AI-driven paperboard tension intelligent control method and system

Abstract

This invention relates to the field of paperboard tension control technology, and more particularly to an AI-driven intelligent tension control method and system for paperboard. By acquiring multimodal sensing data and performing cross-modal semantic alignment, a causal dependency graph is established to identify the transmission path of tension anomalies and map it into a structured knowledge representation. Based on this, a self-evolving knowledge graph is constructed, and counterfactual reasoning constraints are introduced to automatically deduce anomaly transmission patterns to dynamically expand entities and relationships. Then, multi-hop reasoning queries are executed to generate control strategies, and the feedback samples after execution are used to strengthen the counterfactual reasoning constraints to correct the graph topology. This method achieves intelligent identification and adaptive control of tension anomalies, improving the stability and quality of paperboard production.

Description

Technical Field

[0001] This invention relates to the field of paperboard tension control technology, and in particular to an AI-driven intelligent paperboard tension control method and system. Background Technology

[0002] In the paperboard production process, tension control is a critical factor affecting product quality and equipment stability. Current technologies generally employ PID control or fuzzy control algorithms based on physical models, monitoring a single physical quantity (such as tension sensor readings) and combining this with preset thresholds for feedback adjustment. Some advanced systems incorporate machine learning models (such as neural networks or support vector machines) to learn from historical data, predicting tension change trends and adjusting parameters in advance. These methods typically rely on fixed control logic or statically trained models, adjusting parameters during operation through manual experience or offline updates, failing to achieve real-time adaptive optimization.

[0003] Conventional approaches suffer from the following significant drawbacks. Firstly, existing control schemes struggle to effectively integrate multi-source heterogeneous data generated during production. For example, multimodal information such as tension timing signals, equipment vibration frequencies, temperature changes, and motor currents is often processed independently or simply superimposed, neglecting the potential semantic relationships and complementarities between different modal data. This leads to biased judgments about the causes of tension anomalies under complex operating conditions. Secondly, static knowledge bases or models cannot cope with continuously changing production conditions (such as material batch differences, equipment aging, and fluctuations in environmental temperature and humidity). When new types of anomalies emerge, the system cannot automatically identify and expand its knowledge structure, relying solely on manual investigation and retraining. This results in lagging control strategies, impacting production efficiency and product quality consistency. Therefore, a control method capable of dynamically integrating multimodal information, autonomously evolving a knowledge system, and achieving intelligent reasoning is urgently needed. Summary of the Invention

[0004] This invention provides an AI-driven intelligent control method and system for cardboard tension, which can solve the problems in the prior art.

[0005] A first aspect of the present invention provides an AI-driven intelligent control method for cardboard tension, comprising: Multimodal sensing data is acquired during the paperboard production process. Cross-modal semantic alignment is performed on the multimodal sensing data. By establishing a causal dependency graph between tension time-series signals and equipment operating status parameters, multi-level transmission paths of tension abnormal events are identified, and nodes in the multi-level transmission paths are mapped into structured knowledge representations. Based on the structured knowledge representation, a self-evolving knowledge graph is constructed. The self-evolving knowledge graph introduces counterfactual reasoning constraints, automatically infers potential abnormal transmission patterns when new production data is input, and dynamically expands the entity nodes and relation edges of the knowledge graph according to the inference results, so as to realize the continuous autonomous growth of the knowledge base. Multi-hop reasoning queries are performed based on the self-evolving knowledge graph to generate control strategies for the current tension state. The production data after execution is used as feedback samples to correct the topology of the knowledge graph by strengthening the counterfactual reasoning constraints.

[0006] Cross-modal semantic alignment is performed on the multimodal sensing data. By establishing a causal dependency graph between tension time-series signals and equipment operating state parameters, multi-level transmission paths of tension anomaly events are identified, and nodes in the multi-level transmission paths are mapped into structured knowledge representations, including: A cross-modal alignment space is constructed by projecting the time-frequency representation of the tension timing signal and the state evolution trajectory of the device operating state parameters into a unified semantic space, extracting timing synchronization features and state association features, establishing conditional dependencies based on the timing synchronization features and the state association features, and generating an initial causal dependency graph. A reverse causal verification mechanism is introduced into the initial causal dependency graph. By simulating intervention in the device operating state parameters and observing the response changes of the tension time series signal, false causal association edges are eliminated and causal transmission delays are marked to obtain a refined causal dependency graph. Based on the directed path search algorithm in the refined causal dependency graph, multi-level transmission paths from the device state disturbance source to the tension anomaly representation are identified. Semantic parsing is performed on each node in the multi-level transmission path, and the temporal synchronization features and state association features associated with the nodes are converted into structured knowledge representations that include causal role semantics, transmission mechanism descriptions and temporal constraints.

[0007] A reverse causal verification mechanism is introduced into the initial causal dependency graph. By simulating intervention in the device's operating state parameters and observing the response changes of the tension time-series signal, false causal association edges are eliminated and causal propagation delays are labeled, resulting in a refined causal dependency graph including: For each directed edge in the initial causal dependency graph, a counterfactual intervention experiment is constructed. In the counterfactual intervention experiment, a reverse perturbation is applied to the equipment operating state parameter corresponding to the starting node of the directed edge, and the dynamic response trajectory of the tension time series signal corresponding to the ending node of the directed edge is recorded synchronously. By analyzing the causal consistency between the change pattern of the dynamic response trajectory and the action pattern of the reverse perturbation, a causal validity score of the directed edge is generated. Directed edges whose causal validity scores are lower than a preset causal threshold are identified as false causal association edges and removed from the initial causal dependency graph. For directed edges whose causal validity scores meet the preset causal threshold, causal propagation delay is labeled based on the time interval between the peak response time of the dynamic response trajectory and the application time of the reverse perturbation. After completing the false edge removal and delay labeling, a refined causal dependency graph is obtained.

[0008] Based on the structured knowledge representation, a self-evolving knowledge graph is constructed. This self-evolving knowledge graph, by introducing counterfactual reasoning constraints, automatically infers potential anomaly propagation patterns when new production data is input, and dynamically expands the entity nodes and relation edges of the knowledge graph based on the inference results, achieving continuous autonomous growth of the knowledge base, including: The structured knowledge representation is converted into the initial semantic representation of the knowledge graph. A counterfactual reasoning constraint mechanism is introduced into the initial semantic representation. The counterfactual reasoning constraint mechanism applies hypothetical perturbations to entity nodes by constructing a virtual intervention scenario and simulates causal propagation along the relation edges. During the propagation process, the propagation probability and path branches are dynamically adjusted according to the statistical distribution of historical propagation patterns to generate an abnormal propagation hypothesis space containing multiple inference paths. When new production data is input, the abnormal pattern features in the new production data are extracted, and the abnormal pattern features are semantically matched with the inference paths in the abnormal propagation hypothesis space. The target inference paths with a matching degree exceeding a preset matching threshold are selected as the inference results. It is determined whether the entity nodes and relation edges in the inference results already exist in the current knowledge graph. If they do not exist, the entity nodes and relation edges are dynamically added to the knowledge graph to realize the continuous autonomous growth of the knowledge base.

[0009] The counterfactual reasoning constraint mechanism applies hypothetical perturbations to entity nodes by constructing a virtual intervention scenario and simulates causal propagation along relational edges. During propagation, it dynamically adjusts the propagation probability and path branches based on the statistical distribution of historical propagation patterns, generating an anomaly propagation hypothesis space containing multiple inference paths, including: A virtual intervention scenario is constructed for the target entity node in the knowledge graph. In the virtual intervention scenario, a hypothetical perturbation is applied to the target entity node. The hypothetical perturbation generates a counterfactual state by reversing the causal role semantics of the node. Based on the counterfactual state, a causal propagation simulation is initiated along the relation edge. In the propagation simulation, the propagation attenuation coefficient is calculated based on the causal propagation delay and propagation strength of the relation edge. The propagation attenuation coefficient is used to control the diffusion amplitude of the state change to adjacent nodes, forming a set of candidate propagation paths containing node sequences and edge sequences. The transmission pattern samples are extracted from historical production data and grouped and statistically analyzed according to node type to obtain the statistical distribution of historical transmission patterns. At each path branch point in the candidate propagation path set, the node type combination of the current branch is matched with the statistical distribution to calculate the propagation probability. Based on the propagation probability, the path branches are filtered, retaining those with a probability higher than a preset probability threshold and pruning those with a probability lower than the preset probability threshold. All the filtered propagation paths are summarized into an abnormal transmission hypothesis space.

[0010] Perform multi-hop reasoning queries based on the self-evolving knowledge graph to generate a control strategy for the current tension state, and use the executed production data as feedback samples. Strengthen the counterfactual reasoning constraints to correct the topological structure of the knowledge graph, including: The observation data of the current tension state is obtained and converted into query conditions. The starting node that semantically matches the query conditions is located in the self-evolutionary knowledge graph. Starting from the initial node, a multi-hop inference query is executed. During the inference process, a confidence propagation mechanism is constructed. According to the confidence propagation mechanism, the initial confidence of the initial node is attenuated and propagated along the relation edges. At each jump, the confidence attenuation coefficient is calculated based on the causal propagation delay of the relation edges and the historical verification success rate to form the global confidence of the inference path. Based on the global confidence, the optimal inference path is selected and its node sequence and edge sequence are extracted to generate a control strategy for the current tension state. The control strategy is issued and executed, and the production data after execution is collected as a feedback sample. The execution deviation is calculated by comparing the feedback sample with the expected effect of the optimal inference path. When the execution deviation exceeds a preset deviation threshold, a counterfactual verification matrix is ​​constructed. The counterfactual verification matrix performs a virtual deletion experiment on each relation edge in the optimal inference path and calculates the fit between the inference result after deletion and the feedback sample. The relation edge with the most significant improvement in fit is marked as a redundant edge, and the relation edge with the most significant decrease in fit is marked as a missing edge. The topology of the knowledge graph is corrected based on the marking results of the redundant edges and the missing edges.

[0011] The execution deviation is calculated by comparing the feedback sample with the expected effect of the optimal inference path. When the execution deviation exceeds a preset deviation threshold, a counterfactual verification matrix is ​​constructed. The counterfactual verification matrix is ​​calculated by performing a virtual deletion experiment on each relation edge in the optimal inference path and calculating the goodness of fit between the inference result after deletion and the feedback sample, including: Identify the adjustment parameters and expected response timing of each node from the expected effects of the node sequence association in the optimal inference path, and calculate the execution deviation after establishing a synchronous association between the expected response timing and the actual measurement data in the feedback sample based on the timestamp. When the execution deviation exceeds a preset deviation threshold, a counterfactual verification matrix is ​​constructed, and a virtual deletion experiment is performed for each relation edge in the optimal reasoning path. The virtual deletion experiment generates a restricted reasoning path by temporarily blocking the propagation path of the relation edge in the knowledge graph and re-executing the multi-hop reasoning query from the starting node. The expected effect corresponding to the restricted reasoning path is established as the reasoning result after deletion. Calculate the stage-by-stage deviation between the inference result after deletion and the feedback sample under the time-series stage division. Calculate the difference between the stage-by-stage deviation and the execution deviation to obtain the deviation improvement amount of edge deletion. Construct a fitting degree component based on the sign and magnitude of the deviation improvement amount and write it into the corresponding position of the counterfactual verification matrix. After iteratively processing all relation edges to complete the matrix construction, perform multi-weight fusion operation on each row to obtain the comprehensive fitting degree of each edge.

[0012] A second aspect of the present invention provides an AI-driven intelligent control system for cardboard tension, comprising: The knowledge mapping unit is used to acquire multimodal sensing data in the paperboard production process, perform cross-modal semantic alignment on the multimodal sensing data, identify multi-level transmission paths of tension abnormal events by establishing a causal dependency graph between tension time-series signals and equipment operating status parameters, and map the nodes in the multi-level transmission paths into structured knowledge representations. The self-evolving unit is used to construct a self-evolving knowledge graph based on the structured knowledge representation. The self-evolving knowledge graph introduces counterfactual reasoning constraints, automatically infers potential abnormal transmission patterns when new production data is input, and dynamically expands the entity nodes and relation edges of the knowledge graph according to the inference results, so as to realize the continuous autonomous growth of the knowledge base. The regulation feedback unit is used to perform multi-hop reasoning queries based on the self-evolving knowledge graph, generate regulation strategies for the current tension state, and use the production data after execution as feedback samples to correct the topology of the knowledge graph by strengthening the counterfactual reasoning constraints.

[0013] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0015] Cross-modal semantic alignment and causal dependency graph construction of multimodal sensing data significantly improve the accuracy of multi-dimensional signal anomaly identification in the paperboard production process. The multi-level transmission path of tension anomaly events is accurately located, and the fault propagation chain from the source equipment to the end paperboard can be quickly traced. The early warning time is greatly reduced, effectively reducing paper breakage rate and downtime losses.

[0016] After introducing counterfactual reasoning constraints, the self-evolving knowledge graph automatically generates potential abnormal pattern deductions when new production data is input. The entity nodes and relation edges of the knowledge base are dynamically expanded in an unsupervised manner. The system's accuracy in identifying composite fault types that appear for the first time is significantly improved. Rare fault cases can be accumulated autonomously without manual annotation, and the number of abnormal scenarios covered by the knowledge graph continues to grow rapidly.

[0017] The control strategy generated by multi-hop inference query can complete the tension state adjustment in a short time. The feedback sample continuously corrects the topology of the graph by strengthening the counterfactual constraints, so that the system can keep the tension fluctuation range stable within a small range under different speed, humidity and paper type switching. In long-term operation, the risk of model overfitting is greatly reduced, and the generalization ability of the control strategy can achieve seamless transfer across production lines. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the AI-driven intelligent control method for cardboard tension according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating the construction process of a self-evolving knowledge graph based on counterfactual reasoning, according to an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0021] Figure 1 This is a flowchart illustrating the AI-driven intelligent control method for cardboard tension according to an embodiment of the present invention.

[0022] AI-driven intelligent control methods for cardboard tension include: Multimodal sensing data is acquired during the paperboard production process. Cross-modal semantic alignment is performed on the multimodal sensing data. By establishing a causal dependency graph between tension time-series signals and equipment operating status parameters, multi-level transmission paths of tension abnormal events are identified, and nodes in the multi-level transmission paths are mapped into structured knowledge representations. Based on the structured knowledge representation, a self-evolving knowledge graph is constructed. The self-evolving knowledge graph introduces counterfactual reasoning constraints, automatically infers potential abnormal transmission patterns when new production data is input, and dynamically expands the entity nodes and relation edges of the knowledge graph according to the inference results, so as to realize the continuous autonomous growth of the knowledge base. Multi-hop reasoning queries are performed based on the self-evolving knowledge graph to generate control strategies for the current tension state. The production data after execution is used as feedback samples to correct the topology of the knowledge graph by strengthening the counterfactual reasoning constraints.

[0023] In one optional implementation, cross-modal semantic alignment is performed on the multimodal sensing data. By establishing a causal dependency graph between tension time-series signals and equipment operating state parameters, multi-level transmission paths of tension anomaly events are identified, and nodes in the multi-level transmission paths are mapped into structured knowledge representations, including: A cross-modal alignment space is constructed by projecting the time-frequency representation of the tension timing signal and the state evolution trajectory of the device operating state parameters into a unified semantic space, extracting timing synchronization features and state association features, establishing conditional dependencies based on the timing synchronization features and the state association features, and generating an initial causal dependency graph. A reverse causal verification mechanism is introduced into the initial causal dependency graph. By simulating intervention in the device operating state parameters and observing the response changes of the tension time series signal, false causal association edges are eliminated and causal transmission delays are marked to obtain a refined causal dependency graph. Based on the directed path search algorithm in the refined causal dependency graph, multi-level transmission paths from the device state disturbance source to the tension anomaly representation are identified. Semantic parsing is performed on each node in the multi-level transmission path, and the temporal synchronization features and state association features associated with the nodes are converted into structured knowledge representations that include causal role semantics, transmission mechanism descriptions and temporal constraints.

[0024] In the paperboard production process, multimodal sensing data originates from multiple heterogeneous sensing channels, including continuous time-series signals acquired by tension sensors, roll speed pulse sequences recorded by encoders, environmental state quantities output by temperature and humidity probes, and mechanical vibration waveforms captured by vibration accelerometers. These data differ significantly in sampling frequency, dimensional scale, and semantic granularity, making it impossible to directly establish cross-channel correlations. Therefore, it is necessary to construct a cross-modal alignment space, projecting the tension time-series signals and equipment operating state parameters, after feature transformation, onto a unified semantic space.

[0025] A short-time Fourier transform is performed on the tension time-series signal to extract the energy distribution and phase information of each frequency band, forming a time-frequency representation matrix. State evolution trajectories are constructed along the time axis for equipment operating state parameters (such as roller speed, tension roller displacement, steam pressure, etc.). The mean, variance, and rate of change of each state quantity are statistically analyzed using a sliding window to form a state evolution feature vector. The time-frequency representation matrix and the state evolution feature vector are mapped to a semantic embedding space of the same dimension through their respective encoding networks. A contrastive learning loss function is used to shorten the distance between time-stamp-aligned cross-modal sample pairs and widen the distance between unaligned sample pairs, thereby achieving cross-modal semantic alignment. After alignment, two types of core features are extracted from the joint embedding space: time-series synchronization features reflect the coordinated change pattern of tension signals and equipment states in the time dimension, and state association features describe the statistical co-occurrence relationship between specific equipment state combinations and tension amplitude ranges.

[0026] Based on temporal synchronization characteristics and state association characteristics, an initial causal dependency graph is established through conditional independence tests. Specifically, for each pair of variable nodes... and In a given set of conditions In the case of calculating conditional mutual information If the value exceeds the set threshold, the directed edge will be retained. Otherwise, delete. The direction of the directed edges in the initial causal dependency graph is determined by the temporal sequence and the domain prior. The node set covers tension sensor nodes, equipment status nodes, and intermediate process nodes (such as paper web tension status, estimated values ​​of the inter-roller friction coefficient, etc.). The causal dependency graph obtained at this stage contains spurious association edges introduced by common causes or confounding variables, and needs further refinement.

[0027] A reverse causal verification mechanism is introduced into the initial causal dependency graph to verify each candidate causal edge in the graph. Nodes are processed through do-calculus Applying simulated intervention, that is, forcibly adjusting the value distribution of other nodes without changing their natural value distribution. Set as several preset intervention values, observation nodes The corresponding tension timing signal response changes. If in multiple intervention experiments... If a significant monotonic or nonlinear dependency exists between the response change and the intervention value, the causal edge is confirmed as a true causal association; if there is no statistically significant association between the response change and the intervention value, the edge is marked as a spurious association and removed. For the retained true causal edges, the causal propagation delay is further labeled. This time delay is the time interval between the occurrence of a disturbance in the equipment state and the appearance of an observable response to the tension signal. This time delay is estimated using the peak position of the cross-correlation function. After the above elimination and labeling processes, a refined causal dependency graph is obtained, where each directed edge carries both a causal strength weight and a propagation delay attribute.

[0028] Based on a refined causal dependency graph, a directed path search algorithm is employed to identify multi-level transmission paths from equipment state disturbance sources to tension anomaly representation nodes. Nodes with zero in-degree or driven only by external stimuli are defined as disturbance source nodes, while nodes exhibiting abnormal amplitude or spectral characteristics in tension sensing are defined as tension anomaly representation nodes. A depth-first search is used to traverse all directed paths in the refined causal dependency graph, filtering out all simple paths from disturbance sources to tension anomaly representation nodes. These paths are then sorted by path length (i.e., transmission level) and cumulative causal strength, retaining paths with a transmission level not exceeding a preset maximum and a cumulative causal strength exceeding a minimum threshold, thus forming a multi-level transmission path set. These multi-level transmission paths comprehensively depict the transmission mechanism of how equipment state anomalies are amplified step-by-step through intermediate stages, ultimately leading to tension runaway, providing a structured causal chain for subsequent knowledge representation construction.

[0029] Semantic parsing is performed on each node in the multi-level transmission path, converting the temporal synchronization features and state association features associated with the node into structured knowledge representations. Each node's structured knowledge representation contains three semantic dimensions: causal role semantics, transmission mechanism description, and temporal constraints. Causal role semantics describes the node's role in the transmission path, including initial disturbance source, intermediate transmission node, amplification node, or final anomaly representation node. Role determination is based on the node's topological position in the path and its in-degree and out-degree features. The transmission mechanism description converts the state association features associated with the node into natural language or logical predicate forms, such as "roller speed fluctuation exceeds threshold." At that time, the tension is transmitted through the tension roller displacement node, causing the paper web tension to... "Periodic fluctuations occur after milliseconds," among which... This is the threshold for the roller speed fluctuation amplitude. The timing constraints encode the timing synchronization characteristics associated with nodes into delay intervals and duration constraints, ensuring that subsequent inference queries can correctly match abnormal events with timing consistency.

[0030] After completing the semantic parsing of the nodes, the structured knowledge representation of each node is stored in the form of triples. Each triple consists of a node entity, a relation type, and a target entity. The relation types include predefined predicate types such as "causal triggering," "state propagation," and "temporal delay association." The sets of triples from each node in the multi-level propagation path together constitute the structured knowledge representation library, providing initial knowledge seeds for the subsequent construction of the self-evolving knowledge graph. The entire process starts from multimodal perception data and goes through five stages: cross-modal alignment, conditional dependency modeling, reverse causal verification, path search, and node semantic parsing. Finally, it outputs a structured knowledge representation with complete causal semantics, ensuring the causal interpretability and semantic completeness of the knowledge graph in the initialization stage.

[0031] In one optional implementation, a reverse causal verification mechanism is introduced into the initial causal dependency graph. By simulating intervention in the device's operating state parameters and observing the response changes of the tension timing signal, false causal association edges are eliminated and causal propagation delays are labeled, resulting in a refined causal dependency graph, including: For each directed edge in the initial causal dependency graph, a counterfactual intervention experiment is constructed. In the counterfactual intervention experiment, a reverse perturbation is applied to the equipment operating state parameter corresponding to the starting node of the directed edge, and the dynamic response trajectory of the tension time series signal corresponding to the ending node of the directed edge is recorded synchronously. By analyzing the causal consistency between the change pattern of the dynamic response trajectory and the action pattern of the reverse perturbation, a causal validity score of the directed edge is generated. Directed edges whose causal validity scores are lower than a preset causal threshold are identified as false causal association edges and removed from the initial causal dependency graph. For directed edges whose causal validity scores meet the preset causal threshold, causal propagation delay is labeled based on the time interval between the peak response time of the dynamic response trajectory and the application time of the reverse perturbation. After completing the false edge removal and delay labeling, a refined causal dependency graph is obtained.

[0032] After obtaining the initial causal dependency graph, it is necessary to verify the true causal relationship of each directed edge in the graph to eliminate spurious causal associations caused by data covariance, confounding variables, or incidental correlations. The core idea of ​​introducing a reverse causal verification mechanism is: if the causal relationship described by a directed edge is real, then when a reverse perturbation is applied to the device operating state parameters corresponding to the starting node of the edge, the tension time series signal corresponding to the ending node will necessarily exhibit a dynamic response that is opposite to the direction of the forward perturbation and statistically significant; conversely, if the directed edge only reflects statistical correlation at the data level, the response of the tension time series signal will exhibit random fluctuations or irregular changes, failing to form a consistent causal correspondence with the perturbation pattern.

[0033] For each directed edge in the initial causal dependency graph, a counterfactual intervention experiment is constructed one by one. During the experimental design phase, two key elements need to be defined: first, the amplitude and duration of the reverse perturbation applied to the equipment operating state parameters corresponding to the starting node; and second, the time window for synchronously acquiring the tension timing signal corresponding to the ending node. The amplitude of the reverse perturbation should be set with reference to the historical fluctuation range of the parameter during normal production, selecting a perturbation intensity that produces observable effects without affecting production safety. The duration of the perturbation is usually set to several sampling periods to ensure that the tension sensor can capture the complete response process. During the experiment, the equipment operating state parameters corresponding to the starting node are actively intervened according to the preset reverse perturbation scheme, while the tension timing signal corresponding to the ending node is continuously recorded at high frequency, forming a complete dynamic response trajectory.

[0034] The analysis of dynamic response trajectories focuses on two dimensions: the causal consistency between the change pattern and the reverse disturbance action pattern. The extraction of the change pattern includes the response direction, response amplitude, and temporal structure of the response. If the tension time-series signal exhibits a monotonic change trend corresponding to the disturbance direction after the disturbance is applied, and the change amplitude exceeds the signal's noise floor level by several times, then the response direction is considered causally consistent with the disturbance direction. The response amplitude is evaluated using a signal-to-noise ratio (SNR) metric, comparing the change amplitude of the tension signal during the disturbance period with the signal standard deviation during the period without disturbance to quantify the significance of the response. The analysis of the temporal structure focuses on whether the response occurs within a reasonable time window after the disturbance is applied, excluding "premature responses" that occurred before the disturbance is applied, thus avoiding misjudging accidental temporal coincidences as causal relationships.

[0035] Based on the analysis results from the above three dimensions, a causal validity score is generated for each directed edge, denoted as . Its calculation comprehensively considers the consistency score of response direction. Response amplitude significance score and the score for the rationality of the time sequence The final score is obtained by weighting and combining the three factors. ,in , , These are the weighting coefficients for the three dimensions, and the sum of the three is 1. The value is 0 or 1, corresponding to two cases: the response direction is inconsistent with the disturbance direction and the response direction is consistent with the disturbance direction, respectively. Based on the signal-to-noise ratio (SNR) normalization mapping to the range of 0 to 1, the higher the SNR, the better. The closer it is to 1. The value is assigned based on whether the response time falls within the expected time window after the disturbance is applied. If the response time falls within the expected window, then... The value is 1 if the response direction is consistent and the timing is reasonable, and 0 otherwise. The specific value of the weight coefficient can be adjusted according to the sensitivity of each dimension to different production processes. In general cardboard production scenarios, the weight of response direction consistency and timing rationality is usually higher than the weight of amplitude significance, so as to avoid misjudging the real causal edge as a false edge due to the attenuation of response amplitude caused by equipment aging.

[0036] causal validity score Below the preset causality threshold Directed edges are identified as spurious causal relationships and removed from the initial causal dependency graph. The value of the threshold needs to be balanced between the false edge removal rate and the true edge retention rate. An excessively high threshold may lead to the false deletion of some weakly coupled but real causal relationships, while an excessively low threshold will retain more false associations, affecting the inference quality of subsequent knowledge graphs. In practical applications, the threshold can be determined by cross-validation on historical datasets with known causal structures. The reasonable range is usually set between 0.5 and 0.7, and the specific value is fine-tuned according to the sensor accuracy and signal noise level of the production line.

[0037] For causal validity score Satisfying the preset causal threshold The directed edges are further labeled with causal propagation delays. The causal propagation delay is calculated based on the time interval between the peak response time of the dynamic response trajectory and the time of application of the reverse disturbance. Let the disturbance application time be... The moment when the tension timing signal exhibits its maximum offset after the disturbance is Then the causal propagation delay corresponding to the directed edge satisfy In repeated experiments, Due to slight variations in production conditions, the delay may fluctuate. Therefore, the median of the delays obtained from multiple experiments is taken as the final labeled delay for the directed edge to improve the robustness of the labeling results. For cases where the same directed edge exhibits delay differences under different production speeds or tension baseline values, the delay can be labeled separately according to production conditions to form conditional delay attributes, providing a more refined reference for timing alignment in subsequent multi-hop inference.

[0038] After removing spurious edges and annotating causal propagation delays for all directed edges, a refined causal dependency graph is obtained. Compared to the initial causal dependency graph, the refined causal dependency graph is structurally sparser, retaining only true causal relationships verified by reverse causality, with each edge carrying an annotated causal propagation delay attribute. This refinement process effectively reduces the risk of introducing erroneous relationships in subsequent knowledge graph construction stages, while providing a more reliable causal propagation path foundation for the self-evolving knowledge graph to perform counterfactual reasoning. The delay annotation information in the refined causal dependency graph can also support the source analysis of tension anomalies. By tracing the degree of agreement between the response time of each node and the annotated delay along the causal propagation path, the source device node of the anomaly can be quickly located, thus providing a precise intervention target for the generation of control strategies.

[0039] In one optional implementation, a self-evolving knowledge graph is constructed based on the structured knowledge representation. This self-evolving knowledge graph, by introducing counterfactual reasoning constraints, automatically infers potential anomaly propagation patterns when new production data is input, and dynamically expands the entity nodes and relation edges of the knowledge graph based on the inference results, thereby achieving continuous autonomous growth of the knowledge base. The structured knowledge representation is converted into the initial semantic representation of the knowledge graph. A counterfactual reasoning constraint mechanism is introduced into the initial semantic representation. The counterfactual reasoning constraint mechanism applies hypothetical perturbations to entity nodes by constructing a virtual intervention scenario and simulates causal propagation along the relation edges. During the propagation process, the propagation probability and path branches are dynamically adjusted according to the statistical distribution of historical propagation patterns to generate an abnormal propagation hypothesis space containing multiple inference paths. When new production data is input, the abnormal pattern features in the new production data are extracted, and the abnormal pattern features are semantically matched with the inference paths in the abnormal propagation hypothesis space. The target inference paths with a matching degree exceeding a preset matching threshold are selected as the inference results. It is determined whether the entity nodes and relation edges in the inference results already exist in the current knowledge graph. If they do not exist, the entity nodes and relation edges are dynamically added to the knowledge graph to realize the continuous autonomous growth of the knowledge base.

[0040] When converting structured knowledge representations into the initial semantic representation of a knowledge graph, it is necessary to encode the nodes and directed edges in the identified tension anomaly transmission paths. Each node corresponds to a specific equipment operating state parameter or tension signal observation, such as guide roller speed, tension sensor reading, drying temperature, paper web weight, etc.; each directed edge carries attribute information such as causal direction, transmission delay, and causal validity score. The above nodes and edges are formatted according to a unified semantic pattern to form a set of triples, namely, a structured description of "entity-relationship-entity", which constitutes the initial semantic representation of the knowledge graph. On this basis, a state vector is attached to each entity node, recording the typical value range, anomaly trigger threshold, and historical co-occurrence frequency of the node with other nodes in historical production data, providing a statistical basis for subsequent counterfactual reasoning.

[0041] The core of the counterfactual reasoning constraint mechanism lies in constructing a virtual intervention scenario. For a given entity node in the knowledge graph, a hypothetical perturbation is assumed, contradicting historical observations. For example, under normal paper guide roller speed conditions, a sudden deviation is artificially introduced. This hypothetical perturbation is then propagated causally downstream along directed edges. During propagation, the propagation probability of each directed edge is not a fixed value but dynamically adjusted based on the statistical distribution of historical propagation patterns. Specifically, the frequency and conditional probability of upstream node anomalies triggering downstream node anomalies in historical data are statistically analyzed and used as a baseline value for the propagation probability. This baseline value is then weighted and corrected by auxiliary features in the current production environment (such as machine speed, paper width, and ambient humidity) to ensure the propagation probability reflects the causal strength under current conditions. At path branches, if a node connects to multiple downstream directed edges, multi-path parallel deduction is performed based on the propagation probabilities of each edge, generating a tree-structured set of deduction paths.

[0042] Multiple propagation paths collectively constitute the anomaly propagation hypothesis space. Each path in the hypothesis space describes a potential anomaly propagation pattern, including the path's starting node, the sequence of intermediate propagation nodes, the ending node, and the expected anomaly state offset direction for each node. To prevent the hypothesis space from overexpanding and causing a decrease in subsequent matching efficiency, the propagation paths are pruned: if the cumulative propagation probability of a certain path is lower than a preset path retention threshold... If two paths highly overlap in their intermediate node sequences and have the same termination node, they are merged into a single path with higher confidence to avoid redundant inference results consuming storage and computing resources. After pruning and merging, the paths retained in the hypothesis space all possess a certain degree of statistical confidence and can represent anomalous transmission patterns that actually exist or have a high probability of occurring in historical production processes.

[0043] When new production data is input, abnormal pattern features are extracted. This extraction involves two levels: first, signal-level features, which identify abnormal segments from raw sensor data such as tension timing signals, roller speed signals, and temperature signals, and extract statistical descriptors such as abnormal amplitude, duration, and rate of change; second, structural-level features, which infer the propagation order of abnormal events between equipment nodes based on the co-occurrence relationships of abnormalities in multiple sensor signals along the time axis, forming an observed abnormal propagation sequence. The signal-level and structural-level features are combined to form the abnormal pattern feature vector corresponding to the new data, which is used for semantic matching with the inference path in the hypothesis space.

[0044] The semantic matching process employs a path similarity calculation method. The node sequence and relation attributes of each inference path in the hypothesis space are encoded into path embedding vectors. Simultaneously, the anomalous propagation sequences of newly added data are encoded into observation embedding vectors. The cosine similarity between these two vectors is calculated as the matching score, denoted as […]. ,in For the first hypothesis space The index of the derivation path. When Exceeding the preset matching threshold When this happens, the corresponding inference path is selected as the target inference path and used as the inference result for the newly added data. If multiple paths simultaneously meet the matching conditions, then the inference is performed according to... The paths are sorted from highest to lowest, and the path with the highest matching degree is used as the main inference result. The remaining paths that meet the conditions are kept as alternative inference results for path verification in subsequent multi-hop inference.

[0045] To determine whether entity nodes and relation edges in the inference results already exist in the current knowledge graph, uniqueness checks need to be performed on both nodes and edges. For entity nodes, the semantic identifier of the node (i.e., the combination of the device parameter name and the device number to which it belongs) is used as the unique key for querying; for relation edges, the starting node identifier, ending node identifier, and relation type triplet are used as the unique key for querying. If the query result is non-existent, the corresponding entity node or relation edge is marked as an object to be added. When an entity node to be added is written into the knowledge graph, its state vector needs to be initialized synchronously, using the statistical characteristics of the corresponding signal in the newly added production data as the initial value; when a relation edge to be added is written, attributes such as propagation probability, propagation delay, and causal validity score inherited from the target inference path need to be attached to ensure that the new edge is consistent with the existing edge in terms of attribute structure and does not affect the execution of subsequent multi-hop inference.

[0046] After dynamic additions are completed, the knowledge graph's topology undergoes a local expansion. The newly introduced entity nodes and relation edges will participate in subsequent updates and iterations of the counterfactual reasoning constraint mechanism. Each knowledge graph expansion operation triggered by new data is recorded as a version change log, which includes the target inference path index and matching score on which this expansion is based. Meta-information such as the number of newly added nodes and edges facilitates tracing the evolution trajectory of the knowledge graph when strengthening counterfactual reasoning constraints, and assesses the impact of each expansion operation on the overall topological consistency of the knowledge graph. Through this mechanism, the knowledge graph can continuously expand with the continuous input of production data, covering more and more potential anomaly propagation patterns, and achieving continuous autonomous growth of the knowledge base.

[0047] In one optional implementation, the counterfactual reasoning constraint mechanism applies hypothetical perturbations to entity nodes by constructing a virtual intervention scenario and simulates causal propagation along relational edges. During the propagation process, the propagation probability and path branches are dynamically adjusted based on the statistical distribution of historical propagation patterns, generating an anomaly propagation hypothesis space containing multiple inference paths, including: A virtual intervention scenario is constructed for the target entity node in the knowledge graph. In the virtual intervention scenario, a hypothetical perturbation is applied to the target entity node. The hypothetical perturbation generates a counterfactual state by reversing the causal role semantics of the node. Based on the counterfactual state, a causal propagation simulation is initiated along the relation edge. In the propagation simulation, the propagation attenuation coefficient is calculated based on the causal propagation delay and propagation strength of the relation edge. The propagation attenuation coefficient is used to control the diffusion amplitude of the state change to adjacent nodes, forming a set of candidate propagation paths containing node sequences and edge sequences. The transmission pattern samples are extracted from historical production data and grouped and statistically analyzed according to node type to obtain the statistical distribution of historical transmission patterns. At each path branch point in the candidate propagation path set, the node type combination of the current branch is matched with the statistical distribution to calculate the propagation probability. Based on the propagation probability, the path branches are filtered, retaining those with a probability higher than a preset probability threshold and pruning those with a probability lower than the preset probability threshold. All the filtered propagation paths are summarized into an abnormal transmission hypothesis space.

[0048] The core of the counterfactual reasoning constraint mechanism lies in constructing virtual intervention scenarios to simulate the transmission of anomalies that theoretically occur but never happen in real-world production processes, thereby enabling the knowledge graph to have the ability to predict potential failure modes. In the cardboard production scenario, tension anomalies are often caused by cascading transmission between multiple equipment nodes, and simply relying on historical anomaly data cannot cover all transmission paths. Therefore, constructing virtual intervention scenarios for target entity nodes in the knowledge graph is the starting point for generating the anomaly transmission hypothesis space.

[0049] The selection of target entity nodes is determined by both their topological centrality in the knowledge graph and the frequency of historical anomaly associations. Nodes with a high frequency of occurrence in known anomaly events are prioritized, such as tension rollers, guide roller drive units, or paper web humidity sensor nodes. In the virtual intervention scenario, a hypothetical perturbation is applied to the selected target entity node. This perturbation does not directly modify the node's numerical state, but rather generates a counterfactual state by reversing the node's causal role semantics. Causal role semantics refers to the type of role a node plays in the causal dependency graph. For example, a node that plays the role of a "stable tension output source" under normal production conditions will have its semantics switched to "tension fluctuation inducing source" after reversal. This semantic-level reversal operation can simulate the situation where the node's driving method for downstream nodes changes fundamentally under anomaly conditions without destroying the overall structure of the knowledge graph. This makes the inference path closer to the real anomaly transmission mechanism, rather than a simple superposition of numerical perturbations.

[0050] When initiating a causal propagation simulation along relational edges based on counterfactual states, it is necessary to quantitatively control the diffusion amplitude of state changes along each relational edge during the propagation process. This necessitates the introduction of a propagation attenuation coefficient. To characterize the degree of attenuation of state changes as they propagate along a certain relation edge. Causal propagation delay from this side With conduction strength A joint decision, specifically, ,in This is the delay attenuation rate hyperparameter, used to control the sensitivity of conduction delay to propagation attenuation. Conduction strength. Derived from the causal strength estimation of historical data during the knowledge graph construction phase, it reflects the significance of the causal influence between the two nodes connected by the edge; propagation delay. This represents the time span required for a state change to propagate from the source node to the target node. When... When the value is small, it means that the propagation effect of that edge is weak, and state changes are difficult to effectively spread to adjacent nodes; when A value close to 1 indicates high propagation efficiency and short delay for that edge, meaning state changes can be propagated with significant amplitude. This is achieved by calculating the value for each relation edge individually. This allows for precise control of the perturbation amplitude experienced by each node during the propagation simulation, preventing unreasonable large state shifts in distant nodes due to accumulated errors from multiple hops. The propagation simulation proceeds layer by layer along the relational edges, recording the state change and arrival time of each node encountered, ultimately forming a set of candidate propagation paths described by both the node sequence and the edge sequence.

[0051] The candidate propagation path set contains all paths originating from the target entity node and reaching each terminal node via different combinations of relation edges. Due to the complex topology of knowledge graphs and the exponential growth in the number of paths, a probabilistic screening mechanism based on the statistical distribution of historical propagation patterns is needed to prune path branches. Propagation pattern samples are extracted from historical production data and grouped statistically according to node type combinations to obtain the frequency distribution of causal propagation between different node type pairs, denoted as the statistical distribution. Node type combination refers to the pairing of the current node type with the candidate downstream node type at a branch point in a propagation path, such as "tension roller node → speed sensor node" or "humidity sensor node → tension roller node," etc. At each path branch point, the node type combination of the current branch is combined with... Perform matching to obtain the propagation probability of that branch. , It reflects the proportion of the frequency of this type of combination in historical data to the sum of the frequencies of all candidate combinations of the same type of branch point.

[0052] Based on propagation probability Filter the path branches and keep them. Higher than the preset probability threshold The probability branch, pruning is lower than The probability branch. The configuration needs to strike a balance between hypothesis space coverage and computational efficiency: if If the value is too high, many low-probability but actually occurring anomalous propagation paths will be pruned prematurely, leading to an incomplete hypothesis space; if... If the value is too low, too many paths are retained, significantly increasing the computational cost of subsequent multi-hop inference queries. In practical applications, The algorithm can adaptively adjust based on the number of existing entity nodes and the density of relation edges in the current knowledge graph, and appropriately increase the algorithm when the graph is large. To control the path size, appropriately reduce the size when the map size is small. To ensure the richness of the hypothesis space, after layer-by-layer branching and filtering, all propagation paths that satisfy the probability threshold conditions are summarized to form a complete anomaly propagation hypothesis space. Each path in this hypothesis space carries a node sequence, an edge sequence, and the propagation attenuation coefficient of each edge. The overall cumulative propagation probability information of the path provides a structured inference basis for matching abnormal transmission patterns and dynamically expanding the knowledge graph when new production data arrives.

[0053] Throughout the construction of the virtual intervention scenario and the simulation of causal propagation, the generation and propagation of counterfactual states always occur within the semantic constraint framework of the knowledge graph, preventing the creation of transmission paths that contradict known physical laws. For example, the transmission direction of paper tension is constrained by the physical layout of the production line; state changes cannot propagate backward from downstream equipment nodes to upstream raw material supply nodes. These constraints are encoded into the knowledge graph during the direction definition stage of relation edges. Therefore, the propagation simulation process naturally inherits these physical rationality constraints, ensuring the engineering interpretability of each path in the abnormal propagation hypothesis space.

[0054] In one optional implementation, multi-hop reasoning queries are performed based on the self-evolving knowledge graph to generate a control strategy for the current tension state, and the executed production data is used as feedback samples to correct the topology of the knowledge graph by strengthening the counterfactual reasoning constraints, including: The observation data of the current tension state is obtained and converted into query conditions. The starting node that semantically matches the query conditions is located in the self-evolutionary knowledge graph. Starting from the initial node, a multi-hop inference query is executed. During the inference process, a confidence propagation mechanism is constructed. According to the confidence propagation mechanism, the initial confidence of the initial node is attenuated and propagated along the relation edges. At each jump, the confidence attenuation coefficient is calculated based on the causal propagation delay of the relation edges and the historical verification success rate to form the global confidence of the inference path. Based on the global confidence, the optimal inference path is selected and its node sequence and edge sequence are extracted to generate a control strategy for the current tension state. The control strategy is issued and executed, and the production data after execution is collected as a feedback sample. The execution deviation is calculated by comparing the feedback sample with the expected effect of the optimal inference path. When the execution deviation exceeds a preset deviation threshold, a counterfactual verification matrix is ​​constructed. The counterfactual verification matrix performs a virtual deletion experiment on each relation edge in the optimal inference path and calculates the fit between the inference result after deletion and the feedback sample. The relation edge with the most significant improvement in fit is marked as a redundant edge, and the relation edge with the most significant decrease in fit is marked as a missing edge. The topology of the knowledge graph is corrected based on the marking results of the redundant edges and the missing edges.

[0055] When acquiring observational data of the current tension state, real-time measurements need to be collected synchronously from multiple sensing channels such as tension sensors, roller speed encoders, and temperature and humidity meters. The raw data is then normalized, mapping the values ​​of each dimension to a unified semantic space. The process of converting these into query conditions involves extracting feature vectors such as the mean tension, standard deviation of tension fluctuation, and current roller speed deviation rate, and combining them with current operating condition labels (e.g., paper type and specifications, production line speed level) to form a composite query vector. When locating the starting node for semantic matching in the self-evolving knowledge graph, cosine similarity is used to calculate the similarity between the query vector and the embedded representations of each entity node in the graph. The node with the highest similarity exceeding a preset matching threshold is selected as the starting node for multi-hop inference. If multiple candidate starting nodes exist, they are sorted in descending order of similarity, and the top few nodes are used to initiate independent inference queries. The final path is then filtered using global confidence.

[0056] When performing multi-hop inference queries starting from the initial node, the core of the confidence propagation mechanism lies in propagating the initial confidence of the initial node. Propagation decays hop-by-hop along the relation edge. Let the first... The confidence level of the node reached by the jump is Relationship edge The corresponding causal transmission delay is (The meaning of the parameter with the same name in claim 5 is repeated here), relation edge The historical verification success rate is Then the confidence decay coefficient Defined as ,in This is a latency-sensitive hyperparameter used to control the impact of latency on confidence decay. The node confidence is updated for each hop. Global confidence of the inference path The cumulative product of the confidence scores of all nodes on the path, i.e. ,in This represents the total number of hops in the inference path. Historical verification success rate. Derived from the attribute fields of the relation edges in the knowledge graph, it records the proportion of the prediction results of the relation edge that match the actual production data in the historical regulation execution. The initial value is obtained by the causal validity score mapping when constructing the knowledge graph, and is dynamically updated after each feedback correction.

[0057] To prevent exponential expansion of the path space during multi-hop inference, a bundle search strategy is introduced. Each hop retains only a few candidate paths with the highest global confidence for further expansion, while the remaining paths are pruned. Inference termination conditions include: reaching a preset maximum number of hops, the global confidence of the current path falling below a preset cutoff threshold, or the path's terminal node belonging to a control action node type (i.e., a node corresponding to executable parameters such as roller speed adjustment or tension compensation). After inference is complete, the global confidence is selected from all candidate paths. The highest path is taken as the optimal reasoning path. Its node sequence and edge sequence are extracted. The parameter values ​​carried by the control action nodes in the node sequence (such as the roller speed adjustment range and the tension setpoint offset) constitute the control strategy for the current tension state. The nodes are organized into a sequence of instructions that can be issued according to the topological order of the nodes in the path.

[0058] After the control strategy is issued to the production control system for execution, tension timing data, roller speed response data, and other key process parameters are continuously collected within the execution window, and these data are aggregated into feedback samples. The collection time window for feedback samples is dynamically determined based on the maximum causal propagation delay in the optimal inference path, ensuring that all expected responses are fully reflected in the collected data.

[0059] When comparing the feedback samples with the expected results of the optimal inference path, the expected results consist of the predicted output values ​​of each control action node in the inference path, including the expected tension recovery time and the expected tension stability range. Execution deviation. Defined as the root mean square error between the measured tension response curve and the expected response curve in the feedback sample, when Exceeding the preset deviation threshold At that time, the process of constructing the counterfactual verification matrix is ​​triggered.

[0060] The construction of the counterfactual verification matrix takes the set of relation edges in the optimal inference path as the operation object, assuming that there are a total of The relation edge, for the first Relationship edge ( A virtual deletion experiment was conducted: the relation edge was temporarily removed from the in-memory copy of the knowledge graph, and a multi-hop inference query was re-executed to obtain the inference result after deletion. The goodness of fit was then calculated between the expected response curve corresponding to this inference result and the measured curve in the feedback sample. Goodness of fit... The Pearson correlation coefficient is used to measure the correlation, with a range of values ​​of [value missing]. The closer the value is to 1, the better the inference result matches the measured data after deleting that edge. The change in fit of all relation edges is considered. Sort, where This represents the baseline fit between the inference result and the feedback sample when no edges are deleted. The fact that removing this edge improved the fit indicates that the edge introduced interference into the original path. The largest relational edge is marked as a redundant edge; Furthermore, the edge with the largest absolute value indicates that the fit decreased significantly after deletion, indicating that the edge is crucial to the reasoning result, but its current parameter settings are biased, so it is marked as a missing edge (that is, the causal relationship described by the edge is more critical in actual production than recorded in the knowledge graph, and needs to be supplemented or strengthened).

[0061] When correcting the topology of the knowledge graph based on the labeling results of redundant and missing edges, redundant edges are handled by reducing their causal validity score and decreasing their propagation strength attribute value. If an edge is continuously labeled as redundant multiple times and the cumulative decrease exceeds a preset lower limit, the edge is physically deleted from the knowledge graph. Missing edges are handled in two ways: if the missing edge already exists in the knowledge graph but its propagation strength is low, the propagation strength is re-estimated and updated based on the measured response amplitude of the corresponding variable in the feedback samples; if there is no direct relationship between the node pairs corresponding to the missing edge in the knowledge graph, the conditional independence test is re-executed based on the time-series data of the feedback samples. If the test result supports the existence of a causal relationship, a directed relationship edge is added between the two nodes, and its propagation strength, causal propagation delay, and other attributes are initialized. After correction, the embedded representation of the affected nodes is updated to ensure that the confidence propagation mechanism in subsequent multi-hop inference queries can operate correctly based on the latest graph topology, forming a closed-loop iterative mechanism between the generation of control strategies and the self-correction of the knowledge graph.

[0062] In one optional implementation, the execution deviation is calculated by comparing the feedback sample with the expected effect of the optimal inference path. When the execution deviation exceeds a preset deviation threshold, a counterfactual verification matrix is ​​constructed. The counterfactual verification matrix is ​​constructed by performing a virtual deletion experiment on each relation edge in the optimal inference path and calculating the goodness of fit between the inference result after deletion and the feedback sample, including: Identify the adjustment parameters and expected response timing of each node from the expected effects of the node sequence association in the optimal inference path, and calculate the execution deviation after establishing a synchronous association between the expected response timing and the actual measurement data in the feedback sample based on the timestamp. When the execution deviation exceeds a preset deviation threshold, a counterfactual verification matrix is ​​constructed, and a virtual deletion experiment is performed for each relation edge in the optimal reasoning path. The virtual deletion experiment generates a restricted reasoning path by temporarily blocking the propagation path of the relation edge in the knowledge graph and re-executing the multi-hop reasoning query from the starting node. The expected effect corresponding to the restricted reasoning path is established as the reasoning result after deletion. Calculate the stage-by-stage deviation between the inference result after deletion and the feedback sample under the time-series stage division. Calculate the difference between the stage-by-stage deviation and the execution deviation to obtain the deviation improvement amount of edge deletion. Construct a fitting degree component based on the sign and magnitude of the deviation improvement amount and write it into the corresponding position of the counterfactual verification matrix. After iteratively processing all relation edges to complete the matrix construction, perform multi-weight fusion operation on each row to obtain the comprehensive fitting degree of each edge.

[0063] like Figure 2 As shown, the method includes: In the paperboard production process, after the control strategy is executed, it is necessary to compare the tension response curve expected by the optimal inference path with the actual collected feedback samples to quantitatively evaluate the strategy's effectiveness. Adjustment parameters associated with each node are extracted from the node sequence of the optimal inference path, including roller speed setpoints, tension compensation amounts, and drying temperature offsets. Simultaneously, the expected response time series corresponding to each node is read, i.e., the numerical range of tension that should be reached at each moment after the strategy is issued. The expected response time series is then timestamped with the actual tension measurement data in the feedback samples. The time of strategy command issuance is used as the zero-time reference during alignment. A one-to-one correspondence is established between the two curves on the same time axis, sampling point by sampling point, eliminating the misalignment caused by communication delays and sampling frequency differences. After alignment, the execution deviation is calculated using the root mean square error method. ,like Not exceeding the preset deviation threshold If so, then the current optimal reasoning path is considered to accurately describe the production process, and the knowledge graph does not need to be modified; if If this occurs, the counterfactual verification process is triggered, and the matrix construction phase begins.

[0064] The counterfactual verification matrix is ​​constructed by taking all relation edges in the optimal inference path as the operation object, and the total number of edges is denoted as . The number of rows in a matrix and The results are the same, with each row corresponding to a virtual deletion experiment involving one relation edge. This applies to index number... ( For edges with relationships (), the propagation path of that edge is temporarily masked in the knowledge graph. This masking operation is performed only in the memory copy and does not change the persistently stored graph structure, ensuring that the virtual deletion experiments of multiple edges are independent and do not interfere with each other. After masking, starting from the starting node of the optimal inference path, a multi-hop inference query is re-executed on the restricted graph. The first relation edge has been temporarily removed. The inference engine will search for alternative propagation links along other reachable paths. The resulting restricted inference path bypasses the intermediate nodes connected by the deleted edge and will terminate inference at that hop due to the lack of reachable paths. The expected tension response curve corresponding to the restricted inference path is established as the inference result after deletion. This result reflects the inference under assumption #1. In the case where no relational edge exists, the knowledge graph infers the current tension state.

[0065] To precisely measure the actual contribution of each relation edge, the time-series data is divided into several stages. These stages are based on the production process cycle time, such as the switching times between the cardboard entering the pressure roller section, drying section, and winding section. The entire time series is then divided into multiple continuous sub-intervals. Within each time-series stage, the local root mean square error between the deleted inference result and the feedback sample is calculated, resulting in a stage-by-stage deviation sequence. The stage-by-stage deviations within the same stage are then compared with the original execution deviation. The difference between the components at the corresponding stage is used to obtain the deviation improvement amount of edge deletion: if the difference value is positive, it means that the edge deletion was performed at the corresponding stage. If the inference error at this stage is larger than the original result after removing the edge, it indicates that the edge makes a positive contribution to the inference accuracy at this stage. If the difference is negative, it means that the error decreases after removing the edge, indicating that the edge introduced interference or error propagation at this stage. Based on the sign and magnitude of the deviation improvement, a goodness-of-fit component is constructed. When the improvement is positive and the magnitude is large, A higher value reflects the strong supporting effect of the edge on the reasoning result; when the improvement is negative, Take low or negative values ​​and mark the edge as a potential noise edge. Weight the fit components of each stage and write them into the counterfactual verification matrix. The weights of the corresponding column positions of the rows and different time stages can be set differently according to the tension fluctuation range and process importance of that stage. Stages with larger fluctuation ranges are given higher weights to highlight the transmission contribution of the relationship edge under critical operating conditions.

[0066] Loop through all After the matrix is ​​filled with each relation edge, a multi-weighted fusion operation is performed on each row of the matrix to obtain the comprehensive fit of each edge. The fusion operation comprehensively considers three types of weights: the time-series stage weight reflects the importance of each stage in the overall process flow; the deviation magnitude weight assigns a higher weight to stages with larger absolute values ​​of improvement, making the results more sensitive to the significant contribution of transmission; and the historical verification weight introduces the verification success rate of the relation edge in historical inference tasks, giving extra bonuses to edges with long-term stable performance, and avoiding the misclassification of effective edges as noise edges due to single production anomalies. After normalizing the three types of weights, the fit components of each column are weighted and summed to finally obtain the comprehensive fit. This serves as the comprehensive evaluation score for the relation edge in this counterfactual verification. Meanwhile, the baseline fit... The degree of agreement between the inference results and the feedback samples without deleting any edges is determined and used to provide a basis for all... Provide a reference baseline. Calculate the change in fit on each side. , A negative value indicates that removing the edge leads to a decrease in the fit, and that the edge is indispensable to the current inference path; A positive value indicates that removing the edge actually improves the fit, suggesting that the edge is misleading and should have its weight reduced or removed in subsequent knowledge graph revisions.

[0067] In actual engineering deployments, the aforementioned virtual deletion experiment requires... Each relation edge is executed sequentially, and each experiment requires a complete multi-hop inference query on the restricted graph. The computational cost is positively correlated with the path length and the graph size. To improve efficiency, relation edges can be pre-sorted according to their historical contribution, prioritizing edges with lower historical verification success rates. Once an edge is found to have a lower success rate... Experiments on lower-priority edges can be terminated early if the significance threshold is exceeded to reduce unnecessary computation. Furthermore, the granularity of the time-series stage division directly affects the accuracy and computational cost of matrix construction. Too fine a granularity leads to insufficient sample sizes in each stage, increasing the variance of the statistical estimate; too coarse a granularity fails to distinguish the differentiated contributions of relational edges within different process segments. In practice, it is recommended to use process switching events as natural dividing points to ensure a sufficient number of sampling points within each stage to guarantee the reliability of the bias estimation. After completing matrix construction and comprehensive goodness-of-fit calculation, The sequence is passed to the knowledge graph topology correction stage, driving the directional adjustment of the relation edge weights, thereby achieving continuous reinforcement of counterfactual reasoning constraints and closed-loop optimization of the knowledge graph structure.

[0068] A second aspect of the present invention provides an AI-driven intelligent control system for cardboard tension, comprising: The knowledge mapping unit is used to acquire multimodal sensing data in the paperboard production process, perform cross-modal semantic alignment on the multimodal sensing data, identify multi-level transmission paths of tension abnormal events by establishing a causal dependency graph between tension time-series signals and equipment operating status parameters, and map the nodes in the multi-level transmission paths into structured knowledge representations. The self-evolving unit is used to construct a self-evolving knowledge graph based on the structured knowledge representation. The self-evolving knowledge graph introduces counterfactual reasoning constraints, automatically infers potential abnormal transmission patterns when new production data is input, and dynamically expands the entity nodes and relation edges of the knowledge graph according to the inference results, so as to realize the continuous autonomous growth of the knowledge base. The regulation feedback unit is used to perform multi-hop reasoning queries based on the self-evolving knowledge graph, generate regulation strategies for the current tension state, and use the production data after execution as feedback samples to correct the topology of the knowledge graph by strengthening the counterfactual reasoning constraints.

[0069] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0070] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0071] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An AI-driven intelligent control method for paperboard tension, characterized by, include: Multimodal sensing data is acquired during the paperboard production process. Cross-modal semantic alignment is performed on the multimodal sensing data. By establishing a causal dependency graph between tension time-series signals and equipment operating status parameters, multi-level transmission paths of tension abnormal events are identified, and nodes in the multi-level transmission paths are mapped into structured knowledge representations. Based on the structured knowledge representation, a self-evolving knowledge graph is constructed. The self-evolving knowledge graph introduces counterfactual reasoning constraints, automatically infers potential abnormal transmission patterns when new production data is input, and dynamically expands the entity nodes and relation edges of the knowledge graph according to the inference results, so as to realize the continuous autonomous growth of the knowledge base. Multi-hop reasoning queries are performed based on the self-evolving knowledge graph to generate control strategies for the current tension state. The production data after execution is used as feedback samples to correct the topology of the knowledge graph by strengthening the counterfactual reasoning constraints.

2. The method according to claim 1, characterized in that, Cross-modal semantic alignment is performed on the multimodal sensing data. By establishing a causal dependency graph between tension time-series signals and equipment operating state parameters, multi-level transmission paths of tension anomaly events are identified, and nodes in the multi-level transmission paths are mapped into structured knowledge representations, including: A cross-modal alignment space is constructed by projecting the time-frequency representation of the tension timing signal and the state evolution trajectory of the device operating state parameters into a unified semantic space, extracting timing synchronization features and state association features, establishing conditional dependencies based on the timing synchronization features and the state association features, and generating an initial causal dependency graph. A reverse causal verification mechanism is introduced into the initial causal dependency graph. By simulating intervention in the device operating state parameters and observing the response changes of the tension time series signal, false causal association edges are eliminated and causal transmission delays are marked to obtain a refined causal dependency graph. Based on the directed path search algorithm in the refined causal dependency graph, multi-level transmission paths from the device state disturbance source to the tension anomaly representation are identified. Semantic parsing is performed on each node in the multi-level transmission path, and the temporal synchronization features and state association features associated with the nodes are converted into structured knowledge representations that include causal role semantics, transmission mechanism descriptions and temporal constraints.

3. The method according to claim 2, characterized in that, A reverse causal verification mechanism is introduced into the initial causal dependency graph. By simulating intervention in the device's operating state parameters and observing the response changes of the tension time-series signal, false causal association edges are eliminated and causal propagation delays are labeled, resulting in a refined causal dependency graph including: For each directed edge in the initial causal dependency graph, a counterfactual intervention experiment is constructed. In the counterfactual intervention experiment, a reverse perturbation is applied to the equipment operating state parameter corresponding to the starting node of the directed edge, and the dynamic response trajectory of the tension time series signal corresponding to the ending node of the directed edge is recorded synchronously. By analyzing the causal consistency between the change pattern of the dynamic response trajectory and the action pattern of the reverse perturbation, a causal validity score of the directed edge is generated. Directed edges whose causal validity scores are lower than a preset causal threshold are identified as false causal association edges and removed from the initial causal dependency graph. For directed edges whose causal validity scores meet the preset causal threshold, causal propagation delay is labeled based on the time interval between the peak response time of the dynamic response trajectory and the application time of the reverse perturbation. After completing the false edge removal and delay labeling, a refined causal dependency graph is obtained.

4. The method according to claim 1, characterized in that, Based on the structured knowledge representation, a self-evolving knowledge graph is constructed. This self-evolving knowledge graph, by introducing counterfactual reasoning constraints, automatically infers potential anomaly propagation patterns when new production data is input, and dynamically expands the entity nodes and relation edges of the knowledge graph based on the inference results, achieving continuous autonomous growth of the knowledge base, including: The structured knowledge representation is converted into the initial semantic representation of the knowledge graph. A counterfactual reasoning constraint mechanism is introduced into the initial semantic representation. The counterfactual reasoning constraint mechanism applies hypothetical perturbations to entity nodes by constructing a virtual intervention scenario and simulates causal propagation along the relation edges. During the propagation process, the propagation probability and path branches are dynamically adjusted according to the statistical distribution of historical propagation patterns to generate an abnormal propagation hypothesis space containing multiple inference paths. When new production data is input, the abnormal pattern features in the new production data are extracted, and the abnormal pattern features are semantically matched with the inference paths in the abnormal propagation hypothesis space. The target inference paths with a matching degree exceeding a preset matching threshold are selected as the inference results. It is determined whether the entity nodes and relation edges in the inference results already exist in the current knowledge graph. If they do not exist, the entity nodes and relation edges are dynamically added to the knowledge graph to achieve continuous autonomous growth of the knowledge base.

5. The method according to claim 4, characterized in that, The counterfactual reasoning constraint mechanism applies hypothetical perturbations to entity nodes by constructing a virtual intervention scenario and simulates causal propagation along relational edges. During propagation, it dynamically adjusts the propagation probability and path branches based on the statistical distribution of historical propagation patterns, generating an anomaly propagation hypothesis space containing multiple inference paths, including: A virtual intervention scenario is constructed for the target entity node in the knowledge graph. In the virtual intervention scenario, a hypothetical perturbation is applied to the target entity node. The hypothetical perturbation generates a counterfactual state by reversing the causal role semantics of the node. Based on the counterfactual state, a causal propagation simulation is initiated along the relation edge. In the propagation simulation, the propagation attenuation coefficient is calculated based on the causal propagation delay and propagation strength of the relation edge. The propagation attenuation coefficient is used to control the diffusion amplitude of the state change to adjacent nodes, forming a set of candidate propagation paths containing node sequences and edge sequences. The transmission pattern samples are extracted from historical production data and grouped and statistically analyzed according to node type to obtain the statistical distribution of historical transmission patterns. At each path branch point in the candidate propagation path set, the node type combination of the current branch is matched with the statistical distribution to calculate the propagation probability. Based on the propagation probability, the path branches are filtered, retaining those with a probability higher than a preset probability threshold and pruning those with a probability lower than the preset probability threshold. All the filtered propagation paths are summarized into an abnormal transmission hypothesis space.

6. The method according to claim 1, characterized in that, Perform multi-hop reasoning queries based on the self-evolving knowledge graph to generate a control strategy for the current tension state, and use the executed production data as feedback samples. Strengthen the counterfactual reasoning constraints to correct the topological structure of the knowledge graph, including: The observation data of the current tension state is obtained and converted into query conditions. The starting node that semantically matches the query conditions is located in the self-evolutionary knowledge graph. Starting from the initial node, a multi-hop inference query is executed. During the inference process, a confidence propagation mechanism is constructed. According to the confidence propagation mechanism, the initial confidence of the initial node is attenuated and propagated along the relation edges. At each jump, the confidence attenuation coefficient is calculated based on the causal propagation delay of the relation edges and the historical verification success rate to form the global confidence of the inference path. Based on the global confidence, the optimal inference path is selected and its node sequence and edge sequence are extracted to generate a control strategy for the current tension state. The control strategy is issued and executed, and the production data after execution is collected as a feedback sample. The execution deviation is calculated by comparing the feedback sample with the expected effect of the optimal inference path. When the execution deviation exceeds a preset deviation threshold, a counterfactual verification matrix is ​​constructed. The counterfactual verification matrix performs a virtual deletion experiment on each relation edge in the optimal inference path and calculates the fit between the inference result after deletion and the feedback sample. The relation edge with the most significant improvement in fit is marked as a redundant edge, and the relation edge with the most significant decrease in fit is marked as a missing edge. The topology of the knowledge graph is corrected based on the marking results of the redundant edges and the missing edges.

7. The method according to claim 6, characterized in that, The execution deviation is calculated by comparing the feedback sample with the expected effect of the optimal inference path. When the execution deviation exceeds a preset deviation threshold, a counterfactual verification matrix is ​​constructed. The counterfactual verification matrix is ​​calculated by performing a virtual deletion experiment on each relation edge in the optimal inference path and calculating the goodness of fit between the inference result after deletion and the feedback sample, including: Identify the adjustment parameters and expected response timing of each node from the expected effects of the node sequence association in the optimal inference path, and calculate the execution deviation after establishing a synchronous association between the expected response timing and the actual measurement data in the feedback sample based on the timestamp. When the execution deviation exceeds a preset deviation threshold, a counterfactual verification matrix is ​​constructed, and a virtual deletion experiment is performed for each relation edge in the optimal reasoning path. The virtual deletion experiment generates a restricted reasoning path by temporarily blocking the propagation path of the relation edge in the knowledge graph and re-executing the multi-hop reasoning query from the starting node. The expected effect corresponding to the restricted reasoning path is established as the reasoning result after deletion. Calculate the stage-by-stage deviation between the inference result after deletion and the feedback sample under the time-series stage division. Calculate the difference between the stage-by-stage deviation and the execution deviation to obtain the deviation improvement amount of edge deletion. Construct a fitting degree component based on the sign and magnitude of the deviation improvement amount and write it into the corresponding position of the counterfactual verification matrix. After iteratively processing all relation edges to complete the matrix construction, perform multi-weight fusion operation on each row to obtain the comprehensive fitting degree of each edge.

8. An AI-driven intelligent control system for cardboard tension, used to implement the method as described in any one of claims 1-7, characterized in that, include: The knowledge mapping unit is used to acquire multimodal sensing data in the paperboard production process, perform cross-modal semantic alignment on the multimodal sensing data, identify multi-level transmission paths of tension abnormal events by establishing a causal dependency graph between tension time-series signals and equipment operating status parameters, and map the nodes in the multi-level transmission paths into structured knowledge representations. The self-evolving unit is used to construct a self-evolving knowledge graph based on the structured knowledge representation. The self-evolving knowledge graph introduces counterfactual reasoning constraints, automatically infers potential abnormal transmission patterns when new production data is input, and dynamically expands the entity nodes and relation edges of the knowledge graph according to the inference results, so as to realize the continuous autonomous growth of the knowledge base. The regulation feedback unit is used to perform multi-hop reasoning queries based on the self-evolving knowledge graph, generate regulation strategies for the current tension state, and use the production data after execution as feedback samples to correct the topology of the knowledge graph by strengthening the counterfactual reasoning constraints.

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

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images