Intelligent regulation and control method for production energy consumption of artificial board

Abstract

The invention discloses an artificial board production energy consumption intelligent regulation and control method, and belongs to the technical field of industrial production control, and the method comprises the steps: collecting real-time process and energy consumption data of drying and hot pressing processes through a sensor network, and integrating the real-time process and energy consumption data with production batch information; based on mechanism and data driving model weighted fusion, constructing a hybrid prediction model of the drying and hot pressing key indexes; a multi-objective optimization model is established, an improved genetic algorithm is adopted for rolling solution, and optimal process control parameters are output; the parameters are issued and executed, and actual quality indexes are collected for feedback; and dynamically triggering incremental learning of the data driving model according to the prediction error, and updating the hybrid prediction model. According to the method, the problems that in artificial board production, energy consumption is high, quality fluctuation is large, and fixed technological parameters are difficult to adapt to dynamic changes are solved, collaborative optimization of energy consumption and quality and continuous self-adaption of the production process are achieved, production energy consumption is effectively reduced, and product quality is stabilized.

Description

Technical Field

[0001] This invention relates to the field of industrial process control technology, and in particular to a method for intelligent control of energy consumption in the production of engineered wood products. Background Technology

[0002] The production of engineered wood products is a crucial link in the efficient utilization of forestry resources. Its production process mainly includes multiple steps such as drying, laying, hot pressing, and post-treatment. Among these, drying and hot pressing are the core energy-consuming processes and are also key to determining the final physical and mechanical properties and quality of the boards. The drying process removes moisture from the wood raw materials using hot air, consuming a significant portion of energy. Furthermore, the moisture content of the dried material directly affects the stability of the subsequent hot pressing process and the product quality. The hot pressing process shapes the boards and cures the adhesives under high temperature and pressure. The temperature and pressure curves in this process have a decisive impact on the final internal bond strength of the boards, and its energy consumption is also very significant.

[0003] In current industrial production, enterprises generally face the challenge of balancing energy consumption and product quality in order to achieve stable product quality and controllable production costs. Specifically, the main problems are as follows: First, the energy consumption of the production process is high. The heat energy required for drying and the electricity and heat energy required for hot pressing constitute a major part of the production cost, making energy conservation and consumption reduction an urgent need. Second, product quality, especially key indicators such as internal bond strength, is prone to fluctuations due to factors such as batch differences in raw materials, environmental fluctuations, and changes in equipment condition, affecting product qualification rate and consistency. In addition, the production process has complex characteristics such as multiple variables, strong coupling, nonlinearity, and time-varying nature, and there are also mutual influences between the drying and hot pressing processes.

[0004] To address these issues, existing technologies typically employ two main approaches. One approach relies on fixed process parameter specifications. Operators, based on experience or laboratory test results, set a relatively fixed set of drying temperatures, air velocities, and hot-pressing temperature and pressure curves for different types of boards. While simple and easy to implement, this approach lacks flexibility and cannot be dynamically adjusted according to real-time operating conditions. When actual conditions deviate from the preset conditions, it often leads to energy waste or quality defects, making it difficult to achieve optimal operation. The other approach utilizes models based on historical data for optimization. For example, models based on statistical regression or simple neural networks are built to predict quality or energy consumption and attempt to optimize some parameters. However, this type of method has significant limitations. Purely data-driven models heavily rely on large amounts of high-quality historical data, lack predictive ability for new products or new raw materials, have poor interpretability, and are prone to failure and decreased prediction accuracy when operating conditions drift significantly. While models that rely solely on physical mechanisms are highly interpretable, they are often heavily simplified due to the extreme difficulty in providing a completely accurate mathematical description of the complex heat and mass transfer and chemical reactions involved in the production of engineered wood products. This results in limited predictive accuracy in practical applications and an inability to fully characterize the complex nonlinear relationships and unmodeled disturbances in production.

[0005] Therefore, existing technologies for energy consumption and quality control in wood-based panel production generally suffer from poor adaptability, limited optimization dimensions, and an inability to self-improve in long-term operation. This makes it difficult for the production process to maintain an optimal or near-optimal state in terms of both energy consumption and quality, thus restricting further improvement in production efficiency and effectiveness. There is an urgent need for an intelligent control method that can deeply integrate process knowledge and actual data, possess online learning and dynamic optimization capabilities, and robustly cope with production fluctuations. Summary of the Invention

[0006] This invention overcomes the problems of high energy consumption, large quality fluctuations, and difficulty in adapting fixed process parameters to dynamic changes in the production of engineered wood products in the prior art. It provides an intelligent energy consumption control method for engineered wood production, realizing the synergistic optimization of energy consumption and quality and the continuous self-adaptation of the production process, effectively reducing production energy consumption and stabilizing product quality.

[0007] To achieve the above objectives, the present invention adopts the following solution: A method for intelligent control of energy consumption in wood-based panel production includes the following steps: S1: Collect real-time temperature, real-time pressure and real-time energy consumption data of the hot pressing process through a sensor network deployed on the production line, collect real-time temperature, real-time wind speed and real-time energy consumption data of the drying process, and obtain the raw material specification information of the current production batch from the production management system. S2: Based on the heat and mass transfer mechanism, construct the mechanism model of the drying process and the hot pressing process. Based on historical production data, use a long short-term memory network to train the data-driven model of the drying process and the hot pressing process respectively. Then, weightedly fuse the mechanism model and the data-driven model to form a hybrid prediction model of the drying process and a hybrid prediction model of the hot pressing process. S3: Based on the mixed prediction models of the drying process and the hot pressing process, a multi-objective optimization model is established with the optimization objectives of minimizing total energy consumption and maximizing the bonding strength within the product, and with the operating parameter range of each process equipment as constraints. The multi-objective optimization model is solved by a non-dominated sorting genetic algorithm with an elite strategy, and the target control parameters for the next control cycle are output in a rolling manner, including the target temperature and target wind speed of the drying process, and the target temperature and target pressure curves of the hot pressing process. S4: Send the target control parameters to the drying equipment controller and hot press controller of the production line for control execution, and after execution, collect the actual moisture content of the dried outlet material and the actual internal bond strength of the hot-pressed plate. S5: Using the real-time data collected in step S1 as input, and the actual moisture content and actual internal bonding strength collected in step S4 as verification output, calculate the prediction error of the mixing prediction model for the drying process and the mixing prediction model for the hot pressing process. When the prediction error exceeds the set threshold, use the real-time data, actual moisture content, and actual internal bonding strength to perform incremental learning on the long short-term memory network, and update the data-driven model part in the mixing prediction model for the drying process and the mixing prediction model for the hot pressing process.

[0008] Preferably, in step S1, the sensor network includes temperature sensors and hot-wire wind speed sensors respectively installed at the inlet, outlet, and middle of the drying cylinder to collect real-time temperature and real-time wind speed during the drying process; the sensor network also includes thermocouple temperature sensors and pressure sensors installed on each heating plate layer of the hot press to collect real-time temperature and real-time pressure during the hot pressing process; real-time energy consumption data is collected through the power metering module installed on the drive motor of the drying process and the hydraulic system of the hot press, as well as the flow meter installed on the steam heating pipeline; raw material specification information includes wood type, initial moisture content, and target board density; all collected real-time data are assigned a unified timestamp, and are aligned and integrated based on the timestamp and batch switching signal from the production management system to form a time-series data sequence based on production batches; when the real-time data of any sensor continuously exceeds its corresponding normal range within a preset time period, or when a timestamp breaks in the time-series data sequence, the data for that period is marked as invalid data, and an interpolation algorithm based on historical data from the same period is used to fill in the invalid data to generate a continuous and complete time-series data sequence.

[0009] As a preferred option, the method for constructing the mechanism model in step S2 is as follows: A mechanistic model for the drying process is constructed based on the principle of convection drying. Its inputs are the real-time temperature, real-time wind speed, and initial moisture content in the raw material specifications. The output is the predicted moisture content of the material at the drying outlet. A mechanistic model for the hot pressing process is constructed based on the principles of heat conduction and resin curing reaction kinetics. Its inputs are the real-time temperature, real-time pressure curve, and target board density in the raw material specifications. The output is the predicted internal bond strength of the board. When constructing the mechanistic model, an empirical correction coefficient related to the wood pore structure and thermal conductivity is introduced for the wood species in the raw material specifications to calibrate the mass transfer and heat transfer parameters in the mechanistic model. The specific method for training a data-driven model is as follows: Historical time-series data sequences with the same wood species and target board density as the current production batch were extracted from the historical database as training samples. The input features of the drying process data-driven model were the real-time temperature, real-time wind speed sequence, and initial moisture content of the historical drying process, and the output label was the actual moisture content of the material at the corresponding historical drying outlet. The input features of the hot pressing process data-driven model were the real-time temperature, real-time pressure curve sequence, and target board density of the historical hot pressing process, and the output label was the actual internal bond strength of the corresponding historical board. When constructing the input features, a sliding time window method was used to combine the operating condition data from multiple consecutive sampling times into a feature vector to capture the dynamic time-varying characteristics of the production process. The number of hidden layer nodes of the Long Short-Term Memory network was determined by cross-validation based on the number of training samples.

[0010] As a preferred method, the specific approach for weighted fusion to form a mixed prediction model for the drying process and a mixed prediction model for the hot pressing process is as follows: For the mixed prediction model of the drying process, the final predicted moisture content of the dried outlet material is obtained by weighted summation of the predicted output of the drying process mechanism model and the predicted output of the drying process data-driven model; for the mixed prediction model of the hot pressing process, the final predicted internal bond strength of the plate is obtained by weighted summation of the predicted output of the hot pressing process mechanism model and the predicted output of the hot pressing process data-driven model. The weighting coefficients used in the weighted fusion are dynamically calculated based on the predictive performance of the mechanistic model and the data-driven model on recent historical data. The system continuously records the predicted values ​​of the moisture content of the dried outlet material and the internal bonding strength of the board by the two models within the predetermined production cycle in the past, and compares them with the actual moisture content and actual internal bonding strength collected in step S4. The root mean square inverse of the prediction error of each model is calculated and normalized to serve as the dynamic weighting coefficient of the corresponding model under the current working condition. The dynamic weighting coefficients are updated once after each production batch is completed. During the update, the exponential smoothing method is used to merge the new weighting coefficients calculated this time with the old weighting coefficients used in the previous batch.

[0011] Preferably, step S3 specifically includes: A multi-objective optimization model is established, where the first objective function is the predicted total energy consumption within a future control cycle, calculated based on the mixed prediction models for the drying and hot-pressing processes. The predicted total energy consumption is the sum of the predicted energy consumption for the drying and hot-pressing processes. The second objective function is the predicted internal bond strength of the board material, calculated based on the mixed prediction model for the hot-pressing process. Optimization variables include the target temperature of the drying process, the target wind speed of the drying process, the target temperature of the hot-pressing process, and the sequence of pressure setpoints constituting the target pressure curve of the hot-pressing process. Constraints include the operating range of temperature and wind speed of the drying equipment, the operating range of temperature of each heating plate layer of the hot press, the upper and lower limits of the pressure setpoints, and the minimum resin curing time requirement that the pressure curve of the hot-pressing process must meet, determined based on the raw material specifications. A non-dominated sorting genetic algorithm with an elitist strategy is used to solve the multi-objective optimization model. Each individual in the algorithm population represents a combination of values ​​for a set of optimization variables. The predicted total energy consumption and predicted bonding strength of the board corresponding to each individual are evaluated through a mixed prediction model of the drying process and a mixed prediction model of the hot pressing process. Iterative evolution is performed based on non-dominated sorting and crowding calculation. At the beginning of each control cycle, the current real-time temperature, real-time pressure and real-time energy consumption data are obtained as model inputs. The algorithm is executed and the target temperature, target wind speed of the drying process, target temperature of the hot pressing process and target pressure curve of the hot pressing process to be executed in that control cycle are output in a rolling manner.

[0012] As a preferred option, the specific process of using a non-dominated sorting genetic algorithm with an elitist strategy to solve the multi-objective optimization model is as follows: Initialize algorithm parameters, including setting initial values ​​for population size, maximum number of iterations, crossover probability, and mutation probability; randomly generate an initial population, where each individual in the population represents a combination of values ​​for a set of optimization variables using real-number encoding; enter the iteration loop, and in each iteration, use the mixed prediction model for the drying process and the mixed prediction model for the hot pressing process to evaluate the predicted total energy consumption and predicted internal bonding strength of each individual in the current population. All individuals are stratified according to the non-dominated ranking, and then ranked according to the crowding degree within each non-dominated stratum. An elite retention strategy is implemented based on the ranking results, and individuals with higher rankings are directly retained to the next generation of the population. Simulated binary crossover and polynomial mutation operations are performed on individuals in the population to generate offspring individuals. The probabilities of crossover and mutation operations are adjusted using an adaptive mechanism, which is dynamically calculated based on the current iteration number and the dispersion of the fitness of individuals in the population. Determine if the algorithm's termination condition is met. If it is, select the individual with the highest crowding from the final generation's elite individual set, and use the optimized variable values ​​obtained from decoding it as the output of the target temperature, target wind speed, target temperature, and target pressure curves for the drying process, the hot pressing process, and the hot pressing process. If the condition is not met, proceed to the next iteration loop.

[0013] As a preferred option, when establishing a multi-objective optimization model, a third optimization objective function is also included, which is the energy consumption volatility rate. The energy consumption volatility rate is obtained by calculating the sum of squares of the changes in the total predicted energy consumption between adjacent sampling times in a future control cycle. When using a non-dominated sorting genetic algorithm with an elitist strategy to solve the problem, the iterative evolutionary process also includes: In each generation of the population, an elite set of individuals is selected based on non-dominated ranking and crowding calculation. The optimized variable values ​​of all individuals in this elite set are decoded and compared with the historical optimal process parameter set stored in the historical database, which has the lowest actual energy consumption and qualified actual internal bond strength under the same wood species and target board density conditions. The similarity is calculated based on the Euclidean distance between the optimized variable values. If the calculated similarity is lower than a set threshold, it is determined that the current population may be trapped in a local optimum. At this time, the historical optimal process parameter set is injected as a new individual into the next generation of the population to guide the search direction. The termination condition of the iteration process is that the number of iterations reaches the maximum preset number of generations, or the update rate of the elite individual set for several consecutive generations is lower than a preset ratio.

[0014] Preferably, step S4 specifically includes: The target temperature and target air velocity of the drying process, which are output in step S3, are sent to the drying equipment controller; the target temperature and target pressure curves of the hot pressing process are sent to the hot press controller; the drying equipment controller and the hot press controller drive the actuators to adjust the air inlet valve of the drying cylinder, the frequency converter of the induced draft fan, the heating system of the hot press, and the hydraulic proportional valve according to the received target parameters; after the parameter control is executed, the actual moisture content of the dried outlet material is collected by a near-infrared moisture meter installed at the outlet of the drying cylinder, and the actual internal bond strength of the plate is collected by an online mechanical property testing device after the hot-pressed plate has been cooled and cured; During parameter issuance and execution, the actual adjustment values ​​fed back by the drying equipment controller and the hot press controller are monitored in real time, and the actual adjustment values ​​are compared with the issued target parameters to calculate the real-time execution deviation. If any of the real-time execution deviations exceeds the first set threshold for a predetermined duration, the system determines that there is an execution abnormality in the control loop and activates the preset backup control strategy. Based on the mixed prediction model of the drying process and the mixed prediction model of the hot press process, the system recalculates the target temperature, target wind speed, target temperature, and target pressure curves of the drying process, the hot press process, and the target pressure curve of the hot press process within the remaining control cycle, and immediately switches to the recalculated parameter sequence to continue execution.

[0015] Preferably, step S5 specifically includes: For the drying process, the prediction error is set as the absolute value of the difference between the predicted moisture content of the dried material at the outlet output by the hybrid prediction model for the drying process and the actual moisture content collected in step S4. For the hot pressing process, the prediction error is set as the absolute value of the difference between the predicted internal bond strength of the sheet material output by the hybrid prediction model for the hot pressing process and the actual internal bond strength collected in step S4. The threshold for setting the prediction error is dynamically determined based on the statistical distribution of the corresponding prediction errors calculated in multiple consecutive production batches in the past, specifically taken as the preset quantile of the statistical distribution. When the prediction error of the drying process or the hot pressing process exceeds its corresponding dynamic threshold, the incremental learning process of the corresponding model is triggered. After triggering incremental learning, the real-time temperature, real-time pressure, real-time energy consumption data of the current production batch collected in step S1, as well as the raw material specification information, are used as new training samples. The actual moisture content or actual internal bonding strength collected in step S4 is used as the corresponding new sample label. These samples are then merged with some historical samples in the historical training sample library that have the same wood species and target board density to form an incremental training dataset. Using the incremental training dataset, the corresponding long short-term memory network in the drying process data-driven model or the hot-pressing process data-driven model is incrementally trained. During training, the structure above the hidden layer of the network is fixed, and only the weight parameters of the fully connected output layer of the network are updated through backpropagation. After training, the updated data-driven model part replaces the data-driven model part in the original hybrid prediction model, thus completing the model update.

[0016] As a preferred method, when constructing the incremental training dataset, the method for selecting historical samples from the historical training sample database is as follows: Calculate the cosine similarity between the feature vector of the newly added training sample in the current production batch and the feature vector of the historical sample, and select the top K historical samples with the highest similarity, where K is a preset number; at the same time, maintain a historical representative sample pool with a fixed capacity, which stores typical historical samples with low prediction errors for different wood species and target board density combinations; during each incremental learning, samples in the historical representative sample pool that have the same raw material specification information as the current production batch are also added to the incremental training dataset. After updating the weight parameters of the output layer of the Long Short-Term Memory network using the incremental training dataset, its performance is first evaluated on an independent validation dataset. The validation dataset consists of historical samples that did not participate in this incremental training. The evaluation metric is the mean of the prediction error. If the evaluation metric of the updated model is better than that of the model before the update on the original validation dataset, or if the degree of degradation is within an acceptable range, the system confirms that this incremental learning is effective and officially updates the updated model part into the hybrid prediction model for the drying process or the hybrid prediction model for the hot pressing process. Otherwise, the results of this incremental learning are discarded, the original model is retained and continues to run, and this invalid update event is recorded.

[0017] The present invention has at least the following beneficial effects: (1) By constructing a hybrid prediction model that integrates mechanism and data, and establishing a multi-objective rolling optimization model, it can automatically find and track the optimal combination of process parameters in real time, and achieve significant optimization of production energy consumption while ensuring stable product quality; (2) By weightedly fusing a mechanism model with physical interpretability with a data-driven model that can learn complex patterns, and dynamically adjusting the fusion weights based on recent prediction performance, the hybrid prediction model has both theoretical guidance and data adaptability; (3) The multi-objective rolling optimization mechanism can respond to real-time changes in operating conditions; the abnormal execution monitoring and autonomous re-optimization strategy enable the system to quickly adjust and smoothly transition when local equipment is abnormal; and the introduction of energy (3) The auxiliary optimization targets such as consumption volatility further promote the stability of production operation; (4) By setting dynamic thresholds to intelligently trigger incremental learning, and using a carefully designed sample selection strategy to fine-tune part of the network, and finally deploying and updating after strict verification, the hybrid prediction model can safely and robustly adapt to the slow changes of raw materials, equipment and environment, avoid the problem of performance degradation over time, and ensure the reliability of the system's long-term operation; (5) From data collection and integration, to model prediction and optimization decision-making, to control execution and effect feedback, and finally to model evaluation and updating, a complete and autonomous intelligent control loop is formed. It can also learn from each production result and continuously improve itself, truly realizing the intelligent upgrading of the production process. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the principle of the method of the present invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0020] like Figure 1 As shown, the intelligent energy consumption control method for wood-based panel production provided by the present invention includes the following steps: S1: Collect real-time temperature, real-time pressure, and real-time energy consumption data of the hot pressing process through a sensor network deployed on the production line, and collect real-time temperature, real-time wind speed, and real-time energy consumption data of the drying process. At the same time, obtain the raw material specification information of the current production batch from the production management system.

[0021] The sensor network extensively covers the two key energy-consuming processes of drying and hot pressing. The data it collects includes, but is not limited to, real-time temperature, real-time wind speed, and corresponding energy consumption data in the drying process, and real-time temperature, real-time pressure, and corresponding energy consumption data in the hot pressing process. These sensors include non-contact infrared temperature sensors, hot-wire anemometers, pressure transmitters, and energy metering modules, arranged at the inlet, outlet, and middle sections of the drying cylinder, as well as near the heating plates of each layer of the hot press. Simultaneously, it acquires raw material specification information for the current production batch from the production management system, such as wood type, initial moisture content, and target board density. All collected real-time data streams are assigned a unified high-precision timestamp and transmitted to the central processing unit via a data bus. In the central processing unit, based on the timestamp and batch start / stop signals from the production management system, heterogeneous data from different sources are time-aligned and integrated, ultimately forming a time-series data sequence indexed by a single production batch and containing multi-dimensional parameters.

[0022] S2: Based on the heat and mass transfer mechanism, construct the mechanism model of the drying process and the hot pressing process. Based on historical production data, use a long short-term memory network to train the data-driven model of the drying process and the hot pressing process respectively. Then, weightedly fuse the mechanism model and the data-driven model to form a hybrid prediction model of the drying process and a hybrid prediction model of the hot pressing process.

[0023] Based on the fundamental principles of convective heat and mass transfer during drying and the principles of heat conduction and resin curing reaction kinetics during hot pressing, mechanistic models for the drying and hot pressing processes are constructed respectively. The mechanistic models use real-time process parameters (such as temperature, pressure, and wind speed) and raw material specifications as input, and output theoretical predictions of moisture content and internal bond strength by solving built-in physicochemical equations. However, actual production is affected by numerous unmodeled factors; therefore, a data-driven model based on a long short-term memory network is trained in parallel using accumulated historical production data. This network excels at handling time-series data; its training samples come from a historical database, with inputs being a sequence of operating parameters within a past time window, and outputs the corresponding actual moisture content or internal bond strength. During training, the network structure (such as the number of hidden layer nodes, which can be determined through cross-validation, for example, optimizing within the range of 50 to 200 nodes) and learning rate are adjusted to allow the model to learn the complex nonlinear dynamic relationships in the data. Finally, the mechanistic model and the trained data-driven model are weighted and fused to form the final hybrid prediction model for both the drying and hot pressing processes. The fusion weights can be dynamically adjusted based on the recent prediction accuracy of each model. Models with higher accuracy are assigned higher weights. The hybrid model combines physical interpretability and data adaptability, improving the robustness and accuracy of predictions.

[0024] S3: Based on the mixed prediction models of the drying process and the hot pressing process, a multi-objective optimization model is established with the optimization objectives of minimizing total energy consumption and maximizing the bonding strength within the product, and with the operating parameter range of each process equipment as constraints. The multi-objective optimization model is solved by a non-dominated sorting genetic algorithm with an elite strategy, and the target control parameters for the next control cycle are output in a rolling manner, including the target temperature and target wind speed of the drying process, and the target temperature and target pressure curves of the hot pressing process.

[0025] The optimization objectives can be set in two competing directions: first, to minimize the predicted total energy consumption over a future control cycle (covering drying and hot-pressing processes); and second, to maximize the predicted internal bond strength of the hot-pressed sheet (representing product quality). The optimization variables are the key process parameters that need to be controlled, mainly including the target temperature and target air velocity for the drying process, and the target temperature and a time-varying target pressure curve for the hot-pressing process. The solution process needs to be carried out under various constraints, such as the allowable operating range of temperature, pressure, and air velocity of the equipment (e.g., the drying temperature range may be between 120℃ and 180℃, and the hot-pressing pressure range is between 2MPa and 5MPa), and the minimum curing time requirement determined by the resin characteristics. To efficiently solve this complex multi-objective optimization problem, a non-dominated sorting genetic algorithm with an elitist strategy is adopted. This algorithm initializes a population consisting of randomly generated combinations of process parameters, each combination representing a possible solution. In each generation of evolution, the predicted total energy consumption and predicted internal bond strength corresponding to each solution are evaluated using a hybrid prediction model. Then, solutions are screened, crossovered, and mutated based on non-dominated sorting and crowding distance, retaining excellent elitist solutions and exploring new parameter spaces. The optimization is performed on a rolling basis. At the beginning of each control cycle (which can be anywhere from a few minutes to tens of minutes), the optimization calculation is re-executed based on the latest real-time data, and the optimal sequence of control parameters to be executed in that cycle is output, thereby achieving dynamic and adaptive process optimization.

[0026] S4: Send the target control parameters to the drying equipment controller and hot press controller of the production line for control execution, and after execution, collect the actual moisture content of the material at the drying outlet and the actual internal bond strength of the hot-pressed plate.

[0027] The optimized target temperature and wind speed settings for the drying process are transmitted to the programmable logic controller (PLC) or distributed control system (DCS) of the drying equipment via an industrial communication network (such as Profinet or EtherCAT). Similarly, the target temperature and pressure curves for the hot pressing process are transmitted to the dedicated controller for the hot press. Upon receiving the commands, the controller drives the corresponding actuators, such as adjusting the opening of the air inlet valve of the drying drum to control the hot air temperature, adjusting the frequency of the induced draft fan inverter to control the wind speed, adjusting the power of the heating elements of the hot press to control the plate temperature, and controlling the hydraulic proportional valve to achieve the preset pressure curve. After the control parameters are executed, the system collects actual production result data: at the outlet of the drying drum, the actual moisture content of the outlet material is measured online using a near-infrared moisture meter; after the plates have undergone cooling and curing following the hot press, their actual internal bond strength is determined by online mechanical property testing equipment (such as an online internal bond strength tester) or by sampling and laboratory equipment. The actual measured values ​​represent the true product quality status under the current control parameters, providing feedback information for subsequent model verification and updates.

[0028] S5: Using the real-time data collected in step S1 as input, and the actual moisture content and actual internal bonding strength collected in step S4 as verification output, calculate the prediction error of the mixing prediction model for the drying process and the mixing prediction model for the hot pressing process. When the prediction error exceeds the set threshold, use the real-time data, actual moisture content, and actual internal bonding strength to perform incremental learning on the long short-term memory network, and update the data-driven model part in the mixing prediction model for the drying process and the mixing prediction model for the hot pressing process.

[0029] The system calculates the prediction errors for two key indicators: for the drying process, the error is the absolute value of the difference between the predicted moisture content output by the hybrid prediction model and the actual moisture content measured in step S4; for the hot pressing process, the error is the absolute value of the difference between the predicted internal bond strength and the actual internal bond strength. The system sets a dynamic threshold for each error, which is automatically calculated based on the statistical distribution of recent historical errors (e.g., taking the 90th quantile or mean of the error sequence plus several times the standard deviation), rather than a fixed value. When any error exceeds its corresponding dynamic threshold, it indicates that the current operating condition may have drifted or the model may be unsuitable. At this point, the incremental learning process of the data-driven model part of the corresponding hybrid prediction model is triggered. The real-time data of the current batch collected in step S1 and the actual quality data collected in step S4 are used as new training samples, combined with some relevant historical samples selected from the historical database, to form a small-scale incremental training dataset. Then, only the output layer or the last few layers of the Long Short-Term Memory network are fine-tuned during training, while the feature extraction structure of the network front end is fixed, thus quickly adapting to new operating conditions with minimal computational cost. The updated data-driven model was partially replaced by the original hybrid prediction model, completing one iteration of the model update and enabling its predictive capabilities to continuously adapt to changes in the production process.

[0030] Existing technologies often employ fixed process parameters or simple control strategies based on a single model, which struggle to cope with disturbances such as raw material fluctuations and equipment status changes, frequently resulting in high energy consumption and unstable product quality. This method, through the deployment of a comprehensive sensor network and information system integration, achieves real-time perception and fusion of data across the entire production process, enabling refined management. By innovatively fusing mechanistic and data-driven models to construct a hybrid prediction model, it ensures the model's physical rationality under normal operating conditions while enhancing its ability to capture complex nonlinear relationships and unmodeled disturbances, significantly improving the accuracy of predictions for key quality indicators and energy consumption. Employing a multi-objective rolling optimization strategy, it can find and track the optimal combination of process parameters with the lowest energy consumption in real time, while ensuring product quality meets requirements, thus achieving significant energy consumption optimization. The entire system forms a complete closed loop of "perception-prediction-optimization-execution-evaluation-learning," and through an online incremental learning mechanism, the model can continuously improve itself as the production process progresses, exhibiting excellent long-term adaptability. Therefore, this method can not only effectively reduce the unit product energy consumption of wood-based panels and save production costs, but also improve the consistency and stability of product quality, reduce the scrap rate, and enhance the production line's flexible production capacity to cope with changes in various raw materials and production tasks, thereby improving the overall level of intelligence and comprehensive economic benefits of the production process.

[0031] In another technical solution, in step S1, the sensor network includes temperature sensors and hot-wire wind speed sensors respectively installed at the inlet, outlet, and middle of the drying cylinder to collect real-time temperature and real-time wind speed during the drying process; the sensor network also includes thermocouple temperature sensors and pressure sensors installed on each heating plate layer of the hot press to collect real-time temperature and real-time pressure during the hot pressing process; real-time energy consumption data is collected through the power metering module installed on the drive motor of the drying process and the hydraulic system of the hot press, as well as the flow meter installed on the steam heating pipeline; raw material specification information includes wood type, initial moisture content, and target board density; all collected real-time data are assigned a unified timestamp, and are aligned and integrated based on the timestamp and batch switching signal from the production management system to form a time-series data sequence based on production batches; when the real-time data of any sensor continuously exceeds its corresponding normal range within a preset time period, or when a timestamp breaks in the time-series data sequence, the data for that period is marked as invalid data, and an interpolation algorithm based on historical data from the same period is used to fill in the invalid data to generate a continuous and complete time-series data sequence.

[0032] The sensor network strategically covers key physical parameters for the two core energy-consuming processes: drying and hot pressing. For the drying process, to accurately reflect changes in hot air conditions experienced by the material within the drying drum, temperature sensors and hot-wire anemometers are deployed at the material inlet, the central critical reaction zone, and the material outlet. The inlet sensor monitors the initial hot air conditions, the central sensor monitors process stability, and the outlet sensor confirms the final environmental parameters acting on the material. These temperature sensors can be selected from platinum resistance thermometers or thermocouples, and their measurement range typically covers room temperature to 250 degrees Celsius to meet process requirements. The hot-wire anemometers can capture airflow velocity in real time, with a range between 0 and 30 meters per second to accommodate different drying intensities. For the hot pressing process, thermocouple temperature sensors are integrated into each heating plate of the hot press to accurately measure the surface temperature of each plate, ensuring uniform heat transfer during hot pressing. Simultaneously, a high-precision pressure sensor is integrated into the hydraulic system to monitor and record the real-time pressure applied to the slab by the press. By installing a smart energy metering module in the power supply circuit of the drive motor (such as a fan motor) in the drying process, its active energy consumption can be collected in real time; a similar module is installed in the main motor circuit of the hot press hydraulic system. Furthermore, to calculate steam heating costs, vortex flow meters or orifice flow meters are installed on the main steam pipeline leading to the drying drum and the hot press heating system to measure the volumetric or mass flow rate of steam consumed.

[0033] All real-time data collected, whether from physical sensors (temperature, pressure, wind speed, energy consumption readings) or from production management systems (e.g., the type of wood used in the current batch might be pine or poplar, the initial moisture content of the material might be in the range of 30%-60%, and the density of the target board might be 650-750 kg / m³), is automatically assigned a unified millisecond-level timestamp based on a high-precision clock when transmitted to the central data processing unit. Batch information from the production management system (e.g., batch number, start time, end time) provides a logical division of production activities. The system runs a data alignment and integration engine that, based on the unified timestamp, interpolates or resamples sensor data streams with different sampling frequencies, aligning them to the same time grid. Simultaneously, based on batch switching signals (e.g., signals triggered when the production line starts processing a new batch of slabs), the continuous data stream is segmented and labeled into data packets based on production batches. Each data packet contains time-aligned multidimensional time-series data of the entire process from start to finish of the batch. For example, at each point in time, there are corresponding drying temperature, drying wind speed, hot pressing temperature, hot pressing pressure, instantaneous power, cumulative energy consumption, and associated static batch attributes (wood type, target density, etc.).

[0034] The system incorporates real-time data quality monitoring rules. For each sensor, a physically reasonable normal operating range is preset (e.g., the normal range for drying temperature at a certain point might be set between 100℃ and 200℃). The system continuously monitors the readings of each sensor. If a sensor's reading exceeds its normal range for a continuous preset time period (this duration can be adjusted according to process stability, such as 5 seconds, 10 seconds, or 30 seconds), it is determined that the sensor may be drifting, malfunctioning, or experiencing abnormal interference, and the data from that sensor during this time period is marked as invalid. The system checks the integrity of the time-series data sequence after timestamp alignment. If discontinuities or breaks in the timestamps are found in the sequence (i.e., data is not arriving at expected times), this usually indicates communication interruption or data loss, and the data for the missing time period is also marked as invalid. For data marked as invalid, an interpolation algorithm based on historical data from the same period is used. For example, multiple historical batches of data from the same production batch type (same wood, same target density) and the same production stage (e.g., the minute after drying begins) can be searched in a historical database. The average value or trend of the sensor parameter in the same historical period can be calculated, and this calculated trend value can be used to fill the current invalid data points. In this way, a complete time-series data sequence that is continuous in time and reasonable in numerical value can be generated, providing stable data input for modeling and analysis steps. This effectively avoids the risk of the entire system analysis being interrupted or producing misleading results due to single-point data anomalies or temporary loss. Compared with the problems of incomplete sensor deployment, isolated data sources, and lack of effective data cleaning mechanisms in existing technologies, this solution achieves refined and comprehensive monitoring of key process parameters and energy consumption parameters throughout the drying and hot-pressing process by scientifically deploying multiple types of sensors. This overcomes the drawbacks of sparse monitoring points and incomplete parameters in traditional methods. By introducing a unified timestamp and intelligently aligning and integrating it with production batch information, isolated equipment operation data, energy consumption data, and production management information were successfully merged into a structured time-series sequence with a clear production context, solving the problem of data silos. The built-in real-time data quality monitoring and intelligent data filling mechanism greatly enhances the system's robustness to common field interferences (such as momentary sensor failures, signal interference, and communication delays). It can automatically identify and repair data anomalies, ensuring the continuity and reliability of the input data stream and avoiding model prediction distortion or optimization decision errors caused by data quality issues.

[0035] In another technical solution, the method for constructing the mechanism model in step S2 is as follows: A mechanistic model for the drying process is constructed based on the principle of convection drying. Its inputs are the real-time temperature, real-time wind speed, and initial moisture content in the raw material specifications. The output is the predicted moisture content of the material at the drying outlet. A mechanistic model for the hot pressing process is constructed based on the principles of heat conduction and resin curing reaction kinetics. Its inputs are the real-time temperature, real-time pressure curve, and target board density in the raw material specifications. The output is the predicted internal bond strength of the board. When constructing the mechanistic model, an empirical correction coefficient related to the wood pore structure and thermal conductivity is introduced for the wood species in the raw material specifications to calibrate the mass transfer and heat transfer parameters in the mechanistic model. The specific method for training a data-driven model is as follows: Historical time-series data sequences with the same wood species and target board density as the current production batch were extracted from the historical database as training samples. The input features of the drying process data-driven model were the real-time temperature, real-time wind speed sequence, and initial moisture content of the historical drying process, and the output label was the actual moisture content of the material at the corresponding historical drying outlet. The input features of the hot pressing process data-driven model were the real-time temperature, real-time pressure curve sequence, and target board density of the historical hot pressing process, and the output label was the actual internal bond strength of the corresponding historical board. When constructing the input features, a sliding time window method was used to combine the operating condition data from multiple consecutive sampling times into a feature vector to capture the dynamic time-varying characteristics of the production process. The number of hidden layer nodes of the Long Short-Term Memory network was determined by cross-validation based on the number of training samples.

[0036] The mechanistic model for the drying process is based on the principle of convective heat and mass transfer, treating the drying cylinder as a system of heat and mass exchange between hot air and wet material. Its input variables include the real-time temperature, real-time air velocity, and initial moisture content of the raw material obtained from step S1. Internally, the model solves a set of differential equations describing the migration of moisture from the material's interior to its surface and then evaporation into the hot air, ultimately outputting a theoretical prediction of the moisture content of the material at the drying outlet. For the hot-pressing process, the mechanistic model integrates the principles of heat conduction and resin curing reaction kinetics. The model inputs include the real-time temperature, real-time pressure curve, and target board density of the hot-pressing process. This model simulates the process of heat transfer from the hot-pressing plate through the board blank and couples the kinetic equations of the resin's curing and cross-linking reaction under specific temperature and pressure conditions, thereby calculating the predicted internal bond strength of the board. To make the generalized physical equations more closely reflect specific production raw materials, the model introduces empirical correction coefficients related to wood species. Different types of wood (such as pine, eucalyptus, poplar, or mixed materials in different proportions) have inherent differences in pore structure, fiber morphology, and thermal conductivity, directly affecting the heat and mass transfer rates. Therefore, the system maintains an empirical coefficient library, pre-setting a correction coefficient for each common wood species (for example, for wood with high porosity, the mass transfer coefficient correction value is selected between 1.1 and 1.3; for dense wood, it is selected between 0.8 and 1.0). When running the mechanistic model, the system automatically calls the corresponding correction coefficient based on the wood species in the current batch of raw material specifications to calibrate key parameters such as mass transfer coefficient or thermal conductivity within the model, thereby significantly improving the basic prediction accuracy of the mechanistic model under different raw material scenarios.

[0037] The training data is selected from historical production databases, filtering complete time-series data records of historical batches with the same key raw material specifications (i.e., the same wood species and target board density) as the current production batch to be predicted. These historical records constitute the training sample set. For the drying process data-driven model predicting the moisture content at the drying outlet, its input features are constructed as a real-time drying temperature sequence and a real-time wind speed sequence within a historical time window, spliced ​​with the initial moisture content value of the batch of material; its output label (i.e., the learning target) is the actual measured moisture content of the corresponding batch of material at the drying outlet. Similarly, for the hot-pressing process data-driven model predicting the internal bond strength, the input features are a real-time hot-pressing temperature sequence and a real-time pressure curve sequence within a historical time window, spliced ​​with the target board density; the output label is the actual measured internal bond strength of the board. To effectively capture the dynamics of the production process and the temporal dependencies between parameters (e.g., the temperature and pressure state at a previous moment affects the curing process at subsequent moments), a sliding time window method is used when constructing the input features. Specifically, for a given historical point in time, all operating condition parameter values ​​from multiple consecutive sampling points prior to it are arranged chronologically and combined into a multi-dimensional feature vector. This method enables the LSTM (Long Short-Term Memory) network to learn and remember short-term historical operating condition patterns, thereby making more accurate predictions. The network structure itself also needs to be adapted. The setting of the key hyperparameter, the number of hidden layer nodes (such as possible configurations of 64, 128, or 256 nodes), is determined automatically or semi-automatically based on the number of currently available training samples through methods such as cross-validation, in order to achieve a good balance between model capacity and generalization ability.

[0038] The final mixed prediction model for the drying process outputs a predicted moisture content that is the weighted sum of the predicted outputs from the drying process mechanism model and the drying process data-driven model. Similarly, the mixed prediction model for the hot-pressing process outputs a predicted internal bond strength that is the weighted sum of the predicted values ​​from the hot-pressing process mechanism model and the data-driven model. These weighting coefficients are dynamic, calculated based on the relative predictive performance of the two models in recent historical data. The system continuously tracks and records the predicted values ​​for moisture content and internal bond strength from the mechanism model and the data-driven model over a predetermined period or production cycle (e.g., the most recent 5 or 10 production batches). After each batch is completed, step S4 collects the actual moisture content and actual internal bond strength, compares the predicted values ​​of each model with these actual values, and calculates the prediction error of each model for that batch. The calculation of the dynamic weighting coefficients is typically based on the reciprocal of the prediction error or a similar indicator. For example, the root mean square value of the prediction error of each model in recent batches can be calculated, and its reciprocal can be taken to characterize the accuracy (the higher the accuracy, the larger the reciprocal). Subsequently, the two reciprocals are normalized so that their sum equals one. The normalized results are then used as the dynamic weighting coefficients of the mechanistic model and the data-driven model under the current operating conditions. This means that the model with more accurate and reliable recent performance will have a greater say in the current stage of fusion prediction. The dynamic weighting mechanism allows the hybrid model to adaptively adjust its trust in theoretical prior knowledge and actual data patterns. When the operating conditions are stable and the theoretical model fits well, it relies more on the mechanistic model; when unmodeled disturbances or complex nonlinearities occur, it relies more on the patterns learned from the data, thus forming a complementary advantage. Compared with traditional methods that use only a single type of model, this modeling scheme constructs mechanistic models for the drying and hot pressing processes separately, enabling the prediction results to have good interpretability and reasonable trend extrapolation ability under normal conditions. The LSTM-based data-driven model can extract complex patterns, nonlinear relationships, and interference factors that are difficult to describe by equations from massive historical data, greatly enhancing the model's ability to characterize the complexity of actual production. By introducing empirical correction coefficients tied to wood species, the mechanistic model can be quickly and cost-effectively adapted to individual needs, improving its fundamental accuracy when handling various raw materials. A dynamic weighted fusion strategy based on recent predictive performance solves the problem of balancing theoretical and data aspects, forming a resilient model architecture with self-evaluation and self-adjustment capabilities. This hybrid model fusion scheme combines the robustness of the mechanistic model with the flexibility of the data-driven model, fundamentally enhancing the prediction accuracy, environmental adaptability, and long-term reliability of the final drying and hot-pressing process prediction model.

[0039] The specific method for weighted fusion to form a mixed prediction model for the drying process and a mixed prediction model for the hot pressing process is as follows: For the mixed prediction model of the drying process, the final predicted moisture content of the dried outlet material is obtained by weighted summation of the predicted output of the drying process mechanism model and the predicted output of the drying process data-driven model; for the mixed prediction model of the hot pressing process, the final predicted internal bond strength of the plate is obtained by weighted summation of the predicted output of the hot pressing process mechanism model and the predicted output of the hot pressing process data-driven model. The weighting coefficients used in the weighted fusion are dynamically calculated based on the predictive performance of the mechanistic model and the data-driven model on recent historical data. The system continuously records the predicted values ​​of the moisture content of the dried outlet material and the internal bonding strength of the board by the two models within the predetermined production cycle in the past, and compares them with the actual moisture content and actual internal bonding strength collected in step S4. The root mean square inverse of the prediction error of each model is calculated and normalized to serve as the dynamic weighting coefficient of the corresponding model under the current working condition. The dynamic weighting coefficients are updated once after each production batch is completed. During the update, the exponential smoothing method is used to merge the new weighting coefficients calculated this time with the old weighting coefficients used in the previous batch.

[0040] For the drying process, at each point where prediction is needed (e.g., to provide predictions of future operating conditions for optimization algorithms), the system runs the drying process mechanism model and the drying process data-driven model in parallel. The mechanism model calculates a theoretical predicted moisture content value using built-in physical equations based on input real-time temperature, wind speed, initial moisture content, etc.; the data-driven model, based on the same real-time data sequence, outputs a predicted moisture content value based on data patterns through forward propagation of its trained LSTM network. Subsequently, the system weights and sums these two predicted values ​​to obtain the final predicted moisture content of the dried material at the outlet, published by the hybrid prediction model for the drying process. The workflow of the hybrid prediction model for the hot-pressing process is exactly the same. It calculates the theoretical value of the internal bond strength given by the hot-pressing mechanism model and the data-predicted value of the internal bond strength given by the hot-pressing data-driven model in parallel, multiplies each value by its corresponding dynamic weight coefficient, and then sums them to obtain the final predicted internal bond strength of the sheet material. This calculation method ensures that, under any circumstances, the final prediction result is a balance between theoretical priors and data experience, and the magnitude of the weight coefficients directly determines the tendency of this balance. The weighting coefficients are driven by the objective accuracy of the two models in the latest actual production performance. The system maintains a rolling performance evaluation window covering several recently completed production batches (e.g., the window size can be set to 8, 12, or 15 batches to balance response speed and stability). For each completed batch within the window, the system stores the predicted values ​​made by the two models at key points during the batch's production process, as well as the true values ​​of the final actual moisture content and actual internal bond strength of the batch, fed back from step S4. When weight updates are needed, the system retrieves all data within the performance evaluation window for both the drying and hot-pressing processes. For each model (mechanistic model and data-driven model), the root mean square error (RMSE) of the deviation between all predicted values ​​and corresponding actual values ​​within the window period is calculated. The smaller the error, the better the model's recent performance. To convert the error into weights representing confidence, the system calculates the reciprocal of the RMSE for each model. Since a smaller RMSE results in a larger reciprocal, indicating higher model accuracy, a higher weight should be given; therefore, these two reciprocals are normalized. Through this process, the more accurate the model's recent predictions, the larger its calculated normalized weight coefficients will be, and the higher its contribution to the pooled predictions will be. This mechanism ensures that the weight coefficients always reflect the model's latest and proven reliability.

[0041] The system does not directly replace the old coefficients with the newly calculated weights. Instead, it employs an exponential smoothing method for fusion updates. Assuming the old weights used in the previous production batch are known, and the new weights calculated based on the latest performance evaluation window are also available, the system presets a smoothing factor (typically between 0 and 1, such as 0.2, 0.3, or 0.5; smaller values ​​indicate slower and smoother changes). Then, it calculates the final weights for the next batch according to the following logic: New batch weight = Smoothing factor × New calculated weight + (1 - Smoothing factor) × Old weight. This means the new weights are a weighted average of the current calculated values ​​and historically used values. This exponential smoothing update strategy filters out noise. Even if a batch experiences a temporary abnormal increase or decrease in the prediction error of a model due to accidental factors (such as an atypical measurement error or extremely special operating conditions), the smoothing mechanism effectively buffers the impact of this anomaly on the weights, preventing unnecessary large jumps in model contribution and ensuring the stability of the control system. Secondly, it retains certain historical information, making the evolution of weighting coefficients a gradual and continuous process, which is more in line with the characteristic of continuous change in actual production conditions. After each production batch is completed, the system performs such a smooth update, so that the hybrid prediction model's dependence on both the mechanism and data paradigms can adapt to the long-term drift and slow changes in the production process in a robust and gradual manner.

[0042] Compared to strategies using fixed weights or simple switching, dynamic weight calculation enables the model to continuously self-evaluate its performance and automatically adjust its internal structure based on the latest empirical results. This allows components with more accurate predictions to play a greater role, thereby maintaining a consistently high level of overall prediction accuracy. This weight allocation method based on objective data avoids the subjectivity of human intervention, making model fusion more scientific and adaptive. Furthermore, the introduction of an exponentially smoothing weight update strategy effectively solves the decision oscillation problem commonly found in adaptive systems due to short-term data fluctuations. By smoothly combining old and new weights, the system actively adapts to new changes while maintaining necessary inertia, making the prediction signals output to downstream optimization modules more stable and reliable. This avoids frequent and significant adjustments to the optimized process parameters due to sudden changes in upstream model weights, which is crucial for the stable operation of actual production lines.

[0043] In another technical solution, step S3 specifically includes: A multi-objective optimization model is established, where the first objective function is the predicted total energy consumption within a future control cycle, calculated based on the mixed prediction models for the drying and hot-pressing processes. The predicted total energy consumption is the sum of the predicted energy consumption for the drying and hot-pressing processes. The second objective function is the predicted internal bond strength of the board material, calculated based on the mixed prediction model for the hot-pressing process. Optimization variables include the target temperature of the drying process, the target wind speed of the drying process, the target temperature of the hot-pressing process, and the sequence of pressure setpoints constituting the target pressure curve of the hot-pressing process. Constraints include the operating range of temperature and wind speed of the drying equipment, the operating range of temperature of each heating plate layer of the hot press, the upper and lower limits of the pressure setpoints, and the minimum resin curing time requirement that the pressure curve of the hot-pressing process must meet, determined based on the raw material specifications. A non-dominated sorting genetic algorithm with an elitist strategy is used to solve the multi-objective optimization model. Each individual in the algorithm population represents a combination of values ​​for a set of optimization variables. The predicted total energy consumption and predicted bonding strength of the board corresponding to each individual are evaluated through a mixed prediction model of the drying process and a mixed prediction model of the hot pressing process. Iterative evolution is performed based on non-dominated sorting and crowding calculation. At the beginning of each control cycle, the current real-time temperature, real-time pressure and real-time energy consumption data are obtained as model inputs. The algorithm is executed and the target temperature, target wind speed of the drying process, target temperature of the hot pressing process and target pressure curve of the hot pressing process to be executed in that control cycle are output in a rolling manner.

[0044] The first optimization objective function is set to minimize the predicted total energy consumption over a future control cycle. This predicted total energy consumption is not directly measured but calculated using the hybrid prediction model established in step S2. Specifically, the system uses the hybrid prediction model for the drying process to predict the energy consumed in one control cycle under a set of process parameters to be evaluated (such as target temperature and wind speed). Simultaneously, it uses the hybrid prediction model for the hot pressing process to predict the corresponding energy consumption based on another set of process parameters to be evaluated (such as target temperature and pressure curve). Adding these two predicted energy consumption values ​​yields the predicted total energy consumption corresponding to this set of process parameters. The second optimization objective function is set to maximize the predicted internal bond strength of the sheet material. This strength value is directly output by the hybrid prediction model for the hot pressing process, which evaluates the key mechanical performance indicators that the final sheet material product can achieve under given hot pressing process parameters. By simultaneously optimizing energy consumption and quality, a multi-objective optimization problem is constructed. In industrial production, it is often necessary to find the optimal balance between energy saving and high quality, rather than simply sacrificing one to satisfy the other. The search direction needs to be guided towards a truly efficient and high-quality production point. The optimization variables are adjustable parameters directly executed by the control system, mainly comprising four parts: the target temperature of the drying process, the target air velocity of the drying process, the target temperature of the hot pressing process, and a target pressure curve describing how the pressure changes over time during the hot pressing process. This pressure curve is characterized by a series of pressure setpoints arranged in chronological order. For example, target pressure values ​​can be defined at key time points such as the 0th, 30th, and 60th seconds of the hot pressing cycle, and a continuous control curve can be fitted using these points. These variables are strictly constrained by the physical limits of the equipment and process safety requirements. For the drying process, the operating range of temperature and air velocity is determined by the design specifications of the drying equipment; for example, the temperature is typically limited to between 120°C and 180°C, and the air velocity is limited to between 5 m / s and 25 m / s. For the hot pressing process, the operating temperature range of each heating plate layer is also limited, for example, between 140°C and 220°C. The upper and lower limits of the pressure setpoint are determined by the hydraulic system capacity, typically between 2 MPa and 5 MPa. In addition, there are soft constraints based on process principles: according to the resin type and board thickness in the raw material specifications, the system calculates the minimum time required for the resin to fully cure, and the set target pressure curve for the hot pressing process (especially its corresponding high-pressure holding stage) must ensure that the duration meets this minimum curing time requirement to guarantee the most basic product quality. These constraints together constitute a multi-dimensional, feasible process parameter space, within which the optimization algorithm must find the optimal solution.

[0045] The solution engine employed is a non-dominated sorting genetic algorithm with an elitist strategy. Within this algorithm, a population consists of many individuals, each essentially representing a complete and specific set of process parameter values—a particular combination of optimization variable values. At the start of the algorithm's execution, a certain number of such schemes (population size, e.g., 100 or 200 individuals) are randomly generated, followed by an iterative evolutionary process. In each iteration, the algorithm evaluates the merits of each individual (i.e., each scheme) in the current population. The evaluation method involves providing the decoded process parameter values ​​as input to the hybrid prediction models for the drying and hot-pressing processes, respectively. These models quickly calculate the predicted total energy consumption (first objective) and the predicted internal bond strength of the sheet metal (second objective) for one control cycle under these parameters. Each individual thus obtains two objective function values. The algorithm classifies and ranks these individuals according to the non-dominated sorting principle (i.e., Pareto stratification), and within the same stratum, sorts individuals based on crowding to simultaneously promote convergence to the optimal solution set and maintain solution diversity. Through genetic operations such as selection, crossover, and mutation, the population continuously evolves, preserving superior elite individuals. This optimization process is executed on a rolling basis. At the beginning of each control cycle (whose length can be dynamically adjusted according to the process, for example, set to 5 minutes, 10 minutes, or 30 minutes), the system acquires the latest real-time production data as the current state input of the model, and then starts the genetic algorithm to perform a rapid optimization calculation. After the calculation is completed, a suitable solution is selected from the currently found non-dominated optimal solution set (for example, the solution with the highest crowding, representing that this solution has good distinguishability from other solutions in the target space). The target temperature, target wind speed, target temperature, and target pressure curves of the drying process, obtained after decoding, are output as instructions to the controller for execution within the current control cycle. At the end of the cycle, this process is repeated based on new real-time data, thereby realizing closed-loop optimization control that dynamically adjusts the optimization setpoint according to real-time operating conditions. Compared to traditional methods that rely on engineers' experience to manually set fixed parameters or perform single-objective optimization, this solution simultaneously considers two objectives: minimizing energy consumption and maximizing internal binding strength. Utilizing a hybrid predictive model for precise forward-looking evaluation, the system can automatically explore Pareto optimal solutions under complex constraints, identifying combinations of process parameters that significantly reduce energy consumption while ensuring product quality is not compromised or even improved. Alternatively, it can find the operating point with the lowest energy consumption under set quality targets. Specifying optimization variables to directly controllable temperature, wind speed, and pressure curves ensures the optimization results are highly executable. Furthermore, introducing multiple constraints, including equipment physical limits and process chemical requirements, ensures that the optimized solution remains within a safe and feasible range, avoiding theoretically energy-saving solutions that are practically unfeasible or lead to substandard quality.A non-dominated sorting genetic algorithm with an elitist strategy is employed to solve the problem, effectively handling complex optimization problems where the objective function may be nonlinear, non-convex, and coupled between variables, demonstrating powerful global search capabilities. The rolling optimization mechanism enables the system to respond to real-time disturbances such as changes in raw materials and equipment state drift, dynamically adjusting the optimal setpoints.

[0046] The specific process of solving the multi-objective optimization model using a non-dominated sorting genetic algorithm with an elitist strategy is as follows: Initialize algorithm parameters, including setting initial values ​​for population size, maximum number of iterations, crossover probability, and mutation probability; randomly generate an initial population, where each individual in the population represents a combination of values ​​for a set of optimization variables using real-number encoding; enter the iteration loop, and in each iteration, use the mixed prediction model for the drying process and the mixed prediction model for the hot pressing process to evaluate the predicted total energy consumption and predicted internal bonding strength of each individual in the current population. All individuals are stratified according to the non-dominated ranking, and then ranked according to the crowding degree within each non-dominated stratum. An elite retention strategy is implemented based on the ranking results, and individuals with higher rankings are directly retained to the next generation of the population. Simulated binary crossover and polynomial mutation operations are performed on individuals in the population to generate offspring individuals. The probabilities of crossover and mutation operations are adjusted using an adaptive mechanism, which is dynamically calculated based on the current iteration number and the dispersion of the fitness of individuals in the population. Determine if the algorithm's termination condition is met. If it is, select the individual with the highest crowding from the final generation's elite individual set, and use the optimized variable values ​​obtained from decoding it as the output of the target temperature, target wind speed, target temperature, and target pressure curves for the drying process, the hot pressing process, and the hot pressing process. If the condition is not met, proceed to the next iteration loop.

[0047] At the start of each rolling optimization, the algorithm's internal parameters need to be initialized. These parameters include population size (the number of individuals in each generation, such as 100, 150, or 200 individuals), maximum number of iterations (the upper limit of the number of generations the algorithm can evolve, such as 50, 100, or 200 generations to balance solution accuracy and computation time), initial values ​​for crossover probability and mutation probability (these initial probability values ​​are typically selected within the ranges of 0.6 to 0.9 and 0.01 to 0.1, respectively). These parameters collectively determine the algorithm's search capability and efficiency. After the parameters are set, an initial population is created through random generation. Each individual in the population represents a complete set of process parameter schemes. To facilitate genetic operations, continuous variables in these schemes (such as target temperature, target wind speed, and various pressure setpoints) are encoded using real numbers. The chromosome of an individual in the algorithm can be represented as a real number vector, where each element corresponds to a specific value of an optimization variable. For example, an individual can be encoded as [drying temperature = 165.3℃, drying wind speed = 18.2m / s, hot-pressing temperature = 185.7℃, pressure point 1 = 3.2MPa, pressure point 2 = 4.1MPa, ...]. By randomly assigning values ​​within the constraints of each variable, an initial population uniformly distributed within the feasible region can be quickly generated, providing a broad starting point for subsequent evolutionary exploration.

[0048] After the algorithm enters the iterative loop, the primary task in each iteration (i.e., each generation) is to evaluate the fitness of all individuals in the current population. In this multi-objective optimization context, fitness evaluation is achieved by calling a hybrid dry-and-hot-press prediction model to calculate the predicted total energy consumption and predicted intra-group binding strength for each individual's coding scheme. After obtaining the objective values ​​for all individuals, non-dominated sorting is performed: based on the dominance relationships between individuals (i.e., all objectives of one individual are not inferior to those of another, and at least one objective is strictly superior), the entire population is divided into multiple non-dominated levels, with the first level being the Pareto optimal frontier in the current population. Next, within the same non-dominated level, the crowding distance of each individual is calculated. This distance measures the sparsity of an individual in the objective space of its level; individuals with high crowding help maintain solution diversity. The population is sorted according to the non-dominated level (smaller levels are better) and the intra-level crowding (larger levels are better). Subsequently, an elite retention strategy is implemented, directly copying the top-ranked individuals (e.g., the top 20%) from the current population into the next generation to ensure that superior genes are not lost. To explore new possible solutions, the algorithm performs genetic operations on the population (usually including individuals other than elite individuals) to produce offspring. Common operations include simulated binary crossover and polynomial mutation, which respectively apply a certain probability to the mixed perturbation of the codes of two parent individuals and a small random perturbation to the code of a single individual, thus generating new individuals. The crossover and mutation probabilities can employ adaptive mechanisms, such as dynamically calculating and adjusting them based on the current iteration number (later iterations may tend to decrease the mutation probability to strengthen convergence) or the dispersion of the fitness of individuals in the population (increasing the mutation probability when diversity is low), thereby improving the performance of the algorithm at different optimization stages. After each iteration, the algorithm checks whether the preset termination conditions have been met. Common termination conditions include whether the number of iterations has reached the maximum iteration limit set at initialization. If the condition is met, the algorithm stops the evolutionary loop. At this point, the set of elite individuals in the final generation of the population represents the approximate Pareto optimal frontier found by the algorithm. The system needs to select a specific solution from this frontier as the output of this rolling optimization. The selection strategy can be varied. For example, the individual with the highest crowding density can be selected because it has a good distance from other solutions in the target space, usually indicating that its corresponding process parameter combination achieves a good balance between energy consumption and quality, and the parameter settings may be more robust. After selecting an individual, the system decodes its real-valued encoded vector to restore a set of specific optimization variable values: namely, the specific values ​​of the target temperature for the drying process, the target wind speed for the drying process, the target temperature for the hot pressing process, and the specific numerical sequence of each pressure setpoint that constitutes the target pressure curve for the hot pressing process. This complete and quantified set of process parameter settings is the optimal operating instruction planned for the next control cycle by this optimization calculation.If the termination condition is not met, the algorithm merges the newly generated offspring individuals with the retained elite individuals to form a new generation of population, and returns to the next round of iterative evaluation until the termination condition is met, thereby ensuring that each rolling optimization can output a high-quality process parameter scheme that has been fully searched and evaluated.

[0049] Compared to simple search methods or traditional optimization algorithms, this process, based on a non-dominated sorting genetic algorithm with an elitist strategy, uses real-number encoding and random initialization to start the search from a broad feasible region, avoiding initial biases that can lead to local optima. Its core iterative evolutionary mechanism, combining non-dominated sorting and crowding calculation, not only effectively drives the population towards lower energy consumption and higher quality evolution but also cleverly maintains the diversity of the final solution set, allowing the decision-maker (or automatic selection rule) to choose from multiple distinctive optimal solutions. The introduction of the elitist retention strategy ensures that the best solution discovered during evolution is not destroyed by random operations, accelerating convergence and improving the quality of results. The adaptive crossover and mutation probability mechanism gives the algorithm the ability to dynamically adjust exploration and development according to the search progress, encouraging extensive exploration in the early stages and refined development in the later stages, further improving optimization efficiency. The entire process has clear termination logic and output selection mechanism, ensuring that each rolling optimization delivers a clear and executable optimization control instruction within a limited computational time. This complete solution scheme transforms complex multi-objective optimization problems into an automated, streamlined computational task, boasting powerful search capabilities, good convergence, and practical results.

[0050] When establishing a multi-objective optimization model, a third optimization objective function is also included, which is the energy consumption volatility. The energy consumption volatility is obtained by calculating the sum of squares of the changes in the total predicted energy consumption between adjacent sampling times in a future control cycle. When using a non-dominated sorting genetic algorithm with an elitist strategy to solve the problem, the iterative evolutionary process also includes: In each generation of the population, an elite set of individuals is selected based on non-dominated ranking and crowding calculation. The optimized variable values ​​of all individuals in this elite set are decoded and compared with the historical optimal process parameter set stored in the historical database, which has the lowest actual energy consumption and qualified actual internal bond strength under the same wood species and target board density conditions. The similarity is calculated based on the Euclidean distance between the optimized variable values. If the calculated similarity is lower than a set threshold, it is determined that the current population may be trapped in a local optimum. At this time, the historical optimal process parameter set is injected as a new individual into the next generation of the population to guide the search direction. The termination condition of the iteration process is that the number of iterations reaches the maximum preset number of generations, or the update rate of the elite individual set for several consecutive generations is lower than a preset ratio.

[0051] Building upon the original objectives of minimizing total energy consumption and maximizing internal cohesion strength, this scheme further introduces a third optimization objective function: minimizing the volatility of predicted total energy consumption over a future control cycle. Energy consumption volatility is an indicator of the smoothness of energy consumption changes. The main consideration is that excessively drastic and frequent energy consumption fluctuations can not only impact the power supply network and increase the grid's regulatory burden, but may also reflect significant jumps in process parameters, potentially adversely affecting the stability of the production process, equipment lifespan, and the consistency of final product quality. The calculation method for this objective function is as follows: First, using a hybrid prediction model, the instantaneous predicted energy consumption values ​​are calculated at multiple consecutive sampling times (e.g., one point per second or every few seconds) under a specific process parameter scheme within a future control cycle. Then, the changes (differences) between these adjacent sampling times are calculated. Finally, these changes are squared (to highlight the impact of larger fluctuations) and summed; the resulting value is the energy consumption volatility corresponding to this scheme. Using this value as the third objective to minimize means that when searching for energy-saving and high-quality solutions, the optimization algorithm will also tend to select process parameter combinations that make energy consumption output smoother and more stable, thereby guiding the production process towards a more stable and flexible operating state. In the iterative evolution process of the non-dominated sorting genetic algorithm with an elite strategy, in addition to the conventional non-dominated sorting and crowding calculation, an additional historical experience-guided step is added. After each generation of the population selects an elite set of individuals, the system performs the following operations: First, it compares the optimized variable values ​​(i.e., a series of candidate process parameter combinations) decoded from all individuals in this elite set with a "gold standard" parameter set stored in a historical database. This "gold standard" is the historically optimal set of process parameters that, under the same wood species and target board density conditions, has been verified through actual production and has achieved both extremely low actual energy consumption and qualified internal bond strength. The comparison method is to calculate similarity, usually based on the Euclidean distance between the optimized variable values. The smaller the distance, the higher the similarity. The system sets a threshold for this similarity (for example, if the distance exceeds a certain value, such as 10% or 20% of the total distance after normalization of all variables, the similarity is considered low). If the calculation finds that the similarity between all individuals in the current elite set and the historical optimal set is lower than the set threshold (i.e., the distances are all relatively large), this may be a warning signal, indicating that the solution found by the current algorithm population differs greatly from historical success. Although it may perform reasonably well in terms of objective function values, it is highly likely that it is trapped in a local optimum region in the current search space, rather than a globally better region. In order to guide the population out of the possible local optima, the algorithm will inject this historical optimal set of process parameters as a completely new and special individual into the next generation of the population.The injected individual carries past successful experiences, providing the algorithm with a clear, high-quality search direction. This helps the population explore potentially overlooked regions of better solutions, enhancing the algorithm's global optimization capability and convergence reliability.

[0052] Traditional termination criteria often involve reaching a preset maximum number of iterations. This scheme adds a more intelligent termination criterion based on population evolutionary activity. This criterion monitors the update status of the elite individual set over several consecutive generations (e.g., 5, 10, or 15 generations). Here, the "update rate" can be defined as the proportion of newly appearing individuals in the new generation of the elite set compared to the previous generation. If the algorithm is very close to the true Pareto front, population evolution will slow down, the elite set will tend to stabilize, and the update rate will remain at a very low level. Therefore, the system presets a low update rate threshold (e.g., 1%, 2%, or 5%). If the update rate of the elite set is lower than this preset proportion for several consecutive generations, the system can reasonably determine that the algorithm has converged sufficiently, and further iterations are unlikely to produce significantly improved new solutions. In this case, even if the number of iterations has not reached the maximum limit, the algorithm can terminate early, thereby saving unnecessary computation time and improving the response speed of rolling optimization. This condition complements the maximum number of iterations condition, enabling the algorithm to perform a thorough search in complex situations as well as terminate early in simple or fast convergence cases, thus balancing solution accuracy and computational efficiency, making the optimization process more intelligent and efficient.

[0053] This scheme further enhances the comprehensiveness, intelligence, and robustness of the entire optimization control scheme by introducing an energy consumption volatility target and an enhanced algorithm mechanism. The newly added energy consumption volatility minimization target extends the optimization decision-making from focusing solely on results (total energy consumption, final quality) to simultaneously considering process stability. This guides the system to automatically seek process parameters that are not only energy-efficient and of high quality but also achieve stable energy output in production, helping to reduce the impact of production on the power grid, improve the smoothness of equipment operation, and indirectly promote the uniformity of product quality. This achieves optimization from a single economic objective to a multi-dimensional comprehensive performance objective that includes operational stability. The introduction of historical experience guidance and a mechanism to prevent local optima greatly enhances the intelligence level of the optimization algorithm. By introducing the historically optimal parameter set, which has been tested in practice, as heuristic information into the evolutionary process, the algorithm can learn from the successful experiences accumulated by humans over a long period of time while exploring autonomously. This effectively avoids the risk of getting trapped in local optima in complex high-dimensional spaces, increases the probability of finding a globally better or equally high-quality but more stable solution, and makes the optimization results more reliable and in line with engineering practice. The extended, convergence-based termination condition gives the algorithm the ability to determine its own convergence state, enabling adaptive allocation of computing resources. While ensuring optimization quality, it minimizes the computation time required for each rolling optimization and improves the real-time response capability of the entire control system.

[0054] In another technical solution, step S4 specifically includes: The target temperature and target air velocity of the drying process, which are output in step S3, are sent to the drying equipment controller; the target temperature and target pressure curves of the hot pressing process are sent to the hot press controller; the drying equipment controller and the hot press controller drive the actuators to adjust the air inlet valve of the drying cylinder, the frequency converter of the induced draft fan, the heating system of the hot press, and the hydraulic proportional valve according to the received target parameters; after the parameter control is executed, the actual moisture content of the dried outlet material is collected by a near-infrared moisture meter installed at the outlet of the drying cylinder, and the actual internal bond strength of the plate is collected by an online mechanical property testing device after the hot-pressed plate has been cooled and cured; During parameter issuance and execution, the actual adjustment values ​​fed back by the drying equipment controller and the hot press controller are monitored in real time, and the actual adjustment values ​​are compared with the issued target parameters to calculate the real-time execution deviation. If any of the real-time execution deviations exceeds the first set threshold for a predetermined duration, the system determines that there is an execution abnormality in the control loop and activates the preset backup control strategy. Based on the mixed prediction model of the drying process and the mixed prediction model of the hot press process, the system recalculates the target temperature, target wind speed, target temperature, and target pressure curves of the drying process, the hot press process, and the target pressure curve of the hot press process within the remaining control cycle, and immediately switches to the recalculated parameter sequence to continue execution.

[0055] After step S3 outputs the target temperature and target air velocity for the drying process, as well as the target temperature and target pressure curves for the hot pressing process for the next control cycle, these setpoints are precisely distributed through the industrial control system network. The target temperature and target air velocity for the drying process are encapsulated into data packets and sent to the programmable logic controller (PLC) or dedicated drying process controller of the drying equipment via standard industrial communication protocols (such as OPCUA or Modbus TCP). Similarly, the target temperature for the hot pressing process and the well-defined pressure-time curve (consisting of a series of time-pressure coordinate points) are sent to the core controller of the hot press. Upon receiving these instructions, the controller immediately drives its subordinate actuators. For the drying process, the controller adjusts the opening of the steam valve or gas valve on the air inlet pipe of the drying cylinder through analog output modules or bus control, thereby precisely controlling the temperature of the hot air entering the drying cylinder; simultaneously, it adjusts the output frequency of the inverter driving the induced draft fan to change the fan speed, thereby stabilizing the air velocity inside the drying cylinder at the target value. For the hot pressing process, the controller adjusts the power output of the heating elements (such as electric heating tubes or a heat transfer oil circulation system) in the heating plate to quickly reach and maintain the target temperature for each layer. Simultaneously, it controls the pressure output of the hydraulic cylinders through high-precision hydraulic proportional valves or servo valves, ensuring that the cylinders apply corresponding pressure to the slab over time, strictly following the received target pressure curve. Simply executing the instructions does not guarantee the goal has been achieved; the system must verify the actual effect after execution. Therefore, online monitoring equipment is deployed at key nodes in the production line. A near-infrared moisture meter is installed at the material outlet of the drying drum. This instrument uses near-infrared light to irradiate the material passing through the outlet, and based on the different absorption characteristics of materials with different moisture contents for specific wavelengths of light, it measures the actual moisture content of the dried material in real time and non-contactly. This value directly reflects the control quality of the drying process. After the hot pressing process, the pressed slabs need to undergo a cooling and stress balancing curing process. Afterwards, the system uses online mechanical property testing equipment to collect the actual internal bond strength of the slabs. This equipment may employ non-destructive online detection technologies, such as stress wave or ultrasonic technology, to indirectly assess strength. It can also integrate an automated sampling and testing unit, where a robotic arm periodically grasps samples for standard short-cycle laboratory tests and automatically feeds the results back to the system. These actual measurements of moisture content and internal bonding strength are collected in real time and uploaded to a central system for comparison with previous predictions, thereby evaluating the actual effectiveness of the current control round.

[0056] After issuing the command, the system continuously and proactively monitors the execution process, reading in real time the actual adjustment values ​​of each loop from the drying equipment controller and the hot press controller, such as the actual valve opening, the actual output frequency of the frequency converter, the actual temperature of the heating plate, and the actual pressure of the hydraulic system. These actual adjustment values ​​are compared in real time with the previously issued target parameters to calculate the real-time execution deviation for each item (e.g., the difference between the target temperature and the actual temperature, the difference between the target pressure and the actual pressure). The system presets a first set threshold for different types of deviations (e.g., the temperature deviation threshold may be between 2 and 5 degrees Celsius, and the pressure deviation threshold may be between 0.2 and 0.5 MPa for reference) and a predetermined time judgment duration (e.g., 10 seconds, 30 seconds, or 60 seconds). If the real-time execution deviation of any key parameter (such as temperature or pressure) exceeds its first set threshold for a predetermined duration, the system will determine that there is an execution anomaly in that control loop. This may be due to actuator failure, valve jamming, sensor feedback malfunction, or strong external interference. Once an anomaly is detected, a preset intelligent backup control strategy is immediately activated. Using the currently achieved process parameters (i.e., actual temperature, actual pressure, etc.) as new, unchangeable initial conditions, a new round of rolling optimization calculations is performed based on a mixed drying and hot-pressing prediction model. However, this optimization covers the remaining time of the current control cycle. The calculation output is the optimal process parameter adjustment sequence for the remaining time from the current anomaly moment to the end of the current cycle. The system then switches to this newly calculated parameter sequence to continue control.

[0057] This solution ensures that intelligent decisions are accurately and reliably implemented into production actions by clearly defining the complete chain of optimized instruction issuance, actuator driving, and online quality feedback, forming a performance closed loop based on real results. Real-time monitoring and autonomous fault-tolerant control mechanisms during execution endow the system with robust resilience to cope with common equipment anomalies and interferences on-site. This changes the traditional automation system model that either relies entirely on preset program operation (shutting down upon anomalies) or relies entirely on operator intervention. Instead, it enables the system to quickly perform online re-optimization and decision switching based on the current actual state when local execution deviations are detected.

[0058] In another technical solution, step S5 specifically includes: For the drying process, the prediction error is set as the absolute value of the difference between the predicted moisture content of the dried material at the outlet output by the hybrid prediction model for the drying process and the actual moisture content collected in step S4. For the hot pressing process, the prediction error is set as the absolute value of the difference between the predicted internal bond strength of the sheet material output by the hybrid prediction model for the hot pressing process and the actual internal bond strength collected in step S4. The threshold for setting the prediction error is dynamically determined based on the statistical distribution of the corresponding prediction errors calculated in multiple consecutive production batches in the past, specifically taken as the preset quantile of the statistical distribution. When the prediction error of the drying process or the hot pressing process exceeds its corresponding dynamic threshold, the incremental learning process of the corresponding model is triggered. After triggering incremental learning, the real-time temperature, real-time pressure, real-time energy consumption data of the current production batch collected in step S1, as well as the raw material specification information, are used as new training samples. The actual moisture content or actual internal bonding strength collected in step S4 is used as the corresponding new sample label. These samples are then merged with some historical samples in the historical training sample library that have the same wood species and target board density to form an incremental training dataset. Using the incremental training dataset, the corresponding long short-term memory network in the drying process data-driven model or the hot-pressing process data-driven model is incrementally trained. During training, the structure above the hidden layer of the network is fixed, and only the weight parameters of the fully connected output layer of the network are updated through backpropagation. After training, the updated data-driven model part replaces the data-driven model part in the original hybrid prediction model, thus completing the model update.

[0059] After each production batch is completed, the system initiates a formal evaluation of the model's prediction accuracy. For the drying process, the prediction error is defined as the absolute value of the difference between the predicted moisture content of the dried material output by the hybrid prediction model during the production of this batch and the actual moisture content of the batch collected by the near-infrared moisture meter in step S4. For the hot pressing process, the prediction error is the absolute value of the difference between the predicted internal bond strength of the sheet output by the hybrid prediction model and the actual internal bond strength obtained by the online detection equipment. These two absolute errors intuitively reflect the degree of deviation between the model prediction and the actual result. A dynamic threshold method is used to determine whether the error is too large and thus requires triggering model learning. The system continuously tracks the drying or hot pressing prediction errors calculated in multiple consecutive production batches (e.g., the most recent 20, 30, or 50 batches) to form a historical sequence of error values. Statistical analysis is performed on this historical error sequence to understand its distribution pattern (e.g., it may be a normal or skewed distribution). The dynamic threshold is set based on this statistical distribution. Specifically, it can be a high quantile of the distribution, such as the 90th, 95th, or similar high percentile. This means that the threshold will automatically adjust based on recent production performance: if recent model predictions are generally accurate with small errors, the dynamic threshold will be set lower, and the system will have a stricter tolerance for errors; conversely, if recent operating conditions fluctuate greatly and the error distribution is wide, the dynamic threshold will be increased to avoid frequently triggering unnecessary learning within the normal fluctuation range.

[0060] Once the prediction error of the drying or hot-pressing process exceeds its corresponding dynamically set threshold, the system will automatically trigger an incremental learning process for the data-driven model (i.e., LSTM network) for that process. After triggering, the first step is to construct an incremental training dataset. The new training samples naturally come from the recently completed production batch: the complete time-series data of the batch collected in step S1 (real-time temperature, pressure, wind speed, energy consumption, etc.) and raw material specifications are used as input features for a new sample; the actual moisture content (for the drying model) or actual internal bond strength (for the hot-pressing model) of the batch collected in step S4 is used as the label for this new sample. Having only the latest sample is insufficient, as it can easily lead to the model forgetting historical patterns or overfitting to the latest operating conditions. Therefore, the system will select a portion of historical samples from the vast historical training sample library and merge them with the new samples. The selection principle is to choose historical samples with the same wood species and target board density combination as the current batch, because they have the highest process relevance. During the selection process, cosine similarity calculation based on feature vectors (such as the statistical characteristics of operating condition data sequences) can be used to select the top K most similar historical samples (K can be a preset number such as 50, 100, or 200). In addition, the system maintains a fixed-size historical representative sample pool, which stores typical samples that have shown particularly excellent prediction results for various raw material combinations in the past. This incremental learning will also add samples from this pool that match the specifications of the current batch of raw materials. The incremental training dataset constructed in this way includes both the latest feedback information and relevant historical experience and representative patterns.

[0061] After obtaining the incremental training dataset, the system begins incremental learning training on the corresponding Long Short-Term Memory (LSTM) network. To improve learning efficiency, avoid destroying the general features already learned by the network, and reduce the risk of overfitting, a partial fine-tuning strategy is employed. Specifically, the network structure (especially the hidden layers used for feature extraction) is fixed, and its weight parameters remain unchanged during this training. Training (i.e., backpropagation updates) is performed only on the weight parameters of the network's output layer (usually the last fully connected layer). This means that the model primarily learns how to fine-tune the mapping relationship between existing feature extraction capabilities and the latest observed output labels. After training, a data-driven model with updated output layer weights is obtained. To ensure the safety and effectiveness of the update, a validation step is required. The system prepares an independent validation dataset, which consists of samples from the historical sample library that did not participate in this incremental training (e.g., earlier batches). The updated data-driven model is used to make predictions on this validation set, and evaluation metrics such as the mean of the prediction error are calculated. Then, this metric is compared with the performance of the unupdated model on the same validation set (or historical data from the same period). The system only confirms the effectiveness of incremental learning when the updated model's evaluation metrics are better than the original model, or even if there is a slight degradation but it is within a preset, acceptable range. Only after confirmation will the updated data-driven model partially replace the corresponding parts of the original mixed prediction model, completing a model upgrade. If the verification results indicate that the update is ineffective or even results in performance degradation, the system will discard the learning results, retain the original model for continued operation, and record an invalid update event for subsequent analysis.

[0062] Compared to traditional models that require periodic manual retraining or cannot be updated online at all, this solution dynamically sets update thresholds based on recent error statistical distribution. It can autonomously perceive the degradation trend of model performance and trigger learning at appropriate times. This avoids unnecessary redundant updates when the model is still accurate and prevents situations where the model is unaware of significant deviations, achieving precise judgment of update needs. When constructing the incremental training dataset, it cleverly combines the latest batch data, similar historical data, and representative historical samples, ensuring that the learning samples are both timely and retain important historical knowledge, effectively mitigating the "catastrophic forgetting" problem in incremental learning. The strategy of using a fixed feature extraction layer and only fine-tuning the output layer for partial network updates significantly reduces learning complexity and the risk of overfitting. The introduction of a strict independent validation redeployment mechanism sets a safety valve for each model update, ensuring that only proven effective updates are applied, fundamentally eliminating the risk of the entire online model performance collapsing due to a single batch of abnormal data or accidental learning failures.

[0063] When constructing the incremental training dataset, the method for selecting historical samples from the historical training sample library is as follows: Calculate the cosine similarity between the feature vector of the newly added training sample in the current production batch and the feature vector of the historical sample, and select the top K historical samples with the highest similarity, where K is a preset number; at the same time, maintain a historical representative sample pool with a fixed capacity, which stores typical historical samples with low prediction errors for different wood species and target board density combinations; during each incremental learning, samples in the historical representative sample pool that have the same raw material specification information as the current production batch are also added to the incremental training dataset. After updating the weight parameters of the output layer of the Long Short-Term Memory network using the incremental training dataset, its performance is first evaluated on an independent validation dataset. The validation dataset consists of historical samples that did not participate in this incremental training. The evaluation metric is the mean of the prediction error. If the evaluation metric of the updated model is better than that of the model before the update on the original validation dataset, or if the degree of degradation is within an acceptable range, the system confirms that this incremental learning is effective and officially updates the updated model part into the hybrid prediction model for the drying process or the hybrid prediction model for the hot pressing process. Otherwise, the results of this incremental learning are discarded, the original model is retained and continues to run, and this invalid update event is recorded.

[0064] When the system determines that the current production batch needs to be supplied (e.g., the wood species is pine, and the target density is 700 kg / m³), 3 When incremental learning is initiated, the system first calculates the feature vector formed by the newly added batch of data. This feature vector typically contains the statistical characteristics of the batch's time series data (such as mean, variance, and trend), key process parameters, and raw material specification coding information. Subsequently, the system compares this feature vector with all samples in the historical sample library belonging to the same raw material specification category (i.e., all pine wood with a target density of 700 kg / m³). 3 The system compares the feature vectors of the new sample with those of the historical batches. The cosine similarity metric measures the directional closeness between two vectors, with a value between -1 and 1; a value closer to 1 indicates a more similar feature pattern. The system calculates the cosine similarity score between the new sample and each relevant historical sample, and sorts them from highest to lowest score. Then, the system pre-determines a number K (e.g., K can be set to 50, 100, or 200, depending on the size of the historical database and the amount of historical information to be retained) and selects the top K historical samples with the highest similarity scores. These selected samples, because their historical operating characteristics are most similar to the current batch, contain input-output mapping relationships that are most valuable and portable for updating the current model.

[0065] In addition to dynamically selecting samples based on real-time similarity, the system also maintains a fixed-size historical representative sample pool over the long term. This pool is not a complete backup library, but rather a refined collection of the best samples. Its construction logic is as follows: Over long-term operation, the system continuously monitors the prediction errors of all production batches. When certain batches have specific combinations of raw material specifications (such as cedar wood - 650kg / m³), the system will identify the most suitable batches. 3 Poplar - 750kg / m 3 When the model's prediction error is significantly below average (i.e., at a low error level), the system stores this batch of data (including its features and actual results) as a typical success story or high-quality experience template in a representative sample pool. The pool's capacity is fixed (e.g., 5 to 10 of the most representative samples can be retained for each common raw material specification combination, with the total pool capacity potentially on the order of hundreds of samples). When the pool is full, new representative samples may replace relatively older or less representative samples in the pool based on their low prediction error or the uniqueness of the pattern. During each incremental learning iteration, regardless of which historical samples the current batch is highly similar to, the system automatically and unconditionally adds all samples from the representative sample pool that have exactly the same raw material specification information as the current batch to the incremental training dataset. These representative samples represent the "golden moments" or "ideal conditions" when the model achieved the best predictive results for this type of raw material. They help strengthen the most core and essential successful mapping relationship for this type of raw material when the model is fine-tuned, thus playing a key role in stabilizing the core knowledge of the model and preventing "catastrophic forgetting" (i.e. learning new things but forgetting old and important ones) during incremental learning.

[0066] After updating the output layer weights of the Long Short-Term Memory (LSTM) network using the constructed incremental training dataset, the system initiates a separate validation phase. A validation dataset is prepared in advance or temporarily partitioned. This dataset consists of historical samples that did not participate in the training process of this incremental learning in any way (i.e., they are neither newly added batches, nor selected top K similar samples, nor samples from the representative sample pool). They are typically earlier in time or randomly retained portions of historical data, used to simulate unknown tests of the model's generalization ability. During evaluation, the candidate new model is forward-propagated on this validation dataset, and the mean of its prediction error (or combined with other metrics such as root mean square error) is calculated. Subsequently, this evaluation metric is rigorously compared with the performance metrics of the previous model on the same validation dataset (or a comparable, historical benchmark dataset). The update is unconditionally accepted only when the evaluation metric of the candidate new model is clearly superior to that of the previous model. Alternatively, under certain permissible fault-tolerance strategies, even if the new model's metrics deteriorate slightly, as long as the degree of deterioration is within a pre-defined, very limited acceptable range (e.g., the mean error increase does not exceed a relative range of one to three percent), the system may still determine the update is acceptable. This balances the need to adapt to new changes with the need to maintain stability. If the evaluation results indicate that the updated model's performance exceeds the acceptable range, the system will maintain the original model and continue running, recording this attempt as an invalid update event.

[0067] Compared to simply adding the latest data to a historical database for retraining or randomly selecting some historical data, this approach uses cosine similarity filtering to ensure that incremental learning efficiently utilizes the most relevant and valuable historical experience related to the current working conditions. This makes model updates more targeted and efficient, improving the speed and accuracy of learning new knowledge. By introducing and utilizing a pool of representative historical samples, the system effectively resists the catastrophic forgetting problem common in incremental learning, ensuring the continuity and stability of the model's long-term knowledge accumulation. Finally, the rigorous verification and safe deployment process, through independent verification and performance comparison, ensures that every change to the online model is empirically tested, eliminating the risk of performance degradation or even loss of control of the production-guided model due to erroneous learning.

[0068] It should be noted that although the steps are described in a specific order above, this does not mean that they must be performed in that order. In fact, some of these steps can be executed concurrently, or even in a different order, as long as the required functionality is achieved. The number of devices and processing scale described herein are for simplification of the invention; applications, modifications, and variations of this invention will be readily apparent to those skilled in the art.

[0069] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A method for intelligent control of energy consumption in wood-based panel production, characterized in that, Includes the following steps: S1: Collect real-time temperature, real-time pressure and real-time energy consumption data of the hot pressing process through a sensor network deployed on the production line, collect real-time temperature, real-time wind speed and real-time energy consumption data of the drying process, and obtain the raw material specification information of the current production batch from the production management system. S2: Based on the heat and mass transfer mechanism, construct the mechanism model of the drying process and the hot pressing process. Based on historical production data, use a long short-term memory network to train the data-driven model of the drying process and the hot pressing process respectively. Then, weightedly fuse the mechanism model and the data-driven model to form a hybrid prediction model of the drying process and a hybrid prediction model of the hot pressing process. S3: Based on the mixed prediction models of the drying process and the hot pressing process, a multi-objective optimization model is established with the optimization objectives of minimizing total energy consumption and maximizing the bonding strength within the product, and with the operating parameter range of each process equipment as constraints. The multi-objective optimization model is solved by a non-dominated sorting genetic algorithm with an elite strategy, and the target control parameters for the next control cycle are output in a rolling manner, including the target temperature and target wind speed of the drying process, and the target temperature and target pressure curves of the hot pressing process. S4: Send the target control parameters to the drying equipment controller and hot press controller of the production line for control execution, and after execution, collect the actual moisture content of the dried outlet material and the actual internal bond strength of the hot-pressed plate. S5: Using the real-time data collected in step S1 as input, and the actual moisture content and actual internal bonding strength collected in step S4 as verification output, calculate the prediction error of the mixing prediction model for the drying process and the mixing prediction model for the hot pressing process respectively. When the prediction error exceeds the set threshold, use the real-time data, actual moisture content and actual internal bonding strength to perform incremental learning on the long short-term memory network, and update the data-driven model part in the mixing prediction model for the drying process and the mixing prediction model for the hot pressing process.

2. The intelligent energy consumption control method for wood-based panel production according to claim 1, characterized in that, In step S1, the sensor network includes temperature sensors and hot-wire wind speed sensors respectively installed at the inlet, outlet, and middle of the drying cylinder to collect real-time temperature and real-time wind speed during the drying process; the sensor network also includes thermocouple temperature sensors and pressure sensors installed on each heating plate layer of the hot press to collect real-time temperature and real-time pressure during the hot pressing process; real-time energy consumption data is collected through the power metering module installed on the drive motor of the drying process and the hydraulic system of the hot press, as well as the flow meter installed on the steam heating pipeline; the raw material specification information includes wood type, initial moisture content, and target board density; All collected real-time data are assigned a unified timestamp, and are aligned and integrated with batch switching signals from the production management system based on the timestamp to form a time-series data sequence based on production batches. When the real-time data of any sensor continuously exceeds its corresponding normal range within a preset time period, or when timestamp breaks occur in the time-series data sequence, the data for that period is marked as invalid data, and an interpolation algorithm based on historical data from the same period is used to fill in the invalid data to generate a continuous and complete time-series data sequence.

3. The intelligent energy consumption control method for wood-based panel production according to claim 1, characterized in that, In step S2, the method for constructing the mechanism model is as follows: A mechanism model for the drying process is constructed based on the principle of convection drying. Its inputs are the real-time temperature, real-time wind speed, and initial moisture content in the raw material specifications. The output is the predicted moisture content of the material at the drying outlet. A mechanism model for the hot pressing process is constructed based on the principles of heat conduction and resin curing reaction kinetics. Its inputs are the real-time temperature, real-time pressure curve, and target board density in the raw material specifications. The output is the predicted internal bond strength of the board. When constructing the mechanism model, an empirical correction coefficient related to the wood pore structure and thermal conductivity is introduced for the wood type in the raw material specification information to calibrate the mass transfer and heat transfer parameters in the mechanism model. The specific method for training a data-driven model is as follows: Historical time-series data sequences with the same wood species and target board density as the current production batch are extracted from the historical database as training samples. The input features of the drying process data-driven model are the real-time temperature, real-time wind speed sequence, and initial moisture content of the historical drying process, and the output label is the actual moisture content of the material at the corresponding historical drying outlet. The input features of the hot pressing process data-driven model are the real-time temperature, real-time pressure curve sequence, and target board density of the historical hot pressing process, and the output label is the actual internal bond strength of the corresponding historical board. When constructing input features, a sliding time window method is used to combine working condition data from multiple consecutive sampling times into a feature vector to capture the dynamic time-varying characteristics of the production process; the number of hidden layer nodes of the Long Short-Term Memory network is determined by cross-validation based on the number of training samples.

4. The intelligent energy consumption control method for wood-based panel production according to claim 3, characterized in that, The specific method for weighted fusion to form a mixed prediction model for the drying process and a mixed prediction model for the hot pressing process is as follows: For the mixed prediction model of the drying process, the final predicted moisture content of the dried outlet material is obtained by weighted summation of the predicted output of the drying process mechanism model and the predicted output of the drying process data-driven model; for the mixed prediction model of the hot pressing process, the final predicted internal bond strength of the plate is obtained by weighted summation of the predicted output of the hot pressing process mechanism model and the predicted output of the hot pressing process data-driven model. The weighting coefficients used in the weighted fusion are dynamically calculated based on the predictive performance of the mechanistic model and the data-driven model on recent historical data. The system continuously records the predicted values ​​of the moisture content of the dried outlet material and the internal bonding strength of the board by the two models within the predetermined production cycle in the past, and compares them with the actual moisture content and actual internal bonding strength collected in step S4. The root mean square inverse of the prediction error of each model is calculated and normalized to serve as the dynamic weighting coefficient of the corresponding model under the current working condition. The dynamic weighting coefficients are updated once after each production batch is completed. During the update, the exponential smoothing method is used to merge the new weighting coefficients calculated this time with the old weighting coefficients used in the previous batch.

5. The intelligent energy consumption control method for wood-based panel production according to claim 1, characterized in that, Step S3 specifically includes: A multi-objective optimization model is established, where the first objective function is the predicted total energy consumption within a future control cycle, calculated based on the mixed prediction models for the drying and hot-pressing processes. The predicted total energy consumption is the sum of the predicted energy consumption for the drying and hot-pressing processes. The second objective function is the predicted internal bond strength of the board material, calculated based on the mixed prediction model for the hot-pressing process. Optimization variables include the target temperature of the drying process, the target wind speed of the drying process, the target temperature of the hot-pressing process, and the sequence of pressure setpoints constituting the target pressure curve of the hot-pressing process. Constraints include the operating range of temperature and wind speed of the drying equipment, the operating range of temperature of each heating plate layer of the hot press, the upper and lower limits of the pressure setpoints, and the minimum resin curing time requirement that the pressure curve of the hot-pressing process must meet, determined based on the raw material specifications. A non-dominated sorting genetic algorithm with an elitist strategy is used to solve the multi-objective optimization model. Each individual in the algorithm population represents a combination of values ​​for a set of optimization variables. The predicted total energy consumption and predicted bonding strength of the board corresponding to each individual are evaluated through a mixed prediction model of the drying process and a mixed prediction model of the hot pressing process. Iterative evolution is performed based on non-dominated sorting and crowding calculation. At the beginning of each control cycle, the current real-time temperature, real-time pressure and real-time energy consumption data are obtained as model inputs. The algorithm is executed and the target temperature, target wind speed of the drying process, target temperature of the hot pressing process and target pressure curve of the hot pressing process to be executed in that control cycle are output in a rolling manner.

6. The intelligent energy consumption control method for wood-based panel production according to claim 5, characterized in that, The specific process of solving the multi-objective optimization model using a non-dominated sorting genetic algorithm with an elitist strategy is as follows: Initialize algorithm parameters, including setting initial values ​​for population size, maximum number of iterations, crossover probability, and mutation probability; randomly generate an initial population, where each individual in the population represents a combination of values ​​for a set of optimization variables using real-number encoding. Enter the iterative loop. In each iteration, the mixed prediction model of the drying process and the mixed prediction model of the hot pressing process are used to evaluate the predicted total energy consumption and the predicted bonding strength of the plate for each individual in the current population. All individuals are stratified according to the non-dominated ranking, and then ranked according to the crowding degree within each non-dominated stratum. An elite retention strategy is implemented based on the ranking results, and individuals with higher rankings are directly retained to the next generation of the population. Simulated binary crossover and polynomial mutation operations are performed on individuals in the population to generate offspring individuals. The probabilities of crossover and mutation operations are adjusted using an adaptive mechanism, which is dynamically calculated based on the current iteration number and the dispersion of the fitness of individuals in the population. Determine if the algorithm's termination condition is met. If it is, select the individual with the highest crowding from the final generation's elite individual set, and use the optimized variable values ​​obtained from decoding it as the output of the target temperature, target wind speed, target temperature, and target pressure curves for the drying process, the hot pressing process, and the hot pressing process. If the condition is not met, proceed to the next iteration loop.

7. The intelligent energy consumption control method for wood-based panel production according to claim 5, characterized in that, When establishing a multi-objective optimization model, a third optimization objective function is also included, which is the energy consumption volatility rate. The energy consumption volatility rate is obtained by calculating the sum of squares of the changes in the total predicted energy consumption between adjacent sampling times in a future control cycle. When using a non-dominated sorting genetic algorithm with an elitist strategy to solve the problem, the iterative evolutionary process also includes: In each generation of the population, an elite set of individuals is selected based on non-dominated ranking and crowding calculation. The optimized variable values ​​of all individuals in this elite set are decoded and compared with the historical optimal process parameter set stored in the historical database, which has the lowest actual energy consumption and qualified actual internal bond strength under the same wood species and target board density conditions. The similarity is calculated based on the Euclidean distance between the optimized variable values. If the calculated similarity is lower than a set threshold, it is determined that the current population may be trapped in a local optimum. At this time, the historical optimal process parameter set is injected as a new individual into the next generation of the population to guide the search direction. The termination condition of the iteration process is that the number of iterations reaches the maximum preset number of generations, or the update rate of the elite individual set for several consecutive generations is lower than a preset ratio.

8. The intelligent energy consumption control method for wood-based panel production according to claim 1, characterized in that, Step S4 specifically includes: The target temperature and target air velocity of the drying process, which are output in step S3, are sent to the drying equipment controller; the target temperature and target pressure curves of the hot pressing process are sent to the hot press controller; the drying equipment controller and the hot press controller drive the actuators to adjust the air inlet valve of the drying cylinder, the frequency converter of the induced draft fan, the heating system of the hot press, and the hydraulic proportional valve according to the received target parameters; after the parameter control is executed, the actual moisture content of the dried outlet material is collected by a near-infrared moisture meter installed at the outlet of the drying cylinder, and the actual internal bond strength of the plate is collected by an online mechanical property testing device after the hot-pressed plate has been cooled and cured; During parameter issuance and execution, the actual adjustment values ​​fed back by the drying equipment controller and the hot press controller are monitored in real time, and the actual adjustment values ​​are compared with the issued target parameters to calculate the real-time execution deviation. If any of the real-time execution deviations exceeds the first set threshold for a predetermined duration, the system determines that there is an execution abnormality in the control loop and activates the preset backup control strategy. Based on the mixed prediction model of the drying process and the mixed prediction model of the hot press process, the system recalculates the target temperature, target wind speed, target temperature, and target pressure curves of the drying process, the hot press process, and the target pressure curve of the hot press process within the remaining control cycle, and immediately switches to the recalculated parameter sequence to continue execution.

9. The intelligent energy consumption control method for wood-based panel production according to claim 1, characterized in that, Step S5 specifically includes: For the drying process, the prediction error is set as the absolute value of the difference between the predicted moisture content of the dried material at the outlet output by the hybrid prediction model for the drying process and the actual moisture content collected in step S4. For the hot pressing process, the prediction error is set as the absolute value of the difference between the predicted internal bond strength of the sheet material output by the hybrid prediction model for the hot pressing process and the actual internal bond strength collected in step S4. The threshold for setting the prediction error is dynamically determined based on the statistical distribution of the corresponding prediction errors calculated in multiple consecutive production batches in the past, specifically taken as the preset quantile of the statistical distribution. When the prediction error of the drying process or the hot pressing process exceeds its corresponding dynamic threshold, the incremental learning process of the corresponding model is triggered. After triggering incremental learning, the real-time temperature, real-time pressure, real-time energy consumption data of the current production batch collected in step S1, as well as the raw material specification information, are used as new training samples. The actual moisture content or actual internal bonding strength collected in step S4 is used as the corresponding new sample label. These samples are then merged with some historical samples in the historical training sample library that have the same wood species and target board density to form an incremental training dataset. Using the incremental training dataset, the corresponding long short-term memory network in the drying process data-driven model or the hot-pressing process data-driven model is incrementally trained. During training, the structure above the hidden layer of the network is fixed, and only the weight parameters of the fully connected output layer of the network are updated through backpropagation. After training, the updated data-driven model part replaces the data-driven model part in the original hybrid prediction model, thus completing the model update.

10. The intelligent energy consumption control method for wood-based panel production according to claim 9, characterized in that, When constructing the incremental training dataset, the method for selecting historical samples from the historical training sample library is as follows: Calculate the cosine similarity between the feature vector of the newly added training sample in the current production batch and the feature vector of the historical sample, and select the top K historical samples with the highest similarity, where K is a preset number; at the same time, maintain a historical representative sample pool with a fixed capacity, which stores typical historical samples with low prediction errors for different wood species and target board density combinations; during each incremental learning, samples in the historical representative sample pool that have the same raw material specification information as the current production batch are also added to the incremental training dataset. After updating the weight parameters of the output layer of the Long Short-Term Memory network using the incremental training dataset, its performance is first evaluated on an independent validation dataset. The validation dataset consists of historical samples that did not participate in this incremental training; the evaluation metric is the mean of the prediction error; if the evaluation metric of the updated model is better than that of the model before the update on the original validation dataset, or the degree of degradation is within an acceptable range, the system confirms that this incremental learning is effective and officially updates the updated model part into the hybrid prediction model of the drying process or the hybrid prediction model of the hot pressing process; otherwise, the results of this incremental learning are discarded, the original model is retained to continue running, and this invalid update event is recorded.

Images