Cold rolling mill roller micro displacement real-time monitoring and dynamic compensation control early warning method

Abstract

The invention provides a cold rolling mill roller micro displacement real-time monitoring and dynamic compensation control early warning method, which relates to the technical field of monitoring control, and comprises the following steps: monitoring roller temperature field distribution through infrared thermal imaging, establishing a mapping relation between displacement and temperature gradient, calculating thermal deformation and compensating displacement drift; constructing a rolling mill equipment domain knowledge graph and a diagnosis rule base, and performing fault diagnosis by utilizing graph embedding learning; and generating a compensation optimization strategy and establishing fuzzy correlation mapping to calculate an optimal parameter. According to the method, accurate monitoring and compensation of the micro displacement of the roller are achieved, and the equipment fault early warning capacity and the rolling precision are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to monitoring control technology, in particular to a real-time monitoring and dynamic compensation control early warning method for micro-displacement of a cold rolling mill roll. BACKGROUND

[0002] In the steel industry, a cold rolling mill is a key equipment for producing high-quality plates, and the accurate position control of its roll directly affects the thickness precision and surface quality of the products. During rolling, the roll is affected by working conditions such as high temperature, high pressure and high-speed rotation, and is prone to micro-displacement and thermal deformation. The traditional rolling mill control system mainly relies on thickness detection feedback and simple mathematical models for compensation, and is difficult to respond to the dynamic changes of the roll under complex working conditions in real time. With the development of intelligent manufacturing technology, roll displacement monitoring and compensation control technology based on infrared thermal imaging, knowledge graph and deep learning has gradually become a research hotspot.

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

[0004] Existing fault diagnosis methods are mostly based on simple threshold judgment or statistical models, lack deep understanding of equipment knowledge and learning ability, and are difficult to accurately identify the root cause of roll displacement abnormalities under complex working conditions, affecting the accuracy and timeliness of fault early warning.

[0005] Current compensation control strategies usually use fixed parameters or simple adaptive methods, and are difficult to intelligently optimize and adjust according to the dynamic changes of rolling process parameters and equipment states, and cannot achieve optimal dynamic compensation effect, restricting the further improvement of the quality of cold-rolled products. SUMMARY

[0006] The cold rolling mill roll micro-displacement real-time monitoring and dynamic compensation control early warning method provided by the embodiments of the present application can solve the problems in the prior art.

[0007] In a first aspect, the cold rolling mill roll micro-displacement real-time monitoring and dynamic compensation control early warning method is provided, comprising:

[0008] Real-time running data of the cold rolling mill roll is obtained, the temperature field distribution of the roll is monitored in real time by using infrared thermal imaging, a dynamic mapping relationship between the roll displacement and the temperature gradient is established, the thermal deformation amount of the roll under different temperature fields is calculated, the displacement drift caused by the temperature field is compensated, and roll displacement data is obtained;

[0009] constructing a rolling mill equipment domain knowledge graph, constructing a device diagnosis rule base according to the rolling mill equipment domain knowledge graph through ontology modeling technology, and mapping feature data corresponding to the device diagnosis rule base to the rolling mill equipment domain knowledge graph for fault diagnosis by using a graph embedding learning method;

[0010] generating a compensation optimization strategy based on an analysis result of the fault diagnosis, establishing a fuzzy association mapping between the compensation optimization strategy and the real-time running data, calculating optimal compensation parameters according to the fuzzy association mapping, and iteratively optimizing the compensation strategy in combination with the rolling mill equipment domain knowledge graph.

[0011] real-time monitoring of a rolling mill temperature field distribution by using an infrared thermal imaging system, establishing a dynamic mapping relationship between rolling mill displacement and temperature gradient, and calculating a thermal deformation amount of the rolling mill under different temperature fields to compensate for displacement drift caused by the temperature field, so as to obtain rolling mill displacement data including:

[0012] dividing rolling mill surface temperature field distribution data collected by the infrared thermal imaging system into regions according to temperature gradient variation degrees, and arranging a temperature sensor array in the divided regions according to temperature gradient variation rates;

[0013] calculating a temperature gradient variation rate between adjacent monitoring points by using temperature data collected by the temperature sensor array, taking the temperature gradient variation rate as a boundary condition to solve a heat conduction characteristic equation to obtain a main heat conduction path in the rolling mill, and taking the main heat conduction path as a temperature transfer characteristic parameter to establish a dynamic mapping relationship between rolling mill displacement and temperature gradient with the rolling mill surface temperature field distribution data;

[0014] calculating a rolling mill thermal deformation amount according to the dynamic mapping relationship, determining a compensation coefficient according to a spatial distribution characteristic of the main heat conduction path, taking a product of the compensation coefficient and the rolling mill thermal deformation amount as a compensation amount, and obtaining rolling mill displacement data.

[0015] constructing a device diagnosis rule base according to the rolling mill equipment domain knowledge graph through ontology modeling technology, and mapping feature data corresponding to the device diagnosis rule base to the rolling mill equipment domain knowledge graph for fault diagnosis by using a graph embedding learning method includes:

[0016] analyzing hierarchical association rules between components based on the rolling mill equipment domain knowledge graph, constructing a device diagnosis rule base by calculating a probability confidence of the hierarchical association rules, quantitatively evaluating rules in the device diagnosis rule base according to uncertainty, and obtaining a rule reliability index according to an association strength of a rule result probability and rule uncertainty;

[0017] mapping rules in the device diagnosis rule base that meet the rule reliability indicators to a low-dimensional vector space, obtaining an embedding representation of a rule node by performing weight matrix operations on a neighborhood set of the rule node, constructing a diagnosis path from a source node to a target node based on the embedding representation of the rule node, and determining an optimal diagnosis path based on propagation attenuation of edge weights in the diagnosis path;

[0018] classifying the embedding representation of the rule node according to the optimal diagnosis path, obtaining a fault category prediction probability by performing matrix operations, calculating a posterior probability of a fault cause in combination with device operation evidence, correcting the fault category prediction probability according to a coupling relationship between the posterior probability and a causal strength, and implementing diagnosis of a device fault based on the corrected fault category prediction probability.

[0019] obtaining an embedding representation of a rule node by performing weight matrix operations on a neighborhood set of the rule node, constructing a diagnosis path from a source node to a target node based on the embedding representation of the rule node, and determining an optimal diagnosis path based on propagation attenuation of edge weights in the diagnosis path include:

[0020] obtaining a multi-dimensional feature of a rule node and a k-order neighborhood node set of the rule node, calculating a node semantic similarity as an initial weight according to the multi-dimensional feature, adjusting the initial weight according to time sequence co-occurrence information of the rule node and the k-order neighborhood node set in historical diagnosis data, constructing a propagation attenuation function based on a connection relationship of the rule node and the k-order neighborhood node set in a physical topology structure, and performing attenuation on the adjusted weight using the propagation attenuation function to obtain a neighborhood weight matrix of the rule node;

[0021] calculating an attention score between a rule node and a k-order neighborhood node set of the rule node according to a weight distribution of the neighborhood weight matrix, and performing time sequence modeling on a rule node state based on the attention score to obtain an embedding representation of the rule node;

[0022] taking the embedding representation of the rule node as a starting state, determining a node transition probability based on embedding representation similarity between rule nodes, gradually expanding a search space to a target node to construct a candidate diagnosis path set according to the node transition probability, and selecting a path with a maximum propagation attenuation coefficient as an optimal diagnosis path based on embedding representations of adjacent nodes in the candidate diagnosis path set.

[0023] generating a compensation optimization strategy based on an analysis result of the fault diagnosis, establishing a fuzzy association mapping between the compensation optimization strategy and the real-time operation data, and calculating an optimal compensation parameter according to the fuzzy association mapping include:

[0024] The analysis result of fault diagnosis is subjected to fuzzy classification modeling, a fuzzy correlation matrix is constructed, membership values of each fault object are calculated according to the fuzzy correlation matrix, an influence weight of the fault object is determined based on the membership values, the influence weight is compared with a preset weight threshold to obtain a compensation priority, and a local compensation strategy corresponding to each fault object is generated according to the compensation priority;

[0025] The correlation strength between adjacent fault objects in the fuzzy correlation matrix is analyzed by using a dynamic programming method, the correlation strength is combined with a compensation parameter of the local compensation strategy to calculate a propagation influence membership, the compensation parameter of the local compensation strategy is recursively optimized according to the propagation influence membership, and a compensation optimization strategy is generated.

[0026] A fuzzy rule base of dynamic compensation is constructed according to the state characteristics of real-time operation data and the compensation optimization strategy, a comprehensive activation strength of the fuzzy rule is calculated by adaptive fuzzy reasoning, the fuzzy rule base is subjected to barycentric defuzzification by using the comprehensive activation strength, and optimal compensation parameters are obtained.

[0027] A fuzzy rule base of dynamic compensation is constructed according to the state characteristics of real-time operation data and the compensation optimization strategy, a comprehensive activation strength of the fuzzy rule is calculated by adaptive fuzzy reasoning, the fuzzy rule base is subjected to barycentric defuzzification by using the comprehensive activation strength, and optimal compensation parameters are obtained.

[0028] An influence degree matrix between the quantization indexes is constructed according to the state characteristics in the real-time operation data and the quantization indexes of the compensation optimization strategy, the membership distribution of the state characteristic vector and the compensation strategy vector in different fuzzy subspaces is calculated by using the influence degree matrix, and a fuzzy rule base of dynamic compensation is generated based on the membership distribution.

[0029] The initial weight factor of each rule is set based on the influence degree matrix, the initial weight factor of the rule in the fuzzy rule base is subjected to iterative optimization, the optimized weight factor is subjected to fuzzy operator operation with the state characteristic membership value of the corresponding rule in the fuzzy rule base, and a comprehensive activation strength representing the importance of the rule is obtained.

[0030] The comprehensive activation strength is taken as a weight coefficient, barycentric defuzzification operation is performed on the compensation strategy of the rule consequent in the fuzzy rule base to obtain an 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 in which the control error represented by the influence degree matrix is most rapidly decreased, and the optimal solution is reached.

[0031] In a second aspect of the embodiment of the present application, a real-time monitoring and dynamic compensation control and early warning system for micro-displacement of a cold rolling mill roll is provided, which comprises:

[0032] The first unit is used for acquiring real-time running data of a cold rolling mill roll, monitoring a roll temperature field distribution in real time by using infrared thermal imaging, establishing a dynamic mapping relationship between roll displacement and temperature gradient, and calculating a thermal deformation amount of the roll under different temperature fields to compensate displacement drift caused by the temperature field and obtain roll displacement data.

[0033] The second unit is used for constructing a rolling mill equipment domain knowledge graph, constructing a device diagnosis rule base by ontology modeling technology according to the rolling mill equipment domain knowledge graph, and mapping feature data corresponding to the device diagnosis rule base to the rolling mill equipment domain knowledge graph for fault diagnosis by using a graph embedding learning method, and performing deep feature representation learning of fault characteristics by using the roll displacement data.

[0034] The third unit is used for generating a compensation optimization strategy based on an analysis result of the fault diagnosis, establishing a fuzzy association mapping between the compensation optimization strategy and the real-time running data, calculating optimal compensation parameters according to the fuzzy association mapping, and iteratively optimizing the compensation strategy in combination with the rolling mill equipment domain knowledge graph.

[0035] In a third aspect, an electronic device is provided, including:

[0036] a processor;

[0037] a memory for storing processor-executable instructions;

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

[0039] In a fourth aspect, a computer-readable storage medium is provided, which stores computer program instructions, and the computer program instructions are executed by a processor to implement the method described above.

[0040] The present application has the following advantages:

[0041] The present application monitors the roll temperature field distribution in real time by using infrared thermal imaging, establishes a dynamic mapping relationship between roll displacement and temperature gradient, compensates displacement drift caused by the temperature field, effectively improves the accuracy of roll displacement monitoring, and solves the problem that the traditional monitoring method cannot effectively identify and distinguish thermal deformation factors.

[0042] The present application constructs a rolling mill equipment domain knowledge graph, constructs a device diagnosis rule base by ontology modeling technology, performs fault diagnosis by using a graph embedding learning method, performs deep feature representation learning of fault characteristics by using roll displacement data, realizes accurate identification and prediction of rolling mill micro-motion faults, and improves the accuracy and reliability of fault diagnosis.

[0043] The application generates a compensation optimization strategy based on fault diagnosis results, establishes a fuzzy correlation mapping to calculate optimal compensation parameters, iteratively optimizes the compensation strategy in combination with a rolling mill equipment domain knowledge graph, and realizes dynamic compensation control of the rolling mill roll micro-displacement, thereby significantly improving the stability of cold rolling mill production and product quality, and reducing equipment failure rate and maintenance cost. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 A flowchart of a cold rolling mill roll micro-displacement real-time monitoring and dynamic compensation control early warning method according to an embodiment of the application is shown in

[0045] Figure 2 A dynamic compensation strategy optimization flowchart of fault diagnosis according to an embodiment of the application is shown in DETAILED DESCRIPTION

[0046] To make the objectives, technical solutions, and advantages of the embodiments of the application clearer, the technical solutions in the embodiments of the application will be described below in detail with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only a part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the application.

[0047] The technical solutions of the application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes can not be described in some embodiments.

[0048] Figure 1 A flowchart of a cold rolling mill roll micro-displacement real-time monitoring and dynamic compensation control early warning method according to an embodiment of the application is shown in Figure 1 The method comprises the following steps.

[0049] Real-time running data of a cold rolling mill roll is acquired, an infrared thermal imaging is used to monitor a roll temperature field distribution in real time, a dynamic mapping relationship between roll displacement and temperature gradient is established, and the thermal deformation amount of the roll under different temperature fields is calculated to compensate for the displacement drift caused by the temperature field, so as to obtain roll displacement data.

[0050] A rolling mill equipment domain knowledge graph is constructed, a device diagnosis rule base is constructed by ontology modeling technology according to the rolling mill equipment domain knowledge graph, a graph embedding learning method is used to map feature data corresponding to the device diagnosis rule base to the rolling mill equipment domain knowledge graph for fault diagnosis, and deep feature learning of fault characteristics is performed by using the roll displacement data.

[0051] 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 combined with the rolling mill equipment domain knowledge graph.

[0052] In an optional embodiment, the infrared thermal imaging is used to monitor the temperature field distribution of the roll in real time, the dynamic mapping relationship between the roll displacement and the temperature gradient is established, and the thermal deformation of the roll under different temperature fields is calculated to compensate for the displacement drift caused by the temperature field, and the roll displacement data includes:

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

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

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

[0056] The infrared thermal imaging device is configured to continuously monitor the temperature of the roll surface. The infrared thermal imaging device is installed on the side of the rolling mill rack to ensure full coverage scanning of the surface temperature in the entire length range of the roll. The optical lens of the thermal imaging device uses a long-wave infrared band detector, and the wavelength range is set between 8 and 14 microns. The temperature measurement range covers the ambient temperature to 800 degrees Celsius. The spatial resolution of the device is set to at least one pixel per millimeter to ensure that the temperature details of the roll surface can be captured. The sampling frequency of the thermal imaging device is set to 50 frames per second to meet the real-time monitoring requirements under the high-speed operation condition of the rolling mill.

[0057] The temperature data collected by the infrared thermal imaging device is pre-processed, including removing environmental radiation interference, compensating for atmospheric attenuation, and correcting the emissivity variation. The removal of environmental radiation interference is achieved by background temperature compensation, and the background temperature distribution under the rolling mill shutdown state is collected periodically as a reference. The atmospheric attenuation compensation is corrected according to the distance between the device and the rolling mill and the environmental humidity, and the temperature value is corrected using a preset attenuation coefficient table. The emissivity correction considers the influence of the rolling mill surface oxidation degree and material characteristics, and establishes a corresponding relationship table between the emissivity and the surface state.

[0058] The rolling mill surface temperature field distribution data is analyzed to identify areas with significant temperature gradient changes and areas with relatively uniform temperatures. The temperature gradient is calculated using the finite difference method, and the temperature difference between adjacent pixel points is divided by the spatial distance to obtain the gradient value. The entire rolling mill surface is threshold segmented according to the size of the temperature gradient value, the area with a gradient value greater than the high threshold is divided into a high gradient area, the area with a gradient value between the high and low thresholds is divided into a medium gradient area, and the area with a gradient value less than the low threshold is divided into a low gradient area. The threshold setting is based on statistical analysis of historical temperature data to ensure that the area proportion of each region is reasonable.

[0059] The arrangement position of the temperature sensor is determined in each region divided by the temperature gradient, and the sensor arrangement density is proportional to the temperature gradient variation degree of the region. The sensor arrangement interval in the high gradient area is small, usually set to one sensor every 20 mm in the rolling mill circumferential direction and one sensor every 50 mm in the axial direction. The sensor arrangement interval in the medium gradient area is moderate, with one sensor every 40 mm in the circumferential direction and one sensor every 100 mm in the axial direction. The sensor arrangement interval in the low gradient area is large, with one sensor every 80 mm in the circumferential direction and one sensor every 200 mm in the axial direction.

[0060] The temperature sensor array is installed at the specified position on the rolling mill surface, and the sensor is selected from thermocouple or thermal resistance type, with a response time less than 1 second and a measurement accuracy better than ±1 degree Celsius. The sensor is installed in an embedded manner, and the sensor probe is installed by drilling a hole in the rolling mill surface, with a probe depth of 2 to 3 mm below the rolling mill surface, to ensure that the measurement result can reflect the temperature state near the rolling mill surface. Shielded cable is used to connect the sensor and the data acquisition device to avoid the influence of electromagnetic interference on the measurement result.

[0061] Real-time temperature data is collected by the temperature sensor array, and the data acquisition frequency is set to 10 times per second, which is synchronized with the control cycle of the rolling mill. The rate of change of the temperature gradient between adjacent monitoring points is calculated, and the rate of change is calculated using time series analysis method to analyze the temperature gradient difference between adjacent time points. For the adjacent sensors in the circumferential direction of the roll, the temperature difference is calculated and divided by the sensor spacing to obtain the circumferential temperature gradient. For the adjacent sensors in the axial direction of the roll, the temperature difference is calculated and divided by the axial spacing to obtain the axial temperature gradient.

[0062] A numerical calculation model of heat conduction inside the roll is established, and the finite element method is used to discretize the geometric structure of the roll into a large number of small elements. The material properties of each element include density, specific heat capacity, thermal conductivity and other thermal physical parameters, which are determined according to the composition and heat treatment state of the roll material. The boundary conditions are set based on the surface temperature distribution measured by the temperature sensor array and the calculated temperature gradient rate, and the surface temperature is taken as the first type of boundary condition and the surface heat flux is taken as the second type of boundary condition.

[0063] The temperature distribution inside the roll is obtained by solving the heat conduction characteristic equation, and the iterative algorithm is used to approach the steady-state or quasi-steady-state solution step by step. The convergence of the temperature field is monitored during the iteration, and when the temperature difference between two consecutive iterations is less than the set threshold, it is considered to have reached convergence. By analyzing the three-dimensional temperature field distribution obtained by solving, the main heat conduction path inside the roll is identified. The main heat conduction path corresponds to the direction with the largest temperature gradient, indicating the dominant channel of heat transfer.

[0064] The spatial distribution characteristics of the main heat conduction path inside the roll are extracted, including the starting position, ending position, path length, path direction and other geometric parameters. The identification of the heat conduction path uses the temperature gradient vector field analysis method to track the heat transfer trajectory along the direction of the gradient vector. The weight of the path is determined according to the heat flux through the path, and the path with high weight has a greater contribution to the overall thermal deformation of the roll.

[0065] A dynamic mapping relationship between roll displacement and temperature gradient is established, which describes the corresponding relationship between temperature field change and roll deformation. The mapping relationship is established using machine learning method, taking temperature gradient distribution as input feature and roll deformation as output target. The input features include the average temperature gradient, maximum temperature gradient, and temperature gradient variance of each region. The output target includes the radial deformation, axial deformation, and angular deformation of the roll.

[0066] The dynamic mapping relationship model is trained using neural networks or support vector regression machine learning algorithms. The training data is derived from temperature field data collected during historical operation and corresponding roll deformation measurement data. The data samples cover different working conditions and temperature levels. The model training process uses cross-validation methods to evaluate the prediction accuracy and ensure the generalization ability of the model. The trained model can predict the thermal deformation state of the roll based on real-time temperature gradient distribution.

[0067] The established dynamic mapping relationship is used to calculate the roll thermal deformation under the current temperature field conditions. The deformation calculation considers the thermal expansion coefficient and elastic modulus of the roll material. The thermal deformation includes changes in roll diameter, roll length, and roll axis bending deformation. The deformation calculation uses thermal elastic mechanics theory, and the temperature field is applied as a body force load to the roll mechanics model.

[0068] The influence of the spatial distribution characteristics of the main heat conduction path on the thermal deformation is analyzed, and the correlation between the path characteristics and the deformation sensitivity is established. The directionality of the heat conduction path affects the directionality of the deformation. Radial dominant heat conduction paths mainly cause radial deformation, and axial dominant heat conduction paths mainly cause axial deformation. The density distribution of the path affects the spatial distribution of the deformation. The deformation is larger in areas with dense paths and smaller in areas with sparse paths.

[0069] The temperature compensation coefficient is determined based on the spatial distribution characteristics of the main heat conduction path and the geometric parameters of the roll. For areas where radial heat conduction is dominant, the compensation coefficient mainly considers the roll radius and the radial thermal expansion characteristics of the material. For areas where axial heat conduction is dominant, the compensation coefficient mainly considers the roll length and the axial thermal expansion characteristics of the material. The compensation coefficient also needs to consider the constraints of the roll, including the influence of bearing support and the action of rolling force.

[0070] The temperature compensation amount is calculated, which is equal to 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 displacement compensation amount, Y-direction displacement compensation amount, Z-direction displacement compensation amount, and angular displacement compensation amount. The sign of the compensation amount indicates the direction of compensation. A positive value indicates the need for positive compensation, and a negative value indicates the need for negative compensation.

[0071] The calculated temperature compensation amount is applied to the real-time measurement data of the roll displacement to obtain the temperature-compensated roll displacement data. The compensation process uses addition operation to add the original measurement displacement and 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 verification of the compensation effect is evaluated by comparing the stability and repeatability of the displacement data before and after compensation. The compensated data should have better consistency and smaller measurement noise.

[0072] A real-time updating mechanism for temperature compensation is established to dynamically adjust the compensation parameters according to the real-time changes in the temperature field. The updating mechanism includes adaptive adjustment of compensation coefficients and online optimization of mapping relationships, ensuring that the compensation effect can adapt to different working 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 uses an incremental learning method to continuously improve the accuracy of the mapping relationship model using new operating data.

[0073] In an optional embodiment, the device diagnosis rule base is constructed according to the rolling mill equipment domain knowledge graph through ontology modeling technology, and the feature data corresponding to the device diagnosis rule base is mapped onto the rolling mill equipment domain knowledge graph for fault diagnosis using a graph embedding learning method, comprising:

[0074] Based on the hierarchical association rules between components in the rolling mill equipment domain knowledge graph, a device diagnosis rule base is constructed by calculating the probability confidence of the hierarchical association rules, the rules in the device diagnosis rule base are quantitatively evaluated for uncertainty, and a rule reliability index is obtained according to the correlation strength of the rule result probability and the rule uncertainty;

[0075] Rules in the device diagnosis rule base that meet the rule reliability index are mapped to a low-dimensional vector space, the embedding representation of the rule nodes is obtained by weight matrix operation on the neighborhood set of the rule nodes, a diagnosis path from the source node to the target node is constructed based on the embedding representation of the rule nodes, and the optimal diagnosis path is determined based on the propagation attenuation of the edge weights in the diagnosis path;

[0076] According to the optimal diagnosis path, the embedding representation of the rule nodes is classified according to the fault features, the fault category prediction probability is obtained by matrix operation, the posterior probability of the fault cause is calculated combined with the evidence of equipment operation, the fault category prediction probability is corrected according to the coupling relationship between the posterior probability and the causal strength, and the diagnosis of equipment fault is realized based on the corrected fault category prediction probability.

[0077] A device ontology model is constructed based on the rolling mill equipment domain knowledge graph. The ontology model contains structured knowledge representation of core concept categories, concept attributes, and relationships between concepts of the rolling mill equipment. The core concept categories cover main equipment components such as roll components, transmission devices, hydraulic equipment, electrical controllers, lubrication devices, and cooling equipment. Each concept category defines a corresponding attribute set. The attributes of roll components include roll diameter, material type, surface roughness, and wear degree, while the attributes of transmission devices include gear ratio, torque capacity, speed range, and lubrication state. The relationships between concepts describe the logical associations of component connection methods, control dependencies, and energy transmission paths, and express the structured knowledge of the equipment through object attributes and data attributes.

[0078] The hierarchical association rules between the rolling mill equipment components are analyzed, and the hierarchical structure reflects the multi-level organization relationship of the equipment from the sub-components to the whole machine. The bottom layer association rules describe the direct action relationship between the basic components, the middle layer association rules describe the functional coordination relationship between the sub-devices, and the top layer association rules describe the performance association relationship at the whole machine level. The association patterns between the components are extracted by mining the historical operation data and fault records, and the discovery of the association patterns uses the frequent item set mining algorithm and the association rule mining algorithm. The frequent item set reflects the component state combination that often appears at the same time, and the association rule describes the conditional dependency relationship between the component states.

[0079] The probability confidence of the hierarchical association rules is calculated, which quantifies the reliability and prediction accuracy of the rules. The calculation of the confidence is based on the verification statistics of the rules in the historical data, and the frequency of the rule consequent under the condition of the rule antecedent is taken as the confidence value. For the rule with the condition of abnormal roll temperature and excessive vibration amplitude, the total number of cases that simultaneously satisfy the two conditions in the historical data is counted, and the number of cases that actually occur roll failure is counted, and the ratio of the two is the confidence of the rule. The confidence value ranges from zero to one, and the higher the value, the more reliable the rule.

[0080] A device diagnosis rule library containing multiple levels of association rules is constructed, and the rule library uses a hierarchical storage structure to organize diagnosis rules of different levels. Each rule in the rule library contains complete information such as rule identifier, rule antecedent condition, rule consequent conclusion, rule confidence, rule support, etc. The rule antecedent uses a logical expression form to describe the combination of multiple conditions, supporting logical operators such as and, or, and not. The rule consequent describes the diagnosis conclusion such as fault type, fault degree, fault location, etc. The rule library supports dynamic updating, and continuously optimizes and expands the rule content according to new operation data and fault cases.

[0081] The rules in the device diagnosis rule library are quantitatively evaluated for uncertainty, and the uncertainty evaluation considers the applicability changes of the rules under different operating conditions. The quantification of uncertainty uses a combination of fuzzy set theory and probability theory, and the applicable conditions of the rules are represented as fuzzy sets, and the fuzzy membership degree of the rules under the current state is calculated. The noise influence of the measurement data and the accuracy limitation of the sensors are also considered, and the uncertainty range of the rule output is calculated through error propagation analysis. The uncertainty indicators include fuzzy degree, randomness, mixed uncertainty, and other quantitative results in multiple dimensions.

[0082] The correlation strength between the rule result probability and the rule uncertainty is calculated, and the correlation strength reflects the influence degree of uncertainty on the reliability of the rule. The correlation strength is calculated by using the mutual information method to analyze the statistical dependence between the probability distribution of the rule and the uncertainty distribution. The greater the mutual information value is, the stronger the correlation between the two is, and the more significant the influence of the uncertainty of the rule on its reliability is. Through analyzing the probability-uncertainty correlation mode of a large number of rules, a quantitative evaluation model of the correlation strength is established.

[0083] The reliability index of the rule is calculated based on the correlation strength between the rule result probability and the uncertainty, and the reliability index comprehensively considers the prediction accuracy and stability of the rule. The calculation of the reliability index takes the rule confidence as a basic value, and adjusts and corrects the basic value according to the size of the uncertainty. The rules with small uncertainty obtain positive correction, and the reliability index is higher than the basic confidence value. The rules with large uncertainty obtain negative correction, and the reliability index is lower than the basic confidence value. The correction amplitude is proportional to the correlation strength, and the rules with large correlation strength are more significantly affected by the uncertainty.

[0084] The rules meeting the rule reliability index requirement in the rule library of the screening device are screened, and the reliability threshold is set as the judgment standard for rule screening. The rules with reliability index higher than the threshold are retained for subsequent graph embedding learning, and the rules with reliability index lower than the threshold are marked as low-confidence rules or removed from the rule library. The setting of the threshold is based on the balance between the diagnostic accuracy requirement and the rule library size. Too high threshold leads to too few available rules, and too low threshold affects the diagnostic accuracy.

[0085] The screened rules are mapped to a low-dimensional vector space, and the dimension of the vector space is determined according to the complexity of the rules and the computing resources. The mapping process adopts the graph embedding learning method to convert the logical relationship between the rules into the geometric relationship in the vector space. The vector representation of the rules maintains the original semantic similarity and logical association. Semantically similar rules are closer in the vector space, and logically associated rules have specific directional relationships in the vector space.

[0086] A neighborhood set of rule nodes is constructed, which contains other rule nodes directly or indirectly associated with the current rule node. The determination of the neighborhood relationship is based on multiple association types such as logical dependence, semantic similarity, and co-occurrence frequency between rules. The direct neighborhood contains rules that are logically related, and the indirect neighborhood contains rules that are associated through intermediate rules. The size of the neighborhood set is controlled by setting the neighborhood radius or the upper limit of the number of neighbors to avoid excessive neighborhood size leading to high computational complexity.

[0087] A weight matrix of the rule node neighborhood set is calculated, which describes the association strength between the current rule node and each neighborhood node. The calculation of the weight comprehensively considers multiple factors such as semantic similarity between rules, logical correlation degree, historical co-occurrence frequency, etc. The semantic similarity is calculated by comparing the vector representation of the rule text, the logical correlation degree is determined based on the overlap degree of the antecedent and the consequent of the rule, and the historical co-occurrence frequency is obtained by counting the simultaneous activation of the rule in the diagnosis process. The normalization processing of the weight matrix ensures that the sum of the weights of each neighborhood node is equal to one.

[0088] An embedding representation of the rule node is obtained by using the weight matrix operation, which integrates the comprehensive expression of the node's own features and neighborhood information. The embedding calculation adopts the aggregation mechanism of the graph convolution network, which weights the feature vectors of the neighborhood nodes by weighted average, and then linearly combines them with the original features of the current node. The superposition of multi-layer graph convolution enables the node's embedding representation to capture more distant neighborhood information, enhancing the expression ability of the representation.

[0089] A search graph of the diagnosis path is constructed based on the embedding representation of the rule node, and the nodes in the search graph are rule nodes and the edges are the reasoning transfer relationships between rules. The source node corresponds to the initial observed symptom of diagnosis, and the target node corresponds to the final fault diagnosis conclusion. The similarity of the embedding representation is used to calculate the transfer probability between nodes, and the node pairs with high similarity have high transfer probability. A probability path search algorithm is used to find the path from the source node to the target node in the search graph.

[0090] The propagation attenuation of the edge weight in the diagnosis path is calculated, which reflects the step-by-step loss of information in the reasoning chain. The weight of each edge is determined based on the similarity of the embedding representation of the connected nodes, and the higher the similarity, the greater the edge weight. The propagation attenuation of the path is calculated by the product of the weights of all edges on the path, and the product value reflects the information retention degree of the whole path. The path with large propagation attenuation represents high reliability of the reasoning chain and small information loss.

[0091] The path with the maximum propagation attenuation is determined as the optimal diagnosis path, which corresponds to the most reliable fault reasoning chain. The path selection process also considers the influence of path length, and in the case of similar propagation attenuation, the shorter path is preferred. The optimal diagnosis path provides reasoning basis for subsequent fault feature classification and probability calculation.

[0092] According to the optimal diagnosis path, the embedding representation of the rule node is classified into fault features, and the classification process maps the embedding vector to a predefined fault category space. The fault categories include main types such as mechanical fault, electrical fault, hydraulic fault, and control fault, and each type is further subdivided into specific fault modes. The classification adopts a neural network classifier, which takes the embedding representation of the rule node as input features and outputs the predicted probability of each fault category.

[0093] The fault category prediction probability is calculated by a classification matrix operation, and the parameters of the classification matrix are learned from training data. The matrix operation projects the high-dimensional embedding representation into an output space corresponding to the number of fault categories, and each output dimension corresponds to a prediction score of a fault category. The prediction score is normalized by a softmax function to obtain a probability distribution, and the sum of the probabilities of all categories is equal to one.

[0094] The posterior probability of the fault cause is calculated in combination with the evidence of device operation, including sensor measurement data, operation log records, maintenance history information, and other multi-source evidence. The calculation of the posterior probability uses Bayesian inference method to combine the prior probability and the likelihood probability. The prior probability is determined based on historical fault statistical data, and the likelihood probability is calculated based on the matching degree of the current operation evidence and the fault mode.

[0095] The coupling relationship between the posterior probability and the causal strength is analyzed, and the causal strength reflects the causal correlation between the fault cause and the fault phenomenon. The modeling of the coupling relationship considers the adjustment effect of the causal strength on the posterior probability, and the fault cause with high causal strength obtains higher weight of the posterior probability. The coupling coefficient is determined by analyzing the strength distribution of the causal relationship in historical fault cases, and strong causal relationship corresponds to high coupling coefficient, and weak causal relationship corresponds to low coupling coefficient.

[0096] The coupling relationship between the posterior probability and the causal strength is used to correct the fault category prediction probability, and the correction process weights and fuses the prediction probability and the posterior probability. The fusion weight is dynamically adjusted according to the reliability and completeness of the evidence, and the weight of the posterior probability is higher when the evidence is sufficient and reliable, and the weight of the prediction probability is higher when the evidence is insufficient. The corrected probability distribution more accurately reflects the fault state of the current device, and provides a reliable basis for the final diagnosis decision.

[0097] In an optional implementation, an embedding representation of a rule node is obtained by performing a weight matrix operation on a neighborhood set of the rule node, a diagnosis path from a source node to a target node is constructed based on the embedding representation of the rule node, and an optimal diagnosis path is determined based on propagation attenuation of edge weights in the diagnosis path, including:

[0098] A multi-dimensional feature of the rule node and its k-order neighborhood node set is obtained, a node semantic similarity is calculated as an initial weight according to the multi-dimensional feature, and the initial weight is adjusted according to time sequence co-occurrence information of the rule node and the k-order neighborhood node set in historical diagnosis data; a propagation attenuation function is constructed based on a connection relationship of the rule node and the k-order neighborhood node set in a physical topology structure, the adjusted weight is attenuated by using the propagation attenuation function, and a neighborhood weight matrix of the rule node is obtained;

[0099] According to the weight distribution calculation rule of the neighborhood weight matrix, the attention score between the rule node and the k-order neighborhood node set thereof is calculated, and the embedding representation of the rule node is obtained by time series modeling of the rule node state based on the attention score;

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

[0101] A graph structure representation of the rule node is constructed, each diagnosis rule is taken as a node in the graph, and the node contains feature attribute information of the rule. The feature attributes of the rule node include a keyword vector of a condition part of the rule, a semantic representation of a conclusion part, a confidence value of the rule, a support degree statistic of the rule and the like. For each rule node, a k-order neighborhood node set thereof is determined, and the k-order neighborhood refers to all adjacent nodes that can be reached through k edges. The first-order neighborhood is a directly connected node, the second-order neighborhood is a node connected through two edges, and the like until the k-order neighborhood. The determination of the neighborhood node is based on multiple connection criteria such as a logical association relationship between rules, semantic similarity, co-occurrence frequency and the like.

[0102] To solve the problem of exponential growth of the number of k-order neighborhood nodes with k, the following optimization strategies are adopted in the application:

[0103] Multi-level neighborhood sampling: importance sampling is performed on the k-order neighborhood nodes, the sampling rate decreases with the increase of k, the first-order neighborhood: all nodes are reserved; the second-order neighborhood: the sampling rate p2 = min(50 / |N2|, 1), wherein |N2| is the size of the second-order neighborhood; the k-order neighborhood (k>2): the sampling rate p k k = min(50·0.5 (k-2) / |N k |,1). Node importance score: the sampling probability of a node v is proportional to the importance of the node, P(v) = 0.4·sim(v0,v) + 0.3·PR(v) + 0.3·freq(v); wherein 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 a compressed sparse row (CSR) format, and the space complexity is reduced from O(n 2) is reduced to O(nnz), where nnz is the number of non-zero elements. The computation task is layered: offline computation: rule node base embedding vector pre-computation, knowledge graph structure analysis; near real-time computation: state feature update, candidate fault type preliminary screening; real-time computation: active path evaluation, compensation parameter calculation. Dimension reduction: reduce high-dimensional embedding vector to 32-64 dimensions through principal component analysis.

[0104] Through the above optimization measures, the single reasoning time of the system on a standard industrial control computer is controlled within 50 milliseconds, meeting the needs of real-time control of the cold rolling mill.

[0105] In actual testing, an industrial control computer with an Intel Core i5 processor and 8GB of memory runs the system, and the knowledge graph containing more than 5000 nodes is inferred. The average time of the complete inference process is 37.2 milliseconds, the peak memory usage is 128MB, and the CPU utilization is up to 25%. Even in the most complex fault diagnosis scenario, the diagnosis delay is controlled within 100ms, which is much lower than the response time requirement of the cold rolling process (usually 500ms).

[0106] Extract the multi-dimensional feature vector of the rule node and its k-order neighborhood node. The multi-dimensional features include the word frequency statistical features of the rule text, the distributed representation features of the rule semantics, and the structured features of the rule in the knowledge graph. The word frequency statistical features are obtained by counting the frequency of keywords after tokenizing the rule text, and the importance weight of the vocabulary is calculated using the term frequency-inverse document frequency method. The semantic distributed representation is obtained by encoding the rule text using a pre-trained language model to obtain a high-dimensional vector representation, which can capture the deep semantic information of the rule. The structured features reflect the position information and correlation strength of the rule in the entire knowledge system.

[0107] The semantic similarity between the rule node and its k-order neighborhood node is calculated as the initial weight, and the cosine similarity method is used to calculate the semantic similarity. The multi-dimensional feature vectors of the rule node and the neighborhood node 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, and 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 of zero to one as the initial weight, and the mapping method uses linear transformation to convert negative values to positive values close to zero.

[0108] The time sequence co-occurrence information of the acquisition rule node and the k-order neighborhood node in the historical diagnosis data is obtained, and the time sequence co-occurrence information reflects the collaborative appearance mode of the rules in the actual diagnosis process. By analyzing the activation time sequence of the rules in the historical diagnosis cases, the co-occurrence frequency of different rules in a time window is counted. The size of the time window is determined according to the time characteristics of the diagnosis process, and is usually set as the typical duration of the diagnosis process. The co-occurrence frequency is calculated by using a sliding time window method, and the co-occurrence times of the rules are counted by moving the window on the entire historical data time axis.

[0109] The initial weight obtained by the semantic similarity calculation is adjusted by using the time sequence co-occurrence information, and the adjustment method is based on the positive correlation between the co-occurrence frequency and the weight. The weight between the rules with high co-occurrence frequency is positively adjusted, and the weight between the rules with low co-occurrence frequency is negatively adjusted. The adjustment amplitude is related to the statistical significance of the co-occurrence frequency, and the significance is evaluated by a chi-square test method to evaluate the statistical reliability of the co-occurrence mode. The adjusted weight more accurately reflects the association strength between the rules in the actual application.

[0110] The connection relationship between the rule node and the k-order neighborhood node in the physical topology structure is analyzed, and the physical topology structure reflects the logical dependency hierarchy and the reasoning chain between the rules. The directionality and strength of the connection are determined according to the logical implication relationship between the premise conditions and the conclusions of the rules, the rules with the same premise conditions have a strong connection relationship, and the rules with similar conclusions also have a certain connection relationship. The strength of the connection relationship is quantified by analyzing the logical overlap of the rules, and the rule pair with high overlap degree has a strong connection relationship.

[0111] A distance-based propagation attenuation function is constructed, and the attenuation function describes the law that the weight decreases with the increase of the distance between the nodes. The distance is defined as the shortest path length from the rule node to the neighborhood node, and the path length is calculated in terms of the number of edges. The attenuation function adopts an exponential attenuation form, and the attenuation coefficient is determined according to the density and connection characteristics of the rule graph. The farther the distance, the smaller the influence of the neighborhood node on the current rule node, and the distance effect is embodied by the attenuation function.

[0112] The adjusted weight is attenuated by using the propagation attenuation function to obtain the final weight value considering the distance factor. The calculation of the attenuation processing multiplies the adjusted weight by the attenuation coefficient corresponding to the distance, and the attenuation coefficient decreases exponentially with the distance. The attenuation coefficient of the first-order neighborhood node is close to one, and the attenuation coefficient of the k-order neighborhood node is close to zero. The attenuated weight forms a neighborhood weight matrix of the rule node, the rows of the matrix correspond to the current rule node, the columns correspond to the k-order neighborhood nodes thereof, and the matrix elements are the attenuated weight values.

[0113] The attention score between the rule node and its k-order neighborhood node is calculated based on the weight distribution calculation rule of the neighborhood weight matrix, and the attention score reflects the attention degree of the current rule node to different neighborhood nodes. The calculation of the attention score adopts a softmax normalization method, which normalizes the weight values in the corresponding row of the neighborhood weight matrix after exponential transformation. The sum of the normalized attention scores is equal to one, and the larger the score value is, the more important the corresponding neighborhood node is.

[0114] The rule node state is sequentially modeled using the attention score, and the sequential modeling captures the state evolution process of the rule node at different time steps. The state of the rule node is represented as a weighted combination of its feature vector and the feature vector of the neighborhood node, and the weight is the corresponding attention score. The sequential modeling adopts a recurrent neural network architecture, and the node state at the previous time step is used as the input for the current time step calculation. The recurrent update process is repeated for a predefined number of time steps, and the node state at the final time step is used as the embedding representation of the rule node.

[0115] The embedding representation of the rule node integrates the influence of the node's own features and neighborhood information, forming a high-dimensional vector representation. The dimension of the embedding representation is determined according to application requirements and computing resources, and is usually set to tens to hundreds of dimensions. Each dimension of the embedding vector captures the characteristics of the rule node in a certain semantic aspect, and the entire vector can comprehensively represent the semantic and structural characteristics of the rule node.

[0116] When calculating the node transition probability, the present application not only considers the similarity of the embedding vector, but also explicitly models the causal direction of the rule. Specifically, the calculation of the transition probability P(j|i) adopts the following formula:

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

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

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

[0120] The embedding representation of the rule node is taken as the starting state of the diagnostic path search, and the starting state corresponds to the initial symptom or observation result of the diagnostic process. The node transition probability is calculated based on the similarity between the embedding representations of the rule nodes, and the similarity is calculated using the Euclidean distance or cosine similarity method. The nodes with closer distance or higher similarity have higher transition probability, and the transition probability is normalized by the softmax function to ensure the effectiveness of the probability distribution.

[0121] The search space is gradually expanded according to the node transition probability until the target node is reached, and the search process adopts a breadth-first or depth-first strategy. The breadth-first search expands nodes in order of the size of the transition probability, ensuring that high-probability paths are explored first. The depth-first search searches along a single path until the target is reached or a dead end is encountered. During the search process, a candidate path set is maintained, and each path records the complete path information from the starting node to the current node.

[0122] A candidate diagnostic path set from the starting node to the target node is constructed, and the candidate path set contains all diagnostic reasoning chains. Each candidate path represents a diagnostic logic, and the node sequence on the path corresponds to the reasoning steps in the diagnostic process. The generation of candidate paths considers the path length limit to avoid generating excessively long invalid paths. The path length limit is determined according to the complexity of the diagnostic problem and the reasoning depth requirement.

[0123] During the construction of the candidate diagnostic path set, the invention implements a complete loop detection and pruning strategy to ensure that the search process does not fall into a loop. Specifically, a three-color marking method is used for loop detection: the nodes are marked as WHITE, GRAY and BLACK, and when a GRAY state node is encountered during the search process, it indicates that a loop has been detected; a node set S of the current path is maintained, and for each node v to be expanded, it is checked whether v is already in S, and if so, the path is pruned; a hierarchical pruning strategy is implemented, which divides the hierarchy according to the level attribute of the nodes, and only allows transitions from low level to high level or the same level.

[0124] A maximum path length limit MaxLength=log(N)·C is set, where N is the number of nodes in the graph, and C is an empirical coefficient, and paths exceeding this length are automatically pruned; for necessary functional loops, a loop unfolding technique is used to convert the loop into a finite-step feedforward structure, a time step attribute t is added to each node to ensure that t is monotonically increasing; a minimum threshold MinDecay is set for the propagation decay coefficient, and paths below this value are automatically pruned to avoid abnormal decay values.

[0125] For example, in detecting the potential loop of 'roll vibration-bearing damage-lubrication deficiency-friction increase-temperature rise-thermal expansion-gap reduction-vibration exacerbation', the system identifies the repeated occurrence of the 'roll vibration' node through path history records, immediately terminates the further expansion of the path, and records the loop characteristics. For necessary feedback loops, the system is unfolded into 't1-roll vibration→t2-bearing damage→...→t8-vibration exacerbation(t8)' by adding a time step attribute, ensuring that the reasoning path is loop-free in the unfolded time dimension.

[0126] The propagation decay coefficient between adjacent nodes in the candidate path is calculated, which reflects the degree of information preservation during path propagation. The calculation of the decay coefficient is based on the similarity of the embedding representations of adjacent nodes and the logical association strength between nodes, and the node pair with high similarity and strong logical association has a higher propagation decay coefficient. The overall propagation decay coefficient of the path is the product of the propagation decay coefficients of all adjacent node pairs on the path, and the product value reflects the information propagation efficiency of the entire path.

[0127] The path with the maximum propagation decay coefficient is selected as the optimal diagnostic path, which represents the path with the minimum loss during information propagation. The optimal path corresponds to the most reliable diagnostic reasoning chain and can provide the most accurate diagnostic results. The path selection process also considers the length penalty of the path, and in the case of similar propagation decay coefficients, the shorter path is preferred to avoid unnecessary complex reasoning process.

[0128] In selecting the optimal diagnostic path, the present application also considers the influence of path length and introduces an adaptive length penalty coefficient. The length penalty coefficient LP(l) is calculated by the following formula:

[0129] LP(l)=exp(-β·(l-l opt ) 2 );

[0130] Wherein, l is the current path length; l opt is the optimal path length of fault type k, which is obtained by statistical analysis of historical diagnostic data, l opt (k)=μk, μk is the average optimal path length of fault type k; β is the penalty intensity coefficient, β(k)=1 / (σk2 • 2), where σkis the standard deviation of path length for fault type k.

[0131] l opt The and β parameters are automatically determined by offline analysis of historical diagnosis data, avoiding subjective setting. The system continuously updates the parameters of each fault type through Bayesian optimization method:

[0132] P(β|D) ∝ P(D|β) · P(β);

[0133] where D is the diagnosis result data, and P(β) is the prior distribution.

[0134] The final path scoring function considers both the propagation attenuation and the 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 fault' type, the historical diagnosis data shows that its optimal path length mean μ = 4.2 and standard deviation σ = 0.8, then β is calculated as 1 / (0.8 2 · 2) = 0.78. When evaluating a candidate path with 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 length penalty in the path selection process.

[0138] In an alternative embodiment, a compensation optimization strategy is generated based on the analysis result of the fault diagnosis, a fuzzy correlation mapping is established between the compensation optimization strategy and the real-time operation data, and the optimal compensation parameter is calculated according to the fuzzy correlation mapping, comprising:

[0139] The analysis result of the fault diagnosis is fuzzy classified and modeled, a fuzzy correlation matrix is constructed, the membership values of each fault object are calculated according to the fuzzy correlation matrix, the influence weight of the fault object is determined based on the membership values, the compensation priority is obtained by comparing the influence weight with a preset weight threshold, and the local compensation strategy corresponding to each fault object is generated according to the compensation priority;

[0140] The dynamic programming method is used to analyze the correlation strength between adjacent fault objects in the fuzzy correlation matrix, the correlation strength is combined with the compensation parameter of the local compensation strategy to calculate the propagation influence membership, the compensation parameter of the local compensation strategy is recursively optimized according to the propagation influence membership, and a compensation optimization strategy is generated;

[0141] A dynamic compensation fuzzy rule base is constructed according to the state characteristics of real-time operation data and the compensation optimization strategy, the comprehensive activation strength of the fuzzy rule is calculated through adaptive fuzzy reasoning, and the barycenter of the fuzzy rule base is de-fuzzied by using the comprehensive activation strength to obtain optimal compensation parameters.

[0142] As shown in Figure 2 The method comprises the following steps:

[0143] The analysis result data output by the fault diagnosis module is acquired, including the detected fault type identifier, the fault severity level, the fault occurrence position coordinates, the fault duration length and other key information. The fault types are classified according to equipment components, including main categories such as motor faults, sensor faults, controller faults and mechanical transmission faults, and each category is further subdivided into specific fault modes. The fault severity is divided into five levels, and the levels one to five correspond to the evaluation of slight abnormality to serious fault respectively. The fault position information is described in multiple dimensions such as equipment number, component identifier and spatial coordinates, to ensure the accuracy and uniqueness of fault positioning.

[0144] A fuzzy classification model of fault objects is established, and the Gaussian membership function is used to fuzz the fault severity. The center point of the Gaussian function is set at the median position of each severity level, and the standard deviation parameter is determined according to the distribution characteristics of historical fault data. The center point of the membership function of slight fault is set near level one, and the center point of the membership function of serious fault is set near level five, and the membership function of moderate fault covers the middle level range. Through the fuzzing process, the discrete fault levels are converted into continuous membership distribution, which is convenient for subsequent fuzzy reasoning calculation.

[0145] A fuzzy correlation matrix reflecting the mutual relationship between fault objects is constructed, the rows and columns of the matrix correspond to different fault objects, and the matrix elements represent the correlation strength between the corresponding fault objects. The calculation of the correlation strength is based on multiple dimensions such as physical connection relationship, functional dependency relationship and time correlation of the fault objects. The physical connection relationship is determined by analyzing the mechanical structure and electrical connection of the equipment, and the correlation strength of the directly connected components is higher. The functional dependency relationship considers the logical path of fault propagation, and the influence degree of upstream fault on downstream equipment determines the correlation strength. The time correlation is calculated by analyzing the time sequence correlation of different faults in historical fault data.

[0146] The membership value of each fault object in the current state is calculated, which reflects the degree to which the fault object belongs to a certain fault severity level. The membership distribution of the fault object in each severity level is calculated using the currently detected fault parameter value and the pre-established Gaussian membership function. The calculation of the membership value adopts the standard calculation method of the Gaussian function, which normalizes the distance between the fault parameter value and the center point of the membership function. The closer the distance, the higher the membership.

[0147] The influence weight of the fault object is calculated based on the membership value, which comprehensively considers the severity membership of the fault and the importance weight of the fault in the device. The device importance weight is pre-set according to the criticality of the fault object in the overall device operation, and the weight of the core control component is higher than that of the auxiliary component. The maximum membership value of the fault severity is multiplied by the device importance weight to obtain the comprehensive influence weight, which reflects the influence degree of the fault object on the overall device operation.

[0148] A preset weight threshold is set to divide the compensation priority, and the calculated influence weight is compared with the preset threshold to determine the compensation priority of the fault. The preset threshold is set according to the operation requirements and fault tolerance of the device, and three thresholds of high, medium and low are usually set to divide the faults into four priorities. The fault with an influence weight exceeding the high threshold is set as the highest priority and needs to be compensated immediately. The fault with an influence weight between the high threshold and the medium threshold is set as the high priority, the fault with an influence weight between the medium threshold and the low threshold is set as the medium priority, and the fault with an influence weight below the low threshold is set as the low priority.

[0149] According to the compensation priority, a corresponding local compensation strategy is generated for each fault object, which includes compensation mode selection, compensation parameter setting, compensation execution timing and other specific contents. The highest priority fault adopts the active compensation mode, which adjusts the control parameters, switches the standby device, reduces the operation load and other measures for fast compensation. The high priority fault adopts the prediction compensation mode, which adjusts the relevant parameters in advance according to the fault development trend. The medium priority fault adopts the passive compensation mode, which adjusts accordingly when the fault impact appears. The low priority fault adopts the monitoring compensation mode, which continuously monitors the fault state change but does not execute the compensation action temporarily.

[0150] The dynamic programming method is used to analyze the correlation strength between adjacent fault objects in the fuzzy correlation matrix. The state of dynamic programming is defined as the current compensation state of the fault object, and the decision variable is the adjustment amount of the compensation parameter. The state transition equation is established to describe the influence propagation process between fault objects, and the cost function of state transition considers the comprehensive evaluation of compensation effect and compensation cost. The optimal path from the initial state to the target state is found by layer-by-layer recursive calculation, and the decision sequence corresponding to the optimal path is the optimal compensation parameter combination.

[0151] The propagation influence membership degree between fault objects is calculated, which reflects the influence degree of the compensation action of one fault object on the adjacent fault object. The calculation of the propagation influence membership degree combines the correlation strength in the fuzzy correlation matrix and the compensation parameter amplitude of the local compensation strategy. The case of high correlation strength and large compensation parameter amplitude corresponds to a higher propagation influence membership degree, indicating that the compensation action has a greater influence on the adjacent fault object. The propagation influence membership degree is quantified using a trapezoidal membership function, and the function parameters are determined based on statistical analysis of historical compensation effect data.

[0152] The compensation parameters of the local compensation strategy are recursively optimized, and the recursive optimization process considers the propagation influence membership degree between fault objects. The recursive calculation starts from the highest priority fault and propagates the optimization results step by step. The optimization of the compensation parameters of each fault object considers the propagation influence of its directly associated fault objects, and the propagation influence membership degree is used as a weight factor to adjust the original compensation parameters. The recursive process is repeated until the compensation parameters of all fault objects converge to stable values, and the convergence judgment condition is that the parameter change amplitude in two consecutive iterations is less than a set threshold.

[0153] The recursively optimized local compensation strategy is integrated to generate a global compensation optimization strategy, which coordinates the compensation actions of each fault object to avoid conflicts. The integration process uses a weighted average method, and the weights are determined based on the compensation priority and the propagation influence membership degree of the fault objects. Fault objects with high priority and large propagation influence have higher weights, and their compensation strategies dominate in the global strategy.

[0154] Real-time state data during equipment operation is obtained, including key parameters such as sensor measurements, actuator working states, controller output signals, etc. A fuzzy correlation mapping relationship is established between the real-time state data and the compensation optimization strategy, which describes the corresponding relationship between state data changes and compensation strategy adjustments. Fuzzy inference rules are used to express the mapping relationship, with the antecedent being the fuzzy description of the state data and the consequent being the corresponding compensation strategy adjustment suggestion.

[0155] When integrating fuzzy sets and probability theory, the Zadeh possibility-probability consistency principle is used as the theoretical basis. Specifically, when modeling the fuzzy classification of fault objects, first calculate the membership degree of the fault object in each fuzzy set through the Gaussian membership function μA(x), then map the membership degree to the conditional probability through the conversion function P(A|x)=μA(x) / ∑μi(x), where ∑μi(x) is the sum of the membership values of all possible states.

[0156] In constructing the fuzzy association matrix, the Dempster-Shafer evidence theory is used as the uncertainty fusion framework, and the belief function Bel(A) = ∑m(B)(B ⊆ A) and the 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 evidences from different sources is realized by the Dempster combination rule m(C) = [∑m1(A)·m2(B)] / [1-K](A ∩ B = C), which ensures the strict mathematical unification of fuzzy reasoning and probabilistic reasoning.

[0157] For example, for the abnormal fault of roll vibration, the membership degrees of the fault in the'serious' fuzzy subset, the'moderate' fuzzy subset and the'slight' fuzzy subset are 0.85, 0.35 and 0.05 respectively, which are calculated by the Gaussian membership function. Through the normalization conversion, the corresponding conditional probabilities are 0.68, 0.28 and 0.04 respectively. This conversion ensures the strict mathematical conversion of fuzziness and probability, and avoids the theoretical inconsistency caused by direct mixing.

[0158] A fuzzy rule base for dynamic compensation is constructed, which contains a complete set of mapping rules between state characteristics and compensation strategies. The establishment of the rules is based on the mining analysis of expert knowledge and historical operation data, which ensures the completeness and accuracy of the rules. Each rule contains a confidence parameter, which reflects the reliability of the rule and is determined based on the verification results of the rule in historical data.

[0159] The comprehensive activation strength of each fuzzy rule is calculated by adaptive fuzzy reasoning, and the adaptive mechanism dynamically adjusts the activation degree of the rule according to the current operating state. The calculation of the activation strength considers multiple factors such as the matching degree of the rule antecedent and the current state characteristics, the confidence of the rule, and the historical use effect of the rule. The matching degree is determined by calculating the similarity between the antecedent conditions and the current state characteristics, and the higher the similarity, the greater the activation strength.

[0160] The fuzzy rule base is de-fuzzied by the barycenter method using the comprehensive activation strength, and the activated rule consequent is calculated by weighted average. The weight of the weighted average is the comprehensive activation strength of each rule, and the rule with high activation strength contributes more to the final output. The output result of the barycenter de-fuzzification 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 an alternative embodiment, a fuzzy rule base for dynamic compensation is constructed according to the state characteristics of real-time operation data and the compensation optimization strategy, the comprehensive activation strength of the fuzzy rule is calculated by adaptive fuzzy reasoning, and the fuzzy rule base is de-fuzzied by the barycenter method using the comprehensive activation strength, comprising:

[0162] According to the state characteristics in the real-time operation data and the quantitative indicators of the compensation optimization strategy, an influence degree matrix between the quantitative indicators is constructed, the membership degree distribution of the state characteristic vector and the compensation strategy vector in different fuzzy subspaces is calculated by using the influence degree matrix, and a fuzzy rule base of dynamic compensation is generated based on the membership degree distribution;

[0163] An initial weight factor of each rule is set based on the influence degree matrix, the initial weight factor of the rule in the fuzzy rule base is iteratively optimized, a comprehensive activation intensity representing the importance of the rule is obtained by performing fuzzy operator operation on the optimized weight factor and the state characteristic membership value of the corresponding rule in the fuzzy rule base;

[0164] The comprehensive activation intensity is taken as a weight coefficient, a gravity defuzzification operation is performed on the compensation strategy of the rule consequent in the fuzzy rule base to obtain an output value of the optimal compensation strategy, a gradient direction of the optimal compensation strategy is calculated based on the steepest descent method, the output value is iteratively updated along a direction in which a control error represented by the influence degree matrix is most quickly decreased, and the optimal solution is reached.

[0165] Real-time state data in the operation process of the equipment is acquired, including physical parameters such as a working temperature sensor measurement value, a vibration accelerometer output signal, a load current amplitude detected by a current transformer, and a rotating speed signal fed back by an encoder. The original sensor data is converted into a digital signal by an analog-to-digital converter, is subjected to low-pass filtering to remove high-frequency noise interference, and is subjected to smoothing processing by a sliding average algorithm to obtain stable and reliable state characteristic values. Meanwhile, quantitative indicators related to the compensation control strategy are collected, including a voltage amplitude output by a current compensation device, a frequency parameter of a compensation signal, a phase compensation angle, a gain adjustment coefficient and other key control parameters, which reflect the specific execution state of the compensation strategy.

[0166] A correlation relationship matrix between the state characteristics and the quantitative indicators of the compensation strategy is established, and the influence degrees of the state characteristic variables and the compensation strategy parameters are quantified by calculating Pearson correlation coefficients between the state characteristic variables and the compensation strategy parameters. The calculation of the correlation coefficients is based on statistical analysis of historical operation data, and data samples in a long enough time period are collected to ensure the reliability of the statistical results. The absolute values of the correlation coefficients are taken as elements of the influence degree matrix, and the numerical range is normalized to the interval of zero to one. The rows of the influence degree matrix correspond to the state characteristic variables, and the columns correspond to the compensation strategy quantitative indicators. Each element in the matrix represents the influence intensity of the corresponding row variable on the column variable.

[0167] The state characteristic variables are fuzzified by using a triangular membership function, and the value range of each continuous variable is divided into five fuzzy subsets, i.e. low, low-medium, medium, high-medium and high. The vertex positions of the triangular membership function are determined according to the historical statistical distribution of the variable. The vertex of the low subset is set near the minimum value of the variable, the vertex of the high subset is set near the maximum value of the variable, and the vertices of the intermediate subsets are uniformly distributed in the value range. The membership degree of the current state characteristic value in each fuzzy subset is calculated, and the membership degree value reflects the degree to which the value belongs to the corresponding fuzzy subset. The same fuzzification processing method is used for the quantitative index of the compensation strategy, and the corresponding fuzzy subsets and membership functions are established.

[0168] A fuzzy reasoning rule base is established according to the fuzzified state characteristics and compensation strategies, and the rules are expressed in the form of "if-then" conditional statements. The rule antecedent contains one or more combinations of fuzzy variables of state characteristics, and multiple antecedent conditions are connected by the "and" logical operator. The rule consequent is the corresponding compensation strategy output, indicating the compensation measures to be taken when the antecedent conditions are met. A typical rule form is "if the temperature is high and the vibration is medium 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, and the initial value of the weight factor is determined based on the average value of the relevant elements in the influence degree matrix. The average value of the influence degree of each state characteristic in the rule antecedent on each compensation strategy in the rule 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 total weight sum is equal to one. A weight optimization objective function is established, and the objective function takes the sum of the squares of the control errors as the optimization target. The weight factors of each rule are iteratively adjusted by the gradient descent algorithm.

[0170] The direction and step size of the weight adjustment in the gradient descent algorithm 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 weight, and the difference in the objective function value before and after the perturbation is calculated. The approximate partial derivative is obtained by dividing the difference value by the perturbation amount. The gradient direction is a vector composed of the partial derivatives, and the step size is adaptively determined according to the gradient amplitude. A smaller step size is used when the gradient is large to avoid oscillation, and a larger step size is used when the gradient is small to speed up convergence. The weight update process is repeated until the objective function converges or the maximum number of iterations is reached.

[0171] The optimized weight factor is subjected to fuzzy operator operation with the membership values of each state feature in the antecedent of the fuzzy rule, and the comprehensive activation strength of each rule is calculated. The fuzzy operator operation adopts a minimum value operator, and the minimum value of the membership degrees of all state features in the antecedent of the rule is taken as the antecedent activation degree of the rule. The antecedent activation degree is multiplied by the optimized weight factor of the rule to obtain the comprehensive activation strength of the rule. The comprehensive activation strength reflects the importance and contribution of the rule under the current input state, and the greater the value is, the more important the rule is.

[0172] The barycenter solution defuzzification operation is performed to calculate the final compensation strategy output value. The compensation strategy values of the consequents of all rules are multiplied by the comprehensive activation strengths of the corresponding rules, and then summed to obtain the numerator. The comprehensive activation strengths of all rules are summed to obtain the denominator, and the output value of the compensation strategy is obtained by dividing the numerator by the denominator. The calculation method of the weighted average ensures that the rules with greater activation strengths have greater influence on the final output, and realizes intelligent decision-making based on the importance of the rules.

[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 expected control target. The gradient of the error function with respect to the compensation strategy output is calculated by numerical differentiation. The central difference format is adopted to improve the accuracy of the gradient calculation. A small positive and negative perturbation is set near the current output value, and the corresponding error function values are calculated. The gradient value is calculated by using the central difference formula.

[0174] The compensation strategy output value is iteratively updated along the gradient descent direction. The update amount of each iteration is equal to the negative gradient direction multiplied by the learning rate. The learning rate is adaptively adjusted according to the current gradient amplitude and historical convergence. The momentum method is used to accelerate the convergence process and avoid local extrema. The convergence judgment conditions include that the error change in continuous iterations is less than a set threshold, the gradient amplitude is less than a threshold, or the maximum number of iterations is reached. When the convergence conditions are met, the iteration is stopped, and the optimal compensation strategy value is output.

[0175] The calculated optimal compensation strategy value is converted into specific control instructions and sent to the actuator, including voltage regulation instructions, frequency setting instructions, phase adjustment instructions, etc. of the compensator. The control instructions are transmitted to the lower computer controller through the digital communication interface, and the lower computer drives the corresponding actuator to complete the compensation action. At the same time, the weight factor and the influence degree matrix of the rules in the fuzzy rule library are updated, so as to prepare for the calculation of the next control period and realize the dynamic adaptive adjustment of the compensation strategy.

[0176] In a second aspect of the embodiment of the present application, a real-time monitoring and dynamic compensation control and early warning system for micro-displacement of a cold rolling mill roll is provided, comprising:

[0177] The first unit is configured to acquire real-time operation data of a cold rolling mill roll, monitor a roll temperature field distribution in real time by using infrared thermal imaging, establish a dynamic mapping relationship between roll displacement and a temperature gradient, calculate a thermal deformation amount of the roll under different temperature fields, compensate for displacement drift caused by the temperature field, and obtain roll displacement data.

[0178] The second unit is configured to construct a rolling mill equipment domain knowledge graph, construct a device diagnosis rule base by ontology modeling technology according to the rolling mill equipment domain knowledge graph, map feature data corresponding to the device diagnosis rule base to the rolling mill equipment domain knowledge graph for fault diagnosis by using a graph embedding learning method, and perform deep feature representation learning of fault characteristics by using the roll displacement data.

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

[0180] In a third aspect, an electronic device is provided, including:

[0181] a processor;

[0182] a memory for storing processor-executable instructions;

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

[0184] In a fourth aspect, a computer-readable storage medium is provided, which stores computer program instructions, and the computer program instructions are executed by a processor to implement the method described above.

[0185] The present application can be a method, device, system and / or computer program product. The computer program product can include a computer readable storage medium having computer readable program instructions loaded thereon for implementing various aspects of the present application.

[0186] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

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 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. 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, 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: 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 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.

3. 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.

4. The method according to claim 3, 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.

5. 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.

6. The method according to claim 5, 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