High-temperature furnace intelligent control system and method based on machine learning

Abstract

The invention relates to the technical field, in particular to a high-temperature furnace intelligent control system and method based on machine learning, and the method comprises the steps: a data collection module obtains and processes sensor data and outputs synchronous data streams; the twinborn modeling module is used for driving digital twinborn simulation by utilizing a synchronous data stream to obtain physical field data in the furnace; the causal analysis module identifies a key causal chain from the physical field data to generate a causal graph; the long-range prediction module adopts a neural network architecture integrated with a gating memory unit, and a causal diagram is combined to guide an attention mechanism to output a long-period prediction value; the strategy optimization module brings the predicted value into a reward function, and generates a control instruction through reinforcement learning and advanced simulation verification; the robust control module executes the instruction and monitors and predicts uncertainty to realize safe switching; the meta-learning module dynamically optimizes system parameters based on cross-batch performance data. According to the method, the gradient disappearance problem of the recurrent neural network in long-term dependence modeling is solved, and the control precision of the slow dynamic process of the high-temperature furnace is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent adaptive control in industrial process control, and in particular to a high-temperature furnace intelligent control system and method based on machine learning. BACKGROUND

[0002] The high-temperature furnace intelligent control technology based on machine learning trains a model such as a neural network or reinforcement learning through the collection of a large amount of historical operation data, so that the model learns the thermal dynamics law in the furnace. The model analyzes the temperature sensor feedback in real-time operation, predicts the thermal field evolution trend, and automatically optimizes variables such as heating element power or feed rate, so as to achieve temperature stable control and energy efficiency improvement. The core advantage of the model is that it can be iteratively updated from new data, gradually adapting to nonlinear process changes.

[0003] The existing high-temperature furnace intelligent control technology based on machine learning has the following technical pain points. Specifically, in the annealing or sintering process of a high-temperature furnace, the physical and chemical changes of the material are often a slow dynamic process with a delay of several hours or even several tens of hours. When a recurrent neural network is used to establish a controller prediction model, the network needs to learn and remember the complex correlation between process parameters and final product quality across a very long time step. However, the gradient signal is exponentially attenuated or exploded when backtracking along the time sequence in the training process of the recurrent neural network. Gradient vanishing makes it difficult for the network to effectively learn the causal relationship in the long period, so that the model cannot accurately capture the decisive influence of slight power fluctuations in the early heating stage on key indicators such as uniformity of the final product and crystal phase structure. For example, a slight temperature overshoot in the initial stage of furnace heating may have a negative impact on the uniformity of the doping concentration at the end of the process several hours later, but gradient vanishing makes the recurrent neural network controller unable to establish and correct this long-range causal relationship, ultimately leading to a decrease in the yield of the batch product. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application provides a high-temperature furnace intelligent control system and method based on machine learning, which solves the technical problem of inaccurate control of the slow dynamic process of the high-temperature furnace caused by gradient vanishing in long-term dependence modeling of the recurrent neural network.

[0005] To solve the above technical problems, the specific content of the present application is as follows: In a first aspect, the present application provides a high-temperature furnace intelligent control system based on machine learning, which comprises a data acquisition module, a twin modeling module, a causal analysis module, a long-range prediction module, a strategy optimization module, a robust control module, and a meta-learning module. The data acquisition module is configured to obtain sensor readings from a distributed control system of a high-temperature furnace, perform noise reduction and time synchronization processing on the sensor readings, and output a synchronized data stream to the twin modeling module and the long-range prediction module. a twin modeling module configured to receive the synchronized data stream transmitted by the data acquisition module, drive a digital twin of the high-temperature furnace to perform dynamic simulation by taking the synchronized data stream as a boundary condition, obtain physical field data in the furnace, the physical field data in the furnace including virtual temperature measurement point data, and transmit the physical field data in the furnace to the causal analysis module; the causal analysis module configured to receive the physical field data in the furnace transmitted by the twin modeling module, perform causal structure learning on the physical field data in the furnace, identify key early operation variables and time lag effects, generate a causal diagram, and transmit the causal diagram to the long-range prediction module; the long-range prediction module configured to receive the synchronized data stream transmitted by the data acquisition module and the causal diagram transmitted by the causal analysis module, adopt a neural network architecture integrating a gated memory unit, use the causal diagram to guide an attention mechanism of the neural network to process the synchronized data stream, output a long-period prediction value of a product quality index, and transmit the long-period prediction value to the strategy optimization module; the strategy optimization module configured to receive the long-period prediction value transmitted by the long-range prediction module, incorporate the long-period prediction value into a reward function, search for a control strategy through reinforcement learning, send the control strategy to the twin modeling module for advanced simulation verification, receive a simulation verification result returned by the twin modeling module, optimize the control strategy according to the simulation verification result, output a control instruction, and transmit the control instruction to the robust control module; the robust control module configured to receive the control instruction transmitted by the strategy optimization module, convert the control instruction into an actuator set value, simultaneously receive the long-period prediction value transmitted by the long-range prediction module, monitor uncertainty of the long-period prediction value, and switch to a backup control law when the uncertainty exceeds a threshold value; the meta-learning module configured to obtain cross-batch performance data from the long-range prediction module and the strategy optimization module, and dynamically adjust hyperparameters of the long-range prediction module and the strategy optimization module according to the cross-batch performance data.

[0006] Further, the machine learning-based intelligent control system for a high-temperature furnace provided by the application includes the data acquisition module. the signal conditioning unit configured to receive original sensor signals, perform recursive optimal estimation on temperature signals through a Kalman filter, and output denoised signals to the time alignment unit; the time alignment unit configured to receive the denoised signals transmitted by the signal conditioning unit, identify original time stamps of different data streams, perform time axis alignment on low-frequency process event marker points by using a linear interpolation algorithm with the highest sampling frequency as a reference, and output time-aligned data to the data packaging unit; the data packaging unit configured to receive the time-aligned data transmitted by the time alignment unit, package multivariate data at the same time into a structured data frame with a global time stamp, and synchronously write the structured data frame into a real-time cache and a long-term storage.

[0007] Further, the machine learning-based intelligent control system for high-temperature furnace provided by the application, the twin modeling module comprises: The simulation engine is configured to load a three-dimensional geometric model of the high-temperature furnace, take the synchronous data stream transmitted by the data acquisition module as a boundary condition, calculate the temperature field and thermal stress field distribution in the furnace through a computational fluid dynamics solver, and output the simulation result to the virtual data extraction unit; The virtual data extraction unit receives the simulation result transmitted by the simulation engine, extracts virtual temperature measurement point data of the physical sensor installation position from the simulation result, and outputs the virtual temperature measurement point data to the parameter self-correction unit; The parameter self-correction unit receives the virtual temperature measurement point data transmitted by the virtual data extraction unit, compares the virtual temperature measurement point data with the corresponding physical sensor readings, constructs an error function between the virtual temperature measurement point data and the physical sensor readings when the average absolute error is continuously out of limit, calculates the Jacobian matrix of the error function with respect to the thermal physical property parameters of the simulation model, iteratively updates the thermal physical property parameters of the material combined with the damping factor to minimize the error function, and optimizes the thermal physical property parameters of the material in the simulation model.

[0008] Further, the machine learning-based intelligent control system for high-temperature furnace provided by the application, the twin modeling module comprises: The data dimension reduction unit is configured to perform standardization processing on the physical field data in the furnace transmitted by the twin modeling module, and extract characteristic components by principal component analysis on the highly correlated variable group, and output the dimension-reduced characteristic data to the causal diagram construction unit; The causal diagram construction unit receives the dimension-reduced characteristic data transmitted by the data dimension reduction unit, constructs a complete undirected graph with characteristic variables as nodes, removes redundant edges through conditional independence test, performs edge orientation combined with time sequence constraint, generates a directed acyclic graph, and transmits the directed acyclic graph to the lag analysis unit; The lag analysis unit receives the directed acyclic graph transmitted by the causal diagram construction unit, calculates the mutual information function of the edges in the directed acyclic graph, determines the lag time of the maximum causal effect and labels it in the directed acyclic graph, and outputs the directed acyclic graph with the labeled lag time as the causal diagram.

[0009] Further, the machine learning-based intelligent control system for high-temperature furnace provided by the application, in the long-range prediction module: The encoding unit is configured to encode the synchronous data stream transmitted by the data acquisition module by using a sliding window attention mechanism, and only calculate attention weights with adjacent positions within the window at each position, and output the encoded sequence representation to the memory unit; The memory unit receives the encoded sequence representation transmitted by the encoding unit and the causal graph transmitted by the causal analysis module, calculates an update gating signal according to the correlation between the current input and the stored features, persistently saves a long-period key path feature vector identified by the causal graph, and outputs a memory state to the prediction unit; The prediction unit receives the memory state transmitted by the memory unit and the encoded sequence representation transmitted by the encoding unit, performs prediction, and outputs a long-period prediction value of the product quality index. The training of the long-range prediction module adopts a curriculum learning strategy, and short sequence prediction is trained in the initial stage, and the sequence length is gradually lengthened.

[0010] Further, the strategy optimization module of the machine learning-based high-temperature furnace intelligent control system comprises: The state construction unit is configured to combine the real-time sensor readings transmitted by the data acquisition module, the virtual state output by the twin modeling module, and the long-period prediction value transmitted by the long-range prediction module, construct a partially observable Markov decision process state vector, and output the state vector to the strategy verification unit. The strategy verification unit receives the state vector transmitted by the state construction unit, generates a preliminary control strategy sequence according to the state vector, sends the preliminary control strategy sequence to the twin modeling module for advance simulation, receives the simulation verification result including the estimated quality index returned by the twin modeling module, and outputs the simulation verification result to the strategy updating unit. The strategy updating unit receives the simulation verification result transmitted by the strategy verification unit, calculates a comprehensive reward according to the simulation verification result, iteratively updates the strategy network parameters through a proximal policy optimization algorithm, and generates a control instruction.

[0011] Further, the robust control module of the machine learning-based high-temperature furnace intelligent control system comprises: The model predictive controller is configured to receive the control instruction transmitted by the strategy optimization module, take the control instruction as a reference trajectory, solve a constrained quadratic programming problem at each control period, and output a locally optimized control amount to the switching control unit. The uncertainty evaluation unit is configured to receive the long-period prediction value transmitted by the long-range prediction module, generate a plurality of resampling samples by repeatedly randomly sampling the long-period prediction value, calculate the statistics of each resampling sample, calculate the confidence interval width according to the statistical quantity distribution of all resampling samples, and output the confidence interval width data to the switching control unit. The switching control unit receives the local optimization control quantity transmitted by the model predictive controller and the confidence interval width data transmitted by the uncertainty evaluation unit, generates a switching signal when the confidence interval width exceeds a safety threshold, and smoothly transfers the control right from the model predictive controller to the parallel PID controller, otherwise outputs the local optimization control quantity as the actuator set value.

[0012] Further, the meta-learning module of the machine learning-based high-temperature furnace intelligent control system comprises: The metadata collection unit is configured to collect cross-batch performance data including control target deviation, prediction error and policy convergence speed from the running logs of the long-range prediction module and the policy optimization module, and output the collected cross-batch performance data to the task definition unit. The task definition unit receives the cross-batch performance data transmitted by the metadata collection unit, defines different process batches as independent learning tasks based on the cross-batch performance data, and outputs the task definition result to the parameter optimization unit. The parameter optimization unit receives the task definition result transmitted by the task definition unit, adjusts the initial parameters of the neural network of the long-range prediction module and the initial parameters of the policy network of the policy optimization module across tasks through an inner-outer loop optimization framework, and distributes the optimized parameters to the long-range prediction module and the policy optimization module.

[0013] Further, the machine learning-based high-temperature furnace intelligent control system comprises three-layer closed loops formed by bidirectional data flow among the modules in the system, which comprise a perception closed loop, a decision closed loop and an evolution closed loop. In the perception closed loop, the virtual temperature measurement point data is transmitted to the data acquisition module for comparison, and when the comparison error exceeds a threshold, the parameter self-correction process of the twin modeling module is triggered. In the decision closed loop, the generated causal diagram is transmitted to the long-range prediction module to guide the attention mechanism, the long-range prediction module transmits the long-period prediction value to the policy optimization module, the policy optimization module sends the control strategy sequence to the twin modeling module for advanced simulation verification, and the simulation verification result is returned to the policy optimization module for optimizing the control strategy. In the evolution closed loop, the meta-learning module collects cross-batch performance data from the long-range prediction module and the policy optimization module, optimizes the hyperparameters of the long-range prediction module and the policy optimization module based on the cross-batch performance data, and distributes the optimized hyperparameters to the long-range prediction module and the policy optimization module.

[0014] In the second aspect, the machine learning-based high-temperature furnace intelligent control method is applied to the machine learning-based high-temperature furnace intelligent control system, and comprises: Step 1: Obtain sensor readings from the distributed control system of the high-temperature furnace, denoise and time-synchronize the sensor readings, and generate a synchronized data stream; Step 2: Receive the synchronized data stream and use it as a boundary condition to drive the dynamic simulation of the digital twin of the high-temperature furnace, obtaining physical field data in the furnace including virtual temperature measurement point data; Step 3: Receive the physical field data in the furnace and perform causal structure learning on it, identify key early operation variables and time lag effects, and generate a causal diagram; Step 4: Receive the synchronized data stream and the causal diagram, use a neural network architecture with integrated gate memory units, use the causal diagram to guide the attention mechanism of the neural network to process the synchronized data stream, and output long-period prediction values of product quality indicators; Step 5: Receive the long-period prediction values, incorporate them into the reward function, search for control strategies through reinforcement learning, send control strategies to the digital twin for advanced simulation verification, receive simulation verification results, optimize control strategies according to simulation verification results, and generate control instructions; Step 6: Receive the control instructions and convert them into actuator set values, while receiving the long-period prediction values, monitor the uncertainty of the long-period prediction values, and switch to a backup control law when the uncertainty exceeds a threshold; Step 7: Obtain cross-batch performance data from the long-range prediction and strategy optimization process, and dynamically adjust the hyperparameters of long-range prediction and strategy optimization according to the cross-batch performance data.

[0015] The present application has the following advantages: The application effectively solves the gradient disappearance problem in long-term dependence modeling by adopting a neural network architecture integrating a gated memory unit and replacing the existing recurrent neural network with a sliding window attention mechanism; the encoding unit in the long-range prediction module adopts sliding window attention processing of the synchronous data stream, and each position only calculates the attention weight with the adjacent positions within the window, thereby reducing the computational complexity and allowing the model to capture long sequence dependence without relying on gradient backtracking; the memory unit persistently stores the long-period key path feature vector identified by the causal graph generated by the causal analysis module through the gating mechanism, and dynamically updates the memory state using the update gating signal to avoid the attenuation of the gradient in the time series propagation; the causal analysis module identifies key early operation variables and time lag effects from the in-furnace physical field data, generates a causal graph as prior knowledge to guide the attention mechanism to focus on long-period causal relationships, and strengthens the model's perception of key influences on slow dynamic processes; the strategy optimization module performs advanced simulation verification through the twin modeling module, evaluates the long-term effect of the control strategy in the virtual environment, and avoids learning difficulties caused by real system delays; the meta-learning module dynamically adjusts the hyperparameters of the long-range prediction module and the strategy optimization module, enabling the system to adaptively optimize across batches and further improve control accuracy and robustness; such a multi-module collaborative design fundamentally avoids the limitations of recurrent neural networks in long-term dependence modeling, and realizes accurate control of the slow dynamic process of the high-temperature furnace. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the present application, the following will briefly introduce the drawings needed in the embodiments. Obviously, for those skilled in the art, other drawings can also be obtained without creative labor based on the drawings.

[0017] Figure 1 The flowchart of the high-temperature furnace intelligent control method based on machine learning of the present application. DETAILED DESCRIPTION

[0018] In order to make the technical solutions of the present application clearer, the following will combine the specific embodiments of the present application and the corresponding drawings to clearly and completely describe the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application. The following will combine the drawings to specifically describe the present application provided by each embodiment of the present application. In order to better understand the purpose of the present application, the following will further describe the present application in detail.

[0019] In a first aspect, the application provides a machine learning-based intelligent control system for a high-temperature furnace, comprising: a data acquisition module, a twin modeling module, a causal analysis module, a long-range prediction module, a strategy optimization module, a robust control module, and a meta-learning module The data acquisition module is configured to obtain sensor readings from a distributed control system of the high-temperature furnace, perform noise reduction and time synchronization processing on the sensor readings, and output a synchronized data stream to the twin modeling module and the long-range prediction module. The twin modeling module receives the synchronized data stream transmitted by the data acquisition module, uses the synchronized data stream as a boundary condition to drive dynamic simulation of a digital twin of the high-temperature furnace, obtains in-furnace physical field data including virtual temperature measurement point data, and transmits the in-furnace physical field data to the causal analysis module. The causal analysis module receives the in-furnace physical field data transmitted by the twin modeling module, performs causal structure learning on the in-furnace physical field data, identifies key early operation variables and time lag effects, generates a causal graph, and transmits the causal graph to the long-range prediction module. The long-range prediction module receives the synchronized data stream transmitted by the data acquisition module and the causal graph transmitted by the causal analysis module, uses a neural network architecture with integrated gated memory units, uses the causal graph to guide the attention mechanism of the neural network to process the synchronized data stream, outputs long-period prediction values of product quality indicators, and transmits the long-period prediction values to the strategy optimization module. The strategy optimization module receives the long-period prediction values transmitted by the long-range prediction module, incorporates the long-period prediction values into a reward function, searches for a control strategy through reinforcement learning, sends the control strategy to the twin modeling module for advance simulation verification, receives the simulation verification results returned by the twin modeling module, optimizes the control strategy according to the simulation verification results, outputs control instructions, and transmits the control instructions to the robust control module. The robust control module receives the control instructions transmitted by the strategy optimization module, converts the control instructions into actuator set values, simultaneously receives the long-period prediction values transmitted by the long-range prediction module, monitors the uncertainty of the long-period prediction values, and switches to a backup control law when the uncertainty exceeds a threshold. The meta-learning module obtains cross-batch performance data from the long-range prediction module and the strategy optimization module, and dynamically adjusts the hyperparameters of the long-range prediction module and the strategy optimization module according to the cross-batch performance data.

[0020] The data acquisition module obtains sensor readings from the distributed control system of the high-temperature furnace. Firstly, the Kalman filter is used to recursively estimate the temperature signal optimally, suppress the measurement noise, and output the denoised signal. Then, the dynamic time warping algorithm is used to identify the original timestamps of different data streams. Based on the highest sampling frequency, the low-frequency process event marker points are linearly interpolated and aligned to form a unified time axis. Finally, the multivariate data at the same time is packaged into a structured data frame with a global timestamp, and is written into the real-time cache and long-term storage synchronously to generate a synchronous data stream. This process ensures the data quality and consistency, providing reliable input for the subsequent modules.

[0021] The twin modeling module receives the synchronous data stream transmitted by the data acquisition module, and drives the digital twin of the high-temperature furnace to perform dynamic simulation as boundary conditions. The simulation engine loads the three-dimensional geometric model of the high-temperature furnace, calculates the temperature field and thermal stress field distribution in the furnace through the computational fluid dynamics solver, and outputs the simulation results. The virtual data extraction unit extracts the virtual temperature measurement point data at the physical sensor installation position from the simulation results. The parameter self-correction unit compares the virtual temperature measurement point data with the corresponding physical sensor readings. When the average absolute error continuously exceeds the limit, an error function is constructed and the Jacobian matrix is calculated. Combined with the damping factor, the material thermal physical property parameters are iteratively updated to optimize the simulation model. This step realizes the accurate simulation of the physical state in the furnace and the self-correction of the model.

[0022] The causal analysis module receives the in-furnace physical field data transmitted by the twin modeling module. Firstly, the data is standardized, and the principal component analysis is used to extract characteristic components from highly correlated variable groups to reduce the data dimension. The causal graph construction unit constructs a complete undirected graph with characteristic variables as nodes, removes redundant edges through conditional independence test, and generates a directed acyclic graph by edge orientation combined with time sequence constraints. The lag analysis unit calculates the mutual information function of the edges in the directed acyclic graph, determines the lag time of the maximum causal effect, and labels it in the graph, and outputs the causal graph. This process identifies key early operation variables and long-delay causal relationships, providing prior knowledge for the prediction module.

[0023] The long-range prediction module receives the synchronous data stream transmitted by the data acquisition module and the causal graph transmitted by the causal analysis module. The encoding unit uses a sliding window attention mechanism to encode the synchronous data stream, and each position only calculates attention weights with adjacent positions within the window. The output is a sequence representation after encoding. The memory unit calculates the update gate signal according to the relevance of the current input and the stored features, and persistently saves the long-period key path feature vector identified by the causal graph. The output is the memory state. The prediction unit combines the memory state and the encoded sequence to make predictions, and outputs the long-period prediction value of the product quality index. During the training process, a curriculum learning strategy is used to gradually increase the sequence length to improve the adaptability of the model.

[0024] The strategy optimization module receives the long-period prediction value transmitted by the long-range prediction module, the state construction unit combines real-time sensor readings, virtual states and long-period prediction values to construct a partially observable Markov decision process state vector; the strategy verification unit generates a preliminary control strategy sequence according to the state vector, sends it to the twin modeling module for forward simulation, and receives simulation verification results including estimated quality indicators; the strategy updating unit calculates the comprehensive reward according to the simulation verification results, iteratively updates the strategy network parameters through the proximal policy optimization algorithm, and generates control instructions. This step optimizes the control strategy through simulation verification to ensure long-term performance.

[0025] The robust control module receives the control instructions transmitted by the strategy optimization module, and the model predictive controller takes the control instructions as the reference trajectory to solve the constrained quadratic programming problem at each control period and output the local optimization control amount; the uncertainty evaluation unit generates resampled samples by multiple random sampling with replacement on the long-period prediction value, calculates the statistical distribution and obtains the confidence interval width; the switching control unit switches to the PID controller smoothly when the confidence interval width data exceeds the safety threshold, otherwise outputs the local optimization control amount as the actuator set value. This mechanism guarantees the safety and robustness of the control system.

[0026] The meta-learning module collects cross-batch performance data from the running logs of the long-range prediction module and the strategy optimization module, including control target deviation, prediction error and policy convergence speed; the task definition unit defines different process batches as independent learning tasks based on the performance data; the parameter optimization unit adjusts the neural network initial parameters and the policy network initial parameters across tasks through an inner-outer loop optimization framework, and distributes the optimized parameters to the corresponding modules. This process realizes the continuous adaptive optimization of system parameters and improves the overall control performance.

[0027] Each module is closely connected through data flow: the data acquisition module provides synchronous data for twin modeling and long-range prediction; the twin modeling module provides physical field data for causal analysis, and the causal analysis module outputs a causal diagram to guide long-range prediction; the long-range prediction module provides prediction values for strategy optimization, and the strategy optimization module verifies the strategy through twin simulation; the robust control module executes instructions and monitors uncertainty; the meta-learning module optimizes parameters based on global performance. This closed-loop design ensures the coherence and adaptive ability of the system from data perception to decision execution.

[0028] In the data acquisition module, the signal conditioning unit receives the original sensor signal, performs recursive optimal estimation on the temperature signal through the Kalman filter, and outputs the noise-reduced signal to the time alignment unit; the time alignment unit receives the noise-reduced signal, identifies the original time stamp of different data streams, and aligns the time axis of low-frequency process event marker points using linear interpolation algorithm based on the highest sampling frequency, and outputs the time-aligned data to the data packaging unit; the data packaging unit receives the time-aligned data, packages the multivariate data at the same time into a structured data frame with a global time stamp, and synchronously writes the structured data frame into the real-time cache and the long-term storage. This process realizes the unified processing of multi-source heterogeneous data and provides time-synchronized high-quality data input for the subsequent modules.

[0029] In the twin modeling module, the simulation engine loads the high-temperature furnace three-dimensional geometric model, takes the synchronous data stream transmitted by the data acquisition module as the boundary condition, calculates the temperature field and thermal stress field distribution in the furnace through the computational fluid dynamics solver, and outputs the simulation results to the virtual data extraction unit; the virtual data extraction unit receives the simulation results, extracts the virtual temperature measurement point data of the physical sensor installation position from the simulation results, and outputs the virtual temperature measurement point data to the parameter self-correction unit; the parameter self-correction unit receives the virtual temperature measurement point data, compares the virtual temperature measurement point data with the corresponding physical sensor readings, and when the average absolute error is continuously out of limit, constructs an error function and calculates a Jacobian matrix, and iteratively updates the material thermal physical property parameters combined with a damping factor to minimize the error function. This closed-loop correction mechanism ensures the consistency of the simulation accuracy of the digital twin and the physical entity.

[0030] In the causal analysis module, the data dimension reduction unit performs standardization processing on the furnace physical field data transmitted by the twin modeling module, and extracts characteristic components using principal component analysis on highly correlated variable groups, and outputs the dimension-reduced characteristic data to the causal diagram construction unit; the causal diagram construction unit receives the dimension-reduced characteristic data, constructs a complete undirected graph with characteristic variables as nodes, removes redundant edges through conditional independence test, and performs edge orientation combined with time sequence constraints to generate a directed acyclic graph and transmit it to the lag analysis unit; the lag analysis unit receives the directed acyclic graph, calculates the mutual information function of the edges in the directed acyclic graph, determines the lag time of the maximum causal effect and labels it in the graph. This process extracts the causal structure with clear time lag relationship from high-dimensional data.

[0031] In the long-term prediction module, the encoding unit encodes the synchronous data stream using a sliding window attention mechanism, each position only calculates attention weights with adjacent positions within the window, and outputs the encoded sequence representation to the memory unit; the memory unit receives the encoded sequence representation and the causal graph, calculates the update gate signal according to the correlation between the current input and the stored features, persistently saves the long-period key path feature vector identified by the causal graph, and outputs the memory state to the prediction unit; the prediction unit receives the memory state and the encoded sequence representation to predict and output the long-period prediction value. The training process adopts a curriculum learning strategy, gradually extending the sequence length to adapt to long-period dependencies.

[0032] In the strategy optimization module, the state construction unit combines real-time sensor readings, virtual states, and long-period prediction values to construct a partially observable Markov decision process state vector, and outputs the state vector to the strategy verification unit; the strategy verification unit receives the state vector to generate a preliminary control strategy sequence, sends it to the twin modeling module for forward simulation, receives simulation verification results including estimated quality indicators, and outputs them to the strategy update unit; the strategy update unit receives the simulation verification results to calculate the comprehensive reward, iteratively updates the strategy network parameters through a proximal policy optimization algorithm to generate control instructions. This verification and optimization mechanism ensures the long-term effectiveness of the control strategy.

[0033] In the robust control module, the model predictive controller receives the control instructions as the reference trajectory, solves a constrained quadratic programming problem at each control period to output a locally optimized control amount to the switching control unit; the uncertainty assessment unit receives the long-period prediction value to generate resampled samples through multiple random sampling with replacement, calculates the statistical distribution to obtain the confidence interval width data, and outputs them to the switching control unit; the switching control unit receives the locally optimized control amount and the confidence interval width data, switches to the PID controller when the confidence interval width exceeds the safety threshold, otherwise outputs the locally optimized control amount. This design realizes the safe and seamless switching between intelligent control and classic control.

[0034] In the meta-learning module, the meta-data collection unit collects cross-batch performance data including control target deviation, prediction error, and policy convergence speed from the operation log, and outputs the data to the task definition unit; the task definition unit receives the cross-batch performance data, defines different process batches as independent learning tasks, and outputs the task definition results to the parameter optimization unit; the parameter optimization unit receives the task definition results, adjusts the neural network initial parameters and the policy network initial parameters through an inner-outer loop optimization framework, and distributes the optimized parameters to the corresponding modules. This process realizes the continuous self-optimization of system parameters.

[0035] The system realizes collaborative optimization through three layers of closed loops: in the perception closed loop, the comparison of virtual temperature measurement point data and physical sensor readings triggers parameter self-correction; in the decision closed loop, the causal diagram guides the prediction focus, and the twin simulation verifies the long-term effect of the strategy; in the evolution closed loop, cross-batch performance data drives hyperparameter optimization. The three closed loops are coupled through data flow to form an intelligent control system with adaptive capability.

[0036] In a second aspect, referring to Figure 1 The application provides a machine learning-based intelligent control method for a high-temperature furnace, which is applied to a machine learning-based intelligent control system for the high-temperature furnace and includes the following steps: Step 1: Obtain sensor readings from a distributed control system of the high-temperature furnace, perform noise reduction and time synchronization processing on the sensor readings, and generate a synchronous data stream; Step 2: Receive the synchronous data stream and use it as a boundary condition to drive a digital twin of the high-temperature furnace to perform dynamic simulation and obtain furnace physical field data including virtual temperature measurement point data; Step 3: Receive the furnace physical field data, perform causal structure learning thereon, identify key early operation variables and time lag effects, and generate a causal diagram; Step 4: Receive the synchronous data stream and the causal diagram, adopt a neural network architecture integrating a gated memory unit, use the causal diagram to guide the attention mechanism of the neural network to process the synchronous data stream, and output a long-period prediction value of a product quality index; Step 5: Receive the long-period prediction value, incorporate it into a reward function, search for a control strategy through reinforcement learning, send the control strategy to the digital twin for advanced simulation verification, receive the simulation verification result, optimize the control strategy according to the simulation verification result, and generate a control instruction; Step 6: Receive the control instruction and convert it into an actuator set value, simultaneously receive the long-period prediction value, monitor the uncertainty of the long-period prediction value, and switch to a backup control law when the uncertainty exceeds a threshold value; Step 7: Obtain cross-batch performance data from the long-range prediction and strategy optimization process, and dynamically adjust the hyperparameters of the long-range prediction and strategy optimization according to the cross-batch performance data.

[0037] The application solves the gradient disappearance problem by replacing the recurrent neural network architecture with a neural network integrated with a gating memory unit and an attention mechanism. In the long-range prediction module, the encoding unit uses a sliding window attention mechanism to process synchronous data streams, and each position only calculates the attention weight with the adjacent positions within the window, thereby reducing the computational complexity and allowing the model to capture long sequence dependencies without relying on gradient backtracking. The memory unit persists the long-period key path feature vector identified by the causal graph generated by the causal analysis module through the gating mechanism, and dynamically updates the memory state according to the relevance of the current input and the stored features using the gating signal, avoiding the attenuation of the gradient in the time series propagation. The causal analysis module identifies key early operation variables and time lag effects from the in-furnace physical field data, generates a causal graph as prior knowledge to guide the attention mechanism to focus on long-period causal relationships, and strengthens the model's perception of key influences on slow dynamic processes. At the system level, the strategy optimization module performs advanced simulation verification through the twin modeling module to evaluate the long-term effect of the control strategy in the virtual environment, avoiding the learning difficulties caused by the delay of the real system; the meta-learning module dynamically adjusts the hyperparameters of the long-range prediction module and the strategy optimization module, enabling the system to adaptively optimize across batches, further improving the accuracy and robustness of long-period control. This kind of multi-module collaborative design fundamentally avoids the limitations of recurrent neural networks in long-term dependency modeling, and realizes accurate control of the slow dynamic process of high-temperature furnaces.

Claims

1. A machine learning-based intelligent control system for a high-temperature furnace, characterized in that, include: Data acquisition module, twin modeling module, causal analysis module, long-range prediction module, strategy optimization module, robust control module, and meta-learning module. The data acquisition module is configured to acquire sensor readings from the distributed control system of the high-temperature furnace, perform noise reduction and time synchronization processing on the sensor readings, and output a synchronized data stream to the twin modeling module and the long-range prediction module. The twin modeling module receives the synchronous data stream transmitted by the data acquisition module, uses the synchronous data stream as a boundary condition to drive the digital twin of the high-temperature furnace for dynamic simulation, obtains the physical field data inside the furnace, including virtual temperature measurement point data, and transmits the physical field data inside the furnace to the causal analysis module. The causal analysis module receives the furnace physical field data transmitted by the twin modeling module, performs causal structure learning on the furnace physical field data, identifies key early operating variables and time lag effects, generates a causal graph, and transmits the causal graph to the long-range prediction module. The long-range prediction module receives the synchronous data stream transmitted by the data acquisition module and the causal graph transmitted by the causal analysis module. It adopts a neural network architecture with integrated gated memory units, uses the causal graph to guide the attention mechanism of the neural network to process the synchronous data stream, outputs the long-term predicted value of the product quality index, and transmits the long-term predicted value to the strategy optimization module. The strategy optimization module receives long-term prediction values ​​from the long-range prediction module, incorporates these prediction values ​​into the reward function, searches for control strategies through reinforcement learning, sends the control strategies to the twin modeling module for advanced simulation verification, receives the simulation verification results returned by the twin modeling module, optimizes the control strategies based on the simulation verification results, outputs control commands, and transmits the control commands to the robust control module. The robust control module receives control commands from the strategy optimization module, converts the control commands into actuator setpoints, and simultaneously receives long-term prediction values ​​from the long-term prediction module. It monitors the uncertainty of the long-term prediction values ​​and switches to the backup control law when the uncertainty exceeds a threshold. The meta-learning module obtains cross-batch performance data from the long-range prediction module and the strategy optimization module, and dynamically adjusts the hyperparameters of the long-range prediction module and the strategy optimization module based on the cross-batch performance data.

2. The intelligent control system for a high-temperature furnace based on machine learning according to claim 1, characterized in that, The data acquisition module includes: The signal conditioning unit is configured to receive the raw sensor signal, perform recursive optimal estimation of the temperature signal through a Kalman filter, and output the noise-reduced signal to the time alignment unit. The time alignment unit receives the noise-reduced signal transmitted by the signal conditioning unit, identifies the original timestamps of different data streams, uses the highest sampling frequency as a reference, performs time axis alignment on the low-frequency process event markers using a linear interpolation algorithm, and outputs the time-aligned data to the data encapsulation unit. The data encapsulation unit receives time-aligned data transmitted by the time alignment unit, encapsulates multivariate data at the same time into structured data frames with global timestamps, and synchronously writes the structured data frames into the real-time cache and long-term storage.

3. The intelligent control system for a high-temperature furnace based on machine learning according to claim 1, characterized in that, The twin modeling module includes: The simulation engine is configured to load a three-dimensional geometric model of a high-temperature furnace, use the synchronous data stream transmitted by the data acquisition module as boundary conditions, calculate the temperature field and thermal stress field distribution inside the furnace through a computational fluid dynamics solver, and output the simulation results to the virtual data extraction unit. The virtual data extraction unit receives the simulation results transmitted by the simulation engine, extracts the virtual temperature measurement point data of the physical sensor installation location from the simulation results, and outputs the virtual temperature measurement point data to the parameter self-calibration unit. The parameter self-calibration unit receives virtual temperature measurement point data transmitted by the virtual data extraction unit, compares the virtual temperature measurement point data with the corresponding physical sensor readings, and constructs an error function between the virtual temperature measurement point data and the physical sensor readings when the mean absolute error continuously exceeds the limit. It calculates the Jacobian matrix of the error function with respect to the material thermal property parameters of the simulation model, and iteratively updates the material thermal property parameters in combination with the damping factor to minimize the error function and optimize the material thermal property parameters in the simulation model.

4. The intelligent control system for a high-temperature furnace based on machine learning according to claim 1, characterized in that, The causal analysis module includes: The data dimensionality reduction unit is configured to standardize the furnace physical field data transmitted by the twin modeling module, extract feature components from highly correlated variable groups using principal component analysis, and output the dimensionality-reduced feature data to the causal graph construction unit. The cause-effect graph construction unit receives the dimensionality-reduced feature data transmitted by the data dimensionality reduction unit, constructs a completely undirected graph with the feature variables as nodes, removes redundant edges through conditional independence tests, orients edges by combining time order constraints, generates a directed acyclic graph, and transmits the directed acyclic graph to the hysteresis analysis unit. The lag analysis unit receives the directed acyclic graph transmitted by the causal graph construction unit, calculates the mutual information function of the edges in the directed acyclic graph, determines the lag time of the maximum causal effect and marks it in the directed acyclic graph, and outputs the directed acyclic graph marked with lag time as the causal graph.

5. The intelligent control system for a high-temperature furnace based on machine learning according to claim 1, characterized in that, In the long-range prediction module: The encoding unit is configured to use a sliding window attention mechanism to encode the synchronous data stream transmitted by the data acquisition module. Each position only calculates the attention weight with the neighboring positions within the window, and outputs the encoded sequence representation to the memory unit. The memory unit receives the encoded sequence representation transmitted by the encoding unit and the causal graph transmitted by the causal analysis module. It calculates and updates the gating signal based on the correlation between the current input and the stored features, persistently saves the long-period critical path feature vector identified by the causal graph, and outputs the memory state to the prediction unit. The prediction unit receives the memory state transmitted by the memory unit and the encoded sequence representation transmitted by the encoding unit, performs prediction, and outputs long-term predicted values ​​of product quality indicators. The long-range prediction module employs a course-based learning strategy, initially training short-sequence prediction and gradually extending the sequence length.

6. The intelligent control system for a high-temperature furnace based on machine learning according to claim 1, characterized in that, The strategy optimization module includes: The state construction unit is configured to combine real-time sensor readings transmitted by the data acquisition module, virtual states output by the twin modeling module, and long-term predicted values ​​transmitted by the long-range prediction module to construct a partially observable Markov decision process state vector, and output the state vector to the policy verification unit. The strategy verification unit receives the state vector transmitted by the state construction unit, generates a preliminary control strategy sequence based on the state vector, sends the preliminary control strategy sequence to the twin modeling module for advanced simulation, receives the simulation verification results returned by the twin modeling module including the estimated quality index, and outputs the simulation verification results to the strategy update unit. The policy update unit receives the simulation verification results transmitted by the policy verification unit, calculates the comprehensive reward based on the simulation verification results, iteratively updates the policy network parameters through the near-end policy optimization algorithm, and generates control commands.

7. The intelligent control system for a high-temperature furnace based on machine learning according to claim 1, characterized in that, The robust control module includes: The model predictive controller is configured to receive control commands transmitted by the strategy optimization module, use the control commands as a reference trajectory, solve the constrained quadratic programming problem in each control cycle, and output the local optimization control quantity to the switching control unit. The uncertainty assessment unit is configured to receive long-term prediction values ​​transmitted by the long-range prediction module, perform multiple random samplings with replacement on the long-term prediction values ​​to generate multiple resampled samples, calculate the statistic of each resampled sample, calculate the confidence interval width based on the distribution of the statistics of all resampled samples, and output the confidence interval width data to the switching control unit. The switching control unit receives the local optimization control quantity transmitted by the model predictive controller and the confidence interval width data transmitted by the uncertainty assessment unit. When the confidence interval width exceeds the safety threshold, a switching signal is generated to smoothly transfer the control power from the model predictive controller to the parallel PID controller; otherwise, the local optimization control quantity is output as the actuator setpoint.

8. The intelligent control system for a high-temperature furnace based on machine learning according to claim 1, characterized in that, The meta-learning module includes: The metadata collection unit is configured to collect cross-batch performance data from the operation logs of the long-range prediction module and the strategy optimization module. The cross-batch performance data includes control target deviation, prediction error and strategy convergence speed. The collected cross-batch performance data is then output to the task definition unit. The task definition unit receives cross-batch performance data transmitted by the metadata collection unit, defines different process batches as independent learning tasks based on the cross-batch performance data, and outputs the task definition results to the parameter optimization unit. The parameter optimization unit receives the task definition results transmitted by the task definition unit, and adjusts the initial parameters of the neural network of the long-range prediction module and the initial parameters of the policy network of the policy optimization module across tasks through the inner and outer loop optimization framework. The optimized parameters are then distributed to the long-range prediction module and the policy optimization module.

9. The intelligent control system for a high-temperature furnace based on machine learning according to claim 1, characterized in that, The modules within the system form a three-layer closed loop through bidirectional data flow. The three-layer closed loop includes a perception closed loop, a decision closed loop, and an evolution closed loop. In the perception closed loop, the twin modeling module transmits virtual temperature measurement point data to the data acquisition module for comparison. When the comparison error exceeds the threshold, the parameter self-calibration process of the twin modeling module is triggered. In the decision-making closed loop, the causal analysis module transmits the generated causal graph to the long-range prediction module to guide the attention mechanism. The long-range prediction module transmits the long-cycle prediction value to the strategy optimization module. The strategy optimization module sends the control strategy sequence to the twin modeling module for advanced simulation verification and returns the simulation verification results to the strategy optimization module for optimizing the control strategy. In the evolution loop, the meta-learning module collects cross-batch performance data from the long-range prediction module and the policy optimization module, optimizes the hyperparameters of the long-range prediction module and the policy optimization module based on the cross-batch performance data, and distributes the optimized hyperparameters to the long-range prediction module and the policy optimization module.

10. A machine learning-based intelligent control method for high-temperature furnaces, applied to the machine learning-based intelligent control system for high-temperature furnaces as described in any one of claims 1 to 9, characterized in that, include: Step 1: Obtain sensor readings from the distributed control system of the high-temperature furnace, perform noise reduction and time synchronization processing on the sensor readings, and generate a synchronized data stream; Step 2: Receive the synchronous data stream and use it as a digital twin of the high-temperature furnace driven by boundary conditions for dynamic simulation to obtain the furnace internal physical field data including virtual temperature measurement point data. Step 3: Receive the physical field data inside the furnace, perform causal structure learning on it, identify key early operation variables and time lag effects, and generate a causal graph; Step 4: Receive the synchronous data stream and the causal graph, adopt a neural network architecture with integrated gated memory units, use the causal graph to guide the attention mechanism of the neural network to process the synchronous data stream, and output long-term predicted values ​​of product quality indicators. Step 5: Receive the long-cycle predicted value, incorporate it into the reward function, search for a control strategy through reinforcement learning, send the control strategy to the digital twin for advanced simulation verification, receive the simulation verification results, optimize the control strategy based on the simulation verification results, and generate control commands. Step 6: Receive the control command, convert it into actuator setpoint, and simultaneously receive the long-cycle prediction value. Monitor the uncertainty of the long-cycle prediction value and switch to the backup control law when the uncertainty exceeds the threshold. Step 7: Obtain cross-batch performance data from the long-range forecasting and strategy optimization process, and dynamically adjust the hyperparameters of long-range forecasting and strategy optimization based on the cross-batch performance data.

Images