Automatic temperature control method and system for environment-friendly brick roasting kiln

Abstract

The invention provides an automatic temperature control method and system for an environment-friendly brick roasting kiln, and the method comprises the steps: constructing a thermotechnical process knowledge graph skeleton, associating a thermodynamic law with an actual process variable, dynamically correcting a causal structure through multi-batch production data, and forming a multi-version causal path set through combining anti-fact deduction, context causal branches and graph labels; based on a causal atlas and a reinforcement learning mechanism, an intelligent temperature control strategy network is designed, action-causal chain-effect collaborative interpretation is realized, control suggestions with physical feasibility and process context self-adaption are generated, and continuous optimization is performed through closed-loop feedback. And the dynamic collaborative optimization capability of temperature control intellectualization and product quality gain is enhanced.

Description

Technical Field

[0001] This invention relates to the field of process control technology for environmentally friendly brick firing kilns, and in particular to an automatic temperature control method and system for environmentally friendly brick firing kilns. Background Technology

[0002] Currently, in the field of automatic temperature control systems and intelligent temperature control decision-making for environmentally friendly brick firing kilns, mainstream technical solutions largely rely on industrial process automation, expert rule setting, and traditional knowledge graph modeling methods to optimize temperature control parameters. Existing systems typically reflect the relationship between process parameters and product quality through empirical models or mechanism-driven causal networks, using multiple sensors to collect process data, and leveraging basic statistics, regression analysis, or logical rules to achieve process control. These methods can ensure a certain degree of product quality and process stability under normal and stable raw material or process conditions, and are also the basic technical path for the automation of temperature control in industrial kilns. In recent years, with the expansion of the application of artificial intelligence and data-driven methods, the industry has gradually introduced emerging technologies such as machine learning, knowledge graphs, causal reasoning, and reinforcement learning into temperature control systems and process optimization scenarios to realize the logical expression between complex processes. Some research focuses on the automatic discovery of causal relationships, using methods such as Bayesian networks and structural equation modeling to assist in judging the impact of process parameters on product quality; there are also intelligent temperature control strategies based on reinforcement learning, used to dynamically adjust control parameters to optimize system performance. These trends reflect a transformative direction from static modeling to dynamic, self-learning, multi-dimensional data fusion, and causal logic enhancement. However, existing technologies still have significant shortcomings. Specifically, the establishment of causal relationships in traditional knowledge graphs largely relies on expert experience or prior assumptions. The initial modeling stage is often affected by cognitive biases or incomplete domain knowledge, leading to deviations in the logical expression between process parameters and product quality. Static modeling systems struggle to cope with dynamic scenarios such as fluctuations in raw material properties, environmental changes, and adjustments to process conditions. Causal edge weights or paths tend to become rigid, and control rules lack the ability to evolve with data. Furthermore, existing causal structure learning methods are susceptible to overfitting in multivariate, high-noise industrial environments, resulting in insufficient interpretability of generated causal relationships, with some paths even violating physical constraints such as energy conservation and thermal mechanisms. In addition, reinforcement learning techniques lack a closed-loop mechanism for interaction with domain knowledge in practical applications. The temperature control suggestions generated by the policy network cannot be transparently traced, failing to meet the stringent safety and traceability requirements of industrial control systems. These limitations directly reduce the scientific validity and reliability of control rules, impacting product quality improvement and system stability. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, the present invention provides an automatic temperature control method for environmentally friendly brick firing kilns.

[0004] The technical solution of this invention is implemented as follows: an automatic temperature control method for an environmentally friendly brick firing kiln, comprising: S1: Based on the physical and thermal laws of environmentally friendly brick firing kilns, an initial process knowledge graph skeleton is constructed. The initial process knowledge graph skeleton contains the correlation between known variables determined by thermodynamic laws, which serves as the state modeling basis for temperature control decisions. S2: Perform time-series alignment processing on the real-time acquired multidimensional process parameter sequences and corresponding batch product quality inspection results to generate a timestamped observation dataset, which is used to drive the subsequent causal structure learning process; S3: Based on the observed dataset and combined with the physical constraint priors in the initial process knowledge graph skeleton, execute the dynamic causal discovery algorithm to identify the directed causal connections between key variables that significantly affect product quality, and form an updated causal structure model. S4: Based on the updated causal structure model and the counterfactual inference results of historical high-return control trajectories, calculate the confidence score of each causal edge, and create parallel branches with context labels for causal paths whose confidence exceeds a preset threshold and conflict with the existing graph, generating a set of multiple versions of causal paths. S5: Encode the set of multi-version causal paths into the state space of a Markov decision process, define the action space by combining adjustable control parameters, construct a reward function with the prediction of product quality score as the main term and control stability as the secondary term, and form an interpretable reinforcement learning training environment. S6: In the interpretable reinforcement learning training environment, the Transformer architecture policy network enhanced with attention mechanism is used to select actions, output the combination of control parameters and generate attention weight distribution. After the attention weight distribution is mapped to the knowledge graph node, it forms a control suggestion with logical interpretability. S7: Determine whether the expected quality improvement effect corresponding to the control suggestion meets the set threshold. If it does, convert it into a candidate control rule with confidence label and push it to the manual review queue for verification and confirmation. S8: Inject the approved candidate control rules into the temperature control execution module, and at the same time, send the actual production feedback data after the rules are applied back to the observation dataset for the next round of joint optimization iteration of the causal model and policy network.

[0005] The present invention also provides an automatic temperature control system for environmentally friendly brick firing kilns, which uses the above-mentioned automatic temperature control method for environmentally friendly brick firing kilns to achieve automatic temperature control of the kilns.

[0006] The present invention provides an automatic temperature control method for an environmentally friendly brick firing kiln, which has the following beneficial effects: (1) This invention significantly improves the modeling accuracy and adaptability of the temperature control decision system for complex thermal processes by constructing a lightweight process knowledge graph skeleton and integrating a dynamic causal discovery mechanism. In the initialization stage, this invention only introduces basic physical laws as prior constraints (such as the effect of heating rate on thermal stress), avoiding the need to manually set complex and error-prone complete causal chains. During operation, it periodically learns causal structures based on the actual collected multidimensional process parameter sequences and product quality data, using a hybrid PC and GES algorithm to automatically identify the directed dependencies between key variables, and combines a domain knowledge mask matrix to filter invalid connections that violate thermodynamic laws, effectively preventing overfitting and logical paradoxes. For newly discovered high-confidence causal paths, the system does not directly cover the original graph, but retains the historical evolution trajectory in the form of parallel branches with timestamps and context labels, forming a dynamic knowledge structure that supports multi-context reasoning, thereby realizing continuous perception and adaptive updating of changes in the process environment, effectively overcoming the technical bottleneck of static knowledge models lagging behind production reality. (2) This invention introduces an interpretable reinforcement learning framework and a counterfactual verification mechanism, which greatly enhances the reliability, transparency, and engineering usability of the control strategy. The dynamically updated knowledge graph is encoded into a state space, and the action space corresponds to an adjustable combination of process parameters. The reward function comprehensively considers the product quality prediction score and the control stability penalty term, which promotes the strategy to tend towards an efficient and stable operation mode. The strategy network adopts an attention-enhanced Transformer architecture, which can output the attention weight distribution while generating control instructions and map it to the knowledge graph nodes. It automatically generates natural language explanations in the form of "suggesting to increase the target temperature of the preheating zone to reduce the risk of billet cracking", which significantly improves the efficiency of human-machine collaboration and the credibility of decision-making. More importantly, after each strategy optimization, the system uses the high-reward operation trajectory to conduct counterfactual inference: "If the temperature adjustment is not performed, will the density decrease?" The intervention effect is simulated through a causal model to further verify the real effectiveness of the learned strategy, thereby establishing a closed-loop optimization process of "data-driven - strategy generation - causal verification - rule accumulation", ensuring that the recommended actions are not only statistically relevant, but also have physically interpretable causal support. (3) This invention achieves a paradigm shift from a "static rule base" to a "continuously evolving intelligent decision-making system," possessing significant engineering deployment value and cross-scenario generalization potential. This invention abandons the traditional method of relying on manual maintenance of rule tables, instead constructing a knowledge evolution system that can continuously accumulate experience and correct cognitive biases over time, making it particularly suitable for complex industrial scenarios such as environmentally friendly bricks, where raw materials are diverse, process windows are narrow, and quality sensitivity is high. The system periodically outputs candidate control rules with confidence scores, which are then manually reviewed and injected into the execution module, ensuring both automation levels and retaining necessary manual intervention channels, balancing security and flexibility. The entire architecture is lightweight, with causal search and strategy training both completed on edge computing nodes, eliminating the need for a centralized big data platform, making it suitable for deployment in small to medium-sized production lines. Furthermore, its modular structure allows for future expansion to other high-temperature sintering processes (such as ceramics and metallurgy), demonstrating good portability and industrial application prospects. Attached Figure Description

[0007] Figure 1 A flowchart of an automatic temperature control method for an environmentally friendly brick firing kiln according to the present invention; Figure 2 This is a sub-flowchart of an automatic temperature control method for an environmentally friendly brick firing kiln according to the present invention. Figure 3 This is another sub-flowchart of an automatic temperature control method for an environmentally friendly brick firing kiln according to the present invention. Detailed Implementation

[0008] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0009] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0010] like Figure 1 As shown, this invention provides an automatic temperature control method for an environmentally friendly brick firing kiln, specifically including: S1: Based on the physical and thermal laws of environmentally friendly brick firing kilns, an initial process knowledge graph skeleton is constructed. The initial process knowledge graph skeleton contains the correlation between known variables determined by thermodynamic laws, which serves as the state modeling basis for temperature control decisions. S2: Perform time-series alignment processing on the real-time acquired multidimensional process parameter sequences and corresponding batch product quality inspection results to generate a timestamped observation dataset, which is used to drive the subsequent causal structure learning process; S3: Based on the observed dataset and combined with the physical constraint priors in the initial process knowledge graph skeleton, execute the dynamic causal discovery algorithm to identify the directed causal connections between key variables that significantly affect product quality, and form an updated causal structure model. S4: Based on the updated causal structure model and the counterfactual inference results of historical high-return control trajectories, calculate the confidence score of each causal edge, and create parallel branches with context labels for causal paths whose confidence exceeds a preset threshold and conflict with the existing graph, generating a set of multiple versions of causal paths. S5: Encode the set of multi-version causal paths into the state space of a Markov decision process, define the action space by combining adjustable control parameters, construct a reward function with the prediction of product quality score as the main term and control stability as the secondary term, and form an interpretable reinforcement learning training environment. S6: In the interpretable reinforcement learning training environment, the Transformer architecture policy network enhanced with attention mechanism is used to select actions, output the combination of control parameters and generate attention weight distribution. After the attention weight distribution is mapped to the knowledge graph node, it forms a control suggestion with logical interpretability. S7: Determine whether the expected quality improvement effect corresponding to the control suggestion meets the set threshold. If it does, convert it into a candidate control rule with confidence label and push it to the manual review queue for verification and confirmation. S8: Inject the approved candidate control rules into the temperature control execution module, and at the same time, send the actual production feedback data after the rules are applied back to the observation dataset for the next round of joint optimization iteration of the causal model and policy network.

[0011] Step S1: Based on the physical and thermal laws of the environmentally friendly brick firing kiln, an initial process knowledge graph skeleton is constructed. This initial process knowledge graph skeleton contains the relationships between known variables determined by thermodynamic laws, serving as the state modeling basis for temperature control decisions. Specifically, it includes: S1.1: Based on the thermal mechanism analysis of the environmentally friendly brick firing process, identify the set of key physical variables that affect product quality, including temperature distribution, heating rate, heat preservation time, fuel flow rate, exhaust pressure, flue gas oxygen content, product density, compressive strength, thermal stress, etc., as the basis for defining nodes in the initial process knowledge graph. Based on multi-source process monitoring data from environmentally friendly brick firing kilns, a thermal mechanistic analysis method (parameters: kiln zoning structure, combustion process end-state conditions) was used to achieve preliminary enumeration and classification of key physical variables; Furthermore, a correlation weighting analysis algorithm (parameters: Pearson correlation coefficient r threshold 0.7, observation data window length L=150) was used to screen for the significance of variables such as temperature distribution, heating rate, and product compressive strength, and a variable set was obtained. ; Furthermore, through a multivariate heat balance equation solution algorithm (parameter: specific heat capacity) Calorific value This allows for the derivation of the energy chain between fuel flow rate, holding time, and thermal stress, and the generation of a physical correlation verification matrix. ; Furthermore, by using a flue gas composition-based oxygen content inversion model (parameters: CO2 volume fraction, CO volume fraction, excess air coefficient), the inversion calculation of flue gas oxygen content variables is achieved, and an oxygen content time series is generated. ; Furthermore, by using a pressure flow field calculation method (parameters: furnace geometry, flue resistance coefficient), the exhaust pressure variable is numerically estimated, and the deviation is corrected with actual sensor data to generate a corrected exhaust pressure sequence. ; By using variable mapping, the results obtained from the above filtering and calculation... , , , Merge them into a set of key physical variables to achieve more precise definition of nodes in the initial process knowledge graph; For example, in an environmentally friendly brick production line, the kiln is 45m long and divided into four sections: preheating, heating, insulation, and cooling. Each section is equipped with 16 Pt100 platinum resistance temperature sensors. The data acquisition period is set to 10 seconds, and the temperature distribution variable is defined as (…). )to( The heating rate variable was calculated. The holding time was defined as the continuous time the temperature remained stable within the range of 1100℃±5℃ using a threshold method. The fuel flow rate variable was obtained in real-time using a mass flow meter and the fuel calorific value was corrected. The exhaust pressure was obtained using a differential pressure sensor and corrected based on the drag coefficient K=0.85. The flue gas oxygen content was obtained by measuring CO2 and CO content using an infrared gas analyzer, and then calculated using the excess air coefficient λ=1.15 in the combustion chemical equation. Under this condition, the final set of key physical variables identified included 16 temperature node values, heating rate, holding time, fuel flow rate, exhaust pressure, flue gas oxygen content, product density (obtained from laboratory sample testing), compressive strength, and thermal stress (output from finite element simulation). This set effectively supported the deterministic functional relationship modeling of causal edges in subsequent S1.2 and ensured that the nodes of the initial process knowledge graph covered the main physical factors affecting quality. S1.2: Based on the laws of classical thermodynamics and the principles of heat transfer, establish deterministic functional relationships between key variables. For example, derive the relationship between 'heating rate → temperature gradient' using Fourier's law of heat conduction, and model the 'sintering temperature' using the Arrhenius equation. The density-based nonlinear response characteristics form a set of directed causal edge candidates with clear physical meaning; S1.3: The above deterministic functional relationship is symbolically abstracted and converted into a knowledge representation in the form of a triple (subject-relation-object). For example, logical expressions such as 'heating rate - positively correlated with -temperature gradient' and 'holding time - promotes -grain growth' are generated as semantic unit inputs of the initial process knowledge graph. S1.4: Import the symbolized triple semantic units into the graph database to construct a lightweight initial process knowledge graph skeleton, in which each node is named with standardized terms and assigned unit attributes, and each causal edge is labeled with directionality, action type (promotion / inhibition) and theoretical source label, forming a structured state space modeling foundation; The symbolic triple semantic units output from step S1.3 are used as input objects. An attribute-oriented node definition method (parameters: standardized terminology, unit dictionary) is adopted to achieve unique identification of triple subject and object nodes. Each node is assigned a physical quantity unit attribute to ensure the parsing consistency during subsequent state space modeling. Furthermore, through the causal edge attribute expansion algorithm (parameters: directional tag set, action type enumeration, theoretical source code table), the structured annotation of the triple relationship part is realized, generating a set of directed edge attributes containing directional arrows, action type tags (promotion / inhibition) and associated theoretical source labels, and obtaining a causal edge attribute matrix that can be directly stored; Furthermore, an import method based on graph pattern constraints (parameters: node table, edge table, attribute matrix) is adopted to write the above node and causal edge information into the graph database (supporting Cypher or Gremlin syntax), thereby achieving persistent storage of the initial process knowledge graph skeleton and generating a node-edge relationship structure with index optimization to ensure the efficiency of subsequent queries and updates. Furthermore, through the node attribute consistency verification function (parameters: standardized terminology, unit conversion rules), the imported process variable nodes are checked for unit compatibility and name standardization, and nodes that do not conform to the standard are automatically renamed or converted in unit, and a node attribute table that conforms to the thermal engineering standard is output. By using the directed edge topology verification algorithm (parameters: directional constraint rules, action type constraint rules), the results of the previous step are transformed into structured causal graph data that satisfies the physical and thermal operating mechanism, thereby ensuring the executability of the initial state space model. For example, the input triple semantic unit set {'heating rate - positively correlated with -temperature gradient', 'holding time - promotes -grain growth'} is node-normalized, and 'heating rate' is named 'HeatingRate' using a glossary, with units assigned... Name the 'temperature gradient' 'TemperatureGradient' and assign units as follows: The 'heat preservation time' is named 'SoakingTime', and the unit is assigned... 'Grain growth' is named 'GrainGrowth', with null values ​​indicating dimensionless units. A causal edge attribute expansion algorithm is used to label 'heating rate → temperature gradient' with directionality = positive, action type = promoting, and theoretical source = Fourier's law; and 'holding time → grain growth' with directionality = positive, action type = promoting, and theoretical source = material heat treatment experimental data. During graph database import, four records are inserted into the node table and two records into the edge table, and directionality, action, and source labels are indexed in the attribute matrix. During node attribute consistency verification, the system detects that all unit symbols conform to the International System of Units (SI), and no conversion is triggered. Finally, a lightweight initial process knowledge graph skeleton containing four process variable nodes and two directed causal edges with complete attribute labels is formed in the graph database, supporting subsequent dynamic causal discovery processes. S1.5: Perform consistency verification on the completed initial process knowledge graph skeleton, use the domain knowledge mask matrix to detect logical conflicts in potential reverse causal paths (such as 'product color → combustion temperature'), and remove abnormal connections that violate the principles of energy conservation or irreversible processes to ensure that the graph meets the physical realizability requirements of industrial thermal systems.

[0012] Step S2: The real-time acquired multidimensional process parameter sequences and corresponding batch product quality inspection results are time-series aligned to generate a timestamped observation dataset, which drives the subsequent causal structure learning process. Specifically, this includes: S2.1: Obtain real-time stream of multi-dimensional process parameters from the distributed sensor network of the environmentally friendly brick firing kiln. The multi-dimensional process parameters include temperature of each section, oxygen content of flue gas, fuel flow rate, exhaust pressure and negative pressure of the furnace. Based on a high-precision clock synchronization mechanism, timestamp the data of each channel to generate the original time-series data sequence to eliminate the time-series deviation caused by asynchronous sampling. For the real-time data acquisition environment of the distributed sensor network of environmental protection brick firing kiln, multi-dimensional process parameter signals are selected as input objects, including temperature of each section, flue gas oxygen content, fuel flow rate, exhaust pressure and furnace negative pressure. A multi-channel high-speed sampling method (parameters: sampling frequency ≥10Hz, sampling accuracy ±0.1%FS) is adopted to realize the real-time acquisition of physical quantities of distributed sensing nodes, and the analog quantities output by each channel are converted into digital signals by an industrial-grade A / D conversion module to form the original sampling data stream; Furthermore, through a high-precision clock synchronization mechanism (parameters: PTP network timing protocol accuracy ≤1μs, internal RTC drift of temperature sensor ≤0.5ppm), the sampling time of all distributed sensors is uniformly calibrated, and a timestamp strictly aligned with the system master clock is added to the digital sampling record to obtain a multi-channel data packet sequence with millisecond-level time stamp; Furthermore, a sampling queue buffering and timestamp correction algorithm is adopted (parameters: buffer length is twice the window period T, correction threshold). =±2ms), the data stream of each channel is buffered in a first-in-first-out manner, and the data insertion position is adjusted according to the timestamp difference to achieve sampling timing alignment across sensor channels and reduce data offset caused by asynchronous sampling; Furthermore, the inter-channel time drift estimation formula is used.

[0013] in This is the average drift value. This is the sampling timestamp for the i-th channel. For reference, use the same timestamp as the channel. To determine the number of sampling points, calculate the relative drift of each channel and apply linear interpolation correction to ensure that the synchronization of cross-channel events meets the set accuracy. By combining high-precision clock synchronization and time drift correction, the original multidimensional sampling data from the previous step is transformed into an original time series data sequence with timestamp consistency better than 1ms, achieving the expected technical effect of eliminating asynchronous sampling and unifying the time base of observation data. For example, in a distributed network comprising eight zone temperature sensors, two flue gas oxygen analyzers, one fuel flow meter, one exhaust pressure transmitter, and one furnace negative pressure sensor, the sampling frequency for each temperature channel is set to 20Hz, the oxygen analyzer sampling period is 0.5s, and the flow meter and pressure sensor sampling frequency is 10Hz. All sensor nodes are connected to the main control system via Gigabit Ethernet, and IEEE 1588 PTP time synchronization is used to unify the timestamps of all channels to within 0.5μs of the master clock accuracy. During data acquisition, the detected timestamp deviations of each channel are calculated. The average drift value is approximately 0.8ms. After linear interpolation correction, the generated original time series data sequence has a timestamp consistency better than 1ms across all dimensions, ensuring that there is no cross-channel time series offset when the batch quality data and process parameters are initially mapped in S2.2, effectively supporting the accuracy requirements of causal structure learning. S2.2: Obtain the final product quality inspection data for the corresponding production batch. The quality inspection data includes compressive strength, porosity, linear shrinkage rate and appearance defect score. Based on the batch number and the furnace exit time stamp, establish an initial mapping relationship between the batch number and the aforementioned multidimensional process parameter sequence to form a cross-stage associated original paired dataset. S2.3: Based on the dynamic response characteristics of the thermal process, a time alignment window is constructed. For the response delay differences of different physical quantities (such as the τ-second lag between fuel adjustment and temperature change), the Dynamic Time Warping (DTW) algorithm is used to perform nonlinear time-series matching between the process parameter sequence and the quality detection results, and calculate the optimal alignment path to achieve accurate causal attribution between control actions and quality effects. S2.4: Perform data fusion operation on the paired datasets that have completed time-series matching, and perform spatiotemporal alignment and aggregation of the characteristic statistics (mean, slope, variance) of each process parameter at key process nodes (heating section, heat preservation section, cooling section) with the corresponding quality indicators to generate a structured, multi-dimensional observation dataset with unified timestamp labels, which serves as the input data matrix for the dynamic causal discovery algorithm. For cross-stage pairing datasets that have completed time-series matching, a process segment division method (parameters: heating segment, heat preservation segment, cooling segment) is adopted to realize the identification of key process nodes based on the production cycle and construct a segmented index structure for subsequent feature calculation. Based on the segmented index structure, a sliding window statistical algorithm (with the window length set according to the duration of the process segment) is used to calculate the mean characteristics of each process parameter in each process segment, thereby obtaining a central tendency data matrix that reflects the operating status within the segment. Furthermore, by using the least squares regression algorithm (parameters: the sequence of sampling points within the segment and the timestamp index), the slope of each process parameter within the process segment is estimated, and the rate of change within the segment is obtained to characterize the dynamic response characteristics. Furthermore, a variance calculation method (based on the set of sampled values ​​within the segment) is used to obtain the volatility measure of each process parameter within the process segment, forming a stability feature vector, which serves as an important input for quality effect analysis; A spatiotemporal alignment aggregation algorithm (parameters: unified timestamp label, quality indicator sequence) is used to map and match the above mean, slope and variance statistics with the product quality inspection data of the corresponding batch, and merge them into multi-dimensional observation samples on a unified time axis to achieve time synchronization between numerical features and quality indicators. By using feature fusion processing (fusion strategy: feature vector concatenation), the statistical features of each process segment and the corresponding quality indicators are jointly constructed into an input data matrix in a unified format, generating a structured, multi-dimensional observation dataset with a unified timestamp label, which serves as the input to the dynamic causal discovery algorithm. For example, in the data processing of a batch of environmentally friendly brick firing process, the moving average of the sampled temperature sequence during the heating phase was statistically analyzed using a window length of 300 seconds to obtain the average temperature corresponding to the average heating rate. Celsius; the slope is calculated using the least squares method, assuming the sampling timestamp vector is... Second and temperature sequences are The slope is calculated using the least squares regression algorithm based on the Celsius temperature, and the result is approximately... Temperature fluctuations per second (degrees Celsius); variance calculation yields temperature fluctuations of... Squared degrees Celsius. Compare the mean, slope, and variance of this segment with the compressive strength of the corresponding batch (e.g., ...). MPa), porosity (e.g.) Quality indicators such as linear shrinkage rate are timestamped and fused into an observation vector. Ultimately, a multidimensional observation dataset was formed, realizing the accurate spatiotemporal fusion of process segment features and quality indicators, providing high-confidence, structured input for dynamic causal structure learning; S2.5: Perform data quality verification on the generated observation dataset, calculate the time alignment confidence score between each variable, and if the score is lower than the set threshold, trigger the anomaly detection mechanism and start the local realignment process to ensure that the output observation dataset meets the requirements of subsequent causal structure learning for temporal consistency and data integrity.

[0014] like Figure 2 As shown, step S3 involves: based on the observed dataset and combined with the physical constraint priors in the initial process knowledge graph skeleton, executing a dynamic causal discovery algorithm to identify directed causal connections between key variables that significantly affect product quality, thus forming an updated causal structure model. Specifically, this includes: S3.1: Obtain a time-stamped observation dataset generated by aligning the multidimensional process parameter sequence with the corresponding batch product quality inspection results over time. The process parameters include the temperature of each section of the kiln, the oxygen content of the flue gas, the fuel flow rate, the exhaust pressure, and the proportion of combustion air. The product quality indicators include compressive strength, porosity, and geometric deformation. Perform data preprocessing on this observation dataset. Use the sliding window normalization method to standardize the variables of each dimension to eliminate the statistical bias caused by the difference in dimensions. Output a standardized time series tensor as the input data representation for dynamic causal structure learning. S3.2: Based on the thermodynamic law relationships inherent in the initial process knowledge graph skeleton defined in S1, construct a domain knowledge mask matrix. Set known irreversible physical processes (such as 'combustion temperature → flue gas composition') as allowed edges, and prohibit reverse causal assumptions that violate energy conservation or heat transfer directionality (such as 'product color → fuel calorific value'). Use this domain knowledge mask matrix to prune the candidate causal graph space, limit the search range in the dynamic causal discovery process, and output a set of feasible causal structures constrained by physical laws to avoid generating pseudo-causal connections that violate the operating mechanism of industrial thermal systems. S3.3: Parallel execution of the constraint-based PC algorithm and the score-based GES algorithm on the set of feasible causal structures: The PC algorithm eliminates conditionally independent variable pairs layer by layer based on the partial correlation test to identify undirected dependencies; the GES algorithm performs a greedy search on the directed acyclic graph using the Bayesian Information Criterion (BIC) scoring function to find the globally optimal causal ranking; the intermediate results of the two algorithms are fused, and a weighted voting mechanism is used to determine the preliminary causal adjacency matrix, where each directed edge represents a potential causal influence between variables, and its weight is quantified by the joint confidence score to form the preliminary causal structure hypothesis; Based on the set of feasible causal structures obtained after pruning in step S3.2. The constrained PC algorithm (parameters: significance level α=0.05, conditional independence test method is partial correlation test) is used to realize the recursive elimination process of conditional independence between variable pairs and output the undirected dependent skeleton structure that satisfies the constraints. Furthermore, the GES algorithm driven by Bayesian Information Criterion (BIC) scoring (parameter: the penalty term coefficient λ in the scoring function BIC is adaptively set according to the sample size N) is used to achieve a greedy search of the directed acyclic graph space, gradually adding or deleting edges to optimize the score value, and obtaining the optimal directed sorting structure; Furthermore, for the intermediate structures output by the PC algorithm and the GES algorithm respectively, a weighted voting mechanism based on weight vectors is adopted (parameter: weight coefficient). and Based on the information gain contribution, the candidate edge sets of the two are fused to generate a preliminary causal adjacency matrix containing directionality. ; Furthermore, in In the middle, for each directed edge Calculate the joint confidence score The following formula is used for calculation:

[0015] in, and These represent the saliency scores assigned to the edge by the PC algorithm and the GES algorithm, respectively. and These are the weighting coefficients; Through the above fusion processing method, the independent search results of PC and GES in the previous step are transformed into preliminary causal structure assumptions that take into account both constraint consistency and score optimality, thereby improving the robustness of multi-source decision-making with causal connections. For example, in a causal discovery task involving temperature control during the firing of environmentally friendly bricks, the input data includes the distribution of temperature, flue gas oxygen content, fuel flow rate, exhaust pressure, compressive strength, porosity, and other indicators at 12 temperature measurement points in the kiln. The sample period length T = 1800 seconds, the sampling step size Δt = 1 second, and the total sample size N = 1800. The PC algorithm selects a significance level of α = 0.05, uses the Pearson correlation coefficient for partial correlation testing, and calculates the p-value for conditional independence. After removing insignificant variable pairs, an undirected skeleton is output, containing approximately 14 edges. The GES algorithm initializes the feasible edge space of this skeleton as an empty graph and uses the BIC function to evaluate the structure score for each edge addition / deletion, where LL is the log-likelihood value, k is the model degrees of freedom parameter, and λ is set to 1. The search yields a directed acyclic graph containing 12 edges with clearly defined directions. The weighted fusion stage calculates the results based on information gain. =0.55、 =0.45, and the scores of both are calculated as described above. Formula fusion, output A matrix, where each edge is accompanied by a confidence score, for example, the edge 'heating rate → compressive strength'. =0.87, 'Oxygen content in flue gas → Porosity' edge =0.81. The output serves as the basic input for subsequent S3.4 significance verification and simplified causal graph construction. Verification shows that the fusion strategy improves the robustness of causal edge direction determination, while reducing the pseudo-association connections that are prone to occur in the case of single algorithm. S3.4: Perform significance verification on the preliminary causal adjacency matrix, calculate the p-value sequence of each causal edge and perform multiple hypothesis testing correction using the Benjamini-Hochberg procedure, eliminating non-significant connections with a false discovery rate > 0.05; retain the corrected causal edges to form a simplified causal graph, and perform topological comparison with the initial process knowledge graph skeleton to identify newly added or conflicting causal paths; for causal relationships with a confidence level exceeding the preset threshold θ = 0.85 and inconsistent with the original structure, mark them as dynamic causal terms to be verified, serving as candidate inputs to trigger graph evolution; S3.5: Based on the simplified causal graph and the dynamic causal terms to be verified, the causal stability is evaluated by combining the counterfactual inference results of historical high-return control trajectories. The final output is an updated causal structure model containing dynamic correction terms. This model retains physical prior constraints and integrates new causal relationships discovered by data-driven methods, serving as the basic input for state space reconstruction in the temperature control decision module. Based on simplified cause-effect graph With the dynamic causal terms to be verified The Structural Equation Modeling (SEM) algorithm (parameters: model coefficient matrix B, covariance matrix Σ) is used to perform analysis on each term. Counterfactual response prediction under the condition of fixed covariates under causal assumptions; Furthermore, by using the counterfactual intervention operation do(X=x') (parameters: the set value x' of the target variable X, and the covariate set C remaining unchanged), blocking... The information transmission path of the causal edge is specified in the model, and the corresponding path coefficient is adjusted to zero in the structural equation model to generate a control model that removes the causal effect. ; Furthermore, using a control model With the complete model Predict the response value of product quality indicator Y respectively. and Calculate the counterfactual difference index The formula is as follows:

[0016] in, For the predicted quality index value under counterfactual conditions, The predicted quality index value is the value under actual observation conditions; Furthermore, through the difference sign determination algorithm (parameter: Symbols and Symbol contrast, This represents the direction of change in product quality indicators under actual observation conditions (e.g., a positive sign indicates quality improvement, and a negative sign indicates quality decline), thus verifying the validity of the causal direction. Greater than If the signs are consistent, then confidence enhancement processing is performed, as shown in the following formula:

[0017] in, This is the confidence enhancement value. The original confidence level. This is the confidence enhancement coefficient. This is a function to find the maximum value. Through the aforementioned confidence enhancement process, the validated dynamic causal terms are... Integrate into a simplified cause-effect diagram The output includes an updated causal structure model with dynamic correction terms. This ensures that the model retains both physical prior constraints and new causal relationships discovered through data-driven methods, enabling precise reconstruction of the state space of the temperature control decision module. For example, during the operating cycle T=3600 seconds of the environmentally friendly brick firing kiln, the collected process parameters include a preheating zone temperature of 745K, an insulation zone temperature of 1350K, and a fuel flow rate of 0.85 m³ / s. The corresponding batch's observed compressive strength is 48 MPa. Assuming... One of the causal assumptions is "insulation zone temperature → compressive strength". The initial value of the path coefficient in the structural equation model is 0.62. A counterfactual intervention do (insulation zone temperature = 1300K) is performed to block this causal edge, generating a control model. Post-prediction compressive strength =44 MPa, corresponding to the predicted value of the complete model. =48 MPa, then the counterfactual difference =4, and with Same sign. The confidence enhancement value is calculated according to the formula. If the original value is 0.88, the enhanced value is 0.93, exceeding the dynamic threshold of 0.85. Therefore, the causal hypothesis is integrated into... It serves as an important input to the state space of the temperature control decision module in the next cycle, significantly improving the effectiveness and adaptability of rule generation.

[0018] like Figure 3 As shown, step S4 involves calculating the confidence score of each causal edge based on the updated causal structure model and the counterfactual inference results of historical high-return control trajectories. For causal paths with confidence scores exceeding a preset threshold and conflicting with the existing graph, parallel branches with context labels are created, generating a multi-version causal path set. Specifically, this includes: S4.1: Based on the set of directed causal connections in the updated causal structure model and the historical high-return control trajectory database, extract the key variable interaction sequences involved in the successful execution of strategies under similar working conditions, and use them as the benchmark input data for counterfactual reasoning to construct the data context for validating the causal edge. S4.2: For each causal edge, use the counterfactual inference algorithm based on structural equation model to simulate the 'intervention-outcome' comparison scenario under the condition of fixing other covariates: set the intervention condition to block the information transmission path of the causal edge, keep the rest of the structure unchanged, replay the corresponding historical trajectory and predict the product quality index, and calculate the deviation with the predicted value under the original complete model to obtain the functional contribution quantification value of the causal edge. The set of directed causal edges in the updated causal structure model The structural equation modeling (SEM) counterfactual inference algorithm (parameters: model coefficient matrix, covariance matrix, and disturbance term distribution) is used to realize intervention simulation under fixed covariate conditions. Furthermore, by setting intervention conditions on the causal graph, the target causal edge is... The information transmission path is zeroed out, while the remaining structural equations remain unchanged. This enables the solution of the structural equations for a single-sided chain break, and the predicted state vector of the model after intervention is obtained. Furthermore, based on the quality prediction sub-model of the structure equation model after intervention, the corresponding predicted values ​​of product quality indicators are obtained. And through the predicted values ​​of the original complete model. Perform difference calculation; Furthermore, the formula for calculating the absolute value of the deviation is adopted.

[0019] in The deviation value is used to measure the functional contribution of each causal edge to the target variable, and the quantitative result is output as the scoring input. By using a unilateral intervention simulation method based on structural equation modeling, the counterfactual inference results are transformed into causal side function contribution indicators, thereby achieving a quantitative assessment of the effectiveness of causal relationships. For example, in the operational dataset of environmentally friendly brick firing kilns, key causal edges are... The formula is set as "heating rate → compressive strength", and the structural equation coefficients are set as the direct coefficients of heating rate against compressive strength. The covariance of the heating rate with other covariance variables in the covariance matrix is: The disturbance term follows the mean. variance The normal distribution is followed. In the intervention simulation, the coefficient corresponding to the causal edge is set to zero, while keeping the other equations unchanged. The structural equations are solved using the temperature distribution, fuel flow rate, and exhaust pressure from the historical high-return trajectory as fixed covariates to obtain the predicted value of the compressive strength after intervention. MPa, compared with the predicted value of the uninterrupted model Deviation calculation is performed using MPa; The deviation value is given by the formula Calculated MPa, this value is within the allowable range of the compressive strength index and is considered to be of a significant contribution level, proving that the causal edge has a high strength in this working condition context, and is suitable to be marked as a high-confidence causal connection for rule generation; S4.3: Based on the functional contribution metric, statistical significance p-value (from the permutation test), and physical interpretability score (determined by the domain knowledge mask matching degree), a weighted fusion function is used to calculate the comprehensive confidence score of each causal edge. The weight coefficients are determined through offline calibration experiments to ensure that the scoring system takes into account both the strength of data-driven approaches and the consistency of mechanisms. Based on the functional contribution metric, statistical significance test results, and physical interpretability score, a weighted fusion method is used to calculate the comprehensive confidence score of each causal edge. Functional contribution metric method (parameter: (obtained through counterfactual deduction) to analyze the effectiveness of causal edges in product quality changes and generate a numerical contribution index; Furthermore, through the permutation test method (parameter: number of permutations) =1000, significance level α=0.05), to realize the statistical significance verification of the functional contribution, and output the p-value sequence of each causal edge; Furthermore, by employing a physical interpretability scoring method (parameters: domain knowledge mask matrix M and causal direction matching degree s of causal edges), a quantitative assessment of the consistency between causal edges and thermal laws is achieved, resulting in a score. ; Furthermore, through a weighted fusion function Calculate the overall confidence score The fusion function can be expressed as:

[0020] Among them, the weighting coefficient , , Offline calibration experiments were conducted to ensure that the fusion results balance the strength of data-driven approaches with consistency with the underlying mechanisms. The calculated comprehensive confidence score transforms counterfactual inference results, statistical validity, and physical consistency information into data indicators that can be directly used for causal conflict analysis, thereby achieving a quantitative expression of the credibility of the causal structure. For example, under the firing conditions of a certain batch of environmentally friendly bricks, the functional contribution ΔY is obtained by counterfactual simulation as follows: Number of substitution tests The significance p-value was obtained this time. Physical interpretability score Calculated from the matching degree of the domain knowledge mask matrix In the offline calibration experiment, the weighting coefficient was set to... , , Substitute the values ​​into the fusion function to calculate the overall confidence score: The calculation result is Its value is significantly higher than the dynamic threshold set in subsequent steps, thus confirming the high credibility of the causal edge and supporting the priority retention decision in graph conflict resolution. S4.4: Compare the comprehensive confidence score with the preset dynamic threshold (adaptively adjusted based on the 3σ principle of the score distribution within the sliding window), identify the set of newly added causal edges with high confidence and the set of existing edges with low confidence, and activate the graph conflict resolution mechanism for causal path pairs that have directional or existence conflicts between the two. According to the updated causal structure model Compared with the comprehensive confidence score dataset The sliding window statistical analysis method was adopted (parameter: window length). The scoring distribution is monitored in real time (depending on the number of recent production cycles) to provide the basic data support for threshold adjustment. Furthermore, an adaptive threshold calculation algorithm based on the 3σ principle is used (formula parameter: mean score). Standard deviation of rating The dynamic threshold θ is determined using the following formula: ,in Within the sliding window The mean, The corresponding standard deviation ensures that the threshold reflects the fluctuation characteristics of the current rating distribution; Furthermore, the comparison operation module is used to score the overall confidence level of each causal edge. A bidirectional comparison is performed with the dynamic threshold θ to extract a new set of causal edge indices with high confidence scores greater than or equal to θ. Extract the set of original causal edge indices with low confidence values ​​less than θ. This enables the output of edge sets categorized by trust level. Furthermore, through set intersection and mapping matching methods, and The causal edges in the algorithm are matched one by one according to the source node, destination node, and directionality to identify causal path pairs that differ in topological structure or causal direction. This provides input candidates for conflict resolution; The path conflict feature analysis algorithm is used to analyze the matched conflict causal path pairs. Calculate the directional consistency coefficient and structural difference score, and filter out those scores that exceed the internal conflict judgment threshold. The path pairs are entered into the graph conflict resolution mechanism call queue; Through the above threshold evaluation and conflict determination algorithm, the scoring results of the previous step are transformed into structured classification data of the newly added causal set with high confidence and the original set with low confidence, so as to realize the automatic triggering effect of causal edge conflict identification and graph evolution. For example, in the dynamic causal update process of a certain batch of environmentally friendly brick firing kiln, the sliding window length Set to 20 batches, statistics are obtained. The mean μ = 0.82, and the standard deviation σ = 0.045. According to the formula... The calculated dynamic threshold θ = 0.955. Compare all causal edges. Six new causal edges with a score ≥ 0.955 were extracted and entered into the database. The set is used to extract 4 existing causal edges with a score < 0.955. The set was analyzed. Three pairs of nodes were found to have directional conflicts through source-target node mapping, and their difference scores were calculated. All scores were higher than [previous values]. =0.6, indicating a conflict path pair requiring parsing and triggering the creation of subsequent parallel branches. Execution results show that this process can significantly improve the adaptive evolution capability of the causal structure when the raw material ratio changes from low-density clay to high-density clay, providing high-quality conflict identification input for the generation of multi-version causal path sets; S4.5: For the identified conflicting causal path pairs, instead of directly overwriting the original path, a parallel causal branch with context labels is created based on its corresponding working condition labels (such as raw material ratio range, ambient humidity range), timestamp and confidence score. This generates a set containing multiple versions of causal paths, with each path accompanied by an applicable boundary condition description, forming a dynamic knowledge evolution structure that supports context awareness.

[0021] Step S5: Encode the multi-version causal path set into the state space of a Markov decision process, define the action space by combining adjustable control parameters, construct a reward function with the prediction of product quality score as the primary term and control stability as the secondary term, and form an interpretable reinforcement learning training environment. Specifically, this includes: S5.1: Based on each causal branch and its context label in the multi-version causal path set, a graph embedding algorithm is used to perform low-dimensional vector representation of each node (process variable) and directed edge (causal relationship), generating a causal graph embedding vector set containing semantic and structural information, which serves as the initial representation basis of the state space of the Markov decision process. S5.2: Perform temporal context weighting on the causal graph embedding vector set, combine the timestamp of the current batch production stage and the process segment identifier, calculate the activation weight of each causal path under the current working condition, and generate a dynamically weighted state representation tensor to reflect the selective activation mechanism of multiple versions of paths in the knowledge graph under different operating conditions. S5.3: Based on the adjustable set of control parameters in the environmentally friendly brick firing kiln temperature control system (including target temperature setpoints, heating and cooling slopes, heat preservation time, and fuel ratio coefficients for each section), a discretized or continuous parameter combination space is generated and mapped to the action space A of reinforcement learning, ensuring that each action corresponds to a set of feasible real-time control command sequences. Based on the adjustable control parameter set of the environmentally friendly brick firing kiln temperature control system, the parameter domain analysis method is adopted (input: target temperature setpoint for each section, heating and cooling slope, heat preservation time, fuel ratio coefficient) to map physical control variables into a set of operable numerical intervals, and establish upper and lower limit constraint matrices for each parameter to ensure the feasibility and safety of parameter combination. Furthermore, through a parameter discretization algorithm (mode: piecewise linear hierarchical, step size set according to process allowable error), the discretization representation of each continuous control parameter is realized, and an independent hierarchical strategy is adopted for the temperature setpoint of different sections to generate a multi-dimensional parameter discretization grid to form an enumerable parameter combination candidate set. Furthermore, by using a parameter continuous interpolation algorithm (method: spline interpolation, constraint condition: thermal stress change rate does not exceed a set threshold), a continuous transition between discrete combinations is achieved, and a control curve spline coefficient matrix is ​​obtained for dynamic expansion of the action space, ensuring the smoothness of reinforcement learning in the action generation process; Furthermore, a control command mapping function is used to convert the discretized and continuous parameter combinations into a real-time control command sequence required by the kiln execution module. The command content includes temperature zone setpoint adjustment command, fuel ratio adjustment command and heat preservation time control command, ensuring that each action in the action space corresponds to a set of feasible and directly issued execution commands. Furthermore, an action legality verification algorithm (based on safety constraint verification and process rule matching) is used to filter out parameter combinations that are physically unrealizable or exceed safety boundaries, and outputs a set of legal action indexes. This forms the final reinforcement learning action space A; Through the above parameter analysis, discretization, continuousization and instruction mapping processing methods, the set of adjustable control parameters in the previous step is transformed into an action space that is feasible to execute and supports mixed discrete and continuous control, so as to realize the comprehensive coverage and high-precision simulation of the control strategy in the decision training environment. For example, in the environmentally friendly brick firing scenario, the set of adjustable control parameters includes a preheating section target temperature setpoint range of 120℃ to 180℃, a heating rate range of 2℃ / min to 5℃ / min, a holding time of 4h to 8h, and a fuel ratio coefficient of 0.85 to 1.15. A piecewise linear hierarchical method is used, discretizing the preheating section target temperature into 7 levels with a step size of 10℃, the heating rate into 7 levels with a step size of 0.5℃ / min, the holding time into 5 levels with a step size of 1h, and the fuel ratio coefficient into 7 levels with a step size of 0.05, forming a 7×7×5×7 parameter combination grid. Spline interpolation is performed on the combination of heating rate and holding time, setting the thermal stress change rate threshold to 3MPa / h to generate a smooth transition curve between combinations. By executing instruction mapping, the parameter combinations are converted into Modbus TCP instruction frames, including commands to write to the target temperature register, adjust the fuel valve opening, and count the holding time. Using safety constraint checks, actions that cause the furnace pressure to exceed the set limit due to the combination of the heating slope and fuel ratio coefficient are eliminated, resulting in a set of legal action indexes with an action space size of 1120. In reinforcement learning training, each index in the action space corresponds to a complete sequence of control instructions, and the system can call and verify the quality improvement effect and stability performance of the action in a simulation environment; S5.4: The quality prediction model trained using historical process data outputs the expected product quality score after each action is performed in the current state. The L2 norm of the rate of change of the control action is combined as a smoothness penalty term to construct a composite reward function to balance the optimization goals of product quality improvement and control stability. S5.5: The dynamically weighted state representation tensor, the defined action space A, and the composite reward function R are integrated into a unified reinforcement learning environment interface, which is encapsulated into an interpretable reinforcement learning training environment with state transition simulation capabilities. This supports policy networks to perform multi-round trial-and-error learning and attention-guided decision tracing in simulation mode.

[0022] Step S6: In the interpretable reinforcement learning training environment, an attention-enhanced Transformer architecture policy network is used for action selection. While outputting control parameter combinations, an attention weight distribution is generated. This attention weight distribution is mapped to knowledge graph nodes to form control suggestions with logical interpretability. Specifically, this includes: S6.1: The state encoding vector constructed based on the multi-version causal path set initializes the input embedding sequence of the Transformer architecture. The state encoding vector is formed by jointly embedding the source node, causal edge type and target node in each causal path into a fixed-dimensional vector representation to form a structured state input that the policy network can process, and the output is a standardized state embedding matrix. Based on nodes and their context labels in a multi-version causal path set, a node-edge joint embedding method is adopted (parameter: node semantic vector dimension). =64, relation type vector dimension =32), concatenate the source node, causal edge type and target node in sequence and map them to a fixed-dimensional vector through linear transformation to realize the structured encoding of a single causal path; Furthermore, by using a path-level position coding algorithm (parameter: coding period P=100, based on sine / cosine function), the topological position and path length information of the causal path in the graph are embedded into the above structured coding, so that the embedding vector retains the relative position information of the causal chain and the path position enhancement coding matrix is ​​obtained. Furthermore, a sequence construction algorithm is used (parameter: number of multi-version paths). The sequence length L is determined by (Automatically determined), the enhanced coding matrices of each path position are sorted by version label and combined into an embedding sequence of the Transformer architecture input, ensuring that the semantic differences of different version paths are reflected in the sequence order; Furthermore, an embedding normalization method is employed (parameters: mean μ and standard deviation σ are calculated batch-wise) to normalize each dimension of the embedding sequence, thereby achieving consistency in the embedding value range across different batches. The formula is as follows:

[0023] in, The embedded value after normalization transformation. For the original embedded value, The mean of this dimension. This represents the standard deviation of this dimension. Through the above normalization process, the normalized embedding sequence is transformed into a standardized state embedding matrix, thereby achieving the expected technical effect of structural consistency and numerical stability of the policy network input. For example, in batch B123 of the environmentally friendly brick firing kiln temperature control system, the multi-version causal path set contains the number of paths. =5, node semantic vector dimension configured as follows =64, relation type vector dimension configured as follows =32. The source node is "heating rate", the target node is "compressive strength", and the causal edge type is "promoting" relationship. A node-edge joint embedding method is used, mapping the three parts to 64, 32, and 64-dimensional vectors respectively, and then concatenating them into a 160-dimensional path embedding vector. The path position encoding period P is set to 100. Sine / cosine position encoding is performed on the causal chain with a path length of 3, and the results are added element-wise to the embedding vector to obtain the position enhancement encoding matrix. The five position enhancement encoding matrices are sorted from oldest to newest according to the version label, generating an embedding sequence of length L=5. For each dimension of the embedding sequence in batch B123, the mean μ and standard deviation σ are calculated within the batch. The mean is 0.05, and the standard deviation is 0.12, according to the formula... Perform normalization to compress the range of all embedded values ​​to [ The matrix near [1,1] is ultimately a normalized 160×5 matrix, which serves as the input to the Transformer policy network. This implementation significantly improves the stability of cross-path semantic matching in subsequent attention mechanism operations and reduces the interference of numerical range differences between different batches on the decision-making effect. S6.2: Perform multi-head self-attention calculation on the standardized state embedding matrix, extract key dependencies across causal paths using query-key-value (QKV) transformation mechanism, calculate the dynamic association strength between knowledge graph nodes, and generate an attention weight distribution graph with context awareness as a visual intermediate representation for decision-making. S6.3: Based on the attention weight distribution map, action decoding is performed in combination with the adjustable control parameter combination space. A probability distribution action selection vector is generated through the position feedforward network and the softmax output layer. The optimal control parameter combination is sampled from it, including the target temperature setpoint, heating and cooling slope and heat preservation time of each section. The output is a candidate set of specific executable control instructions. S6.4: The attention weight distribution map is reverse-mapped to the nodes and causal edges of the original process knowledge graph to identify the key variable path that contributes the most to the current action selection, and generate control explanation statements in natural language form, such as 'It is recommended to increase the target temperature of the preheating zone to improve the compressive strength'. The output is a logically interpretable control suggestion with causal traceability tagging. S6.5: Perform consistency verification on the control suggestions and their corresponding control instruction candidate sets, verify the safety of their operation boundaries based on the historical high-return trajectory database, and determine whether they conflict with the prior physical constraints. If the verification is passed, they are marked as reliable recommendation results for subsequent rule transformation processes.

[0024] Step S7: Determine whether the expected quality improvement effect corresponding to the control suggestion meets the set threshold. If it does, convert it into a candidate control rule with confidence level labeling and push it to the manual review queue for verification and confirmation. Specifically, this includes: S7.1: Based on the control proposal output by the interpretable reinforcement learning strategy network and its corresponding attention weight distribution, obtain the set of knowledge graph nodes and associated causal edges that the proposal is applied to. Using the causal strength parameters and timestamp labels stored in the multi-version causal path set constructed in the previous steps, calculate the expected product quality gain ΔQ of the control action in the current process context, as the input basis for subsequent threshold judgment. Based on the control suggestions output by the interpretable reinforcement learning policy network and their corresponding attention weight distribution, a node mapping parsing method (parameters: node ID set, weight threshold) is used to select the set of knowledge graph nodes and their associated causal edges that have the greatest impact on the control suggestions from the attention weight distribution; Furthermore, by using the causal path index retrieval method (parameters: node set, associated causal edge), the selected node set is matched with the multi-version causal path set constructed in the previous steps, and the corresponding causal strength parameters and timestamp labels are extracted to obtain the path subset bound to the context information. Furthermore, by using a process context feature fusion method (parameters: causal strength parameter, operating condition label vector), the dynamic causal edge strength is weighted and combined with the current process environment parameters to generate a process state-sensitive control action feature vector. Furthermore, a product quality gain prediction algorithm (parameters: control action feature vector, quality influence model coefficients) is employed to predict the expected product quality gain of the current control action. Quantitative calculations are performed, and the prediction model is constructed based on a linear combination and interaction terms of the structural equation model. Its calculation formula is as follows:

[0025] in, The weight coefficient for the i-th causal feature is... This is the output value of the quality response function corresponding to the i-th feature; Furthermore, by using a time-weighted smoothing method (parameters: ΔQ, timestamp label), the instantaneous predicted gain value is smoothly calculated within a time window to eliminate short-term fluctuations and obtain an expected quality gain index with enhanced stability. Through the combined algorithm of causal analysis, path retrieval, feature fusion, gain calculation and smoothing, the control suggestions from the previous step are transformed into expected product quality gains with clear quantitative indicators. This provides a reliable input basis for subsequent threshold determination; For example, in the preheating zone temperature control optimization scenario of a certain environmentally friendly brick firing kiln, the policy network outputs a control suggestion to increase the target temperature of the preheating zone to 850K. Attention weight distribution analysis reveals a set of nodes involved, including "preheating zone temperature distribution," "temperature rise slope," and the associated causal edge "temperature rise slope → compressive strength." The current strength parameter of this causal edge in the multi-version causal path set is 0.72, and the timestamp label corresponds to the current batch raw material moisture content of 3.5%. The path subset obtained through causal path indexing contains three strongly correlated edges with strength vectors {0.72, 0.65, 0.58}. These are weighted and combined with process state characteristics (raw material moisture content 3.5%, flue gas oxygen content 8%), mapped to the quality impact model coefficient set {1.2, 0.9, 1.0}, and substituted into the formula:

[0026] get The expected quality gain index value. After smoothing by sliding time window to remove the influence of batch-to-batch fluctuations, the smoothed ΔQ index is 2.01, which is significantly higher than the baseline value of 1.45 under similar operating conditions without optimization strategy, demonstrating that this control action has a significant effect on improving the compressive strength of the product in the current process context; S7.2: Based on the actual quality feedback data under the same or similar working conditions in historical batches, the expected product quality gain calculated in S7.1 is corrected for deviation. The weighted moving average method is used to fuse the actual values ​​of n recent similar trajectories to generate a corrected expected quality improvement score, so as to reduce the uncertainty of single model prediction. S7.3: Compare the corrected expected quality improvement score with the preset quality improvement threshold. If the corrected expected quality improvement score is greater than or equal to the quality improvement threshold, the control suggestion is deemed to have the value of being converted into a rule and enters the rule encapsulation process; otherwise, the suggestion is discarded and recorded in the inefficient policy log library for subsequent negative sample training optimization of the policy network. S7.4: For control suggestions that pass the threshold screening, a structured candidate control rule template is generated by combining the multi-version causal path context labels of its source, the counterfactual inference confidence score, and the combination of control parameters to be adjusted for this action. The template contains a five-tuple of 'trigger condition-causal path ID-action instruction-expected effect-quality improvement score', and adds a timestamp and batch identifier to form a complete rule instance. S7.5: Inject the generated candidate control rule instances with confidence labels into the manual review queue management system. Present their causal logic links and interpretable explanatory text through a visual interface for process experts to conduct semantic consistency and engineering feasibility reviews. This completes the key transition from automated generation to human-machine collaborative confirmation, ensuring that the new rules meet production safety and process specification requirements.

[0027] Step S8: Inject the approved candidate control rules into the temperature control execution module, and simultaneously send the actual production feedback data after rule application back to the observation dataset for the next round of joint optimization iteration of the causal model and policy network. Specifically, this includes: S8.1: Obtain candidate control rules that have passed the manual review queue. The candidate control rules include combinations of control parameters with confidence labels and their corresponding multi-version causal path context labels, which serve as input conditions for rule injection. Sort the candidate rules based on the rule priority scheduling mechanism to determine their effective order in the temperature control execution module. S8.2: Map the sorted candidate control rules to the setpoint adjustment interface of the temperature control execution module to generate the target temperature curve, heating / cooling slope command and heat preservation time parameter package for the corresponding kiln zone; use the industrial communication protocol Modbus TCP to encapsulate and transmit the parameter package to complete the dynamic injection of control rules into the execution layer; S8.3: During the rule execution process, multi-point temperature monitoring sequences, fuel flow sensor data, and exhaust pressure feedback signals from the temperature acquisition module are collected in real time to form a process operation data stream during rule application; the process operation data stream is timestamped and outlier filtered to generate structured actual production feedback data; S8.4: Perform cross-domain correlation matching between actual production feedback data and the product quality inspection results of the corresponding batch to generate a closed-loop feedback sample with a rule source identifier; calculate the actual quality improvement gain and stability index of the control rule based on the closed-loop feedback sample, as a quantitative output for evaluating the effectiveness of the rule; S8.5: The closed-loop feedback sample and its evaluation results are fed back and appended to the global observation dataset to update the input of the dynamic causal discovery algorithm for the next cycle; at the same time, the multi-version causal path branch activated by the sample is marked as 'verified' to support the incremental recalculation of the causal edge confidence score, and realize the joint online optimization of knowledge graph and policy network.

[0028] The present invention also provides an automatic temperature control system for environmentally friendly brick firing kilns, which uses the above-mentioned automatic temperature control method for environmentally friendly brick firing kilns to achieve automatic temperature control of the kilns.

[0029] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0030] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. An automatic temperature control method for an environmentally friendly brick firing kiln, characterized in that, Includes the following steps: S1: Based on the physical and thermal laws of environmentally friendly brick firing kilns, construct the initial process knowledge graph framework; S2: Perform time-series alignment processing on the real-time acquired multidimensional process parameter sequences and the corresponding batch product quality inspection results to generate an observation dataset with timestamps; S3: Based on the observed dataset and combined with the physical constraint priors in the initial process knowledge graph skeleton, execute the dynamic causal discovery algorithm to identify the directed causal connections between key variables that significantly affect product quality, and form an updated causal structure model. S4: Based on the updated causal structure model and the counterfactual inference results of historical high-return control trajectories, calculate the confidence score of each causal edge, and create parallel branches with context labels for causal paths whose confidence exceeds a preset threshold and conflict with the existing graph, generating a set of multiple versions of causal paths. S5: Encode the set of multi-version causal paths into the state space of a Markov decision process, define the action space by combining adjustable control parameters, construct a reward function with the prediction of product quality score as the main term and control stability as the secondary term, and form an interpretable reinforcement learning training environment. S6: In the interpretable reinforcement learning training environment, the Transformer architecture policy network enhanced with attention mechanism is used to select actions, output control parameter combinations and generate attention weight distributions, and then the attention weight distributions are mapped to knowledge graph nodes to form control suggestions.

2. The automatic temperature control method for an environmentally friendly brick firing kiln according to claim 1, characterized in that, Following step S6, the following is also included: S7: Determine whether the expected quality improvement effect corresponding to the control suggestion meets the set threshold. If it does, convert it into a candidate control rule with confidence label and push it to the manual review queue for verification and confirmation. S8: Inject the approved candidate control rules into the temperature control execution module, and at the same time, send the actual production feedback data after the rules are applied back to the observation dataset for the next round of joint optimization iteration of the causal model and policy network.

3. The automatic temperature control method for an environmentally friendly brick firing kiln according to claim 1, characterized in that, Step S1 specifically includes: Based on the thermal mechanism analysis of the environmentally friendly brick firing process, a set of key physical variables affecting product quality is identified, which serves as the basis for defining nodes in the initial process knowledge graph. Based on the laws of classical thermodynamics and the principles of heat transfer, a deterministic functional relationship between key variables is established, forming a candidate set of directed causal edges; The deterministic functional relationship is symbolically abstracted and converted into a knowledge representation in the form of triples; The symbolized triple semantic units are imported into the graph database to construct the initial process knowledge graph skeleton. A consistency check is performed on the completed initial process knowledge graph skeleton. The domain knowledge mask matrix is ​​used to detect logical conflicts in potential reverse causal paths and to remove abnormal connections that violate the principles of energy conservation or irreversible processes.

4. The automatic temperature control method for an environmentally friendly brick firing kiln according to claim 3, characterized in that, Each node in the initial process knowledge graph skeleton is named with standardized terminology and assigned a unit attribute. Each causal edge is labeled with directionality, action type, and theoretical source tag. The action type includes promotion or inhibition.

5. The automatic temperature control method for an environmentally friendly brick firing kiln according to claim 1, characterized in that, Step S2 specifically includes: The system acquires real-time streams of multi-dimensional process parameters from a distributed sensor network of an environmentally friendly brick firing kiln, and timestamps the data from each channel based on a high-precision clock synchronization mechanism to generate the original time-series data sequence. Obtain the final product quality inspection data of the corresponding production batch, and establish an initial mapping relationship between the batch number and the exit time stamp and the original time series data sequence to form an original paired dataset that is associated across stages; A time alignment window is constructed based on the dynamic response characteristics of the thermal process. To address the differences in response delays of different physical quantities, the Dynamic Time Warping (DTW) algorithm is used to perform nonlinear time-series matching between the original time-series data sequence and the quality detection results, and to calculate the optimal alignment path. Perform data fusion on the paired datasets that have completed time-series matching, and aggregate the feature statistics of each process parameter at key process nodes with the corresponding quality indicators in a spatiotemporal alignment to generate a multidimensional observation dataset. The data quality of the multidimensional observation dataset is verified, and the time alignment confidence score between each variable is calculated. If the score is lower than the set threshold, the anomaly detection mechanism is triggered and the local realignment process is started.

6. The automatic temperature control method for an environmentally friendly brick firing kiln according to claim 1, characterized in that, Step S3 specifically includes: A multidimensional observation dataset is obtained, and the dataset is preprocessed. The sliding window normalization method is used to standardize the scale of each dimension variable, and the standardized time series tensor is output. Based on the thermodynamic law relationships inherent in the defined initial process knowledge graph skeleton, a domain knowledge mask matrix is ​​constructed. The domain knowledge mask matrix is ​​then used to prune the candidate causal graph space, outputting a set of feasible causal structures. On the set of feasible causal structures, the constraint-based PC algorithm and the score-based GES algorithm are executed in parallel. The intermediate results output by the PC algorithm and the GES algorithm are fused, and a weighted voting mechanism is used to determine the preliminary causal adjacency matrix. The significance of the preliminary causal adjacency matrix is ​​verified, the p-value sequence of each causal edge is calculated and multiple hypothesis testing is performed to correct it, a simplified causal graph is constructed, and it is compared with the initial process knowledge graph skeleton to identify new or conflicting causal paths. Causal relationships with confidence exceeding a preset threshold and inconsistent with the original structure are marked as dynamic causal items to be verified. Based on the simplified causal graph and the dynamic causal terms to be verified, the causal stability is evaluated by combining the counterfactual inference results of historical high-return control trajectories, and the updated causal structure model containing dynamic correction terms is output.

7. The automatic temperature control method for an environmentally friendly brick firing kiln according to claim 6, characterized in that, The PC algorithm identifies undirected dependencies by eliminating conditionally independent variable pairs layer by layer based on the partial correlation test; the GES algorithm performs a greedy search on the directed acyclic graph using the Bayesian information criterion scoring function to find the globally optimal causal ranking.

8. The automatic temperature control method for an environmentally friendly brick firing kiln according to claim 1, characterized in that, Step S4 further includes fusing historical high-return trajectories, intervening in the simulation through structural equation modeling, quantifying the functional contribution of each causal edge to the product quality indicators, combining statistical significance and mask physical consistency scores, and calculating the comprehensive confidence level through weighted fusion. When a new causal path that exceeds the dynamic threshold conflicts with the original path, context labeling branches are used to retain multiple versions of the causal path.

9. The automatic temperature control method for an environmentally friendly brick firing kiln according to claim 1, characterized in that, Step S5 specifically includes: Based on each causal branch and its context label in the multi-version causal path set, a graph embedding algorithm is used to generate a causal graph embedding vector set by representing each node and directed edge in low dimension. The causal graph embedding vector set is subjected to temporal context weighting processing. Combined with the timestamp of the current batch production stage and the process segment identifier, the activation weight of each causal path under the current working condition is calculated to generate a dynamically weighted state representation tensor. Based on the adjustable set of control parameters in the temperature control system of environmentally friendly brick firing kiln, a discretized or continuous parameter combination space is generated and mapped to the action space of reinforcement learning. The quality prediction model trained using historical process data outputs the expected product quality score after each action is performed in the current state, and combines the L2 norm of the rate of change of the control action as a smoothness penalty term to construct a composite reward function. The dynamically weighted state representation tensor, the action space, and the composite reward function are integrated into a unified reinforcement learning environment interface and encapsulated as an interpretable reinforcement learning training environment.

10. An automatic temperature control system for an environmentally friendly brick firing kiln, characterized in that: The automatic temperature control method for environmentally friendly brick firing kilns as described in any one of claims 1-9 is used for automatic temperature control of environmentally friendly brick firing kilns.

Images