algorithm, the scanned data is reconstructed in three dimensions to extract the thickness distribution and surface morphology parameters of the coating layer. The above inversion process is performed independently for each spatial location, and finally, a vector from the observation data is established. To physical parameter vector The inversion mapping relationship yields the spatial distribution of physical parameters covering the entire adhesive coating area.

[0028] From the obtained physical parameter field, each parameter is classified and extracted according to its physical properties. The parameters related to the coating thickness and surface three-dimensional morphology obtained from the inversion are organized into geometric parameter tensors. As a geometric feature component, where and These represent the spatial height and width of the adhesive coating image, respectively. The number of channels represents the geometric parameters; the inverted temperature field, thermal conductivity, and other thermally relevant parameters are organized into a thermal parameter tensor. As a thermal characteristic component, The number of channels represents the thermal parameters; the optical response parameters such as absorption coefficient, scattering coefficient, and refractive index obtained from the inversion are organized into a reflection parameter tensor. As a reflection characteristic component, The number of channels represents the reflection parameter. The three types of characteristic components are physically independent and characterize the essential properties of coating defects from three dimensions: spatial morphology, thermal field distribution, and optical response.

[0029] For the three types of feature components mentioned above, three structurally independent variational autoencoders (VAEs) are constructed, denoted as follows: , and Each corresponds to an encoding and decoding task for geometric, thermal, and reflection feature components. Taking geometric feature components as an example... For example, input it encoder network The encoder consists of multiple convolutional and fully connected layers, which extract and compress features from the input tensor layer by layer, and finally output probability distribution parameters, including the mean vector. With log-variance vector ,in The embedding dimension of the geometric subspace. Thermal eigencomponents. through encoder Process and output the mean vector. With log-variance vector , Embedding dimension of thermal subspace; reflective feature components through encoder Process and output the mean vector. With log-variance vector , The embedding dimension of the reflection subspace is defined. The kernel size, number of layers, and hidden layer dimension of each encoder network are independently designed based on the spatial resolution and physical complexity of the corresponding feature components to ensure that each subspace has sufficient expressive power.

[0030] After obtaining the probability distribution parameters of each encoder output, a reparameterization technique is used to sample each distribution to ensure that the gradient can successfully pass through the sampling operation during backpropagation. For the geometric feature components, noise vectors are sampled from the standard normal distribution. ,pass Obtain the geometric feature embedding vector ,in This represents element-wise multiplication. Depend on The result is obtained by taking the square root of the exponent. Similarly, the thermal feature embedding vector is obtained by sampling the thermal feature components. The reflection feature embedding vector is obtained by sampling the reflection feature components. ,in This is the corresponding independent sampling noise vector.

[0031] , and The three subspaces—geometric, thermal, and reflection—are physically distinct and independent within the feature embedding space. These subspaces do not overlap in dimensionality and together constitute a complete physically decoupled feature embedding space. During the training phase, each VAE, in addition to minimizing its own reconstruction loss (i.e., the mean square error between the decoder output and the corresponding input feature components), applies a KL divergence regularization term to constrain the probability distribution of each subspace to approximate a standard normal distribution, thereby ensuring the continuity and sampleability of the embedding space.

[0032] To further enhance the physical decoupling among the three types of embedding vectors, independence constraints are introduced into the training objective. Specifically, for , , The mutual information estimate between each pair of samples in the batch is calculated and added as a penalty term to the total loss function. By minimizing the mutual information loss, redundant dependencies between different physical property feature components are suppressed, making the embedding vectors in the geometric, thermal, and reflection subspaces as orthogonal as possible in a statistical sense. After the above joint training, the obtained physically decoupled feature embedding vectors can clearly separate the feature information of the coating defect in the three dimensions of geometric morphology, thermal anomaly, and optical anomaly, providing a high-quality physically interpretable feature representation for subsequent defect analysis based on causal relationship maps.

[0033] In one optional implementation, an independence constraint is imposed on each feature component in the feature embedding space to ensure that feature components with different physical properties are orthogonally distributed, and redundant dependencies between feature components are suppressed by minimizing mutual information loss, resulting in a physically decoupled feature embedding vector including: Calculate the covariance matrix between each pair of the geometric feature embedding vector, the thermal feature embedding vector, and the reflection feature embedding vector, and construct the orthogonality constraint loss. The orthogonality constraint loss quantifies the degree of deviation of the off-diagonal elements in the covariance matrix. For all pairwise combinations of the geometric feature embedding vector, the thermal feature embedding vector, and the reflection feature embedding vector, a corresponding joint probability distribution estimation network and a marginal probability distribution estimation network are constructed respectively. Based on the product of the joint probability distribution output by each joint probability distribution estimation network and the marginal probability distribution output by the corresponding marginal probability distribution estimation network, the mutual information between each feature embedding vector pair is calculated, and the mutual information is summed to obtain the total mutual information loss. A joint optimization objective is constructed, which includes a weighted combination of the orthogonality constraint loss and the total mutual information loss. The network parameters of the encoders of the geometric variational autoencoder, the thermal variational autoencoder, and the reflection variational autoencoder are updated by minimizing the joint optimization objective, and the decoupled geometric feature embedding vector, thermal feature embedding vector, and reflection feature embedding vector are obtained as the feature embedding vectors of the physical decoupling.

[0034] like Figure 2 As shown, the method includes: In obtaining geometric feature embedding vectors Thermal feature embedding vector With reflection feature embedding vector Next, independence constraints need to be applied to the feature embedding space to ensure that the feature components of the three types of physical properties are orthogonally distributed, thereby eliminating aliasing interference between different physical mechanisms. In the process of applying glue to plywood, the geometric shape, thermal conduction behavior and optical reflection characteristics are physically independent properties. However, due to the cross-response during sensor acquisition, the directly extracted features often carry redundant cross-modal information, which needs to be eliminated through explicit decoupling constraints.

[0035] The orthogonality constraint loss is constructed based on the covariance matrix. For three pairs of feature embedding vectors... , and Calculate the cross-feature covariance matrix between each pair. For example, let the number of samples in a batch be... After zero-mean normalization of the geometric and thermal feature embedding vectors of all samples within a batch, the cross-modal covariance matrix is ​​calculated. ,in Embedding dimensions into a geometric subspace Let the thermal subspace embedding dimension be defined. If the two types of features are completely decoupled, then... Each element should approach zero. Orthogonality constraint loss. The deviation of off-diagonal elements is quantified by summing the Frobenius norms of all cross-modal covariance matrices, specifically expressed as follows: ,in The cross-modal covariance matrix between the geometric feature embedding vector and the reflection feature embedding vector. Let be the cross-modal covariance matrix between the thermal feature embedding vector and the reflection feature embedding vector. This loss directly penalizes the linear correlation between the feature components of different physical attributes, driving the encoder to gradually make the three types of feature components tend to be orthogonally distributed in the embedding space during training.

[0036] Relying solely on orthogonality constraints can only eliminate linear correlations, but cannot capture the potential nonlinear dependencies between feature components. Therefore, a loss term based on mutual information estimation is further introduced to suppress nonlinear redundant dependencies between feature components. For all pairwise combinations of the three pairs of feature embedding vectors, corresponding joint probability distribution estimation networks and marginal probability distribution estimation networks are constructed respectively. For example, joint probability distribution estimation network Receive the concatenated vector As input, the output is a joint distribution Log density estimation; marginal probability distribution estimation network To each and marginal distribution and The log density estimates of the two marginal distributions are then summed to obtain the log density estimate of the product of the marginal distributions. The difference between the two estimates is the mutual information between the feature pairs. The estimated value is obtained by using a variational lower bound method similar to MINE (Mutual Information Neural Estimation), which provides a differentiable approximate estimate of the mutual information through a neural network. This allows the mutual information loss to impose gradient constraints on the encoder parameters during backpropagation.

[0037] Calculate the mutual information for each of the three pairs of feature embedding vectors. , and Then, the sum of the three is obtained to obtain the total mutual information loss. ,in The mutual information between the geometric feature embedding vector and the reflection feature embedding vector. Let be the mutual information between the thermal feature embedding vector and the reflection feature embedding vector. The goal of minimizing the total mutual information loss is to drive the three feature components to statistically become independent, eliminating not only linear correlations but also suppressing higher-order nonlinear dependencies, thus achieving more thorough physical decoupling. In actual training, the joint probability distribution estimation network and the marginal probability distribution estimation network adopt a lightweight multilayer perceptron structure, with the hidden layer dimension set to twice the feature embedding dimension, and LeakyReLU chosen as the activation function to ensure the stability of the gradient flow.

[0038] Joint optimization objective The orthogonality constraint loss and the total mutual information loss are combined in a weighted manner, expressed as: ,in These are the weighting coefficients for the orthogonality constraint loss. These are the weight coefficients for the total mutual information loss. The two weight coefficients are adjusted in the early stages of training based on the decoupling effect on the validation set, typically by... Set between 0.1 and 1.0. The value is set between 0.01 and 0.1 to balance the constraint strength of linear and nonlinear decoupling. In the overall training objective, Together with the reconstruction loss of each variational autoencoder and the KL divergence loss, they form a complete loss function, which is jointly optimized by the Adam optimizer.

[0039] Regarding the parameter update mechanism, by minimizing... The network parameters of the encoders for the geometric variational autoencoder, thermal variational autoencoder, and reflection variational autoencoder are updated synchronously. The gradients of the three encoders are calculated independently during backpropagation, but they share the cross-modal constraint signal from the decoupling loss. Specifically, the gradient of the orthogonality constraint loss with respect to the covariance matrix is ​​simultaneously propagated to the two encoders that generate the corresponding feature embedding vectors, while the gradient of the mutual information loss is backpropagated to the corresponding encoder parameters through the output difference between the joint probability distribution estimation network and the marginal probability distribution estimation network. After sufficient training, the three encoders output decoupled geometric feature embedding vectors, thermal feature embedding vectors, and reflection feature embedding vectors, respectively. The three types of feature components are orthogonally distributed and statistically independent in the embedding space, together constituting physically decoupled feature embedding vectors, providing a high-quality feature foundation for subsequent causal effect analysis and defect identification.

[0040] In practical applications, the decoupling effect can be quantitatively evaluated by calculating the correlation coefficient matrix between each pair of the three types of feature embedding vectors after training. If the absolute value of each cross-modal correlation coefficient is lower than a preset threshold (e.g., 0.05), the decoupling constraint is considered to be fully effective. For different batches of plywood glued samples, due to fluctuations in glued process parameters, the distribution of the three types of feature components has a certain offset. Therefore, during the online inference stage, it is necessary to periodically fine-tune the joint probability distribution estimation network and the marginal probability distribution estimation network to maintain the accuracy of mutual information estimation and ensure that the physically decoupled feature embedding vectors have stable independence under different production conditions.

[0041] In one optional implementation, a causal relationship map is constructed to characterize the causal relationship between process parameters and defect generation mechanisms. The causal effect strength of each process parameter on defect generation is calculated based on the causal relationship map, including: Identify the independent and dependent variable nodes from the process parameters; Based on the change sequence of the physically decoupled feature embedding vector in the time series, where the time series is a physical decoupled feature sequence of more than 3 consecutive frames, the temporal dependency relationship between the independent variable node and the dependent variable node is calculated, and based on the temporal dependency relationship, it is determined whether the change of the independent variable node precedes the change of the dependent variable node in time. For independent variable nodes and dependent variable nodes that satisfy the time-series prior conditions, a conditional independence test mechanism is constructed. By introducing a set of control variables and calculating the difference in conditional probability distributions, the mechanism identifies whether there is a direct causal path between the independent variable nodes and the dependent variable nodes. Using the independent variable nodes and the dependent variable nodes as vertices in the graph structure, and the identified direct causal paths as directed edges in the graph structure, a causal relationship graph is constructed from the process parameter nodes to the defect type nodes. For each directed edge in the causal relationship graph, a numerical perturbation is applied to the process parameters corresponding to the starting node, and the response change is calculated when the numerical perturbation propagates to the defect type corresponding to the ending node. The ratio of the change in response to the numerical perturbation is taken as the causal effect strength corresponding to the directed edge.

[0042] From the collected set of process parameters, independent and dependent variable nodes are distinguished. Process parameters typically include process quantities that can be actively controlled or passively recorded, such as coating speed, coating pressure, adhesive temperature, coating thickness, and ambient humidity. These quantities are defined as independent variable nodes. Defect types include observable quality anomalies such as bubble defects, adhesive leakage defects, and uneven thickness defects. These quantities are defined as dependent variable nodes. The independent and dependent variable nodes together constitute the vertex set of the subsequent graph structure. The physical meaning of each node must strictly correspond to the role of each quantity in the actual process flow to avoid misclassifying intermediate state quantities.

[0043] After identifying the node types, the temporal dependency between independent and dependent variable nodes is quantitatively assessed by utilizing the change sequence formed by the physically decoupled feature embedding vectors on the time axis. Specifically, for any pair of independent variable nodes... With dependent variable node The feature embedding values ​​at multiple consecutive sampling times are extracted to form a time series. and ,in The total number of sampling times. Indicates the first Independent variable node at time 1 Feature embedding values, Indicates the first The dependent variable node at time 1 The feature embedding values ​​are used. An autoregressive model including lag terms is constructed using the Granger causality test framework. This is achieved by comparing the feature embedding values. Lagging terms and those not included Lag term time pair The prediction error, to determine Does historical information affect The current value has a statistically significant predictive gain. If the test results show... The change preceded in time If the changes are significant and have significant predictive power, then the node is considered to meet the temporal prior conditions and can proceed to the subsequent causal path identification stage.

[0044] For node pairs that meet the temporal prior conditions, a conditional independence test mechanism is further constructed to eliminate spurious correlation interference. In actual process environments, there are common confounding variables (such as adhesive viscosity) between adhesive application speed and adhesive leakage defects, resulting in a statistical correlation between the two but not a direct causal relationship. Therefore, for each candidate node pair... Select a set of control variables from the remaining process parameters to act as confounding or mediating variables. Calculate in a given under conditions and The difference in conditional probability distributions. If and In control The conditional probability distribution remains statistically correlated if the difference exceeds a preset threshold. If the two are found to have a direct causal path, then it is determined that there is a direct causal path between them; otherwise, it is considered that the correlation is caused by... The variables involved are either mediated or confused and do not constitute a direct causal relationship. Conditional independence tests can be performed using partial correlation coefficient tests or kernel-based conditional independence tests; the specific choice depends on the data distribution characteristics and sample size.

[0045] After determining the temporal dependencies and verifying the conditional independence of all candidate node pairs, the causal relationships of the node pairs that passed the double screening are encoded into a directed graph structure. Starting with the process parameter node and ending with the defect type node, each confirmed direct causal path corresponds to a directed edge in the graph, thus constructing a complete causal relationship graph. ,in It is the set of all nodes (including independent variable nodes and dependent variable nodes). It is the set of all directed edges. The topological structure of the graph intuitively reflects the transmission path of process parameters to the defect generation mechanism. For example, the glue application pressure node points to both the bubble defect node and the thickness unevenness defect node, indicating that abnormal pressure can induce two different types of quality problems.

[0046] Based on the constructed causal relationship graph, the causal effect strength corresponding to each directed edge is quantitatively calculated. This is applied to the graph from the process parameter node... Pointer to defect type node Directed edge The intensity of causal effects was estimated using an intervention simulation method. Specifically, while keeping the values ​​of other process parameter nodes constant, the following steps were taken: The applied amplitude is The numerical perturbation was observed to propagate along the causal path to... The resulting change in response Intensity of causal effect Defined as the ratio of the change in response to the numerical disturbance, i.e. ,in Indicates process parameter nodes For defect type nodes The strength of the causal effect To be applied to The numerical perturbation amplitude, for The change in response under this perturbation. To improve the robustness of the estimation, the average of the ratios under multiple different perturbation amplitudes can be taken to reduce the random error introduced by a single perturbation. When When the value is large, it indicates that the process parameter has a strong driving effect on the generation of the corresponding defect type, and should be given higher attention weight in the subsequent feature subspace identification stage; when When the value is close to zero, it indicates that the causal influence corresponding to the directed edge is negligible, and its contribution ratio can be reduced during the feature matching stage.

[0047] To ensure the physical rationality of the intervention simulation, the amplitude of the numerical perturbation... The process parameters must be selected within the practically feasible range to avoid exceeding physical constraints and causing distortion of simulation results. Furthermore, in cases where multi-hop paths exist in the graph (i.e., process parameters indirectly affect defect types through intermediate nodes), direct and indirect causal effects must be distinguished. Direct causal effects are only considered... arrive The single-step directed edge contribution, while the total causal effect requires traversing all directed edges. arrive The directed path is summed by multiplying the effects along each path. In practical applications, depending on the accuracy requirements of the defect identification task, either the direct causal effect strength or the total causal effect strength can be used as the weighting basis for subsequent feature association matching. Through the complete process described above, the causal effect strength value corresponding to each directed edge is finally obtained, providing a quantitative basis for subsequent association matching of physically decoupled feature embedding vectors with causal effect strengths and generation of defect-sensitive feature maps.

[0048] In one optional implementation, the physically decoupled feature embedding vector is correlated and matched with the causal effect strength to identify the feature subspace causally affected by specific process parameters and generate a defect-sensitive feature map characterizing the spatial distribution of defects. Extracting connected regions from the defect-sensitive feature map as candidate defect regions includes: For the causal paths of process parameters and defect types in the causal relationship graph where the intensity of the causal effect exceeds a preset causal effect threshold, the parameter identifiers corresponding to the process parameters and the defect identifiers corresponding to the defect types are extracted to form strong causal association pairs. For each strong causal relationship pair, extract the feature embedding vector corresponding to the process parameter in terms of physical properties from the feature embedding vector of the physical decoupling as the feature vector to be analyzed, and calculate the response intensity distribution of the feature vector to be analyzed in each dimension of the feature embedding space; Based on the response intensity distribution, dimensions with response intensity exceeding a preset response threshold are selected to form a sensitive dimension set, and the feature subspace corresponding to the sensitive dimension set is marked as a feature subspace causally affected by the process parameters. Spatial back projection is performed on the feature embedding vectors in the feature subspace, and the feature embedding vectors are mapped back to the spatial coordinate system of the original coated surface to obtain the correspondence between spatial position and feature response value; Based on the correspondence, spatial locations where the feature response value exceeds the response threshold are marked in the two-dimensional space of the coated surface. The marked areas are then subjected to connectivity analysis and adjacent marked areas are merged to generate a defect-sensitive feature map characterizing the spatial distribution of the defect type. The regions marked in the defect-sensitive feature map are extracted as the defect candidate regions.

[0049] After constructing the physical decoupling feature embedding vectors and establishing the causal relationship map, the two need to be correlated and matched to locate the feature subspace truly driven by specific process parameters, and finally generate defect candidate regions in the two-dimensional space of the coated surface. The core of this process is that not all feature dimensions have a substantial causal relationship with defect generation; only those feature dimensions corresponding to process parameters that have significant causal effect strength in the causal relationship map have real discriminative value for defect identification.

[0050] For causal relationship mapping Each of the process parameter nodes Pointer to defect type node Directed edge Read the corresponding causal effect strength and with a preset causal effect threshold Comparison. When At that time, it was assumed that there was a strong causal relationship between the process parameter and the corresponding defect type, and the process parameter was... Parameter identifiers and defect types The defect identifiers together constitute a strong causal relationship. Threshold The settings need to be calibrated based on prior knowledge of the actual production process, and can usually be determined by statistical analysis of the perturbation test results of various process parameters in historical defect data. For the same defect type... In cases involving multiple strongly causally related process parameters, it is necessary to construct independent strongly causal pairs for all process parameters that meet the conditions, and then perform feature subspace analysis on each one to avoid the causal effects of different process parameters masking each other.

[0051] For each strong causal pair According to process parameters The physical attribute category is determined, and the component corresponding to the physical attribute is extracted from the feature embedding vector of the physical decoupling as the feature vector to be analyzed. Specifically, if the process parameters For parameters that are geometric (such as adhesive thickness and adhesive width), extract the geometric feature embedding vector. As If it belongs to thermal parameters (such as curing temperature, drying rate, etc.), then extract the thermal feature embedding vector. If it belongs to the reflection type parameters (such as adhesive gloss, surface reflectivity, etc.), then extract the reflection feature embedding vector. This correspondence based on physical attributes ensures the physical consistency of feature extraction and avoids causal attribution bias caused by the mixing of features across physical domains.

[0052] Obtain the feature vector to be analyzed Then, the response intensity distribution in each dimension of the feature embedding space is calculated. Let the dimension of the feature embedding space be . , for the Response intensity in each dimension It is measured by the normalized result of the difference between the absolute value of the feature value of this dimension and the mean within the batch, i.e. ,in For the eigenvector to be analyzed in the th... The value of dimension, This represents the mean of this dimension within the batch. This represents the standard deviation of this dimension within the batch. This normalization process eliminates the influence of differences in the dimensions of different dimensions, making the response strengths of each dimension comparable.

[0053] Based on the response intensity distribution across various dimensions, a preset response threshold is set. , will satisfy Dimension index Collect them to form a set of sensitive dimensions. Sensitive Dimension Set The spanned subspace is then labeled as being subject to process parameters. Feature subspace of causal influence In practice, An adaptive approach can be adopted, for example, by taking the mean of the response intensity of all dimensions plus a multiple of the standard deviation, thereby dynamically adjusting the threshold according to the overall response distribution of the current sample and enhancing the adaptability of the method to different coating conditions.

[0054] After labeling the feature subspace, it is necessary to map the information embedded in the feature space back to the two-dimensional spatial coordinate system of the original coated surface. This process is called spatial backprojection. Spatial backprojection utilizes the inverse mapping relationship of the variational autoencoder decoder to map the feature subspace... The feature embedding vectors in the image are projected onto the spatial coordinates of the original image. The feature response value corresponding to each spatial location is obtained above. In practical implementation, the idea of ​​gradient-weighted activation mapping can be used to calculate the gradient contribution of each dimension of the feature subspace to the original input space, and the product of the gradient magnitude and the feature activation value is accumulated to the corresponding spatial location, thereby forming a continuous response value distribution map. During the backprojection process, coordinate alignment must be maintained, that is, the scaling ratio between the spatial resolution of the feature map and the resolution of the original coated image must be accurately recorded to ensure the accuracy of subsequent spatial location annotation.

[0055] Obtain spatial location and characteristic response values After establishing the correspondence, in the two-dimensional space of the adhesive surface, the conditions that satisfy... The spatial locations of the pixels are marked to form the initial response activation region. Since the defect areas on the actual coated surface are often spatially continuous, and the initial marking results contain small voids or breaks, connectivity analysis of the marked regions is necessary. The connectivity analysis uses the eight-connectivity criterion, grouping spatially adjacent marked pixels that all meet the response threshold conditions into the same connected region. For areas smaller than a preset minimum area threshold... Isolated connected regions are considered noise interference and are removed; regions with a spatial distance less than a preset merging distance threshold are also excluded. The adjacent connected regions are merged into a complete defect candidate region to address the spatial fragmentation of the defect region caused by uneven glue application.

[0056] After connectivity analysis and region merging, the resulting set of labeled regions constitutes the defect type. The defect-sensitive feature map is generated. Each labeled region in the defect-sensitive feature map not only records its spatial boundary coordinates but also stores the corresponding average feature response value and associated defect type identifier, providing structured defect candidate region information for subsequent time-series enhancement processing and prototype matching. When multiple strong causal pairs exist, their respective defect-sensitive feature maps are generated. The labeled regions in all defect-sensitive feature maps are then extracted into the final defect candidate region set by taking the union of all the regions. This ensures that different types of defects induced by different process parameters can be effectively covered, avoiding defect omissions caused by single causal path analysis.

[0057] In one optional implementation, a temporal prediction task and a temporal reconstruction task are constructed. The temporal prediction task predicts the feature evolution state of the defect candidate region at the next time step, and the temporal reconstruction task reconstructs the defect region features of the occluded portion at the current time step, including: Defect candidate region samples are extracted from historically collected adhesive surface data and arranged in chronological order to form a time-series training set. For each time-series sample in the time-series training set, the physical decoupling feature embedding vectors corresponding to the defect candidate regions at the current time and the next adjacent time are extracted as the current state features and target evolution features, respectively. A temporal predictive encoder and a temporal predictive decoder are constructed. The temporal predictive encoder receives the current state features and outputs a temporal encoded representation. The temporal predictive decoder receives the temporal encoded representation and outputs the predicted feature evolution state. The network parameters of the temporal predictive encoder and the temporal predictive decoder are updated by calculating the reconstruction error between the predicted feature evolution state and the target evolution feature; For each time-series sample in the time-series training set, a spatial occlusion operation is applied to the feature embedding vector of the physically decoupled defect candidate region at the current time to generate partially visible features that include the occluded region and the visible region. A temporal reconstruction encoder and a temporal reconstruction decoder are constructed. The temporal reconstruction encoder receives the partially visible features and outputs an occlusion-aware encoded representation. The temporal reconstruction decoder receives the occlusion-aware encoded representation and outputs the reconstructed complete features. The network parameters of the temporal reconstruction encoder and the temporal reconstruction decoder are updated by calculating the reconstruction error between the reconstructed complete features and the original complete features.

[0058] When extracting continuous time-series defect candidate region samples from historically collected glued surface data, it is necessary to ensure the continuity and integrity of the time-series samples in the temporal dimension. Specifically, for multiple frames of data collected during the glued process of the same plywood, they are arranged in chronological order according to the collection timestamps, and continuous frame segments containing defect candidate regions are extracted. Each segment constitutes a time-series training sample. Each sample in the time-series training set contains defect candidate region features from at least two adjacent time points, ensuring that subsequent time-series prediction tasks can establish a mapping relationship from the current time to the next time point. To avoid the interference of collection noise on temporal consistency, the original collected data is time-aligned before constructing the time-series training set to keep the time interval between adjacent frames uniform, thereby ensuring that the time-series samples can truly reflect the evolution of defect regions during the glued process.

[0059] For each time-series sample in the training set, extract the physical decoupling feature embedding vectors corresponding to the candidate defect regions at the current time and the next adjacent time step. The feature embedding vector at the current time step is denoted as the current state feature. The feature embedding vector at the next time step is denoted as the target evolution feature. The physically decoupled feature embedding vector here originates from the concatenated result of the variational autoencoder encoding the geometric, thermal, and reflection feature components in the preceding steps. It has already undergone orthogonal distribution constraints in the feature embedding space and possesses physical property independence. and As the input-output pair for the time-series prediction task, it provides supervision signals for the training of the time-series prediction encoder and the time-series prediction decoder.

[0060] The temporal predictive encoder employs a sequence modeling structure based on a self-attention mechanism, receiving current state features. As input, the dependencies between the dimensions within the feature embedding vector are captured by a multi-head attention layer, and the output is a temporal encoded representation. The temporal prediction decoder uses As a condition, the predicted feature evolution state is output through a multi-layer feedforward network and residual connection structure. Reconstruction error in time series prediction tasks The mean squared error is used to measure the difference between the predicted feature evolution state and the target evolution feature. The calculation method is as follows: ,in Let represent the Euclidean norm. By minimizing... Backpropagation updates all network parameters of the temporal predictive encoder and temporal predictive decoder, driving the model to learn the feature evolution of defect candidate regions in the time dimension.

[0061] When constructing the temporal reconstruction task, for each temporal sample in the temporal training set, a spatial occlusion operation is applied to the physically decoupled feature embedding vector of the defect candidate region at the current time step. The specific implementation of the spatial occlusion operation is as follows: the feature embedding vector is divided into several non-overlapping local blocks in the spatial dimension, and occlusion is performed according to a preset ratio. Randomly select a portion of local blocks and set them to zero to generate partially visible features that include both occluded and visible areas. Occlusion ratio The value of is typically set between 0.3 and 0.6, ensuring that the visible region contains sufficient contextual information to support reconstruction, while also making the occluded region challenging enough to facilitate the model's learning of the complete feature distribution. The random selection of occlusion locations is performed independently in each iteration, enhancing the generalization ability of the temporal reconstruction task.

[0062] The temporal reconstruction encoder employs a mask-aware attention mechanism, receiving partially visible features. As input, by explicitly distinguishing between visible and occluded locations during attention computation, the encoding process relies solely on feature information from the visible region for contextual modeling, outputting an occlusion-aware encoded representation. The temporal reconstruction decoder uses... As input, through upsampling and feature fusion operations, the output is the reconstructed complete features. Reconstruction error in time-series reconstruction tasks The calculation is performed only at the corresponding location in the occluded area to avoid interference from direct copying of the visible area in the loss calculation. The calculation method is as follows: ,in A set of indices representing the location of occlusion. The total number of occluded locations. Indicates the reconstruction features in the first The value of each occlusion position, This represents the true value of the original complete feature at the corresponding position. By minimizing... Backpropagation updates all network parameters of the temporal reconstruction encoder and temporal reconstruction decoder, enabling the model to infer the complete defect feature distribution of the occluded region from the locally visible features.

[0063] The temporal prediction and temporal reconstruction tasks share some parameters of the underlying feature extraction layer during training to achieve complementary supervision between the two tasks. The temporal prediction task focuses on modeling the dynamic evolution trend of defect features along the time axis, while the temporal reconstruction task focuses on modeling the structural integrity of defect features in the spatial dimension. Joint training of the two tasks enables the encoder to simultaneously perceive the temporal evolution and spatial distribution patterns of defects, thereby forming a defect feature representation with stronger temporal consistency in the feature embedding space. In each training iteration, the same batch of temporal samples are simultaneously calculated... and Through weighted summation The joint optimization objective is obtained, where and These are the loss weight coefficients for the temporal prediction task and the temporal reconstruction task, respectively, used to balance the contribution of the two tasks to the final feature learning.

[0064] After training, the temporal prediction encoder and the temporal reconstruction encoder encode the feature embedding vectors of the defect candidate regions respectively, to obtain the temporal prediction encoding representation. Occlusion-aware coding representation The two are fused into a temporally enhanced defect feature embedding vector through a concatenation operation, which is then used in subsequent multi-granularity prototype matching steps. This fusion method ensures that the temporally enhanced features simultaneously carry defect evolution trend information and spatial integrity information, improving the robustness and accuracy of subsequent defect identification steps.

[0065] Both the temporal predictive encoder and the temporal reconstruction encoder are single-layer Transformer encoders with 4 heads and 128 hidden layer dimensions.

[0066] In one optional implementation, the reconstruction errors of the temporal prediction task and the temporal reconstruction task are jointly optimized, and a temporal consistency feature embedding vector of the evolution law of defects in the time dimension is learned to obtain a temporally enhanced defect feature embedding vector, including: The reconstruction error between the predicted feature evolution state and the target evolution feature is extracted from the time series prediction task and used as the time series prediction loss; The reconstruction error between the reconstructed complete features and the original complete features is extracted from the temporal reconstruction task and used as the temporal reconstruction loss. A joint optimization objective function is constructed, which sums the time series prediction loss and the time series reconstruction loss by weighting them. The contribution of the time series prediction task and the time series reconstruction task to the joint optimization is controlled by adjusting the weight coefficients. Based on the joint optimization objective function, the network parameters of the temporal prediction encoder, the temporal prediction decoder, the temporal reconstruction encoder, and the temporal reconstruction decoder are updated synchronously, so that the temporal prediction task and the temporal reconstruction task share the underlying feature representation space; After the joint optimization objective function converges, the temporal coding representation output by the temporal prediction encoder and the occlusion-aware coding representation output by the temporal reconstruction encoder are extracted and merged into a unified temporal consistency feature embedding vector through feature fusion operation; The temporal consistency feature embedding vector and the physically decoupled feature embedding vector are concatenated to generate a temporally enhanced defect feature embedding vector.

[0067] After constructing the temporal prediction and temporal reconstruction tasks, a joint optimization mechanism is needed to integrate the learning signals from both tasks. This allows the encoder network to simultaneously capture the temporal evolution of defects and the integrity of their spatial structure within the same feature representation space. The reconstruction errors from the two tasks are treated as independent loss signals in the joint optimization, and their calculation method directly determines the quality of the final temporally consistent feature embedding vector.

[0068] Temporal prediction loss originates from the predicted feature evolution state in the temporal prediction task. With target evolution characteristics Reconstruction error between Specifically, Measuring the performance of a temporal predictive decoder given current state features Under the given conditions, the accuracy of predicting the feature distribution of the candidate defect region at the next time step is considered. When the defect region undergoes morphological evolution between adjacent frames during the adhesive coating process, this loss drives the temporal coding representation output by the temporal predictive encoder. It can fully encode the dynamic evolution trend of defects, enabling the network to learn the inherent laws of the evolution patterns of adhesive coating defects, such as expansion, contraction, or displacement over time.

[0069] The temporal reconstruction loss originates from the complete features reconstructed in the temporal reconstruction task. With original complete features Reconstruction error between The loss only applies to the set of indices at the occlusion locations. The calculations are performed at the corresponding spatial location, meaning that constraints are imposed only on the feature reconstruction quality of the occluded region, rather than on the global features. This design enables the occlusion-aware coding representation output by the temporal reconstruction encoder to... It can learn the spatial structure prior of the defect region, in some visible features Based on this, the feature distribution of the occluded part is inferred, thereby enhancing the robustness of the feature representation to local occlusion interference.

[0070] Joint optimization objective function Time series prediction loss With time series reconstruction loss Perform a weighted summation, using the loss weight coefficients. and The contribution of each task to the joint optimization objective is controlled separately, i.e. .when When the value is relatively large, the network focuses more on learning the temporal evolution of defects; when... When the weights are relatively large, the network focuses more on learning the spatial structural integrity representation of defects. In practical applications, the two weight coefficients can be adjusted according to the specific characteristics of the adhesive application process. For example, for scenarios with fast adhesive application speeds and significant changes in defect morphology, the weights can be appropriately increased. To enhance timing prediction capabilities; for detection environments with significant occlusion interference, appropriately increase... To improve the robustness of reconstruction.

[0071] Based on joint optimization objective function The backpropagation algorithm synchronously updates all network parameters of the temporal predictive encoder, temporal predictive decoder, temporal reconstruction encoder, and temporal reconstruction decoder. During the same optimization process, the four network components receive gradient signals from both tasks, allowing the temporal predictive encoder and temporal reconstruction encoder to influence each other during parameter updates, gradually forming a shared low-level feature representation space. The core significance of this sharing mechanism lies in the fact that the dynamic evolution information learned by the temporal prediction task and the static structural information learned by the temporal reconstruction task complement each other at the low-level representation level. This ensures that the features extracted by the encoder contain both the dynamic information of defect evolution over time and the structural information of defect spatial morphology, thereby significantly improving the discriminative power of the feature representation. During training, monitoring... When the curve of the parameter change no longer decreases significantly over several consecutive training rounds, it is determined that the joint optimization objective function has converged, and parameter updates are stopped.

[0072] After the joint optimization objective function converges, the output representations are extracted from the temporal predictive encoder and the temporal reconstruction encoder, respectively. The temporal predictive encoder receives the current state features... Post-output timing code representation This vector encodes the evolutionary trend information of the defect in the time dimension; the temporal reconstruction encoder shows visible features in the receiving part. Post-output occlusion-aware coding representation This vector encodes the structural integrity information of the defect in the spatial dimension. and The two encoded representations are merged into a unified temporally consistent feature embedding vector through feature fusion. The feature fusion operation uses vector concatenation, directly concatenating the two encoded representations along the feature dimension, preserving their complete information without compression. This ensures that the temporally consistent feature embedding vector retains both dynamic evolutionary information and static structural information, which are complementary semantic levels.

[0073] The temporally consistent feature embedding vector is concatenated with the physically decoupled feature embedding vector to generate a temporally enhanced defect feature embedding vector. The physically decoupled feature embedding vector contains elements from the geometric feature embedding vector. Thermal feature embedding vector With reflection feature embedding vector The multimodal feature information obtained after physical decoupling represents the static multimodal description of the coated surface in terms of physical properties. The temporal consistency feature embedding vector represents the dynamic evolution description of the defect in terms of time. By concatenating the two, the temporally enhanced defect feature embedding vector further incorporates temporal dynamic information on the basis of physical multimodal features. This enables the subsequent prototype matching stage to simultaneously utilize differences in physical properties and temporal evolution to distinguish defect types, effectively improving the recognition accuracy of defect types that are similar in appearance but have different evolutionary patterns. The concatenation operation is performed in the feature dimension without modifying any feature values, ensuring that the two types of feature information coexist in the vector space in a non-interfering manner, providing information-rich and structurally clear feature inputs for subsequent multi-granularity prototype matching.

[0074] After obtaining the temporally enhanced defect feature embedding vector, it is necessary to perform distance measurement and matching with a multi-granularity prototype feature library to ultimately complete the defect type identification. The multi-granularity prototype feature library pre-stores typical feature representations of various defects at different abstraction levels, including three levels: fine-grained prototype, medium-grained prototype, and coarse-grained prototype, which correspond to the subtle difference features, common features of categories, and major category distinguishing features of defects, respectively.

[0075] For the temporally enhanced defect feature embedding vector, the Euclidean distance or cosine similarity between it and the prototype feature vectors at each level in the multi-granularity prototype feature library are calculated, resulting in three sets of distance measurement results. For the fine-grained level, the distance measurement identifies the specific defect subtype most similar to the current defect candidate region feature; for the medium-grained level, the defect category is identified; and for the coarse-grained level, it is determined whether it belongs to a macroscopic category such as surface defect or internal defect. The matching results of the three levels are weighted and fused according to preset confidence weights. When the confidence of the fine-grained matching exceeds a threshold, the fine-grained identification result is directly adopted; otherwise, a voting decision is made by combining the medium-grained and coarse-grained matching results to comprehensively determine the final defect identification result. Through the multi-granularity prototype matching mechanism, accurate identification of common defect types is ensured, while the generalization ability to identify rare defect types is improved, significantly enhancing the overall identification accuracy and robustness.

[0076] A second aspect of the present invention provides a deep learning-based intelligent identification system for adhesive defects in plywood, comprising: The physical decoupling unit is used to perform frame-by-frame parameter inversion on continuous time-series multi-source data of the glued surface of plywood, separate geometric feature components, thermal feature components and reflection feature components, and map them to the feature embedding space through variational autoencoder. In the feature embedding space, independence constraints are applied to each feature component to make the feature components of different physical properties orthogonally distributed, and the redundant dependence between feature components is suppressed by minimizing mutual information loss, so as to obtain the physical decoupling feature embedding vector. The causal effect unit is used to construct a causal relationship map characterizing the causal relationship between process parameters and defect generation mechanisms, and to calculate the causal effect intensity of each process parameter on defect generation based on the causal relationship map. Candidate region units are used to associate and match the physical decoupled feature embedding vectors with the causal effect intensity, identify the feature subspace causally affected by specific process parameters and generate a defect-sensitive feature map characterizing the spatial distribution of defects, and extract connected regions from the defect-sensitive feature map as defect candidate regions. The temporal task unit is used to construct a temporal prediction task and a temporal reconstruction task. The temporal prediction task predicts the feature evolution state of the defect candidate region at the next time step, and the temporal reconstruction task reconstructs the feature of the defect region of the occluded part at the current time step. The temporal enhancement unit is used to jointly optimize the reconstruction error of the temporal prediction task and the temporal reconstruction task, learn the temporal consistency feature embedding vector of the evolution law of defects in the time dimension, and obtain the defect feature embedding vector of temporal enhancement. The defect identification unit is used to calculate the distance metric between the time-enhanced defect feature embedding vector and the fine-grained, medium-grained, and coarse-grained prototype feature vectors in the multi-grained prototype feature library. It fuses the matching results of different granularities according to the confidence weight and performs voting decision to obtain the defect identification result. The defect identification result and the corresponding abnormal process parameters are sent to the glue coating machine control system to realize the online control of the process parameters.

[0077] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0078] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0079] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0080] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions 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. A method for intelligent identification of glue defects in plywood based on deep learning, characterized in that, include: Frame-by-frame parameter inversion is performed on the continuous time-series multi-source data of the glued surface of plywood to separate geometric feature components, thermal feature components and reflection feature components. These are then mapped to the feature embedding space by a variational autoencoder. In the feature embedding space, independence constraints are applied to each feature component to make the feature components of different physical properties orthogonally distributed. Redundant dependencies between feature components are suppressed by minimizing mutual information loss, resulting in a physically decoupled feature embedding vector. Construct a causal relationship map characterizing the causal relationship between process parameters and defect generation mechanisms, and calculate the causal effect intensity of each process parameter on defect generation based on the causal relationship map; The physical decoupled feature embedding vector is associated and matched with the causal effect intensity to identify the feature subspace affected by the causal influence of specific process parameters and generate a defect-sensitive feature map characterizing the spatial distribution of defects. Connected regions are extracted from the defect-sensitive feature map as defect candidate regions. A temporal prediction task and a temporal reconstruction task are constructed. The temporal prediction task predicts the feature evolution state of the defect candidate region at the next time step, and the temporal reconstruction task reconstructs the feature of the defect region of the occluded part at the current time step. By jointly optimizing the reconstruction error of the time-series prediction task and the time-series reconstruction task, and learning the temporal consistency feature embedding vector of the evolution law of defects in the time dimension, a time-enhanced defect feature embedding vector is obtained. The distance metric between the temporally enhanced defect feature embedding vector and the fine-grained, medium-grained, and coarse-grained prototype feature vectors in the multi-granularity prototype feature library is calculated. The matching results of different granularities are fused according to the confidence weight and a voting decision is made to obtain the defect identification result. The defect identification result and the corresponding abnormal process parameters are sent to the glue coating machine control system to realize the online control of the process parameters.

2. The method according to claim 1, characterized in that, Frame-by-frame parameter inversion is performed on continuous temporal multi-source data of the glued surface of plywood to separate geometric feature components, thermal feature components, and reflection feature components. These components are then mapped to the feature embedding space via a variational autoencoder, including: Based on the thermal conduction model and optical reflection model of the adhesive material, the physical parameters of the data vector at each spatial location in the multi-source data are solved, and an inversion mapping relationship from the observed data vector to the physical parameter vector is established to obtain the spatially distributed physical parameter field. The geometric parameter tensor characterizing the spatial distribution of the coating thickness is extracted from the physical parameter field as the geometric feature component, the thermal parameter tensor characterizing the spatial distribution of the temperature field of the coating material is extracted as the thermal feature component, and the reflection parameter tensor characterizing the spatial distribution of the surface optical response is extracted as the reflection feature component. For the geometric feature components, the thermal feature components, and the reflection feature components, corresponding variational autoencoders are constructed respectively, and each feature component is input into the corresponding encoder. Each encoder outputs probability distribution parameters containing a mean vector and a variance vector. The probability distribution parameters output by each encoder are reparameterized and sampled to obtain geometric feature embedding vector, thermal feature embedding vector and reflection feature embedding vector. The geometric feature embedding vector, the thermal feature embedding vector and the reflection feature embedding vector are respectively located in the geometric subspace, thermal subspace and reflection subspace of the feature embedding space.

3. The method according to claim 2, characterized in that, In the feature embedding space, independence constraints are applied to each feature component to ensure that feature components of different physical properties are orthogonally distributed. Redundant dependencies between feature components are suppressed by minimizing mutual information loss, resulting in a physically decoupled feature embedding vector comprising: Calculate the covariance matrix between each pair of the geometric feature embedding vector, the thermal feature embedding vector, and the reflection feature embedding vector, and construct the orthogonality constraint loss. The orthogonality constraint loss quantifies the degree of deviation of the off-diagonal elements in the covariance matrix. For all pairwise combinations of the geometric feature embedding vector, the thermal feature embedding vector, and the reflection feature embedding vector, a corresponding joint probability distribution estimation network and a marginal probability distribution estimation network are constructed respectively. Based on the product of the joint probability distribution output by each joint probability distribution estimation network and the marginal probability distribution output by the corresponding marginal probability distribution estimation network, the mutual information between each feature embedding vector pair is calculated, and the mutual information is summed to obtain the total mutual information loss. A joint optimization objective is constructed, which includes a weighted combination of the orthogonality constraint loss and the total mutual information loss. The network parameters of the encoders of the geometric variational autoencoder, the thermal variational autoencoder, and the reflection variational autoencoder are updated by minimizing the joint optimization objective, and the decoupled geometric feature embedding vector, thermal feature embedding vector, and reflection feature embedding vector are obtained as the feature embedding vectors of the physical decoupling.

4. The method according to claim 1, characterized in that, Construct a causal relationship map characterizing the causal relationship between process parameters and defect generation mechanisms, and calculate the causal effect strength of each process parameter on defect generation based on the causal relationship map, including: Identify the independent and dependent variable nodes from the process parameters; Based on the change sequence of the physical decoupled feature embedding vector in the time series, where the time series is a physical decoupled feature sequence of more than 3 consecutive frames, the temporal dependency relationship between the independent variable node and the dependent variable node is calculated, and based on the temporal dependency relationship, it is determined whether the change of the independent variable node precedes the change of the dependent variable node in time. For independent variable nodes and dependent variable nodes that satisfy the time-series prior conditions, a conditional independence test mechanism is constructed. By introducing a set of control variables and calculating the difference in conditional probability distributions, the mechanism identifies whether there is a direct causal path between the independent variable nodes and the dependent variable nodes. Using the independent variable nodes and the dependent variable nodes as vertices in the graph structure, and the identified direct causal paths as directed edges in the graph structure, a causal relationship graph is constructed from the process parameter nodes to the defect type nodes. For each directed edge in the causal relationship graph, a numerical perturbation is simulated on the process parameters corresponding to the starting node, and the response change generated when the numerical perturbation propagates to the defect type corresponding to the ending node is calculated; the ratio of the response change to the numerical perturbation is taken as the causal effect strength corresponding to the directed edge.

5. The method according to claim 1, characterized in that, Associating and matching the physically decoupled feature embedding vector with the causal effect strength, identifying the feature subspace causally affected by specific process parameters, and generating a defect-sensitive feature map characterizing the spatial distribution of defects, extracting connected regions as defect candidate regions from the defect-sensitive feature map includes: For the causal paths of process parameters and defect types in the causal relationship graph where the intensity of the causal effect exceeds a preset causal effect threshold, the parameter identifiers corresponding to the process parameters and the defect identifiers corresponding to the defect types are extracted to form strong causal association pairs. For each strong causal relationship pair, extract the feature embedding vector corresponding to the process parameter in terms of physical properties from the feature embedding vector of the physical decoupling as the feature vector to be analyzed, and calculate the response intensity distribution of the feature vector to be analyzed in each dimension of the feature embedding space; Based on the response intensity distribution, dimensions with response intensity exceeding a preset response threshold are selected to form a sensitive dimension set, and the feature subspace corresponding to the sensitive dimension set is marked as a feature subspace causally affected by the process parameters. Spatial back projection is performed on the feature embedding vectors in the feature subspace, and the feature embedding vectors are mapped back to the spatial coordinate system of the original coated surface to obtain the correspondence between spatial position and feature response value; Based on the correspondence, spatial locations where the feature response value exceeds the response threshold are marked in the two-dimensional space of the coated surface. The marked areas are then subjected to connectivity analysis and adjacent marked areas are merged to generate a defect-sensitive feature map characterizing the spatial distribution of the defect type. The regions marked in the defect-sensitive feature map are extracted as the defect candidate regions.

6. The method according to claim 1, characterized in that, A temporal prediction task and a temporal reconstruction task are constructed. The temporal prediction task predicts the feature evolution state of the defect candidate region at the next time step, and the temporal reconstruction task reconstructs the feature of the occluded defect region at the current time step, including: Defect candidate region samples are extracted from historically collected adhesive surface data and arranged in chronological order to form a time-series training set. For each time-series sample in the time-series training set, the physical decoupling feature embedding vectors corresponding to the defect candidate regions at the current time and the next adjacent time are extracted as the current state features and target evolution features, respectively. A temporal predictive encoder and a temporal predictive decoder are constructed. The temporal predictive encoder receives the current state features and outputs a temporal encoded representation. The temporal predictive decoder receives the temporal encoded representation and outputs the predicted feature evolution state. The network parameters of the temporal predictive encoder and the temporal predictive decoder are updated by calculating the reconstruction error between the predicted feature evolution state and the target evolution feature; For each time-series sample in the time-series training set, a spatial occlusion operation is applied to the feature embedding vector of the physically decoupled defect candidate region at the current time to generate partially visible features that include the occluded region and the visible region. A temporal reconstruction encoder and a temporal reconstruction decoder are constructed. The temporal reconstruction encoder receives the partially visible features and outputs an occlusion-aware encoded representation. The temporal reconstruction decoder receives the occlusion-aware encoded representation and outputs the reconstructed complete features. The network parameters of the temporal reconstruction encoder and the temporal reconstruction decoder are updated by calculating the reconstruction error between the reconstructed complete features and the original complete features.

7. The method according to claim 6, characterized in that, By jointly optimizing the reconstruction errors of the temporal prediction task and the temporal reconstruction task, and learning the temporal consistency feature embedding vector of the evolution law of defects in the time dimension, the temporal enhanced defect feature embedding vector is obtained, including: The reconstruction error between the predicted feature evolution state and the target evolution feature is extracted from the time series prediction task and used as the time series prediction loss; The reconstruction error between the reconstructed complete features and the original complete features is extracted from the temporal reconstruction task and used as the temporal reconstruction loss. A joint optimization objective function is constructed, which sums the time series prediction loss and the time series reconstruction loss by weighting them. The contribution of the time series prediction task and the time series reconstruction task to the joint optimization is controlled by adjusting the weight coefficients. Based on the joint optimization objective function, the network parameters of the temporal prediction encoder, the temporal prediction decoder, the temporal reconstruction encoder, and the temporal reconstruction decoder are updated synchronously, so that the temporal prediction task and the temporal reconstruction task share the underlying feature representation space; After the joint optimization objective function converges, the temporal coding representation output by the temporal prediction encoder and the occlusion-aware coding representation output by the temporal reconstruction encoder are extracted and merged into a unified temporal consistency feature embedding vector through feature fusion operation; The temporal consistency feature embedding vector and the physically decoupled feature embedding vector are concatenated to generate a temporally enhanced defect feature embedding vector.

8. A deep learning-based intelligent identification system for plywood adhesive defects, used to implement the method as described in any one of claims 1-7, characterized in that, include: The physical decoupling unit is used to perform frame-by-frame parameter inversion on continuous time-series multi-source data of the glued surface of plywood, separate geometric feature components, thermal feature components and reflection feature components, and map them to the feature embedding space through variational autoencoder. In the feature embedding space, independence constraints are applied to each feature component to make the feature components of different physical properties orthogonally distributed, and the redundant dependence between feature components is suppressed by minimizing mutual information loss, so as to obtain the physical decoupling feature embedding vector. The causal effect unit is used to construct a causal relationship map characterizing the causal relationship between process parameters and defect generation mechanisms, and to calculate the causal effect intensity of each process parameter on defect generation based on the causal relationship map. Candidate region units are used to associate and match the physical decoupled feature embedding vectors with the causal effect intensity, identify the feature subspace causally affected by specific process parameters and generate a defect-sensitive feature map characterizing the spatial distribution of defects, and extract connected regions from the defect-sensitive feature map as defect candidate regions. The temporal task unit is used to construct a temporal prediction task and a temporal reconstruction task. The temporal prediction task predicts the feature evolution state of the defect candidate region at the next time step, and the temporal reconstruction task reconstructs the feature of the defect region of the occluded part at the current time step. The temporal enhancement unit is used to jointly optimize the reconstruction error of the temporal prediction task and the temporal reconstruction task, learn the temporal consistency feature embedding vector of the evolution law of defects in the time dimension, and obtain the defect feature embedding vector of temporal enhancement. The defect identification unit is used to calculate the distance metric between the time-enhanced defect feature embedding vector and the fine-grained, medium-grained, and coarse-grained prototype feature vectors in the multi-grained prototype feature library. It fuses the matching results of different granularities according to the confidence weight and performs voting decision to obtain the defect identification result. The defect identification result and the corresponding abnormal process parameters are sent to the glue coating machine control system to realize the online control of the process parameters.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke 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 the processor, they implement the method described in any one of claims 1 to 7.

Images