Environment-friendly compression-resistant artificial panel performance prediction system based on multi-modal data analysis

Abstract

The invention discloses an environment-friendly compression-resistant artificial panel performance prediction system based on multi-modal data analysis, and particularly relates to the technical field of industrial intelligent monitoring and prediction, and the system comprises an early sensing module, a causal fusion module, a physical enhancement prediction module and a self-adaptive learning module. According to the method, multi-modal early signals in production are synchronously collected, core features are discovered and identified by using causality and fused, then a prediction model with enhanced physical mechanism is constructed, an output result and uncertainty are output, and finally, continuous optimization of the model is realized through adaptive learning. According to the method, the problems of prediction lag, poor generalization, unreasonable physics, static model degradation and the like are solved, and early, accurate and reliable prediction and lifelong learning optimization of the plate strength are realized.

Description

A Performance Prediction System for Environmentally Friendly Compression-Resistant Artificial Boards Based on Multimodal Data Analysis Technical Field

[0001] This invention relates to the field of industrial intelligent monitoring and prediction technology, and more specifically, to a performance prediction system for environmentally friendly, pressure-resistant engineered wood panels based on multimodal data analysis. Background Technology

[0002] Currently, in the production process of environmentally friendly engineered wood panels, the monitoring and prediction of compressive strength, a key quality indicator, mainly relies on two types of technologies. The first type is the traditional manual sampling and destructive testing method in the laboratory, which involves taking samples at the end of the production line and sending them to the laboratory for testing according to standard procedures. This method provides accurate and authoritative results, but it suffers from significant time lag and cannot provide real-time quality assessment and process control for panels in production. The second type is online monitoring and prediction technology, which can be further divided into several paths: First, based on monitoring by a single or a few sensors, predictions are made by establishing empirical regression models of these process parameters and the final strength. However, such models ignore key internal processes such as the evolution of the panel's internal microstructure and the curing kinetics of the adhesive, resulting in limited prediction accuracy and reliability. Second, more advanced sensors are used to collect signals during the production process, and machine learning models are used for data-driven strength prediction. These methods can acquire information earlier, but they usually simply splice together multi-source heterogeneous signals or process them through "black box" models, failing to delve into the stable causal relationship between signals and performance. The models have poor interpretability, and their generalization ability decreases when raw materials, formulations, or processes change, making them prone to prediction failure.

[0003] However, in practical use, it still has some shortcomings. For example, existing online prediction methods mostly focus on later or single-dimensional signals, lacking collaborative perception and deep fusion of multi-physical fields and multi-modal early precursor signals in the entire "paving-hot pressing" process, resulting in delayed prediction timing and failure to achieve true early warning and intervention. Most data-driven models are correlational rather than causal, and are prone to learning spurious correlations or environmental noise in the data. When production conditions change, the model's prediction performance is unstable and its generalization ability is weak. Purely data-driven models lack the constraints of materials science and mechanical principles, and their prediction results sometimes violate basic physical laws, resulting in insufficient credibility in critical decision-making. Existing systems are usually static and lack continuous self-learning and adaptive mechanisms. After model deployment, performance will gradually degrade over time due to equipment aging, process updates, etc., making it difficult to achieve lifelong learning and continuous optimization, and resulting in high long-term maintenance costs. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, the present invention provides an environmentally friendly and pressure-resistant artificial board performance prediction system based on multimodal data analysis, which solves the problems mentioned in the background art through the following solutions.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a performance prediction system for environmentally friendly, compression-resistant engineered wood panels based on multimodal data analysis, comprising: an early sensing module: deployed in the laying and hot-pressing processes of panel production, used to synchronously collect multimodal early precursor signals related to the final compressive strength of the panel through a multi-physics field sensor array, and to preprocess and extract features from the signals, outputting a multimodal feature set and early performance prediction values; a causal fusion module: used to receive the multimodal feature set, identify a core feature subset with a causal relationship to compressive strength from the feature set through a causal discovery algorithm, and construct a causal-enhanced fusion feature representation based on a causal-guided attention mechanism; and a physical enhancement prediction module: The system receives the causal-enhanced fusion feature representation by embedding a materials science mechanism model into a neural network in the form of soft constraints or structural priors to construct a prediction model with enhanced physical information. It then uses multi-granular dynamic integration and a Bayesian framework to output the final predicted compressive strength value and its uncertainty quantification result. The adaptive learning module receives the multimodal feature set corresponding to the new production batch, the predicted compressive strength value, and its uncertainty quantification result. After evaluating the effectiveness and ensuring safe ingestion of the new batch data, it updates the prediction model parameters through an incremental learning algorithm and dynamically updates the causal graph and the core feature subset based on performance and data distribution monitoring results, thereby achieving continuous optimization and lifelong learning of the system.

[0006] The technical effects and advantages of this invention are as follows: It enables early and accurate early warning and empowers real-time process control: By deploying a multi-physics field sensor array, this invention captures multi-modal precursor signals that are strongly correlated with the final compressive strength in real time during the paving and hot pressing process. This advances the performance prediction time point from "finished product inspection" after production is completed to "during production". Once the performance deviation trend is predicted, the process parameters can be fed back and adjusted in real time, thereby reducing scrap and improving the first pass rate and production efficiency.

[0007] Integrating causal reasoning enhances model interpretability and generalization: The model exhibits stronger robustness and generalization ability when facing changes in production conditions, and the prediction results are more interpretable, providing direct causal insights for process optimization. Embedding physical mechanisms ensures reasonable and reliable prediction results: By constructing a physical information neural network, knowledge from materials science fields such as composite material mixing laws, curing kinetics, and porous material strength models is embedded into the model as soft constraints or structural priors, avoiding physically impossible and absurd predictions that may arise from purely data-driven models, thus improving the credibility and reliability of prediction results in key quality decisions. Building a lifelong learning system enables adaptive evolution: The system can safely and continuously update the model incrementally using new production data, while preventing "catastrophic forgetting" of learned knowledge. It can proactively monitor and diagnose drift in data distribution or performance relationships and make targeted responses. Providing uncertainty quantification and comprehensive decision support: Employing a Bayesian framework and a multi-granularity dynamic integration strategy, it not only outputs point prediction values ​​but also quantifies the uncertainty of predictions and provides confidence intervals. At the same time, the dynamic integration mechanism can adaptively select the most reliable prediction source under different operating conditions. Attached Figure Description

[0008] Figure 1 is a schematic diagram of the overall structure of the present invention.

[0009] Figure 2 is a schematic diagram of the early sensing module structure of the present invention.

[0010] Figure 3 is a schematic diagram of the causal fusion module structure of the present invention.

[0011] Figure 4 is a schematic diagram of the physical enhancement prediction module of the present invention.

[0012] Figure 5 is a schematic diagram of the adaptive learning module structure of the present invention. Detailed Implementation

[0013] 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.

[0014] Referring to Figures 1-5, an environmentally friendly, pressure-resistant engineered wood panel performance prediction system based on multimodal data analysis includes: an early perception module, a causal fusion module, a physical enhancement prediction module, and an adaptive learning module.

[0015] The early sensing module: During the key process stages of board production (laying, hot pressing), by deploying a multi-physics field sensor array, it captures early precursor signals that are strongly correlated with the final compressive strength in real time, and advances the performance prediction time point from "after finished product inspection" to "during production", so as to realize online, dynamic and early performance trend prediction.

[0016] S101: Synchronous Acquisition and Alignment of Multi-Source Heterogeneous Data. A high-speed structured light 3D scanner is installed above the paving unit outlet to ensure complete coverage of the paving area and collect 3D point cloud data of fibers / rubber fragments after paving.

[0017] A high-sensitivity piezoelectric sensor array and distributed fiber optic temperature / pressure sensors are embedded in the upper and lower pressure plates of the hot press in a grid pattern. The sensor spacing is set to 50-100mm according to the size of the hot press, enabling full-area monitoring of the slab. The piezoelectric sensor array collects micro-stress wave signals (voltage signals) generated inside the slab due to fiber breakage and adhesive layer cracking. Distributed fiber optic sensors synchronously collect real-time temperature data of different regions of the slab during the hot pressing process. ) and pressure data ( ).

[0018] A non-contact dielectric measurement probe and a short-wave near-infrared spectroscopy probe are installed on the side of the hot press, aligned with the monitoring area at the edge of the slab, ensuring that the distance between the probe and the slab surface is consistently maintained at 5-10 cm; the dielectric measurement probe collects the dielectric loss factor during the adhesive curing process. and dielectric constant Near-infrared spectral probes collect near-infrared spectral data (absorbance) during the adhesive curing process. , (Wavelength).

[0019] Data synchronization: All 3D point cloud data, micro-stress wave signals, temperature / pressure data, dielectric parameter data and near-infrared spectral data collected by all sensors are synchronized at the hardware level through a unified high-precision industrial timing synchronization controller with a synchronization accuracy of ≤1μs, ensuring that multiple types of data from different physical locations and different sampling rates have a unified timestamp.

[0020] S102: Low-noise preprocessing and feature generation of raw signals: Parallel processing of three types of core precursor signals, extracting physical / chemical features directly related to the performance of the board material through signal analysis: A. Internal micro-stress wave signal processing flow: Noise reduction: Acquired raw voltage signal First, the signal is filtered by bandpass filtering to remove low-frequency noise such as mechanical vibrations from the equipment, while retaining the acoustic emission signal corresponding to the microscopic damage.

[0021] Event detection and segmentation: Using short-time average zero-crossing rate (reflecting signal frequency characteristics) and energy thresholding (threshold set at 3 times the historical noise energy), burst-type acoustic emission signal segments representing microscopic damage events such as fiber breakage and adhesive layer cracking are automatically detected and segmented. .

[0022] Feature extraction: for each event Perform Fast Fourier Transform (FFT) and Wavelet Packet Transform to calculate its key features, forming the "fingerprint" vector of the event. : in, The wavelet packet energy spectrum can effectively distinguish different damage modes such as fiber fracture and adhesive layer cracking.

[0023] Timing statistics: Within each process time window (e.g., every 10 seconds), the macroscopic timing characteristics such as the count rate, cumulative energy, and average frequency of all acoustic emission events within that window are statistically analyzed to form a stress wave evolution curve synchronized with the hot pressing process, which intuitively reflects the dynamic development of internal damage in the slab.

[0024] B. Dielectric-Near-Infrared Signal Processing Stream during Curing: Dielectric Signal Analysis: Real-time Acquisition of Dielectric Loss Factor The curve directly reflects the activity level of the adhesive molecular chain movement. During the cross-linking and curing of molecular chains, Characteristic peaks will appear. Extract the peak temperature. Area under the curve These characteristics are related to the activation energy and total heat of the curing reaction, respectively, and directly determine the bonding strength of the adhesive layer.

[0025] Near-infrared spectral analysis: The acquired near-infrared spectra are preprocessed using the second derivative to eliminate baseline drift and scattering interference, and then the characteristic absorption peak areas of specific functional groups are monitored. The decay rate is the rate at which the curing reaction is most intense; the faster the decay, the more vigorous the curing reaction.

[0026] Estimation of fusion curing degree: Establishment , True curing degree obtained from benchmark DSC test Multiple nonlinear regression model between: in , To synchronously collect hot-pressing temperature and pressure data, and use these as input variables to correct for environmental influences in the model, the final online output is a high-confidence real-time estimate of the degree of curing. and its first derivative.

[0027] C. 3D Image Processing Flow of Paving Morphology: Point Cloud Generation and Denoising: Point cloud data acquired by a 3D scanner is statistically filtered to remove outliers and retain the true 3D coordinate information of fibers / fragments.

[0028] 3D density field reconstruction: The point cloud is voxelized, the point density within each voxel is calculated, and a 3D density distribution matrix after paving is generated. Quantify the uniformity of paving.

[0029] Topological feature calculation: For the binarized 3D structure, a 3D image analysis algorithm is applied: Pore network analysis: The connectivity, average pore size, and pore tortuosity of the pores are calculated to reflect the mass transfer and stress transmission paths within the slab; Fiber orientation tensor: By calculating the direction vectors of fibers / fragments in a local region, the principal orientation direction and the degree of anisotropy are obtained. The more uniform the orientation, the more stable the compressive strength; Surface area to volume ratio: The total surface area available for adhesive bonding between fibers is quantified, which directly affects the interfacial bonding strength between the adhesive layer and the fibers.

[0030] S103: Early Performance Index Calculation and Interface Output. The processed features are input into a lightweight early prediction sub-model. This model is trained using historical production data and learns to output early compressive strength predictions based on all currently collected precursor signals before hot pressing is completed. The processed multimodal feature set and early prediction values ​​are output synchronously to provide standardized input data for the subsequent causal fusion module.

[0031] The causal fusion module identifies and focuses on the core feature set that has a stable causal relationship with compressive strength from the multimodal features output by the early perception module, and constructs a feature fusion representation that is resistant to interference, interpretable, and has strong generalization ability, providing high-quality input for the final performance prediction model.

[0032] S201: Data Preparation and Causal Discovery. Gather historical production batch data to construct a structured dataset. .in, It is a high-dimensional vector containing all output features of the early sensing module, specifically including: fiber basic features: length, moisture content; process parameter features: hot pressing temperature, pressure, time; acoustic emission statistical features: count rate of micro-stress wave events, cumulative energy, average frequency, etc.; curing process features: peak curing rate, final degree of curing, peak temperature of dielectric loss factor, etc.; three-dimensional morphological features: pore connectivity, orientation anisotropy, surface area to volume ratio, etc.; the dimension is usually 50-100. It is the corresponding laboratory-measured compressive strength, which is used as the target variable for causal analysis.

[0033] Preliminary Causal Graph Learning: Applying a constraint-based causal discovery algorithm to uncover causal relationships between features and performance: Initialization: Constructing a completely undirected graph where nodes represent all features. and performance indicators There is an undirected edge between any two nodes.

[0034] Conditional independence test: For each pair of adjacent nodes in the graph Given all possible sets of conditions Next, perform a statistical independence test: if Then remove the edge. .

[0035] Directed edges: Assign direction to the remaining undirected edges using directional rules, for example, if a path exists. and and If conditions are independent, then Assuming collision nodes are defined and edge orientations follow the above pattern, a directed acyclic graph is ultimately obtained. It describes the potential causal network between features and performance, as well as between features themselves.

[0036] S202: Causal Feature Screening and Enhanced Key Causal Path Identification: In Causal Graphs In the process, identify all features that point from the input features to the target. The directed paths are identified. The "causal flow strength" of each path is calculated, for example, by calculating the product of the average causal effects of each edge on the path. The larger the absolute value of the product, the more significant the impact of the path on performance, and it is preferred to retain it.

[0037] Constructing a subset of causal features: Direct causal parent nodes: Filtering out those in China as The feature set of the direct parent node These characteristics are considered to be important for It has the most direct impact.

[0038] Important mediating variables: This involves identifying a set of key mediating nodes located on critical causal paths that connect multiple upstream causes to y. These variables are the core links in the causal chain and cannot be omitted.

[0039] Fusion Core Set: Defining the Causal Core Feature Set Typically, the dimensions are 10-20, and the system will assign higher fusion weights to these features to reduce the interference of non-causal features.

[0040] S203: Constructing a Causal-Oriented Fusion Encoding Network: Injecting the aforementioned causal knowledge into the neural network fusion process to ensure the model focuses on meaningful causal relationships: Design a multi-branch encoding network. Each modality's data first passes through its own encoder, mapping the original features to a primary feature representation with uniform dimensionality. ( (For modal indexing).

[0041] Causal-guided attention fusion layer: integrating causal core feature sets Each feature in A set of causal query vectors is formed by mapping the vectors to their corresponding primary feature representation dimensions through a fully connected layer. It is used to locate information related to causal features in each modality.

[0042] Using a cross-modal attention mechanism, these causal query vectors { Go to "query" all primary modal features. } Calculate attention and aggregate information: here, Calculate using dot product similarity. Indicates the first The causal feature for the first Attention level for each modality (value range 0-1). It is the corresponding cross-modal aggregation information that highlights the contribution of causal correlation features.

[0043] Ultimately, all This is concatenated with a general global attention fusion feature to form a causal-enhanced fusion feature representation with a dimension of 64×(K+1) (K being the number of causal core features). .

[0044] Causal regularization training: A causal regularization term is explicitly added to the total loss function of the training prediction model to guide the model to learn causal invariant features. Encourage models to make more use of causal core sets. The features in the model are achieved by applying L1 penalties to the network connection weights corresponding to non-core features. Set to 0.01 to balance sparsity and performance; Invariant learning loss Set to 0.1. Divide the data into multiple environments based on different operating conditions; this loss term encourages the model to learn causal feature representations. and The relationship between them should remain consistent across all environments, with the goal of: ,at the same time in For the environment The MSE loss, along with the objective function, prompts the model to discard those methods that only work in specific environments. It retains the characteristics of correlation (pseudo-correlation), preserves the causal invariance characteristics that are stable across environments, and improves generalization ability.

[0045] S204: Online Inference and Feature Contribution Analysis. When the system predicts new samples, it uses attribution methods (such as ensemble gradients) to calculate the contribution of each feature in the causal core set. Contribution to feature fusion results By quantifying contributions, the interpretability of the feature fusion process is achieved, ultimately outputting standardized causal enhanced fusion features. This provides input to the physics-enhanced prediction module.

[0046] The physical enhancement prediction module uses knowledge from the fields of materials science and mechanics as soft constraints or structural priors, explicitly embedding it into the neural network architecture and training process to construct a physical information-enhanced "grey box" prediction model. This achieves an optimal balance between high accuracy driven by data and rationality driven by mechanism, and quantifies the uncertainty of the prediction.

[0047] S301: Formalization and Embedding Strategy Design of Physical Knowledge: Identifying key physical and empirical models affecting the compressive strength of environmentally friendly engineered wood products, such as: the mixing law of composite materials: predicting macroscopic elastic modulus. The relationship between the fiber and the adhesive reflects the synergistic effect between the fiber and the adhesive: in , Volume fraction of fibers and adhesives ( ), For fiber elastic modulus, This refers to the elastic modulus of the adhesive after it has cured.

[0048] Curing kinetics model: describes the degree of curing With time ,temperature The relationship (autocatalytic model) reflects the curing reaction law of adhesives: in , Pre-exponential factor, The activation energy was obtained through DSC experimental fitting. The gas constant is... represents the order of the autocatalytic reaction, and the fitting parameters.

[0049] Empirical formulas for the strength of porous materials: such as the Gibson-Ashby model, which calculates compressive strength. With relative density Related: in For the strength of dense materials, The density of the dense material is determined by fitting experimental data.

[0050] Embedding Strategy Selection: Considering both model complexity and the rigor of physical laws, the following embedding strategies are adopted: Soft Constraints (Loss Function Term): The residuals of the composite material mixing law and curing kinetics model are added to the loss function as regularization terms (PINN concept), allowing for small deviations; Hard Constraints (Network Structure): When designing the network output layer, strict constraints are enforced. Equal conservation relationships are maintained to avoid physically impossible outcomes; prior guidance (initialization or intermediate supervision): the network output is pre-trained and initialized using the prediction results of the Gibson-Ashby model to improve the model's convergence speed.

[0051] S302: Physical Information Neural Network (PINN) Construction and Training Network Architecture Design: Building a Multi-Output Neural Network , where: input Features output by the causal fusion module This feature set integrates multimodal core information such as micro-stress waves, curing process, and three-dimensional morphology collected by the early sensing module; the output includes: the final predicted compressive strength. and a series of physical intermediate variables, such as the predicted fiber volume fraction. Matrix modulus Curing degree curve parameters The output range of all physical intermediate variables is limited to a reasonable physical range (e.g., ).

[0052] Definition of composite loss function: The total loss function consists of three parts, forcing the network to simultaneously fit the data and physical laws: Data loss The mean square error between the predicted intensity and the actual intensity directly guarantees the prediction accuracy. Physical loss : Quantify the deviation between the network output and the laws of physics Set to 0.05: in The macroscopic elasticity modulus predicted by the network. The auxiliary loss is obtained by differentiating the curing degree curve output by the network. Supervision loss (MSE) for physical intermediate variables with measurement data. Setting it to 0.02 improves the accuracy of physical variable prediction and indirectly ensures the rationality of intensity prediction.

[0053] Training process: The Adam optimizer was used, with a learning rate of 0.001 and a decay rate of 0.99. Automatic differentiation techniques were employed to calculate the impact of the physical equation residuals on the network parameters. The gradient is used to achieve end-to-end training. This is achieved by adjusting the hyperparameters. and Balancing data fit with physical consistency.

[0054] S303: Multi-granularity collaborative prediction and dynamic integration: To further improve prediction robustness, a multi-granularity parallel prediction channel is designed, which integrates the advantages of each channel through dynamic integration: Parallel prediction channel: Channel A (Physically enhanced PINN): Output It combines the high precision of data-driven approaches with the rationality of mechanism-driven approaches, making it suitable for most conventional operating conditions; Channel B (pure data-driven model): such as a deep fully connected network, output... It focuses on data fitting accuracy and is suitable for operating conditions with sufficient historical data; Channel C (Lightweight Mechanism Model): Based on simplified empirical formulas, it outputs... It has a fast calculation speed (≤1ms / sample) and ensures robustness under extreme conditions.

[0055] Dynamic weighted adaptive ensemble: Assigning dynamic weights related to the input sample to each channel. (The weights sum to 1), and the weights are determined by the "confidence assessment function". calculate: in Feedback Channel Reliability of the current sample: The smaller the physical residual of PINN, the lower the historical error of the data model in the feature space neighborhood (K nearest neighbors, K=20), and the smaller the deviation of the input parameters of the mechanistic model from the design range, the higher the corresponding confidence score.

[0056] Final prediction: The dynamic integration mechanism can adaptively select the most reliable prediction source under specific working conditions. Compared with a single model, the prediction accuracy is improved by ≥5%, and the robustness under extreme working conditions is improved by ≥10%.

[0057] S304: Uncertainty Quantification Based on Bayesian Framework: To quantify the reliability of prediction results, a Bayesian framework is introduced to model uncertainty: Bayesian neural network implementation: In the PINN channel, the hidden layer weights are replaced with Gaussian distribution, and Monte Carlo Dropout (MCDO) is used as an approximate inference method to model model uncertainty through the probability distribution of weights.

[0058] Predicted distribution sampling: for new input samples The predicted sample set is obtained through T=50 independent forward propagations. The distribution of the sample set can approximate the posterior distribution of the predicted values.

[0059] Uncertainty Calculation: Final Predicted Value: Taking the sample mean, balancing randomness and accuracy: Cognitive uncertainty: Measured using sample standard deviation. Total Uncertainty and Confidence Intervals: Integrating Cognitive Uncertainty and Inherent Noise Calculate the total uncertainty and output the 95% confidence interval. Set an uncertainty threshold; when the confidence interval width exceeds this threshold, a model adjustment signal is triggered.

[0060] The adaptive learning module establishes an automated, closed-loop, adaptive lifelong learning system for the model, enabling predictive performance to evolve over time rather than degrade.

[0061] S401: Safe Intake and Evaluation Triggering Conditions for New Sample Streams: When a new production batch is completed and its laboratory standard compressive strength test value is obtained. Afterwards, the system automatically triggers the learning process to ensure that new data is incorporated into the model in a timely manner.

[0062] Sample validity test: Outlier filtering: Calculate the previous predicted value for this batch. Compared with the true value residual If the residual difference is abnormally large, the sample is marked as a "suspicious sample" and will not be used directly for training. Instead, an anomaly detection mechanism will be triggered. The detection directions include: abnormal sensor data acquisition in the early sensing module, laboratory testing errors, production process fluctuations, or new defect patterns. After the detection is completed, a decision will be made on whether to include it in the training set.

[0063] Confidence assessment: Uncertainty when the system backtracks the predictions for this batch. If the confidence interval width of the prediction uncertainty is >2MPa, it indicates that the model is not well-suited to this type of working condition. This sample is considered a high-value sample and should be prioritized for training to supplement information on new working conditions.

[0064] Data buffer pool: Validated Yes, it is stored in a fixed-size First-In-First-Out (FIFO) buffer pool B. The buffer pool data contains the processed features corresponding to the complete multimodal data collected by the early perception module in this batch, ensuring that the buffer pool data can fully reflect the characteristics of recent production conditions, avoiding low training efficiency due to excessive historical data, and ensuring that the model can quickly adapt to recent changes in operating conditions.

[0065] S402: Incremental Update Based on Elastic Weight Consolidation: To avoid catastrophic forgetting, the Elastic Weight Consolidation (EWC) method is used for incremental model updates: Parameter Importance Calculation (Offline / Periodic): The system maintains a parameter importance matrix. Its elements Each parameter in the model was quantified. The importance of all historical tasks already learned; key parameters correspond to the core of historical knowledge and require special protection.

[0066] Using the EWC method, Approximately the elements on the diagonal of the Fisher Information Matrix (FIM), calculated periodically after the (k-1)th round of training (e.g., every 100 new samples): in It is a sample set representing historical knowledge (stratified sampling from historical archives to ensure coverage of all typical working conditions, with a sample size ≥ 2000), each sample... All of them contain multimodal features corresponding to the early perception module. It is the current optimal parameter. The partial derivative is calculated by automatic differentiation. The larger the diagonal element of FIM, the more important the parameter is to historical knowledge.

[0067] Incremental learning loss function: When the number of samples in buffer pool B reaches a threshold (e.g., 100), the model is updated using this new data. The optimized total loss function LEWC is defined as follows: First item The first term is the prediction loss (MSE) for the new batch of data, ensuring that the model learns the correlation between multimodal features and compressive strength under new working conditions; the second term is the EWC regularization term. It is the balance coefficient. It affects the parameters. Deviating from its historical best value The severity of the penalty is determined by the degree of punishment, and the intensity of the penalty is related to the importance of the parameter. Proportional.

[0068] Security Update Strategy: To avoid online service anomalies caused by update failures, a security update strategy of "sandbox verification + dual evaluation" is adopted: updates are performed in an isolated "sandbox" model copy, without affecting the current model of the online service; after the update, performance is evaluated simultaneously on both the historical validation set and the new data validation set; the update is only submitted to the online service model to replace the original model if the prediction accuracy of both validation sets does not decrease (the decrease is <3%); otherwise, the update is rolled back and marked as "conflicting", triggering further operational analysis or model adjustment processes.

[0069] S403: Active Detection and Diagnosis of Prediction Drift: Real-time monitoring of model prediction performance, timely detection and response to data / concept drift: Monitoring metrics: Error distribution monitoring: Prediction error calculated using a sliding window. Calculate the mean of the error within this window. (Reflecting systematic bias) and variance (Reflects error fluctuations) and compared with the historical average error during stable periods. and variance Compare them.

[0070] Feature distribution monitoring: Simultaneously monitor input features The distribution changes are analyzed, with a focus on the core modal features corresponding to the early sensing modules. The Wasserstein distance or maximum mean difference (MMD) between the features within the current sliding window and the historical baseline distribution is calculated.

[0071] Drift detection: For the error distribution, the two-sample Kolmogorov-Smirnov (KS) test is used to compare the error distribution of the recent window. Error distribution with long-term steady period : like If the threshold is exceeded (e.g., 0.05, corresponding to a 95% confidence level), then concept drift is determined to have occurred (i.e., (Changed); for the feature distribution, if the Wasserstein distance consistently exceeds the threshold, then data drift is determined to have occurred (i.e., ...). (Changes have occurred) may originate from changes in operating conditions monitored by the early sensing module, such as shifts in three-dimensional morphological features caused by batch changes of raw materials, or fluctuations in temperature data caused by equipment aging.

[0072] Root Cause Analysis and Response: Primarily due to data drift: Trigger the cache pool expansion mechanism to accelerate the model's adaptation to new operating conditions, while increasing the frequency of feature distribution monitoring, focusing on investigating the sensor stability of early sensing modules; Primarily due to concept drift: Initiate the causal discovery process to update the causal graph. and core feature set Furthermore, during incremental learning, the weight of recent data is temporarily increased to accelerate the model's learning of new causal relationships.

[0073] S404: Knowledge Base Update and Model Version Management Causal Knowledge Base Evolution: If the model's prediction accuracy improves under both new and old operating conditions after incremental learning (e.g., an overall accuracy improvement of ≥5%), or if drift diagnosis reveals a new stable pattern, the system automatically initiates the causal discovery process, using a fusion dataset of new and historical data to update the causal graph in the causal fusion module. and core feature set This ensures that causal knowledge evolves in sync with actual production conditions.

[0074] Model Snapshots and Versioning: Each successful security update generates a model version snapshot with a timestamp and metadata, including: main training data distribution, performance metrics, update time, and hyperparameter settings. Model rollback and A / B testing are supported to ensure system stability.

[0075] Meta-learner optimization: Dynamic ensemble weight network in the physics-enhanced prediction module It is also a learnable module. The system regularly (e.g., quarterly) fine-tunes the meta-learner using data from the most recent 1000 batches based on the ensemble prediction performance, optimizing its confidence evaluation function and improving the accuracy and robustness of the ensemble prediction.

[0076] After each batch of production is completed, a "System Adaptive Status Report" is output, and a "Monthly Adaptive Summary Report" is output monthly. The report content includes: Learning Status: the status of the most recent update (success / failure / conflict), the current model version number, the sample size and remaining capacity of the buffer pool; Performance Health: the statistics of the current prediction error (mean, variance), the comparison results with the historical stable period, whether there is drift (yes / no) and the type of drift (data drift / concept drift), and the severity of drift (mild / moderate / severe); Knowledge Freshness: the last update time of the causal graph, the time span of the data in the buffer pool (e.g., covering the most recent 3 months), whether it contains new working condition data (yes / no); Action Recommendations: targeted optimization suggestions are given, such as "It is recommended to collect more winter raw material data", "The current model has low prediction confidence for new process A, it is recommended to conduct 10-20 batches of targeted experiments to supplement data", "Mild data drift was detected, it is recommended to calibrate the hot pressing temperature sensor".

[0077] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. Finally: The above description is only a preferred embodiment of this invention and is not intended to limit this invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of this invention should be included within the protection scope of this invention.

Claims

1. A performance prediction system for environmentally friendly, compression-resistant engineered wood panels based on multimodal data analysis, characterized in that, include: Early Sensing Module: Deployed during the laying and hot-pressing processes of board production, it synchronously collects multimodal early precursor signals related to the final compressive strength of the board through a multi-physics field sensor array, preprocesses and extracts features from the signals, and outputs a multimodal feature set and early performance prediction values. Causal Fusion Module: Receives the multimodal feature set, identifies a subset of core features causally related to compressive strength from the feature set using a causal discovery algorithm, and constructs a causally enhanced fusion feature representation based on a causal-guided attention mechanism. Physical Enhancement Prediction Module: Receives the causally enhanced fusion feature representation, embeds a materials science mechanism model into a neural network in the form of soft constraints or structural priors, constructs a physically enhanced prediction model, and outputs the final compressive strength prediction value and its uncertainty quantification results using multi-granularity dynamic integration and a Bayesian framework. Adaptive learning module: It is used to receive the multimodal feature set corresponding to the new production batch, the predicted compressive strength value and its uncertainty quantification result, and after evaluating the effectiveness and safely ingesting the new batch data, it updates the prediction model parameters through incremental learning algorithm, and dynamically updates the causal graph and the core feature subset according to the performance and data distribution monitoring results, thereby realizing the continuous optimization and lifelong learning of the system.

2. The environmentally friendly, compression-resistant engineered wood panel performance prediction system based on multimodal data analysis according to claim 1, characterized in that, The multiphysics sensing array includes: a high-speed structured light 3D scanner deployed above the paving unit outlet for acquiring 3D point cloud data of fibers / rubber scraps after paving; a piezoelectric sensor array and distributed fiber optic temperature / pressure sensors embedded in the upper and lower pressure plates of the hot press for acquiring micro-stress wave signals inside the slab and temperature and pressure data across the entire area; and a non-contact dielectric measurement probe and a short-wave near-infrared spectral probe deployed on the side of the hot press for simultaneously acquiring dielectric parameters and near-infrared spectral data during the adhesive curing process; all sensors are hardware-level synchronized through a high-precision industrial timing synchronization controller to ensure that multi-source heterogeneous data have a unified timestamp.

3. The environmentally friendly, compression-resistant engineered wood panel performance prediction system based on multimodal data analysis according to claim 1, characterized in that, The preprocessing and feature extraction of the multimodal early precursor signals include: an internal micro-stress wave signal processing stream, which performs noise reduction, event detection and segmentation on the voltage signals collected by the piezoelectric sensor array, extracts acoustic emission signal segments representing micro-damage events, and generates feature vectors containing energy, frequency and wavelet packet energy spectrum through transformation calculation, while statistically analyzing the macroscopic temporal characteristics of events within a set process time window; a dielectric-near-infrared signal processing stream for the curing process, which analyzes the dielectric parameters and spectral data collected by the dielectric measurement probe and near-infrared spectral probe, extracts the peak temperature, area under the curve and characteristic absorption peak area decay rate reflecting the adhesive curing reaction, and fuses the synchronously collected temperature and pressure data, and estimates the real-time degree of curing online through a multivariate nonlinear regression model; and a three-dimensional image processing stream for the paving morphology, which performs noise reduction and voxelization on the three-dimensional point cloud data collected by the high-speed structured light three-dimensional scanner, reconstructs the three-dimensional density distribution matrix, and applies three-dimensional image analysis algorithms to calculate the pore network connectivity, fiber orientation anisotropy and surface area-to-volume ratio topological features.

4. The environmentally friendly, compression-resistant engineered wood panel performance prediction system based on multimodal data analysis according to claim 3, characterized in that, The internal micro-stress wave signal processing flow includes: bandpass filtering of the acquired raw voltage signal to filter out low-frequency noise caused by mechanical vibration of the equipment and retain the acoustic emission signal corresponding to the micro-damage event; using the short-time average zero-crossing rate and energy threshold method to automatically detect and segment burst-type acoustic emission signal segments representing fiber breakage or adhesive layer cracking, wherein the energy threshold is set to 3 times the historical noise energy; performing fast Fourier transform and wavelet packet transform on each segmented event signal segment to calculate its key features and form a feature vector containing signal energy, peak frequency, duration and energy spectrum based on wavelet packet transform; within a preset process time window, statistically analyzing the count rate, cumulative energy and average frequency of all acoustic emission events within the window to form a macroscopic time-series evolution curve synchronized with the hot pressing process.

5. The environmentally friendly, compression-resistant engineered wood panel performance prediction system based on multimodal data analysis according to claim 3, characterized in that, The dielectric-near-infrared signal processing flow of the curing process includes: analyzing the dielectric loss factor curve acquired in real time by the dielectric measurement probe to extract the peak temperature and area under the curve characteristics reflecting the curing reaction of the adhesive; performing second derivative preprocessing on the spectral data acquired by the near-infrared spectral probe to eliminate baseline drift and scattering interference, and monitoring the decay rate of the absorption peak area of ​​specific functional groups; fusing the analyzed dielectric characteristics, spectral characteristics, and synchronously acquired hot-pressing temperature and pressure data to establish a multivariate nonlinear regression model between the actual degree of curing obtained by the benchmark differential scanning calorimetry test, and outputting the estimated value of the real-time degree of curing and its first derivative online.

6. The environmentally friendly, compression-resistant engineered wood panel performance prediction system based on multimodal data analysis according to claim 3, characterized in that, The 3D image processing flow for the paving morphology includes: performing statistical filtering on the 3D point cloud data acquired by the high-speed structured light 3D scanner to remove outlier noise and retain the true 3D coordinate information of the fibers / rubber scraps; performing voxelization on the denoised point cloud data, and generating a 3D density distribution matrix after paving by calculating the point density within each voxel unit to quantify the uniformity of the paving; performing binarization on the 3D density distribution matrix, and applying a 3D image analysis algorithm to calculate the pore network topology features including pore connectivity, average pore diameter, and pore tortuosity, as well as the fiber orientation tensor reflecting the degree of anisotropy in the fiber / rubber scrap distribution direction, and quantifying the ratio of the total surface area to volume available for adhesive bonding between fibers.

7. The environmentally friendly, compression-resistant engineered wood panel performance prediction system based on multimodal data analysis according to claim 1, characterized in that, The selection of the core feature subset includes: compiling a historical dataset containing all features output by the early perception module and laboratory-measured compressive strength; applying a constraint-based causal discovery algorithm to perform conditional independence testing and edge orientation on the historical dataset to construct a directed acyclic graph (DAG) describing the potential causal relationship between features and compressive strength; identifying all directed paths from input features to compressive strength in the DAG, and estimating the causal flow intensity of each path based on the product of the causal effects of each edge; and selecting, based on the DAG and the causal flow intensity, a feature set that serves as the direct parent node of compressive strength and a feature set of important intermediate nodes located on paths with high causal flow intensity, the union of which constitutes the core feature subset.

8. The environmentally friendly, compression-resistant engineered wood panel performance prediction system based on multimodal data analysis according to claim 1, characterized in that, The construction of the prediction model includes: designing a multi-output physical information neural network, whose input is the causal enhancement fusion feature representation; the output of the neural network includes: the final predicted compressive strength, and physical intermediate variables including fiber volume fraction, matrix elastic modulus, and curing curve parameters, and the output values ​​of the physical intermediate variables are constrained within a preset reasonable physical range; training the neural network using a composite loss function, which includes at least: a data loss term to minimize the error between the predicted compressive strength and the laboratory measured value; and a physical loss term to quantify the deviation between the physical intermediate variables output by the network and the predefined materials science mechanism equation, which includes at least the composite material mixing law equation and the curing kinetics equation; and through end-to-end gradient optimization, ensuring that the neural network, while fitting the data, satisfies the constraints of the materials science mechanism in its internal representation and output.

9. The environmentally friendly, compression-resistant engineered wood panel performance prediction system based on multimodal data analysis according to claim 1, characterized in that, The updating of the prediction model parameters includes: collecting new batches of data pairs that have passed validity checks, where the input is the multimodal features processed by the early perception module, and the label is the laboratory measured compressive strength; periodically calculating the Fisher information matrix of the model parameters based on historical training data, where the diagonal elements are used to quantify the importance of each model parameter to the learned knowledge; constructing an incremental learning loss function, which minimizes the prediction loss of the new batch of data while adding a regularization term to penalize the deviation of important model parameters from the historical optimal value; updating the parameters in isolated model replicas, and deploying the updated model to the online service only when the prediction accuracy on both the historical validation set and the new data validation set has not decreased.

Images