Real-time monitoring and dynamic compensation control and early warning method for micro-displacement of cold rolling mill roll

Abstract

The present application provides a cold rolling mill roller micro-displacement real-time monitoring and dynamic compensation control early warning method, relates to the technical field of monitoring and control, and comprises the following steps: monitoring the temperature field distribution of the roller through infrared thermal imaging, establishing the mapping relationship between displacement and temperature gradient, calculating the thermal deformation amount and compensating the displacement drift; constructing a rolling mill equipment domain knowledge graph and a diagnosis rule library, and using graph embedding learning for fault diagnosis; generating a compensation optimization strategy and establishing a fuzzy association mapping to calculate the optimal parameters. The present application realizes accurate monitoring and compensation of the roller micro-displacement, and improves the equipment fault early warning capability and rolling precision.

Description

Technical Field

[0001] This invention relates to monitoring and control technology, and more particularly to a method for real-time monitoring and dynamic compensation control and early warning of micro-displacement of rolls in cold rolling mills. Background Technology

[0002] In the steel industry, cold rolling mills are key equipment for producing high-quality sheet metal, and the precise position control of their rolls directly affects the thickness accuracy and surface quality of the product. During the rolling process, the rolls are subjected to high temperature, high pressure, and high-speed rotation, making them prone to fretting displacement and thermal deformation. Traditional rolling mill control systems mainly rely on thickness detection feedback and simple mathematical models for compensation, which are insufficient to cope with the dynamic changes of the rolls under complex operating conditions in real time. With the development of intelligent manufacturing technology, roll displacement monitoring and compensation control technology based on infrared thermal imaging, knowledge graphs, and deep learning has gradually become a research hotspot.

[0003] Traditional monitoring methods are difficult to effectively obtain the temperature field distribution on the roll surface and cannot establish an accurate mapping relationship between temperature gradient and roll micro-displacement, resulting in insufficient accuracy of thermal deformation compensation and affecting the thickness control accuracy of the sheet.

[0004] Existing fault diagnosis methods are mostly based on simple threshold judgments or statistical models, lacking a deep understanding and learning ability of equipment knowledge. They are difficult to accurately identify the root cause of abnormal roll displacement under complex working conditions, affecting the accuracy and timeliness of fault warnings.

[0005] Current compensation control strategies typically employ fixed parameters or simple adaptive methods, which are difficult to intelligently optimize and adjust based on dynamic changes in rolling process parameters and equipment status. This makes it impossible to achieve optimal dynamic compensation results and hinders further improvement in the quality of cold-rolled products. Summary of the Invention

[0006] This invention provides a method for real-time monitoring and dynamic compensation control and early warning of micro-displacement of cold rolling mill rolls, which can solve the problems in the prior art.

[0007] A first aspect of the present invention provides a method for real-time monitoring and dynamic compensation control and early warning of micro-displacement of rolls in a cold rolling mill, comprising:

[0008] Real-time operating data of cold rolling mill rolls are obtained, and the temperature field distribution of rolls is monitored in real time using infrared thermal imaging. A dynamic mapping relationship between roll displacement and temperature gradient is established, and the thermal deformation of rolls under different temperature fields is calculated. Displacement drift caused by temperature field is compensated to obtain roll displacement data.

[0009] A knowledge graph of the rolling mill equipment domain is constructed. Based on the knowledge graph, an equipment diagnosis rule base is constructed using ontology modeling technology. A graph embedding learning method is used to map the feature data corresponding to the equipment diagnosis rule base onto the knowledge graph of the rolling mill equipment domain for fault diagnosis. The roll displacement data is used for deep representation learning of fault features.

[0010] Based on the analysis results of the fault diagnosis, a compensation optimization strategy is generated. The compensation optimization strategy is then linked to the real-time operating data to establish a fuzzy correlation mapping. The optimal compensation parameters are calculated based on the fuzzy correlation mapping. Finally, the compensation strategy is iteratively optimized using the knowledge graph of the rolling mill equipment domain.

[0011] Infrared thermal imaging was used to monitor the temperature field distribution of the rolls in real time, establishing a dynamic mapping relationship between roll displacement and temperature gradient. The thermal deformation of the rolls under different temperature fields was calculated, and displacement drift caused by the temperature field was compensated for. The resulting roll displacement data included:

[0012] The temperature field distribution data of the roll surface collected by the infrared thermal imaging system is divided into regions according to the degree of temperature gradient change, and a temperature sensor array is arranged in the divided regions according to the rate of temperature gradient change.

[0013] The temperature gradient change rate between adjacent monitoring points is calculated using the temperature data collected by the temperature sensor array. The temperature gradient change rate is used as a boundary condition to solve the heat conduction characteristic equation to obtain the main heat conduction path inside the roll. The main heat conduction path is used as a temperature transfer characteristic parameter to establish a dynamic mapping relationship between roll displacement and temperature gradient with the temperature field distribution data on the roll surface.

[0014] The roll thermal deformation is calculated based on the dynamic mapping relationship. The compensation coefficient is determined based on the spatial distribution characteristics of the main heat conduction path. The product of the compensation coefficient and the roll thermal deformation is used as the compensation amount to obtain the roll displacement data.

[0015] Based on the knowledge graph of the rolling mill equipment domain, an equipment diagnosis rule base is constructed using ontology modeling technology. Then, a graph embedding learning method is used to map the feature data corresponding to the equipment diagnosis rule base onto the rolling mill equipment domain knowledge graph for fault diagnosis, including:

[0016] Based on the hierarchical association rules between the knowledge graph analysis components of the rolling mill equipment domain, an equipment diagnosis rule base is constructed by calculating the probability confidence of the hierarchical association rules. The uncertainty of the rules in the equipment diagnosis rule base is quantitatively evaluated, and the rule reliability index is obtained based on the correlation strength between the probability of the rule result and the uncertainty of the rule.

[0017] The rules in the device diagnostic rule base that meet the rule reliability index are mapped to a low-dimensional vector space. The embedding representation of the rule node is obtained by performing a weight matrix operation on the neighborhood set of the rule node. A diagnostic path from the source node to the target node is constructed based on the embedding representation of the rule node. The optimal diagnostic path is determined based on the propagation decay of the edge weights in the diagnostic path.

[0018] The embedded representation of the rule node is classified according to the optimal diagnostic path. The predicted probability of the fault category is obtained by the classification matrix operation. The posterior probability of the fault cause is calculated by combining the equipment operation evidence. The predicted probability of the fault category is corrected according to the coupling relationship between the posterior probability and the causal strength. The equipment fault diagnosis is realized based on the corrected predicted probability of the fault category.

[0019] The embedding representation of a regular node is obtained by performing a weight matrix operation on the neighborhood set of the regular nodes. A diagnostic path from the source node to the target node is constructed based on the embedding representation of the regular nodes. The optimal diagnostic path is determined based on the propagation decay of the edge weights in the diagnostic path, including:

[0020] A multidimensional feature set of a rule node and its k-order neighboring node set is obtained. The semantic similarity of the nodes is calculated based on the multidimensional feature set as the initial weight. The initial weight is adjusted based on the temporal co-occurrence information of the rule node and the k-order neighboring node set in historical diagnostic data. A propagation decay function is constructed based on the connection relationship between the rule node and the k-order neighboring node set in the physical topology. The adjusted weight is decayed using the propagation decay function to obtain the neighborhood weight matrix of the rule node.

[0021] The attention score between the regular node and its k-order neighbor node set is calculated based on the weight distribution of the neighborhood weight matrix. The embedded representation of the regular node is obtained by performing temporal modeling on the state of the regular node based on the attention score.

[0022] The embedding representation of the rule node is used as the starting state. The node transition probability is determined based on the similarity of the embedding representations between rule nodes. The search space is gradually expanded to the target node according to the node transition probability to construct a set of candidate diagnostic paths. The path with the largest propagation attenuation coefficient is selected as the optimal diagnostic path based on the embedding representations of adjacent nodes in the set of candidate diagnostic paths.

[0023] Based on the analysis results of the fault diagnosis, a compensation optimization strategy is generated. A fuzzy correlation mapping is established between the compensation optimization strategy and the real-time operating data. The optimal compensation parameters are calculated based on the fuzzy correlation mapping, including:

[0024] The analysis results of fault diagnosis are used to perform fuzzy classification modeling to construct a fuzzy correlation matrix; the membership degree value of each fault object is calculated based on the fuzzy correlation matrix, the influence weight of the fault object is determined based on the membership degree value, the influence weight is compared with a preset weight threshold to obtain the compensation priority, and the local compensation strategy corresponding to each fault object is generated based on the compensation priority.

[0025] The association strength between adjacent fault objects in the fuzzy association matrix is ​​analyzed using dynamic programming. The association strength is combined with the compensation parameters of the local compensation strategy to calculate the membership degree of the propagation influence. The compensation parameters of the local compensation strategy are recursively optimized based on the membership degree of the propagation influence to generate a compensation optimization strategy.

[0026] Based on the state characteristics of real-time running data and the compensation optimization strategy, a dynamic compensation fuzzy rule base is constructed. The comprehensive activation intensity of the fuzzy rules is calculated through adaptive fuzzy inference. The fuzzy rule base is then defuzzified using the comprehensive activation intensity to obtain the optimal compensation parameters.

[0027] Based on the state characteristics of real-time operating data and the aforementioned compensation optimization strategy, a dynamic compensation fuzzy rule base is constructed. The comprehensive activation intensity of the fuzzy rules is calculated through adaptive fuzzy inference. The centroid defuzzification of the fuzzy rule base is then performed using the comprehensive activation intensity, including:

[0028] Based on the state characteristics and quantitative indicators of the compensation optimization strategy in the real-time running data, an influence matrix between the quantitative indicators is constructed. The influence matrix is ​​used to calculate the membership distribution of the state feature vector and the compensation strategy vector in different fuzzy subspaces. Based on the membership distribution, a fuzzy rule base for dynamic compensation is generated.

[0029] Based on the influence matrix, the initial weight factors of each rule are set, the initial weight factors of the rules in the fuzzy rule base are iteratively optimized, and the optimized weight factors are used to perform fuzzy operator operations with the state feature membership values ​​of the corresponding rules in the fuzzy rule base to obtain the comprehensive activation intensity that represents the importance of the rule.

[0030] Using the comprehensive activation intensity as a weighting coefficient, the centroid defuzzification operation is performed on the compensation strategy of the rule consequent in the fuzzy rule base to obtain the output value of the optimal compensation strategy. The gradient direction of the optimal compensation strategy is calculated based on the steepest descent method. The output value is iteratively updated along the direction that makes the control error represented by the influence matrix decrease the fastest until the optimal solution is reached.

[0031] A second aspect of the present invention provides a real-time monitoring and dynamic compensation control early warning system for the micro-displacement of rolls in a cold rolling mill, comprising:

[0032] The first unit is used to acquire real-time operating data of the cold rolling mill rolls. It uses infrared thermal imaging to monitor the temperature field distribution of the rolls in real time, establishes a dynamic mapping relationship between roll displacement and temperature gradient, calculates the thermal deformation of the rolls under different temperature fields, compensates for the displacement drift caused by the temperature field, and obtains the roll displacement data.

[0033] The second unit is used to construct a knowledge graph of the rolling mill equipment domain. Based on the knowledge graph of the rolling mill equipment domain, an equipment diagnosis rule base is constructed using ontology modeling technology. The feature data corresponding to the equipment diagnosis rule base is mapped to the knowledge graph of the rolling mill equipment domain using graph embedding learning method to perform fault diagnosis. The roll displacement data is used to perform deep representation learning of fault features.

[0034] The third unit is used to generate a compensation optimization strategy based on the analysis results of the fault diagnosis, establish a fuzzy correlation mapping between the compensation optimization strategy and the real-time operation data, calculate the optimal compensation parameters according to the fuzzy correlation mapping, and iteratively optimize the compensation strategy in combination with the knowledge graph of the rolling mill equipment domain.

[0035] A third aspect of the present invention provides an electronic device, comprising:

[0036] processor;

[0037] Memory used to store processor-executable instructions;

[0038] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0039] 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.

[0040] The beneficial effects of this application are as follows:

[0041] This invention uses infrared thermal imaging to monitor the temperature field distribution of the roll in real time, establishes a dynamic mapping relationship between roll displacement and temperature gradient, and compensates for displacement drift caused by temperature field, effectively improving the accuracy of roll displacement monitoring and solving the problem that traditional monitoring methods cannot effectively identify and distinguish thermal deformation factors.

[0042] This invention constructs a knowledge graph of rolling mill equipment domain, builds an equipment diagnosis rule base through ontology modeling technology, uses graph embedding learning method for fault diagnosis, and utilizes roll displacement data for deep characterization learning of fault features, thereby achieving accurate identification and prediction of rolling mill micro-motion faults and improving the accuracy and reliability of fault diagnosis.

[0043] This invention generates a compensation optimization strategy based on fault diagnosis results, establishes a fuzzy correlation mapping to calculate the optimal compensation parameters, and combines the knowledge graph of the rolling mill equipment domain to iteratively optimize the compensation strategy, thereby realizing dynamic compensation control of the micro-displacement of the rolls. This significantly improves the stability of cold rolling mill production and product quality, and reduces equipment failure rate and maintenance costs. Attached Figure Description

[0044] Figure 1 This is a flowchart illustrating the real-time monitoring and dynamic compensation control early warning method for micro-displacement of cold rolling mill rolls according to an embodiment of the present invention.

[0045] Figure 2 This is a flowchart illustrating the optimization of the dynamic compensation strategy for fault diagnosis in an embodiment of the present invention. Detailed Implementation

[0046] 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.

[0047] 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.

[0048] Figure 1 This is a flowchart illustrating the real-time monitoring and dynamic compensation control early warning method for micro-displacement of cold rolling mill rolls according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0049] Real-time operating data of cold rolling mill rolls are obtained, and the temperature field distribution of rolls is monitored in real time using infrared thermal imaging. A dynamic mapping relationship between roll displacement and temperature gradient is established, and the thermal deformation of rolls under different temperature fields is calculated. Displacement drift caused by temperature field is compensated to obtain roll displacement data.

[0050] A knowledge graph of the rolling mill equipment domain is constructed. Based on the knowledge graph, an equipment diagnosis rule base is constructed using ontology modeling technology. A graph embedding learning method is used to map the feature data corresponding to the equipment diagnosis rule base onto the knowledge graph of the rolling mill equipment domain for fault diagnosis. The roll displacement data is used for deep representation learning of fault features.

[0051] Based on the analysis results of the fault diagnosis, a compensation optimization strategy is generated. The compensation optimization strategy is then linked to the real-time operating data to establish a fuzzy correlation mapping. The optimal compensation parameters are calculated based on the fuzzy correlation mapping. Finally, the compensation strategy is iteratively optimized using the knowledge graph of the rolling mill equipment domain.

[0052] In one optional implementation, infrared thermal imaging is used to monitor the temperature field distribution of the roll in real time, establish a dynamic mapping relationship between roll displacement and temperature gradient, calculate the thermal deformation of the roll under different temperature fields, compensate for the displacement drift caused by the temperature field, and obtain roll displacement data including:

[0053] The temperature field distribution data of the roll surface collected by the infrared thermal imaging system is divided into regions according to the degree of temperature gradient change, and a temperature sensor array is arranged in the divided regions according to the rate of temperature gradient change.

[0054] The temperature gradient change rate between adjacent monitoring points is calculated using the temperature data collected by the temperature sensor array. The temperature gradient change rate is used as a boundary condition to solve the heat conduction characteristic equation to obtain the main heat conduction path inside the roll. The main heat conduction path is used as a temperature transfer characteristic parameter to establish a dynamic mapping relationship between roll displacement and temperature gradient with the temperature field distribution data on the roll surface.

[0055] The roll thermal deformation is calculated based on the dynamic mapping relationship. The compensation coefficient is determined based on the spatial distribution characteristics of the main heat conduction path. The product of the compensation coefficient and the roll thermal deformation is used as the compensation amount to obtain the roll displacement data.

[0056] Infrared thermal imaging equipment is configured to continuously monitor the surface temperature of the rolls. The equipment is mounted on the side of the mill stand to ensure full coverage scanning of the surface temperature along the entire length of the roll. The optical lens of the thermal imaging equipment uses a long-wave infrared detector with a wavelength range of 8 to 14 micrometers, covering a temperature measurement range from ambient temperature to 800 degrees Celsius. The spatial resolution of the equipment is set to include at least one pixel per millimeter to ensure the capture of detailed temperature changes on the roll surface. The sampling frequency of the thermal imaging equipment is set to 50 frames per second to meet the real-time monitoring requirements under high-speed mill operation.

[0057] The temperature data collected by the infrared thermal imaging equipment undergoes preprocessing, including steps such as removing environmental radiation interference, compensating for atmospheric attenuation, and correcting for emissivity variations. Environmental radiation interference removal is achieved through background temperature compensation, with background temperature distribution periodically collected as a reference under roll shutdown conditions. Atmospheric attenuation compensation is performed based on the distance between the equipment and the rolls and ambient humidity, using a pre-set attenuation coefficient table to correct temperature values. Emissivity correction considers the influence of the roll surface oxidation degree and material properties, establishing a correspondence table between emissivity and surface condition.

[0058] Analyzing the temperature field distribution data of the roll surface, regions with significant temperature gradient changes and relatively uniform temperatures were identified. The temperature gradient was calculated using the finite difference method, dividing the temperature difference between adjacent pixels by the spatial distance to obtain the gradient value. The entire roll surface was then segmented according to the magnitude of the temperature gradient value: regions with gradient values ​​greater than a high threshold were classified as high-gradient regions, those between high and low thresholds as medium-gradient regions, and those less than a low threshold as low-gradient regions. The thresholds were set based on statistical analysis of historical temperature data to ensure a reasonable area ratio for each region.

[0059] The placement of temperature sensors is determined within each region defined by the temperature gradient. The sensor density is directly proportional to the degree of temperature gradient change in the region. In high-gradient regions, the sensor spacing is small, typically one sensor every 20 mm circumferentially and one every 50 mm axially. In medium-gradient regions, the sensor spacing is moderate, with sensors every 40 mm circumferentially and 100 mm axially. In low-gradient regions, the sensor spacing is large, with sensors every 80 mm circumferentially and 200 mm axially.

[0060] A temperature sensor array is installed at a designated location on the roll surface. The sensors are thermocouples or resistance temperature detectors (RTDs) with a response time of less than 1 second and a measurement accuracy better than ±1 degree Celsius. The sensors are installed using an embedded method, with holes drilled into the roll surface to mount the sensor probes. The probe depth is controlled to be 2 to 3 millimeters below the roll surface to ensure that the measurement results reflect the temperature state near the roll surface. Shielded cables are used to connect the sensors to the data acquisition equipment to avoid the influence of electromagnetic interference on the measurement results.

[0061] Real-time temperature data is collected using a temperature sensor array at a frequency of 10 times per second, synchronized with the mill's control cycle. The rate of change of temperature gradient between adjacent monitoring points is calculated using time series analysis, analyzing the temperature gradient difference between adjacent time points. For adjacent sensors along the roll's circumference, the circumferential temperature gradient is obtained by dividing the temperature difference by the sensor spacing. For adjacent sensors along the roll's axial direction, the axial temperature gradient is obtained by dividing the temperature difference by the axial spacing.

[0062] A numerical model for heat conduction inside the roll was established, and the geometric structure of the roll was discretized into a large number of small elements using the finite element method. The material properties of each element include thermophysical parameters such as density, specific heat capacity, and thermal conductivity, which are determined based on the composition of the roll material and its heat treatment state. The boundary conditions were set based on the surface temperature distribution measured by the temperature sensor array and the calculated rate of change of the temperature gradient. The surface temperature was used as the first type of boundary condition, and the surface heat flux density was used as the second type of boundary condition.

[0063] The temperature distribution inside the roll is obtained by solving the characteristic equation of heat conduction. An iterative algorithm is used to solve the characteristic equation, progressively approximating the steady-state or quasi-steady-state solution. The convergence of the temperature field is monitored during the iteration process; convergence is considered achieved when the temperature difference between two consecutive iterations is less than a set threshold. By analyzing the obtained three-dimensional temperature field distribution, the main heat conduction paths inside the roll are identified. The main heat conduction paths correspond to the directions with the largest temperature gradients, representing the dominant channels for heat transfer.

[0064] The spatial distribution characteristics of the main heat conduction paths inside the roll are extracted, including geometric parameters such as the starting and ending points of the paths, path length, and path direction. The heat conduction paths are identified using temperature gradient vector field analysis, tracing the heat transfer trajectory along the direction of the gradient vector. The weight of each path is determined based on the magnitude of the heat flux density passing through it; paths with higher heat flux densities have higher weights and contribute more to the overall thermal deformation of the roll.

[0065] A dynamic mapping relationship between roll displacement and temperature gradient is established, describing the correspondence between temperature field changes and roll deformation. This mapping relationship is established using machine learning methods, with temperature gradient distribution as input features and roll deformation as the output target. Input features include statistical quantities such as the average temperature gradient, maximum temperature gradient, and temperature gradient variance for each region. The output target includes multiple components such as radial deformation, axial deformation, and angular deformation of the roll.

[0066] The dynamic mapping relationship model is trained using machine learning algorithms such as neural networks or support vector regression. Training data comes from temperature field data and corresponding roll deformation measurement data collected during historical operation, covering different working conditions and temperature levels. Cross-validation is used during model training to evaluate prediction accuracy and ensure the model's generalization ability. The trained model can predict the thermal deformation state of the rolls based on real-time temperature gradient distribution.

[0067] The thermal deformation of the roll under the current temperature field conditions is calculated using the established dynamic mapping relationship. The calculation of deformation considers material properties such as the coefficient of thermal expansion and the modulus of elasticity of the roll material. The thermal deformation includes multiple components such as the change in roll diameter, the change in roll length, and the bending deformation of the roll axis. The deformation calculation adopts the thermoelastic mechanics theory, applying the temperature field as a body force load to the mechanical model of the roll.

[0068] This study analyzes the influence of the spatial distribution characteristics of major heat conduction paths on thermal deformation and establishes the correlation between path characteristics and deformation sensitivity. The directionality of the heat conduction path affects the directionality of deformation; radially dominant heat conduction paths mainly cause radial deformation, while axially dominant heat conduction paths mainly cause axial deformation. The density distribution of the paths affects the spatial distribution of deformation; areas with dense paths experience greater deformation, while areas with sparse paths experience less deformation.

[0069] The temperature compensation coefficient is determined based on the spatial distribution characteristics of the main heat conduction paths and the geometric parameters of the rolls. For regions dominated by radial heat conduction, the compensation coefficient primarily considers the roll radius and the radial thermal expansion characteristics of the material. For regions dominated by axial heat conduction, the compensation coefficient primarily considers the roll length and the axial thermal expansion characteristics of the material. The compensation coefficient also needs to consider the constraints on the rolls, including the influence of bearing supports and the effect of rolling forces.

[0070] The temperature compensation amount is calculated as the product of the compensation coefficient and the roll thermal deformation. The compensation amount is calculated for each component of the roll displacement, including X-direction, Y-direction, Z-direction, and angular displacement compensation. The sign of the compensation amount indicates the direction of compensation; a positive value indicates positive compensation and a negative value indicates reverse compensation.

[0071] The calculated temperature compensation amount is applied to the real-time measurement data of the roll displacement to obtain temperature-compensated roll displacement data. The compensation process uses an additive operation, adding the original measured displacement to the temperature compensation amount to obtain the compensated displacement value. The compensated displacement data eliminates the influence of temperature field changes on displacement measurement and can more accurately reflect the actual position state of the roll. The compensation effect is verified by comparing the stability and repeatability of the displacement data before and after compensation. The compensated data should have better consistency and lower measurement noise.

[0072] A real-time temperature compensation update mechanism is established to dynamically adjust compensation parameters based on real-time changes in the temperature field. This update mechanism includes adaptive adjustment of the compensation coefficients and online optimization of the mapping relationship, ensuring that the compensation effect can adapt to different operating conditions and temperature environments. Adaptive adjustment is based on feedback evaluation of the compensation effect, automatically adjusting relevant parameters when the compensation effect deviates from expectations. Online optimization employs an incremental learning method, continuously improving the accuracy of the mapping relationship model using new operational data.

[0073] In one optional implementation, an equipment diagnosis rule base is constructed based on the rolling mill equipment domain knowledge graph using ontology modeling technology, and a graph embedding learning method is used to map the feature data corresponding to the equipment diagnosis rule base onto the rolling mill equipment domain knowledge graph for fault diagnosis, including:

[0074] Based on the hierarchical association rules between the knowledge graph analysis components of the rolling mill equipment domain, an equipment diagnosis rule base is constructed by calculating the probability confidence of the hierarchical association rules. The uncertainty of the rules in the equipment diagnosis rule base is quantitatively evaluated, and the rule reliability index is obtained based on the correlation strength between the probability of the rule result and the uncertainty of the rule.

[0075] The rules in the device diagnostic rule base that meet the rule reliability index are mapped to a low-dimensional vector space. The embedding representation of the rule node is obtained by performing a weight matrix operation on the neighborhood set of the rule node. A diagnostic path from the source node to the target node is constructed based on the embedding representation of the rule node. The optimal diagnostic path is determined based on the propagation decay of the edge weights in the diagnostic path.

[0076] The embedded representation of the rule node is classified according to the optimal diagnostic path. The predicted probability of the fault category is obtained by the classification matrix operation. The posterior probability of the fault cause is calculated by combining the equipment operation evidence. The predicted probability of the fault category is corrected according to the coupling relationship between the posterior probability and the causal strength. The equipment fault diagnosis is realized based on the corrected predicted probability of the fault category.

[0077] A knowledge graph of rolling mill equipment domain is used to construct an equipment ontology model. This ontology model includes structured knowledge representations of core concept categories, concept attributes, and relationships between concepts. Core concept categories cover major equipment components such as roll parts, transmission devices, hydraulic equipment, electrical controllers, lubrication devices, and cooling equipment. Each concept category defines a corresponding set of attributes. Attributes for roll parts include roll diameter, material type, surface roughness, and wear degree; attributes for transmission devices include gear ratio, torque capacity, speed range, and lubrication status. Relationships between concepts describe the logical connections between components, control dependencies, energy transfer paths, etc., fully expressing the structured knowledge of the equipment through object attributes and data attributes.

[0078] This study analyzes the hierarchical association rules among rolling mill equipment components. The hierarchical structure reflects the multi-level organizational relationships of the equipment from sub-components to the entire machine. Bottom-level association rules describe the direct interactions between basic components, middle-level association rules describe the functional coordination relationships between sub-units, and top-level association rules describe the performance relationships at the machine-wide level. Association patterns between components are extracted by mining historical operating data and fault records. The discovery of association patterns employs frequent itemset mining and association rule mining algorithms. Frequent itemsets reflect combinations of component states that frequently occur simultaneously, while association rules describe the conditional dependencies between component states.

[0079] The probability confidence score of hierarchical association rules is calculated, and this score measures the reliability and predictive accuracy of the rules. The confidence score is calculated based on the rule's validation statistics in historical data, using the frequency of occurrence of the consequent under the condition of the antecedent as the confidence score value. For a rule where the conditions are abnormal roll temperature and excessive vibration amplitude, the total number of cases where both conditions are met simultaneously in historical data is counted, and then the number of cases where roll failure actually occurred is counted. The ratio of these two values ​​is the confidence score of the rule. The confidence score value ranges from zero to one; a higher value indicates a more reliable rule.

[0080] A device diagnostic rule base containing multi-level association rules is constructed. The rule base employs a hierarchical storage structure to organize diagnostic rules at different levels. Each rule in the rule base includes complete information such as a rule identifier, rule antecedent conditions, rule consequent conclusion, rule confidence, and rule support. The rule antecedent describes a combination of multiple conditions in the form of a logical expression, supporting logical operators such as AND, OR, and NOT. The rule consequent describes the corresponding diagnostic conclusion, including the fault type, fault severity, and fault location. The rule base supports dynamic updates, continuously optimizing and expanding the rule content based on new operational data and fault cases.

[0081] Uncertainty quantification assessment is performed on the rules in the equipment diagnostic rule base, taking into account the changes in the applicability of the rules under different operating conditions. Uncertainty quantification employs a combination of fuzzy set theory and probability theory, representing the applicability conditions of the rules as fuzzy sets and calculating the fuzzy membership degree of the rule in the current state. Simultaneously, the noise impact of measurement data and sensor accuracy limitations are considered, and the uncertainty range of the rule output is calculated through error propagation analysis. Uncertainty indices include quantification results across multiple dimensions such as fuzziness, randomness, and mixed uncertainty of the rules.

[0082] The correlation strength between the probability of a rule's outcome and its uncertainty is calculated, reflecting the degree to which uncertainty affects the reliability of the rule. The correlation strength is calculated using the mutual information method, analyzing the statistical dependence between the rule's probability distribution and the uncertainty distribution. A larger mutual information value indicates a stronger correlation and a more significant impact of rule uncertainty on its reliability. By analyzing probability-uncertainty correlation patterns of a large number of rules, a quantitative evaluation model for the correlation strength is established.

[0083] The reliability index of a rule is calculated based on the correlation strength between the probability and uncertainty of the rule's outcome. This reliability index comprehensively considers both the predictive accuracy and stability of the rule. The calculation uses the rule's confidence level as a base value and adjusts it according to the magnitude of the uncertainty. Rules with low uncertainty receive a positive correction, resulting in a reliability index higher than the base confidence value; rules with high uncertainty receive a negative correction, resulting in a reliability index lower than the base confidence value. The magnitude of the correction is proportional to the correlation strength; rules with stronger correlations are more significantly affected by uncertainty.

[0084] Rules in the device diagnostic rule base that meet the rule reliability index requirements are selected, and a reliability threshold is set as the criterion for rule selection. Rules with a reliability index higher than the threshold are retained for subsequent graph embedding learning, while rules with a reliability index lower than the threshold are marked as low-reliability rules or removed from the rule base. The threshold setting is based on a balance between diagnostic accuracy requirements and the size of the rule base; too high a threshold results in too few usable rules, while too low a threshold affects diagnostic accuracy.

[0085] The filtered rules are mapped to a low-dimensional vector space, the dimension of which is determined by the complexity of the rules and computational resources. The mapping process employs graph embedding learning, transforming the logical relationships between rules into geometric relationships in the vector space. The vector representations of the rules preserve their original semantic similarity and logical connections; semantically similar rules are closer together in the vector space, while rules with strong logical connections exhibit specific directional relationships.

[0086] A neighborhood set is constructed for each rule node. This set includes other rule nodes that are directly or indirectly related to the current rule node. Neighborhood relationships are determined based on various association types, such as logical dependencies, semantic similarity, and co-occurrence frequency. Direct neighbors contain logically related rules, while indirect neighbors contain rules linked through intermediate rules. The size of the neighborhood set is controlled by setting a neighborhood radius or an upper limit on the number of neighbors to prevent excessive computational complexity due to an overly large neighborhood.

[0087] The weight matrix of the neighborhood set of a rule node is calculated, describing the association strength between the current rule node and each neighboring node. The weight calculation comprehensively considers multiple factors such as semantic similarity, logical relevance, and historical co-occurrence frequency among rules. Semantic similarity is calculated by comparing the vector representations of the rule text; logical relevance is determined based on the overlap of antecedents and consequents of the rules; and historical co-occurrence frequency is obtained by statistically analyzing the simultaneous activation of rules during the diagnostic process. The normalization process of the weight matrix ensures that the sum of the weights of all neighboring nodes equals one.

[0088] The embedding representation of regular nodes is obtained through weight matrix operations. This embedding representation integrates the node's own features and neighborhood information. The embedding computation employs the aggregation mechanism of a graph convolutional network, which weights and averages the feature vectors of neighboring nodes according to their weights, and then linearly combines them with the original features of the current node. The stacking of multiple layers of graph convolutions enables the node's embedding representation to capture information from a greater distance in the neighborhood, enhancing the expressive power of the representation.

[0089] A search graph for diagnostic paths is constructed based on the embedding representation of rule nodes. Nodes in the search graph are rule nodes, and edges represent the reasoning and transition relationships between rules. The source node corresponds to the initial observed symptoms for diagnosis, and the target node corresponds to the final fault diagnosis conclusion. The transition probability between nodes is calculated using the similarity of the embedding representation; node pairs with high similarity have a higher transition probability. A probabilistic path search algorithm is then used to find the path from the source node to the target node in the search graph.

[0090] The propagation decay of edge weights in the diagnostic path is calculated, reflecting the gradual loss of information in the inference chain. The weight of each edge is determined based on the similarity of the embedding representations of the connected nodes; the higher the similarity, the larger the edge weight. The propagation decay of the path is calculated by multiplying the weights of all edges on the path, and the product value reflects the degree of information preservation throughout the path. Paths with large propagation decay indicate high reliability of the inference chain and minimal information loss.

[0091] The path with the greatest propagation attenuation is determined as the optimal diagnostic path, corresponding to the most reliable fault reasoning chain. The path selection process also considers the impact of path length, prioritizing shorter paths when propagation attenuation is similar. The optimal diagnostic path provides the reasoning basis for subsequent fault feature classification and probability calculation.

[0092] Fault feature classification is performed on the embedded representations of regular nodes based on the optimal diagnostic path. The classification process maps the embedded vectors to a predefined fault category space. Fault categories include major types such as mechanical faults, electrical faults, hydraulic faults, and control faults, each further subdivided into specific fault modes. A neural network classifier is used for classification, taking the embedded representations of regular nodes as input features and outputting the predicted probability of each fault category.

[0093] The predicted probability of fault categories is calculated through classification matrix operations. The parameters of the classification matrix are learned from the training data. Matrix operations project the high-dimensional embedding representation onto the output space corresponding to the number of fault categories, with each output dimension corresponding to a predicted score for a fault category. The predicted scores are normalized using a softmax function to obtain a probability distribution, and the sum of the probabilities of all categories equals one.

[0094] The posterior probability of the fault cause is calculated by combining equipment operational evidence, which includes multi-source evidence such as sensor measurement data, operation log records, and maintenance history information. The posterior probability is calculated using Bayesian inference, combining prior probability and likelihood probability. Prior probability is determined based on historical fault statistics, while likelihood probability is calculated based on the degree of matching between current operational evidence and the fault mode.

[0095] This paper analyzes the coupling relationship between posterior probability and causal strength, where causal strength reflects the degree of causal association between the cause and phenomenon of a failure. The modeling of this coupling relationship considers the moderating effect of causal strength on posterior probability; failure causes with higher causal strength receive higher posterior probability weights. The coupling coefficient is determined by analyzing the distribution of causal strength in historical failure cases; strong causal relationships correspond to high coupling coefficients, and weak causal relationships correspond to low coupling coefficients.

[0096] The predicted probability of fault categories is corrected by leveraging the coupling relationship between posterior probability and causal strength. The correction process involves a weighted fusion of the predicted and posterior probabilities. The fusion weights are dynamically adjusted based on the reliability and completeness of the evidence; the posterior probability has a higher weight when the evidence is sufficient and reliable, while the predicted probability has a higher weight when the evidence is insufficient. The corrected probability distribution more accurately reflects the current fault state of the equipment, providing a reliable basis for the final diagnostic decision.

[0097] In one optional implementation, the embedding representation of a regular node is obtained by performing a weight matrix operation on the neighborhood set of the regular nodes. A diagnostic path from the source node to the target node is constructed based on the embedding representation of the regular nodes. The optimal diagnostic path is determined based on the propagation attenuation of the edge weights in the diagnostic path, including:

[0098] A multidimensional feature set of a rule node and its k-order neighboring node set is obtained. The semantic similarity of the nodes is calculated based on the multidimensional feature set as the initial weight. The initial weight is adjusted based on the temporal co-occurrence information of the rule node and the k-order neighboring node set in historical diagnostic data. A propagation decay function is constructed based on the connection relationship between the rule node and the k-order neighboring node set in the physical topology. The adjusted weight is decayed using the propagation decay function to obtain the neighborhood weight matrix of the rule node.

[0099] The attention score between the regular node and its k-order neighbor node set is calculated based on the weight distribution of the neighborhood weight matrix. The embedded representation of the regular node is obtained by performing temporal modeling on the state of the regular node based on the attention score.

[0100] The embedding representation of the rule node is used as the starting state. The node transition probability is determined based on the similarity of the embedding representations between rule nodes. The search space is gradually expanded to the target node according to the node transition probability to construct a set of candidate diagnostic paths. The path with the largest propagation attenuation coefficient is selected as the optimal diagnostic path based on the embedding representations of adjacent nodes in the set of candidate diagnostic paths.

[0101] A graph structure representation of rule nodes is constructed, with each diagnostic rule treated as a node in the graph. Each node contains feature attribute information about the rule. These feature attributes include multi-dimensional information such as the keyword vector of the rule's condition part, the semantic representation of the conclusion part, the rule's confidence score, and the rule's support statistics. For each rule node, its k-order neighborhood node set is determined. The k-order neighborhood refers to all adjacent nodes reachable through k edges. First-order neighbors are directly connected nodes, second-order neighbors are nodes connected by two edges, and so on up to the k-order neighborhood. The determination of neighborhood nodes is based on multiple connection criteria, including logical relationships between rules, semantic similarity, and co-occurrence frequency.

[0102] To address the computational complexity issue caused by the exponential growth of the number of k-order neighborhood nodes with respect to k, this invention employs the following optimization strategy:

[0103] Multi-level neighborhood sampling: Importance sampling is performed on nodes in the k-th order neighborhood. The sampling rate decreases as k increases. First-order neighborhood: all nodes are retained; Second-order neighborhood: sampling rate p2 = min(50 / |N2|, 1), where |N2| is the size of the second-order neighborhood; K-th order neighborhood (k>2): sampling rate p k =min(50·0.5 (k-2) / |N k |,1). Node importance scoring: The sampling probability of node v is proportional to its importance, P(v)=0.4·sim(v0,v)+0.3·PR(v)+0.3·freq(v); where sim is the similarity, PR is the PageRank value, and freq is the historical activation frequency. Sparse matrix optimization: The weight matrix is ​​stored in Compressed Sparse Row (CSR) format, reducing the space complexity from O(n... 2The computational complexity is reduced to O(nnz), where nnz is the number of non-zero elements. The computational tasks are layered: offline computation: pre-computation of basic embedding vectors for rule nodes, knowledge graph structure analysis; near real-time computation: state feature update, initial screening of candidate fault types; real-time computation: active path evaluation, compensation parameter calculation. Dimensionality reduction: high-dimensional embedding vectors are reduced to 32-64 dimensions through principal component analysis.

[0104] Through the above optimization measures, the system's single inference time on a standard industrial control computer is controlled within 50 milliseconds, meeting the requirements for real-time control of cold rolling mills.

[0105] In actual testing, this system was run on an industrial control computer with an Intel Core i5 processor and 8GB of memory. It performed inference on a knowledge graph containing over 5000 nodes. The average time for the complete inference process was 37.2 milliseconds, with a peak memory usage of 128MB and a maximum CPU utilization of 25%. Even in the most complex fault diagnosis scenarios, the diagnostic latency was kept below 100ms, far lower than the response time requirements of the cold rolling process (typically 500ms).

[0106] Multidimensional feature vectors are extracted from rule nodes and their k-order neighbor nodes. These multidimensional features include word frequency statistics of the rule text, distributed semantic representation features of the rule, and structured features of the rule in the knowledge graph. Word frequency statistics are obtained by counting the frequency of keywords after word segmentation of the rule text, and the importance weights of words are calculated using the word frequency-inverse document frequency method. The distributed semantic representation is obtained by encoding the rule text using a pre-trained language model to obtain a high-dimensional vector representation that captures the deep semantic information of the rule. Structured features reflect the positional information and relational strength of the rule within the entire knowledge system.

[0107] The semantic similarity between a rule node and its k-order neighbor nodes is used as the initial weight. The semantic similarity is calculated using the cosine similarity method. The multidimensional feature vectors of the rule node and its neighbor nodes are normalized, and the inner product of the two normalized vectors is calculated to obtain the cosine similarity value. The cosine similarity value ranges from -1 to +1; the closer the value is to +1, the higher the semantic similarity, and the closer the value is to -1, the greater the semantic difference. The cosine similarity value is mapped to the interval between zero and one as the initial weight. The mapping method uses a linear transformation to convert negative values ​​into positive values ​​close to zero.

[0108] This study obtains the temporal co-occurrence information of rule nodes and their k-order neighbor nodes in historical diagnostic data. This information reflects the collaborative occurrence patterns of rules during the actual diagnostic process. By analyzing the activation time series of rules in historical diagnostic cases, the co-occurrence frequency of different rules within a time window is statistically analyzed. The size of the time window is determined based on the temporal characteristics of the diagnostic process, typically set to the typical duration of the diagnostic process. The co-occurrence frequency is calculated using a sliding time window method, moving the window across the entire historical data timeline to count the number of times rules co-occur.

[0109] The initial weights calculated from semantic similarity are adjusted using temporal co-occurrence information, based on the positive correlation between co-occurrence frequency and weight. Weights between rule pairs with high co-occurrence frequency are positively adjusted, while weights between rule pairs with low co-occurrence frequency are negatively adjusted. The adjustment magnitude is correlated with the statistical significance of the co-occurrence frequency, and the statistical reliability of the co-occurrence pattern is assessed using a chi-square test. The adjusted weights more accurately reflect the strength of association between rules in practical applications.

[0110] This analysis examines the connections between rule nodes and their k-order neighbor nodes within the physical topology, which reflects the logical dependencies and reasoning chains between rules. The directionality and strength of these connections are determined based on the logical implications between the rule's preconditions and conclusions. Rules with similar preconditions have stronger connections, as do rules with similar conclusions. The strength of these connections is quantified by analyzing the logical overlap between rules; rules with high overlap have stronger connections.

[0111] A distance-based propagation decay function is constructed, which describes how the weight decreases as the distance between nodes increases. Distance is defined as the shortest path length from a regular node to its neighboring nodes, and the path length is calculated using the number of edges. The decay function adopts an exponential decay form, and the decay coefficient is determined based on the density and connectivity characteristics of the regular graph. The greater the distance between neighboring nodes, the smaller their influence on the current regular node; this distance effect is reflected in the decay function.

[0112] The adjusted weights are attenuated using a propagation attenuation function to obtain the final weight values ​​that take distance into account. The attenuation calculation multiplies the adjusted weights by a attenuation coefficient corresponding to the distance, with the coefficient decreasing exponentially with distance. The attenuation coefficient for first-order neighbors is close to one, and for k-order neighbors it is close to zero. The attenuated weights form the neighborhood weight matrix of the regular nodes, where rows correspond to the current regular node, columns correspond to its k-order neighbors, and matrix elements are the attenuated weight values.

[0113] The attention score between a rule node and its k-th order neighbors is calculated based on the weight distribution of the neighborhood weight matrix. The attention score reflects the degree of attention the current rule node pays to different neighbors. The attention score is calculated using the softmax normalization method, which normalizes the weight values ​​of the corresponding rows in the neighborhood weight matrix after an exponential transformation. The sum of the normalized attention scores equals one; a larger score indicates a more important neighbor.

[0114] Attention scores are used to perform temporal modeling of the state of regular nodes, capturing the state evolution of regular nodes at different time steps. The state of a regular node is represented as a weighted combination of its feature vector and the feature vectors of its neighboring nodes, with the weights being the corresponding attention scores. Temporal modeling employs a recurrent neural network architecture, using the node state of the previous time step as the input for the computation of the current time step. The cyclic update process repeats for a predefined number of time steps, and the node state of the final time step serves as the embedded representation of the regular node.

[0115] The embedding representation of a regular node integrates the node's own features and neighborhood information, forming a high-dimensional vector representation. The dimension of the embedding representation is determined based on application requirements and computational resources, typically ranging from tens to hundreds of dimensions. Each dimension of the embedding vector captures a semantic feature of the regular node, and the entire vector comprehensively represents the semantic and structural characteristics of the regular node.

[0116] When calculating the node transition probability, this invention not only considers the similarity of the embedding vectors but also explicitly models the causal direction of the rule. Specifically, the transition probability P(j|i) is calculated using the following formula:

[0117] P(j|i)=softmax(sim(v_ i ,v_ j )·dir(i,j)·α(i,j));

[0118] Where, sim(v) i ,v j ) is the cosine similarity of the node embedding vectors; dir(i,j) is the causal direction consistency function, which takes the value of 1 when the transition direction from node i to node j is consistent with the predefined causal direction in the knowledge graph, and takes the value of exp(-λ·θ) when they are inconsistent, where θ is the path length between the two nodes in the causal graph and λ is the directional penalty coefficient; α(i,j) is the modulation factor based on domain knowledge.

[0119] By introducing a causal direction consistency function, the diagnostic reasoning process is ensured to strictly follow the causal direction from cause to effect, effectively avoiding the problem of causal reversal. For example, in the rolling mill knowledge graph, there is a causal relationship between the nodes 'insufficient bearing lubrication' and 'increased bearing temperature', with the direction 'insufficient bearing lubrication → increased bearing temperature'. When calculating the transition probability from 'insufficient bearing lubrication' to 'increased bearing temperature', dir(i,j)=1; while when calculating the reverse transition probability, dir(i,j)=exp(-0.8·1)≈0.45, significantly reducing the transition probability that violates the causal direction and ensuring that the reasoning path conforms to the physical causal relationship.

[0120] The embedded representations of regular nodes are used as the starting state for diagnostic path search, corresponding to the initial symptoms or observations in the diagnostic process. Node transition probabilities are calculated based on the similarity of the embedded representations between regular nodes, using Euclidean distance or cosine similarity methods. Node pairs that are closer or have higher similarity have higher transition probabilities. The transition probabilities are normalized using a softmax function to ensure the effectiveness of the probability distribution.

[0121] The search space is expanded incrementally based on node transition probabilities until the target node is reached. The search process employs either a breadth-first or depth-first strategy. Breadth-first search expands nodes sequentially according to their transition probabilities, ensuring that high-probability paths are explored first. Depth-first search delves deeper along a single path until the target is reached or a dead end is encountered. A set of candidate paths is maintained during the search, with each path recording complete path information from the starting node to the current node.

[0122] A set of candidate diagnostic paths is constructed from the starting node to the target node. This set contains all diagnostic reasoning chains. Each candidate path represents a diagnostic logic, and the sequence of nodes on the path corresponds to the reasoning steps in the diagnostic process. The generation of candidate paths takes into account path length constraints to avoid generating excessively long and invalid paths. The path length constraint is determined based on the complexity of the diagnostic problem and the required depth of reasoning.

[0123] In constructing the candidate diagnostic path set, this invention implements a complete loop detection and pruning strategy to ensure that the search process does not get stuck in a loop. Specifically, a three-color labeling method is used for loop detection: nodes are marked as unvisited (WHITE), being visited (GRAY), and visited (BLACK). When a GRAY node is encountered during the search, it indicates that a loop has been detected. The set of nodes S for the current path is maintained. For each node v to be expanded, it is checked whether v is already in S. If it is, the path is pruned. A hierarchical pruning strategy is implemented, which divides the nodes into levels according to their hierarchical attributes, allowing only transfers from lower levels to higher levels or within the same level.

[0124] Set the maximum path length limit MaxLength=log(N)·C, where N is the number of nodes in the graph and C is an empirical coefficient. Paths exceeding this length are automatically pruned. For necessary functional loops, loop unrolling technology is used to convert the loop into a finite-step feedforward structure. Each node is given a time step attribute t to ensure that t is monotonically increasing. Set a minimum threshold MinDecay for the propagation attenuation coefficient. Paths below this value are automatically pruned to avoid abnormal attenuation values.

[0125] For example, when detecting a potential loop of 'roll vibration - bearing damage - insufficient lubrication - increased friction - increased temperature - thermal expansion - reduced clearance - intensified vibration', the system identifies the repeated occurrence of the 'roll vibration' node through path history records, immediately terminates further expansion of the path, and records the loop characteristics. For necessary feedback loops, the system expands them into 't1-roll vibration → t2-bearing damage → ... → t8-intensified vibration (t8)' by adding a time step attribute, ensuring that the inference path is acyclic in the expanded time dimension.

[0126] The propagation attenuation coefficient between adjacent nodes in the candidate path is calculated. This coefficient reflects the degree to which information is preserved during path propagation. The calculation of the attenuation coefficient is based on the similarity of the embedding representations of adjacent nodes and the strength of the logical association between nodes. Node pairs with high similarity and strong logical association have higher propagation attenuation coefficients. The overall propagation attenuation coefficient of the path is the product of the propagation attenuation coefficients of all adjacent node pairs on the path; this product reflects the information propagation efficiency of the entire path.

[0127] The path with the maximum propagation attenuation coefficient is selected as the optimal diagnostic path. The maximum propagation attenuation coefficient indicates that the path suffers the least loss during information propagation. The optimal path corresponds to the most reliable diagnostic reasoning chain, providing the most accurate diagnostic results. The path selection process also considers path length penalties; when propagation attenuation coefficients are similar, shorter paths are preferred to avoid unnecessary complex reasoning processes.

[0128] When selecting the optimal diagnostic path, this invention also considers the impact of path length and introduces an adaptive length penalty coefficient. The length penalty coefficient LP(l) is calculated using the following formula:

[0129] LP(l)=exp(-β·(ll opt ) 2 );

[0130] Where l is the current path length; opt The optimal path length for fault type k is obtained through statistical analysis of historical diagnostic data. opt (k) = μk, where μk is the average optimal path length for fault type k; β is the penalty intensity coefficient, β(k) = 1 / (σk)2 ·2), where σk is the standard deviation of the path length for fault type k.

[0131] Different types of faults opt The β parameter is automatically determined through offline analysis of historical diagnostic data, avoiding subjective settings. The system continuously updates the parameters for each fault type using a Bayesian optimization method.

[0132] P(β|D)∝P(D|β)·P(β);

[0133] Where D represents the diagnostic results data, and P(β) represents the prior distribution.

[0134] The final path scoring function takes into account both propagation attenuation and length penalty:

[0135] S(path) = TD(path)·LP(l);

[0136] Where TD(path) is the propagation attenuation coefficient of the path.

[0137] For example, for the 'bearing failure' type, historical diagnostic data shows that its optimal path length has a mean of μ=4.2 and a standard deviation of σ=0.8, then β is calculated as 1 / (0.8). 2 ·2)=0.78. When evaluating a candidate path of length 5, its length penalty coefficient LP(5)=exp(-0.78·(5-4.2) 2 =exp(-0.78·0.64)=exp(-0.5)≈0.61. This coefficient is multiplied by the propagation attenuation coefficient of the path to obtain the comprehensive score of the path, ensuring the objective quantification of the length penalty in the path selection process.

[0138] In one optional implementation, a compensation optimization strategy is generated based on the analysis results of the fault diagnosis. A fuzzy correlation mapping is established between the compensation optimization strategy and the real-time operating data. The optimal compensation parameters are calculated based on the fuzzy correlation mapping, including:

[0139] The analysis results of fault diagnosis are used to perform fuzzy classification modeling to construct a fuzzy correlation matrix; the membership degree value of each fault object is calculated based on the fuzzy correlation matrix, the influence weight of the fault object is determined based on the membership degree value, the influence weight is compared with a preset weight threshold to obtain the compensation priority, and the local compensation strategy corresponding to each fault object is generated based on the compensation priority.

[0140] The association strength between adjacent fault objects in the fuzzy association matrix is ​​analyzed using dynamic programming. The association strength is combined with the compensation parameters of the local compensation strategy to calculate the membership degree of the propagation influence. The compensation parameters of the local compensation strategy are recursively optimized based on the membership degree of the propagation influence to generate a compensation optimization strategy.

[0141] Based on the state characteristics of real-time running data and the compensation optimization strategy, a dynamic compensation fuzzy rule base is constructed. The comprehensive activation intensity of the fuzzy rules is calculated through adaptive fuzzy inference. The fuzzy rule base is then defuzzified using the comprehensive activation intensity to obtain the optimal compensation parameters.

[0142] like Figure 2 As shown, the method includes:

[0143] The system acquires analysis results from the fault diagnosis module, including key information such as the detected fault type identifier, fault severity level, fault location coordinates, and fault duration. Fault types are categorized by equipment component, including main categories such as motor faults, sensor faults, controller faults, and mechanical transmission faults, with each category further subdivided into specific fault modes. Fault severity is assessed using a five-level scale, ranging from minor anomalies to severe faults, corresponding to numerical ratings from one to five. Fault location information is described using multiple dimensions, including equipment number, component identifier, and spatial coordinates, ensuring the accuracy and uniqueness of fault location.

[0144] A fuzzy classification model for fault objects is established. Gaussian membership functions are used to fuzzify the severity of faults. The center point of the Gaussian function is set at the median of each severity level, and the standard deviation parameter is determined based on the distribution characteristics of historical fault data. The center point of the membership function for minor faults is set near level one, the center point for severe faults is set near level five, and the membership function for moderate faults covers the intermediate level range. Fuzzification transforms discrete fault levels into a continuous membership degree distribution, facilitating subsequent fuzzy inference calculations.

[0145] A fuzzy correlation matrix reflecting the relationships between faulty objects is constructed. Rows and columns of the matrix correspond to different faulty objects, and matrix elements represent the correlation strength between corresponding faulty objects. The correlation strength is calculated based on multiple dimensions, including physical connections, functional dependencies, and temporal correlations. Physical connections are determined by analyzing the mechanical structure and electrical connections of the equipment; directly connected components have higher correlation strengths. Functional dependencies consider the logical path of fault propagation; the degree of impact of upstream faults on downstream equipment determines the correlation strength. Temporal correlations are calculated by analyzing the correlation of the occurrence time series of different faults in historical fault data.

[0146] Calculate the membership degree value of each faulty object in its current state. The membership degree value reflects the degree to which the faulty object belongs to a certain fault severity level. Using the currently detected fault parameter values ​​and a pre-established Gaussian membership function, calculate the membership degree distribution of the faulty object at each severity level. The membership degree value is calculated using the standard Gaussian function calculation method, normalizing the distance between the fault parameter value and the center point of the membership function; the closer the distance, the higher the membership degree.

[0147] The impact weight of a faulty object is calculated based on its membership degree value. This impact weight comprehensively considers both the severity of the fault's membership degree and its importance within the equipment. The equipment importance weight is pre-set based on the faulty object's criticality in the overall equipment operation, with core control components having a higher weight than auxiliary components. The comprehensive impact weight is obtained by multiplying the maximum membership degree value of the fault's severity by the equipment importance weight. This weight reflects the degree of impact of the faulty object on the overall equipment operation.

[0148] Preset weight thresholds are used to classify compensation priorities. The calculated impact weights are compared with the preset thresholds to determine the compensation priority of the fault. The preset thresholds are set according to the equipment's operating requirements and fault tolerance, typically using high, medium, and low thresholds to classify faults into four priorities. Faults with impact weights exceeding the high threshold are assigned the highest priority and require immediate compensation. Faults with impact weights between the high and medium thresholds are assigned high priority, faults with impact weights between the medium and low thresholds are assigned medium priority, and faults with impact weights below the low threshold are assigned low priority.

[0149] Based on the compensation priority, corresponding local compensation strategies are generated for each fault object. These strategies include specific details such as compensation method selection, compensation parameter settings, and compensation execution sequence. High-priority faults are compensated proactively by adjusting control parameters, switching to backup equipment, and reducing operating load. High-priority faults are compensated predictively by adjusting relevant parameters in advance based on fault development trends. Medium-priority faults are compensated passively by adjusting as the fault's impact becomes apparent. Low-priority faults are compensated by monitoring, continuously monitoring fault status changes but not performing compensation actions immediately.

[0150] Dynamic programming is used to analyze the correlation strength between adjacent faulty objects in a fuzzy correlation matrix. The state in dynamic programming is defined as the current compensation state of the faulty object, and the decision variables are the adjustments to the compensation parameters. A state transition equation is established to describe the propagation process of influence between faulty objects, and the cost function of the state transition considers a comprehensive evaluation of the compensation effect and cost. The optimal path from the initial state to the target state is found through layer-by-layer recursive calculation, and the decision sequence corresponding to the optimal path is the optimal combination of compensation parameters.

[0151] The membership degree of propagation influence between faulty objects is calculated. This membership degree reflects the degree to which a compensation action of one faulty object affects its neighboring faulty objects. The calculation of propagation influence membership degree combines the association strength in the fuzzy association matrix and the magnitude of the compensation parameter of the local compensation strategy. Higher association strength and larger compensation parameter magnitudes correspond to higher propagation influence membership degrees, indicating that the compensation action has a greater impact on neighboring faulty objects. The propagation influence membership degree is quantified using a trapezoidal membership function, with the function parameters determined based on statistical analysis of historical compensation effect data.

[0152] The compensation parameters of the local compensation strategy are recursively optimized, taking into account the membership degree of the propagation effect between fault objects. The recursive calculation starts with the highest priority fault and propagates the optimization results downwards level by level. The compensation parameter optimization for each fault object considers the propagation effect of its directly related fault objects, using the membership degree of the propagation effect as a weighting factor to adjust the original compensation parameters. The recursive process is repeated until the compensation parameters of all fault objects converge to stable values. The convergence criterion is that the parameter change is less than a set threshold in two consecutive iterations.

[0153] The recursively optimized local compensation strategies are integrated to generate a global compensation optimization strategy. This global strategy coordinates the compensation actions of each faulty object to avoid conflicts. The integration process uses a weighted average method, with weights determined based on the compensation priority and propagation influence membership of the faulty object. Faulty objects with higher priority and greater propagation influence receive higher weights, and their compensation strategies dominate the global strategy.

[0154] Acquire real-time status data during equipment operation, including key parameters such as sensor measurements, actuator operating status, and controller output signals. Establish a fuzzy correlation mapping between the real-time status data and compensation optimization strategies. This mapping describes the correspondence between changes in status data and adjustments to the compensation strategy. Fuzzy inference rules are used to express the mapping relationship, where the antecedent of the rule is a fuzzy description of the status data, and the consequent is the corresponding compensation strategy adjustment suggestion.

[0155] When integrating fuzzy sets and probability theory, the Zadeh probability-probability consistency principle is adopted as the theoretical basis. Specifically, when performing fuzzy classification modeling on fault objects, the membership degree value of the fault object in each fuzzy set is first calculated using the Gaussian membership function μA(x). Then, the membership degree is mapped to conditional probability using the transformation function P(A|x)=μA(x) / ∑μi(x), where ∑μi(x) is the sum of the membership degree values ​​of all possible states.

[0156] When constructing the fuzzy correlation matrix, Dempster-Shafer evidence theory is adopted as the uncertainty fusion framework. A confidence function Bel(A) = ∑m(B)(B⊆A) and a likelihood function Pl(A) = 1 - Bel(Ā) = ∑m(B)(B∩A≠∅) are introduced, where m(B) is the basic probability assignment function. The fusion of evidence from different sources is achieved through the Dempster combination rule m(C) = [∑m1(A)·m2(B)] / [1-K](A∩B=C), ensuring a rigorous mathematical unity between fuzzy reasoning and probabilistic reasoning.

[0157] For example, for abnormal roll vibration faults, the membership degree calculated using Gaussian membership functions is 0.85 in the 'severe' fuzzy subset, 0.35 in the 'medium' fuzzy subset, and 0.05 in the 'slight' fuzzy subset. Through normalization transformation, the corresponding conditional probabilities are 0.68, 0.28, and 0.04, respectively. This transformation ensures a rigorous mathematical conversion between fuzziness and probability, avoiding theoretical inconsistencies caused by direct mixing.

[0158] A dynamic compensation fuzzy rule base is constructed, containing a complete set of mapping rules between state features and compensation strategies. Rule creation is based on expert knowledge and the mining and analysis of historical operational data to ensure the completeness and accuracy of the rules. Each rule includes a confidence parameter, reflecting the rule's reliability, which is determined based on the rule's validation results in historical data.

[0159] The comprehensive activation intensity of each fuzzy rule is calculated through adaptive fuzzy inference. The adaptive mechanism dynamically adjusts the activation level of the rule based on the current running state. The calculation of activation intensity considers multiple factors, including the matching degree between the rule's antecedent and the current state, the rule's confidence, and the rule's historical usage effect. The matching degree is determined by calculating the similarity between the antecedent conditions and the features of the current state; the higher the similarity, the greater the activation intensity.

[0160] The centroid defuzzification method utilizes the comprehensive activation intensity to perform centroid defuzzification on the fuzzy rule base. It calculates a weighted average of the activated rule consequents. The weight of the weighted average is the comprehensive activation intensity of each rule; rules with higher activation intensity contribute more to the final output. The output of centroid defuzzification is the optimal value of each compensation parameter, including the specific values ​​of key parameters such as compensation amplitude, compensation frequency, and compensation phase.

[0161] In one optional implementation, a dynamically compensated fuzzy rule base is constructed based on the state characteristics of real-time running data and the compensation optimization strategy. The comprehensive activation intensity of the fuzzy rules is calculated through adaptive fuzzy inference. The centroid defuzzification of the fuzzy rule base is then performed using the comprehensive activation intensity, including:

[0162] Based on the state characteristics and quantitative indicators of the compensation optimization strategy in the real-time running data, an influence matrix between the quantitative indicators is constructed. The influence matrix is ​​used to calculate the membership distribution of the state feature vector and the compensation strategy vector in different fuzzy subspaces. Based on the membership distribution, a fuzzy rule base for dynamic compensation is generated.

[0163] Based on the influence matrix, the initial weight factors of each rule are set, the initial weight factors of the rules in the fuzzy rule base are iteratively optimized, and the optimized weight factors are used to perform fuzzy operator operations with the state feature membership values ​​of the corresponding rules in the fuzzy rule base to obtain the comprehensive activation intensity that represents the importance of the rule.

[0164] Using the comprehensive activation intensity as a weighting coefficient, the centroid defuzzification operation is performed on the compensation strategy of the rule consequent in the fuzzy rule base to obtain the output value of the optimal compensation strategy. The gradient direction of the optimal compensation strategy is calculated based on the steepest descent method. The output value is iteratively updated along the direction that makes the control error represented by the influence matrix decrease the fastest until the optimal solution is reached.

[0165] Real-time status data during equipment operation is acquired, including physical parameters such as operating temperature sensor measurements, vibration accelerometer output signals, load current amplitude detected by current transformers, and encoder feedback speed signals. This raw sensor data is converted into digital signals via an analog-to-digital converter, filtered to remove high-frequency noise interference, and then smoothed using a moving average algorithm to obtain stable and reliable status characteristic values. Simultaneously, quantitative indicators related to the compensation control strategy are collected, including key control parameters such as the current compensator output voltage amplitude, compensation signal frequency parameters, phase compensation angle, and gain adjustment coefficient. These parameters reflect the specific execution status of the compensation strategy.

[0166] A correlation matrix is ​​established between state characteristics and quantitative indicators of compensation strategies. The influence of each state characteristic variable and each compensation strategy parameter is quantified by calculating the Pearson correlation coefficient. The correlation coefficient calculation is based on statistical analysis of historical operating data, collecting data samples over a sufficiently long period to ensure the reliability of the statistical results. The absolute values ​​of the correlation coefficients are used as elements of the influence matrix, with the values ​​normalized to the interval between zero and one. The rows of the influence matrix correspond to state characteristic variables, and the columns correspond to quantitative indicators of compensation strategies. Each element in the matrix represents the strength of the influence of the corresponding row variable on the column variable.

[0167] A triangular membership function is used to fuzzify the state feature variables, dividing the value range of each continuous variable into five fuzzy subsets: low, low-medium, medium, high-medium, and high. The positions of the vertices of the triangular membership functions are determined based on the historical statistical distribution of the variables. The vertices of the low subset are set near the minimum value of the variable, the vertices of the high subset are set near the maximum value of the variable, and the vertices of the medium subset are evenly distributed within the value range. The membership degree of the current state feature value in each fuzzy subset is calculated; the membership degree reflects the degree to which the value belongs to the corresponding fuzzy subset. The same fuzzification method is used for the quantitative indicators of the compensation strategy, establishing corresponding fuzzy subsets and membership functions.

[0168] A fuzzy inference rule base is established based on the fuzzified state characteristics and compensation strategies, expressing rules in the form of "if-then" conditional statements. The antecedent of a rule contains a combination of one or more fuzzy variables representing state characteristics, connected by the "and" logical connector. The consequent of a rule is the corresponding compensation strategy output, indicating the compensation measure to be taken when the antecedent conditions are met. A typical rule form is "If the temperature is high, the vibration is moderate, and the current is low, then the compensation amplitude is large and the compensation frequency is high." A complete rule set is established based on expert experience and historical data analysis results to ensure coverage of all input state combinations.

[0169] An initial weight factor is assigned to each rule in the fuzzy rule base. The initial value of the weight factor is determined based on the average value of the relevant elements in the influence matrix. The average influence of each state feature in the rule's antecedent on each compensation strategy in the rule's consequent is calculated, and this average value is used as the initial weight of the rule. The initial weight factors of all rules are normalized to ensure that the sum of the weights equals one. An objective function for weight optimization is established, with the sum of squared control errors as the optimization objective. The weight factors of each rule are iteratively adjusted using a gradient descent algorithm.

[0170] In the gradient descent algorithm, the direction and step size of weight adjustments are calculated. The gradient direction is obtained by calculating the partial derivatives of the objective function with respect to each weight factor using numerical differentiation. A small perturbation is applied to the current weights, and the difference in the objective function value before and after the perturbation is calculated. The difference is divided by the perturbation amount to obtain an approximate partial derivative. The gradient direction is a vector composed of the partial derivatives. The step size is adaptively determined based on the gradient magnitude; a smaller step size is used when the gradient is large to avoid oscillations, and a larger step size is used when the gradient is small to accelerate convergence. The weight update process is repeated until the objective function converges or the maximum number of iterations is reached.

[0171] The optimized weighting factors are used in fuzzy operator operations with the membership values ​​of each state feature in the antecedent of the fuzzy rule to calculate the comprehensive activation intensity of each rule. The fuzzy operator operation uses the minimum value operator, taking the minimum value of the membership degrees of all state features in the antecedent of the rule as the antecedent activation degree. The antecedent activation degree is then multiplied by the optimized weighting factor of the rule to obtain the comprehensive activation intensity of the rule. The comprehensive activation intensity reflects the importance and contribution of the rule in the current input state; a higher value indicates a more important rule.

[0172] The final compensation strategy output value is calculated by performing centroid defuzzification. The numerator is the sum of the compensation strategy values ​​of each rule consequent multiplied by the overall activation intensity of the corresponding rule. The denominator is the sum of the overall activation intensities of all rules, and the numerator is divided by the denominator to obtain the compensation strategy output value. This weighted average calculation method ensures that rules with higher activation intensities have a greater impact on the final output, achieving intelligent decision-making based on rule importance.

[0173] A control error function is established to further optimize the compensation strategy output. The error function is defined as the square of the deviation between the actual control effect and the desired control objective. The gradient of the error function with respect to the compensation strategy output is calculated using numerical differentiation, with the central difference scheme employed to improve accuracy. Small positive and negative perturbations are introduced near the current output value, and the corresponding error function values ​​are calculated for each. The gradient value is then calculated using the central difference formula.

[0174] The compensation strategy output value is iteratively updated along the gradient descent direction, with each iteration's update equal to the negative gradient direction multiplied by the learning rate. The learning rate is adaptively adjusted based on the current gradient magnitude and historical convergence data, employing momentum to accelerate the convergence process and avoid local maxima. Convergence criteria are set, including error change less than a set threshold during consecutive iterations, gradient magnitude less than a threshold, or reaching the maximum number of iterations. Iteration stops when the convergence criteria are met, and the optimal compensation strategy value is output.

[0175] The calculated optimal compensation strategy values ​​are converted into specific control commands and sent to the actuators, including voltage adjustment commands, frequency setting commands, and phase adjustment commands for the compensator. These control commands are transmitted to the lower-level controller via a digital communication interface. The lower-level controller then drives the corresponding actuators to complete the compensation action according to the commands. Simultaneously, the weight factors and influence matrices of the rules in the fuzzy rule base are updated to prepare for the calculation in the next control cycle, achieving dynamic adaptive adjustment of the compensation strategy.

[0176] A second aspect of the present invention provides a real-time monitoring and dynamic compensation control early warning system for the micro-displacement of rolls in a cold rolling mill, comprising:

[0177] The first unit is used to acquire real-time operating data of the cold rolling mill rolls. It uses infrared thermal imaging to monitor the temperature field distribution of the rolls in real time, establishes a dynamic mapping relationship between roll displacement and temperature gradient, calculates the thermal deformation of the rolls under different temperature fields, compensates for the displacement drift caused by the temperature field, and obtains the roll displacement data.

[0178] The second unit is used to construct a knowledge graph of the rolling mill equipment domain. Based on the knowledge graph of the rolling mill equipment domain, an equipment diagnosis rule base is constructed using ontology modeling technology. The feature data corresponding to the equipment diagnosis rule base is mapped to the knowledge graph of the rolling mill equipment domain using graph embedding learning method to perform fault diagnosis. The roll displacement data is used to perform deep representation learning of fault features.

[0179] The third unit is used to generate a compensation optimization strategy based on the analysis results of the fault diagnosis, establish a fuzzy correlation mapping between the compensation optimization strategy and the real-time operation data, calculate the optimal compensation parameters according to the fuzzy correlation mapping, and iteratively optimize the compensation strategy in combination with the knowledge graph of the rolling mill equipment domain.

[0180] A third aspect of the present invention provides an electronic device, comprising:

[0181] processor;

[0182] Memory used to store processor-executable instructions;

[0183] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0184] 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.

[0185] 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.

[0186] 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. A method for real-time monitoring and dynamic compensation control and early warning of micro-displacement of cold rolling mill rolls, characterized in that, include: Real-time operating data of cold rolling mill rolls is acquired. Infrared thermal imaging is used to monitor the roll temperature field distribution in real time, establishing a dynamic mapping relationship between roll displacement and temperature gradient. The thermal deformation of the rolls under different temperature fields is calculated, and displacement drift caused by the temperature field is compensated for, resulting in roll displacement data, including: The temperature field distribution data of the roll surface collected by the infrared thermal imaging system is divided into regions according to the degree of temperature gradient change, and a temperature sensor array is arranged in the divided regions according to the rate of temperature gradient change. The temperature gradient change rate between adjacent monitoring points is calculated using the temperature data collected by the temperature sensor array. The temperature gradient change rate is used as a boundary condition to solve the heat conduction characteristic equation to obtain the main heat conduction path inside the roll. The main heat conduction path is used as a temperature transfer characteristic parameter to establish a dynamic mapping relationship between roll displacement and temperature gradient with the temperature field distribution data on the roll surface. The roll thermal deformation is calculated based on the dynamic mapping relationship. The compensation coefficient is determined by the spatial distribution characteristics of the main heat conduction path. The product of the compensation coefficient and the roll thermal deformation is used as the compensation amount to obtain the roll displacement data. A knowledge graph of the rolling mill equipment domain is constructed. Based on the knowledge graph, an equipment diagnosis rule base is constructed using ontology modeling technology. A graph embedding learning method is used to map the feature data corresponding to the equipment diagnosis rule base onto the knowledge graph of the rolling mill equipment domain for fault diagnosis. The roll displacement data is used for deep representation learning of fault features. Based on the analysis results of the fault diagnosis, a compensation optimization strategy is generated. The compensation optimization strategy is then linked to the real-time operating data to establish a fuzzy correlation mapping. The optimal compensation parameters are calculated based on the fuzzy correlation mapping. Finally, the compensation strategy is iteratively optimized using the knowledge graph of the rolling mill equipment domain.

2. The method according to claim 1, characterized in that, Based on the knowledge graph of the rolling mill equipment domain, an equipment diagnosis rule base is constructed using ontology modeling technology. Then, a graph embedding learning method is used to map the feature data corresponding to the equipment diagnosis rule base onto the rolling mill equipment domain knowledge graph for fault diagnosis, including: Based on the hierarchical association rules between the knowledge graph analysis components of the rolling mill equipment domain, an equipment diagnosis rule base is constructed by calculating the probability confidence of the hierarchical association rules. The uncertainty of the rules in the equipment diagnosis rule base is quantitatively evaluated, and the rule reliability index is obtained based on the correlation strength between the probability of the rule result and the uncertainty of the rule. The rules in the device diagnostic rule base that meet the rule reliability index are mapped to a low-dimensional vector space. The embedding representation of the rule node is obtained by performing a weight matrix operation on the neighborhood set of the rule node. A diagnostic path from the source node to the target node is constructed based on the embedding representation of the rule node. The optimal diagnostic path is determined based on the propagation decay of the edge weights in the diagnostic path. The embedded representation of the rule node is classified according to the optimal diagnostic path. The predicted probability of the fault category is obtained by the classification matrix operation. The posterior probability of the fault cause is calculated by combining the equipment operation evidence. The predicted probability of the fault category is corrected according to the coupling relationship between the posterior probability and the causal strength. The equipment fault diagnosis is realized based on the corrected predicted probability of the fault category.

3. The method according to claim 2, characterized in that, The embedding representation of a regular node is obtained by performing a weight matrix operation on the neighborhood set of the regular nodes. A diagnostic path from the source node to the target node is constructed based on the embedding representation of the regular nodes. The optimal diagnostic path is determined based on the propagation decay of the edge weights in the diagnostic path, including: A multidimensional feature set of a rule node and its k-order neighboring node set is obtained. The semantic similarity of the nodes is calculated based on the multidimensional feature set as the initial weight. The initial weight is adjusted based on the temporal co-occurrence information of the rule node and the k-order neighboring node set in historical diagnostic data. A propagation decay function is constructed based on the connection relationship between the rule node and the k-order neighboring node set in the physical topology. The adjusted weight is decayed using the propagation decay function to obtain the neighborhood weight matrix of the rule node. The attention score between the regular node and its k-order neighbor node set is calculated based on the weight distribution of the neighborhood weight matrix. The embedded representation of the regular node is obtained by performing temporal modeling on the state of the regular node based on the attention score. The embedding representation of the rule node is used as the starting state. The node transition probability is determined based on the similarity of the embedding representations between rule nodes. The search space is gradually expanded to the target node according to the node transition probability to construct a set of candidate diagnostic paths. The path with the largest propagation attenuation coefficient is selected as the optimal diagnostic path based on the embedding representations of adjacent nodes in the set of candidate diagnostic paths.

4. The method according to claim 1, characterized in that, Based on the analysis results of the fault diagnosis, a compensation optimization strategy is generated. A fuzzy correlation mapping is established between the compensation optimization strategy and the real-time operating data. The optimal compensation parameters are calculated based on the fuzzy correlation mapping, including: The analysis results of fault diagnosis are used to perform fuzzy classification modeling to construct a fuzzy correlation matrix; the membership degree value of each fault object is calculated based on the fuzzy correlation matrix, the influence weight of the fault object is determined based on the membership degree value, the influence weight is compared with a preset weight threshold to obtain the compensation priority, and the local compensation strategy corresponding to each fault object is generated based on the compensation priority. The association strength between adjacent fault objects in the fuzzy association matrix is ​​analyzed using dynamic programming. The association strength is combined with the compensation parameters of the local compensation strategy to calculate the membership degree of the propagation influence. The compensation parameters of the local compensation strategy are recursively optimized based on the membership degree of the propagation influence to generate a compensation optimization strategy. Based on the state characteristics of real-time running data and the compensation optimization strategy, a dynamic compensation fuzzy rule base is constructed. The comprehensive activation intensity of the fuzzy rules is calculated through adaptive fuzzy inference. The fuzzy rule base is then defuzzified using the comprehensive activation intensity to obtain the optimal compensation parameters.

5. The method according to claim 4, characterized in that, Based on the state characteristics of real-time operating data and the aforementioned compensation optimization strategy, a dynamic compensation fuzzy rule base is constructed. The comprehensive activation intensity of the fuzzy rules is calculated through adaptive fuzzy inference. The centroid defuzzification of the fuzzy rule base is then performed using the comprehensive activation intensity, including: Based on the state characteristics and quantitative indicators of the compensation optimization strategy in the real-time running data, an influence matrix between the quantitative indicators is constructed. The influence matrix is ​​used to calculate the membership distribution of the state feature vector and the compensation strategy vector in different fuzzy subspaces. Based on the membership distribution, a fuzzy rule base for dynamic compensation is generated. Based on the influence matrix, the initial weight factors of each rule are set, the initial weight factors of the rules in the fuzzy rule base are iteratively optimized, and the optimized weight factors are used to perform fuzzy operator operations with the state feature membership values ​​of the corresponding rules in the fuzzy rule base to obtain the comprehensive activation intensity that represents the importance of the rule. Using the comprehensive activation intensity as a weighting coefficient, the centroid defuzzification operation is performed on the compensation strategy of the rule consequent in the fuzzy rule base to obtain the output value of the optimal compensation strategy. The gradient direction of the optimal compensation strategy is calculated based on the steepest descent method. The output value is iteratively updated along the direction that makes the control error represented by the influence matrix decrease the fastest until the optimal solution is reached.

Images