A fault early warning method and system for copper strip rolling production

Abstract

This invention provides a fault early warning method and system for copper strip rolling production. The method involves collecting rolling status signals from multiple monitoring locations during the copper strip rolling process; constructing a multimodal correlation map from the rolling status signal set to generate a rolling status correlation map containing correlation parameters between signals; generating a deviation evolution sequence based on the rolling status correlation map to obtain a set of deviation evolution sequences characterizing the changes in abnormal features over rolling time; modeling the risk transmission path of the deviation evolution sequence set to generate a risk transmission network model containing potential fault propagation paths and impact strengths; and outputting a fault early warning command for copper strip rolling production based on the risk transmission network model. The fault early warning command includes a fault type identifier and risk diffusion control parameters. This invention enables accurate identification, propagation path tracing, and impact range assessment of faults during copper strip rolling production.

Description

Technical Field

[0001] This invention relates to the field of fault early warning, and more specifically, to a fault early warning method and system for copper strip rolling production. Background Technology

[0002] In the field of copper strip rolling production, fault early warning is a key technical means to ensure production continuity and improve product quality. Its core lies in monitoring and analyzing physical signals during the rolling process to identify potential faults and take timely intervention measures. Currently, this is typically achieved by deploying multiple types of sensors to collect various physical signals during the rolling process, including vibration, temperature, and rolling force. Traditional data analysis models are then used to process these signals, establishing a correspondence between signals and fault types to achieve early warning functionality. However, this approach has significant limitations in practical applications. Due to the complex dynamic correlations between various physical signals during copper strip rolling, traditional models struggle to accurately depict the changing patterns of these correlations with rolling conditions. This results in insufficient ability to identify early, subtle fault characteristics. Furthermore, existing methods mostly focus on fault type identification, failing to effectively reveal the evolution path of faults from occurrence to propagation and the impact of different fault modes on the rolling process. Consequently, the generated early warning commands are difficult to directly guide the optimization and adjustment of rolling process parameters, failing to meet the refined fault early warning requirements of high-precision copper strip rolling production. Summary of the Invention

[0003] In view of this, embodiments of this application provide a fault early warning method and system for copper strip rolling production.

[0004] According to one aspect of the present invention, a fault early warning method for copper strip rolling production is provided. The method includes: collecting a set of rolling state signals from multiple monitoring locations during the copper strip rolling process; constructing a multimodal correlation graph from the set of rolling state signals to generate a rolling state correlation graph containing correlation parameters between signals, wherein nodes in the rolling state correlation graph represent signal types and edges represent the dynamic correlation strength between different types of signals; generating a deviation evolution sequence based on the rolling state correlation graph to obtain a set of deviation evolution sequences characterizing the changes of abnormal features with rolling time, wherein each sequence element in the deviation evolution sequence set corresponds to an abnormal feature description of a rolling period; modeling a risk transmission path from the set of deviation evolution sequences to generate a risk transmission network model containing potential fault propagation paths and influence strengths; and outputting a fault early warning command for copper strip rolling production based on the risk transmission network model, wherein the fault early warning command includes a fault type identifier and risk diffusion control parameters.

[0005] According to another aspect of the present invention, a fault warning system is provided, comprising: a processor; and a memory, wherein the memory stores a computer program that, when executed by the processor, causes the processor to perform the method described above.

[0006] The fault early warning method for copper strip rolling production provided by this invention collects rolling state signal sets from multiple monitoring locations during the copper strip rolling process. It then constructs a multimodal correlation map of the rolling state signal sets, generating a rolling state correlation map containing correlation parameters between signals. Based on the rolling state correlation map, it generates a set of deviation evolution sequences characterizing the changes in abnormal features over rolling time. Finally, it models the risk transmission path of the deviation evolution sequence set, generating a risk transmission network model containing potential fault propagation paths and impact strengths. Based on the risk transmission network model, it outputs fault early warning commands containing fault type identifiers and risk diffusion control parameters. This method transforms multimodal rolling state signals into a topologically structured correlation network. Through dynamic deviation evolution analysis, it captures the changing patterns of abnormal features over time. The constructed risk transmission network model can identify fault propagation paths and quantify impact strengths. The generated early warning commands not only contain fault type information but also provide risk diffusion parameters that can be used for control decisions. This enables accurate identification of faults, tracing of propagation paths, and assessment of the impact range during copper strip rolling production, providing comprehensive decision support for rolling process optimization. Attached Figure Description

[0007] Figure 1 A schematic diagram illustrating the application scenarios provided in the embodiments of this application; Figure 2 A schematic diagram illustrating the implementation process of a fault early warning method for copper strip rolling production provided in this application embodiment; Figure 3 This is a schematic diagram of the composition structure of a fault early warning device provided in an embodiment of this application; Figure 4 This is a schematic diagram of the hardware entity of a fault early warning system provided in an embodiment of this application. Detailed Implementation

[0008] The fault early warning method for copper strip rolling production provided in this application embodiment can be applied to, for example... Figure 1In the application environment shown, the sensing device 102 communicates with the fault warning system 104 via a network. The data storage system can store the data that the fault warning system 104 needs to process. The data storage system can be integrated into the fault warning system 104, or it can be located in the cloud or on other network servers. When a fault warning is needed, the fault warning system 104 can obtain a set of rolling status signals from the sensing device 102. The sensing device 102 can be various sensors, and the fault warning system 104 can be the computer system of the copper strip rolling production line, or a standalone server deployed in the cloud, or a server cluster consisting of multiple servers.

[0009] Please refer to Figure 2 The fault early warning method for copper strip rolling production provided in this application embodiment is applied to a fault early warning system, and specifically includes the following steps: Step S100: Collect a set of rolling status signals from multiple monitoring locations during the copper strip rolling process.

[0010] In the copper strip rolling process, rolling status signals reflect various physical phenomena and equipment operating conditions. Monitoring locations include, but are not limited to, roll bearings, motor housings, copper strip surfaces, and rolling oil pipes. Different monitoring locations correspond to different types of rolling status signals. For example, at the roll bearings, vibration and temperature signals are primarily collected. Vibration signals reflect the smoothness of the roll bearing's operation; abnormal vibration may indicate wear, loosening, or other problems. Temperature signals reflect the bearing's heat generation; excessively high temperatures may be due to poor lubrication or excessive load. Temperature and current sensors are installed on the motor housing. Temperature signals monitor the motor's heat generation, while current signals reflect the motor's load; abnormal current fluctuations may indicate a motor malfunction.

[0011] For the copper strip surface, stress sensors collect surface stress signals, which reflect the stress conditions experienced during rolling. Rolling force signals are acquired through pressure sensors installed on the rolling equipment, directly affecting the thickness and shape of the copper strip. Shape signals are collected using specialized shape detection equipment to assess the flatness and shape accuracy of the copper strip. Temperature and flow sensors are installed on the rolling oil pipeline to collect the temperature and flow signals of the rolling oil, respectively. The temperature of the rolling oil affects its lubrication and cooling performance, while the flow rate determines whether the rolling oil can function effectively.

[0012] Step S200: Construct a multimodal correlation map for the set of rolling state signals to generate a rolling state correlation map containing correlation parameters between signals. The nodes of the rolling state correlation map represent signal types, and the edges represent the dynamic correlation strength between different types of signals.

[0013] The construction of multimodal correlation maps aims to uncover potential correlations between different types of signals in a set of rolling state signals. Correlation parameters are used as quantitative indicators to describe the degree and characteristics of correlation between signals.

[0014] The rolling state correlation diagram graphically illustrates the relationships between signals. Nodes represent different signal types, such as vibration signals, temperature signals, and rolling force signals. Each node has corresponding attributes, including the location of the signal source and the characteristics of the propagation medium. The location of the signal source clarifies the location where the signal is generated, while the characteristics of the propagation medium describe the properties of the medium through which the signal propagates, such as the damping and elasticity of the medium.

[0015] The edges represent the dynamic correlation strength between different types of signals. This correlation strength is not fixed but dynamically adjusts with changes in rolling conditions. For example, at higher rolling speeds, the correlation strength between the rolling force signal and the vibration signal may increase; while at lower rolling oil temperatures, the correlation between the rolling oil temperature signal and the rolling force signal may be more pronounced.

[0016] Constructing multimodal correlation maps can help to more comprehensively understand the interactions between various factors during copper strip rolling. By analyzing the nodes and edges in the map, potential causal relationships and cooperative change patterns between signals can be discovered. This correlation map can provide a basis for fault early warning; when abnormal changes occur in the correlation between certain signals, early warning signals can be issued in a timely manner.

[0017] As one implementation method, step S200 can be implemented as the following steps S210~S250: Step S210: The rolling state signal set is divided into layers according to the monitoring position and physical field type to obtain multi-level signal subsets. The layer division of the multi-level signal subsets is based on the structural topology of the rolling equipment and the signal propagation path.

[0018] The structural topology of rolling mills describes the spatial positions and connections between its various components. For example, a rolling mill consists of components such as rolls, stands, and drive systems, and these components have clearly defined structural relationships. Signal propagation paths describe how signals are transmitted from one location to another within the equipment.

[0019] Based on this information, the rolling status signal set is divided into different levels. This hierarchical division helps to more clearly analyze the correlations between signals, as signals at different levels have different characteristics and functions. For example, signals at the equipment body level mainly reflect the operating status of the equipment itself, while signals at the rolling process level are directly related to the rolling quality of the copper strip.

[0020] Each level in the multi-level signal subset contains signals related to the monitoring location and physical field type. There are certain relationships between levels; changes in the signal at a higher level may affect the signal at a lower level, and vice versa. This hierarchical division simplifies and categorizes complex signal sets.

[0021] As one implementation method, step S210 can be implemented as the following steps S211~S215: Step S211: By analyzing the equipment structure layout of the copper strip rolling production line, determine the spatial position relationship of key equipment such as rolling mill, tension roller, guide roller, and coiler, as well as the distribution of signal monitoring points.

[0022] The equipment layout of a copper strip rolling production line can be obtained by acquiring the production line's design drawings and related technical documents, as well as through on-site measurements. The rolling mill is the core equipment in copper strip rolling; its position and operating status directly affect the rolling quality of the copper strip. Tension rollers are used to adjust the tension of the copper strip, ensuring the stability of the rolling process; the accuracy of their position is crucial for tension control. Guide rollers guide the running direction of the copper strip; adjustments to their position and angle affect the flatness and shape accuracy of the copper strip. The coiler is used to coil the rolled copper strip; its position and running speed are closely related to the coordination of the rolling process.

[0023] Step S212: Construct a three-level signal acquisition architecture based on the equipment structure layout. The first level is the equipment body monitoring layer, which includes vibration and temperature signals installed on the roll bearings and motor housing. The second level is the rolling process monitoring layer, which includes process signals such as copper strip surface stress, rolling force, and strip shape. The third level is the environmental correlation layer, which includes environmental signals such as rolling oil temperature and workshop temperature and humidity.

[0024] The first level, the equipment body monitoring layer, primarily focuses on the operational status of the equipment itself. Vibration and temperature signals of the roll bearings reflect the bearing's health condition. Abnormal vibration may be caused by bearing wear, poor lubrication, etc., while excessively high temperatures may be due to overload, poor heat dissipation, etc. Vibration and temperature signals of the motor housing are equally important; the normal operation of the motor is crucial to the stability of the entire rolling process. By monitoring the motor's vibration and temperature, potential motor malfunctions can be detected in a timely manner, preventing equipment damage and production interruptions. The second level, the rolling process monitoring layer, is directly related to the rolling quality of the copper strip. The copper strip surface stress signal reflects the stress on the copper strip during the rolling process. The rolling force signal is a key parameter for controlling the thickness and shape of the copper strip; by precisely controlling the rolling force, copper strips that meet quality requirements can be produced. The strip shape signal is used to evaluate the flatness and shape accuracy of the copper strip. The third level, the environmental correlation layer, considers the impact of environmental factors on the rolling process. The rolling oil temperature has a significant impact on the lubrication performance and cooling effect of the rolling oil. A suitable rolling oil temperature ensures good lubrication between the rolls and the copper strip, reducing wear and improving rolling quality. Changes in workshop temperature and humidity affect the physical properties of the copper strip and the operational stability of the equipment. For example, excessive humidity may cause the copper strip surface to rust, affecting product quality.

[0025] Step S213: Classify and aggregate the rolling status signal set according to the three-level signal acquisition architecture, merge signals of the same level and type of physical field into signal subgroups, and generate equipment body monitoring signal group, rolling process monitoring signal group and environmental related signal group.

[0026] Based on the three-tiered signal acquisition architecture, signals are classified according to tier and physical field type. Signals from the same tier and type of physical field share similar properties and characteristics; merging them into subgroups facilitates subsequent analysis and processing. For example, at the equipment body monitoring tier, all vibration signals are merged into one subgroup, and all temperature signals into another. This allows for centralized analysis of signals of the same type, making it easier to detect abnormal changes in the signals.

[0027] The equipment body monitoring signal group includes all signal subgroups in the equipment body monitoring layer, such as the vibration signal subgroup for the roll bearing and the temperature signal subgroup for the motor housing. This signal group is mainly used to monitor the operating status of the equipment and promptly detect potential equipment failures.

[0028] The rolling process monitoring signal group encompasses various signal subgroups within the rolling process monitoring layer, such as the copper strip surface stress signal subgroup, rolling force signal subgroup, and strip shape signal subgroup. This signal group is directly related to the rolling quality of the copper strip. By analyzing these signals, rolling parameters can be adjusted in real time to ensure the quality of the copper strip.

[0029] The environmental correlation signal group includes signal subgroups from the environmental correlation layer, such as the rolling oil temperature signal subgroup and the workshop temperature and humidity signal subgroup. This signal group is used to assess the impact of environmental factors on the rolling process so that appropriate measures can be taken for adjustment.

[0030] By classifying and grouping signals, complex signal sets can be effectively organized and managed, improving the efficiency and accuracy of signal processing.

[0031] Step S214: Perform internal signal consistency verification on each signal group. By comparing the time series trends of the same physical field signals within the same level, eliminate abnormal signal samples whose trend deviation exceeds the preset allowable range, and retain the valid signal units within the signal group.

[0032] Signals from the same physical field within the same level should exhibit similar time-series trends because they reflect different aspects of the same physical phenomenon. For example, in the vibration signal subgroup of the equipment body monitoring layer, vibration signals at different locations should have similar fluctuation trends. If the time-series trend of a vibration signal differs significantly from other signals, it may indicate an anomaly in that signal. This anomaly could be caused by sensor malfunction, signal transmission interference, or localized equipment failure.

[0033] The preset allowable range is a threshold determined based on a large amount of experimental data and experience. In practical applications, the normal fluctuation range of the signal can be calculated by statistically analyzing historical data of the same physical field signal within the same level, and then the preset allowable range can be determined based on this range.

[0034] There are various methods to compare the time series trends of signals from the same physical field within the same level. A common method is to calculate the correlation coefficient between the signals. The closer the correlation coefficient is to 1, the more similar the trends of the signals; the closer the correlation coefficient is to 0, the greater the difference in trends between the signals. If the correlation coefficient of a signal with other signals is lower than a preset threshold, the signal is considered an anomalous signal sample.

[0035] Another approach is to use the Dynamic Time Warping (DTW) algorithm, which compares the similarity of two time series without considering strict alignment of the signal time axes. By calculating the DTW distance between signals, if the DTW distance of a signal from other signals exceeds a preset allowable range, it is identified as an anomalous signal sample.

[0036] Step S215: Arrange the verified equipment body monitoring signal group, rolling process monitoring signal group and environmental related signal group in hierarchical order to generate a multi-level signal subset containing hierarchical identifiers and physical field attributes. The hierarchical identifiers are used to distinguish the acquisition level to which the signal belongs, and the physical field attributes are used to describe the physical characteristics of the signal.

[0037] Arranging the verified signal groups in hierarchical order is to construct an ordered multi-level signal subset. The hierarchical order is as follows: equipment body monitoring layer, rolling process monitoring layer, and environmental correlation layer.

[0038] Hierarchical identifiers help quickly distinguish the acquisition level to which a signal belongs. For example, the number 1 can represent the equipment body monitoring level, the number 2 the rolling process monitoring level, and the number 3 the environmental correlation level. In subsequent analysis and processing, hierarchical identifiers can facilitate the classification and processing of signals from different levels.

[0039] Physical field attributes are used to describe the essential physical characteristics of a signal. For example, for a vibration signal, physical field attributes may include the frequency, amplitude, and phase of the vibration; for a temperature signal, physical field attributes may include the temperature value and rate of change. The generated multi-level signal subset is a structured dataset containing signal groups at each level, along with their corresponding level identifiers and physical field attributes. This dataset can serve as input for subsequent cross-level feature association processing, providing a solid data foundation for further exploring potential correlations between signals.

[0040] Step S220: Perform cross-level feature association processing on the multi-level signal subset, construct the causal relationship between upper and lower level signals through the inter-level signal transfer function, and generate a hierarchical association feature matrix. The row vectors of the hierarchical association feature matrix correspond to the upper level signal features, the column vectors correspond to the lower level signal features, and the matrix elements represent the inter-level signal transfer efficiency.

[0041] Inter-level signal transfer functions describe the patterns and characteristics of signal transmission from upper to lower levels. They can be obtained through methods such as experimental data fitting and theoretical modeling. For example, between the equipment body monitoring layer and the rolling process monitoring layer, the vibration signal of the roll bearing may be transmitted to the rolling force signal through the mill stand. By analyzing a large amount of experimental data, a transfer function from the vibration signal to the rolling force signal can be established, which describes how changes in the vibration signal affect the rolling force signal.

[0042] Causal relationships refer to the causal connections between signals at different levels. Changes in signals at higher levels will cause corresponding changes in signals at lower levels, and vice versa. By establishing causal relationships, we can gain a deeper understanding of the interactions between equipment operating conditions, the rolling process, and environmental factors.

[0043] A hierarchical correlation feature matrix is ​​a matrix used to represent the correlation between signals at different levels. The row vectors of the matrix correspond to the features of the upper-level signals, and the column vectors correspond to the features of the lower-level signals. The matrix elements represent the signal transmission efficiency between levels, that is, the degree of influence of the upper-level signal features on the lower-level signal features. Signal transmission efficiency can be obtained through experimental measurement or theoretical calculation. For example, by changing a certain feature of the upper-level signal and observing the corresponding change in the lower-level signal, the signal transmission efficiency can be calculated.

[0044] The generation of hierarchical correlation feature matrices helps to more intuitively analyze the correlation between signals at different levels. By analyzing the matrix elements, it is possible to identify which upper-level signal features have a greater impact on lower-level signal features and which have a smaller impact. For example, if a vibration signal feature in the equipment body monitoring layer is found to have a significant impact on the rolling force signal in the rolling process monitoring layer, then when an abnormality in this vibration signal is detected, changes in the rolling force signal can be monitored in a timely manner, providing early warning of potential rolling quality problems.

[0045] Step S230: Calculate the mutual dependence of signal types within the same level and the transmission attenuation coefficient of cross-level signals based on the hierarchical correlation feature matrix, and generate a multi-dimensional correlation strength index. The mutual dependence is used to describe the cooperative change law of signals within the same level, and the transmission attenuation coefficient is used to describe the energy loss characteristics of signals propagating between levels.

[0046] The interdependence of signal types within the same level reflects the degree of coordinated change between different signals at the same level. For example, in the rolling process monitoring layer, the surface stress signal of copper strip and the rolling force signal may have a strong interdependence. When the rolling force increases, the surface stress of copper strip will also increase accordingly. This coordinated change can be quantified by the interdependence.

[0047] One method for calculating cross-dependency is to use mutual information theory. Mutual information is an indicator of the dependency between two random variables, obtained by calculating the difference between the joint entropy and the marginal entropy of the two variables. In practical applications, signals within the same level can be treated as random variables, and their time series data can be analyzed to calculate the mutual information value. The larger the mutual information value, the stronger the cross-dependency between the signals. Another method is to use the Granger causality test. This test can determine whether one signal is a Granger cause of another signal, that is, whether a change in one signal can predict a change in another signal. By performing the Granger causality test on signals within the same level, the causal relationship and cross-dependency between the signals can be obtained.

[0048] The transmission attenuation coefficient of cross-level signals describes the energy loss characteristics of signals propagating between levels. Because signals are affected by various factors during propagation, such as damping and reflection of the propagation medium, the signal strength gradually attenuates. For example, when the vibration signal of a rolling mill bearing is transmitted to the rolling force signal of the monitoring layer in the rolling process, it passes through components such as the mill stand, which attenuate the vibration signal to some extent.

[0049] The transmission attenuation coefficient can be obtained through experimental measurement or theoretical calculation. In experiments, signals can be simultaneously acquired at different monitoring points, and the changes in signal intensity can be compared to calculate the attenuation ratio. Theoretical calculations, based on physical models of signal propagation and considering factors such as the characteristics of the propagation medium and the propagation distance, derive the formula for calculating the transmission attenuation coefficient.

[0050] The multidimensional correlation strength index is an index generated by comprehensively considering multiple factors such as interdependence and propagation attenuation coefficient, which can comprehensively describe the correlation strength between signals. The dimensionality of the index matches the number of levels and physical field types of the signal subset.

[0051] As one implementation method, step S230 can be implemented as the following steps S231~S236: Step S231: Perform row standardization on the hierarchical correlation feature matrix so that the sum of the elements in each row is 1, to obtain the probabilistic hierarchical correlation matrix, where the matrix elements represent the probability of the influence of the upper-level signal features on the lower-level signal features.

[0052] Row standardization is a crucial step in converting the hierarchical association feature matrix into a probabilistic hierarchical association matrix. The elements in the hierarchical association feature matrix represent the signal transmission efficiency between hierarchies, but the range of values ​​for these elements may differ, making direct comparison and analysis difficult.

[0053] The specific method for row standardization is to divide each element of the hierarchical association feature matrix by the sum of the elements in that row. This is done to ensure that the sum of the elements in each row is 1, thereby converting the matrix elements into probability values. For example, for a row in the hierarchical association feature matrix with elements a1, a2, and a3, the row sum is S = a1 + a2 + a3.

[0054] The elements in the probabilistic hierarchical correlation matrix represent the probability of the influence of upper-level signal features on lower-level signal features.

[0055] Step S232: Within the same level, the signal type is treated as a random variable, and the mutual information value is calculated by analyzing the joint probability distribution of the signal time series. The mutual information value is described by the difference between the joint entropy and the marginal entropy, and is used to quantify the nonlinear dependence between signals in the same level.

[0056] Treating signal types as random variables within the same level is a processing method based on information theory. A random variable is a variable whose value is random under certain conditions. Signal types within the same level may have different values ​​at different times, therefore they can be regarded as random variables.

[0057] The joint probability distribution of a signal time series describes the probability of multiple signals taking values ​​simultaneously. By analyzing the joint probability distribution, we can understand the joint variation patterns between different signal types within the same level. For example, in the rolling process monitoring layer, the joint probability distribution of the copper strip surface stress signal and the rolling force signal can reflect the probability of these two signals simultaneously taking different values. Mutual information is an indicator used to measure the dependency between two random variables. It is described by the difference between joint entropy and marginal entropy. Joint entropy is a measure of the joint uncertainty of multiple random variables, while marginal entropy is a measure of the uncertainty of a single random variable. The larger the mutual information value, the stronger the dependency between the two random variables.

[0058] One method for calculating mutual information is discretization. First, the signal time series is discretized, dividing the continuous signal values ​​into several intervals, each corresponding to a discrete value. Then, the frequency of each discrete value is counted, and the joint probability distribution and marginal probability distributions are calculated. Finally, the mutual information value is calculated according to the definition of mutual information. Another method is kernel density estimation. This method can directly process continuous signal time series without discretization. Through kernel density estimation, the probability density function of the signal can be obtained, and then the joint probability distribution and marginal probability distributions can be calculated, ultimately yielding the mutual information value. Mutual information can quantify the nonlinear dependencies between signals at the same level. Unlike linear correlation coefficients, mutual information can capture nonlinear relationships between signals, thus having a greater advantage in analyzing complex signal correlations.

[0059] Step S233: Normalize the mutual information values ​​to map them to a range of 0 to 1 to obtain the cross-dependency of signals at the same level. The normalization process is based on a linear transformation of the mutual information value range of all signal pairs within the same level.

[0060] Normalization is performed by linearly transforming the range of mutual information values ​​for all signal pairs within the same level. Specifically, the minimum and maximum mutual information values ​​for all signal pairs within the same level are first identified, denoted as min and max, respectively. For a given signal pair's mutual information value I, its normalized mutual dependency D can be calculated using the following formula: D = (I - min) / (max - min). Through this linear transformation, the mutual information value is mapped to a value between 0 and 1. When the mutual dependency is 0, it indicates that there is no dependency between the two signals; when the mutual dependency is 1, it indicates that there is a strong dependency between the two signals.

[0061] Step S234: Between different levels, an attenuation model is constructed based on the physical characteristics of the signal propagation path. The attenuation model considers the damping coefficient of the propagation medium, the propagation distance, and the interface reflection characteristics. By simulating the energy change process of the signal propagating from the upper node to the lower node, the attenuation ratio of the signal strength is calculated.

[0062] Between different levels, signal propagation is affected by various factors, causing signal strength to gradually attenuate. The damping coefficient of the propagation medium refers to the medium's resistance to signal propagation. Different propagation media have different damping coefficients; the larger the damping coefficient, the greater the energy loss of the signal during propagation. The farther the signal travels, the greater its strength attenuation. This is because the signal continuously interacts with the medium during propagation, gradually dissipating energy.

[0063] Interface reflection characteristics refer to the reflection of signals at the interface between different media. When a signal propagates from one medium to another, it will be reflected at the interface, and some of the signal energy will be reflected back, resulting in signal attenuation.

[0064] Based on these physical characteristics, attenuation models can be constructed to simulate the energy change process of a signal propagating from an upper node to a lower node. One attenuation model is based on the wave equation. This model considers factors such as signal propagation speed, damping coefficient, and propagation distance. By solving the wave equation, the energy change of the signal during propagation can be obtained. Another approach is to use empirical models. Through a large amount of experimental data, the relationship between signal strength attenuation and factors such as propagation medium, propagation distance, and interface reflection is summarized, and an empirical formula is established. For example, the signal strength under different propagation distances and media can be measured experimentally, and then an attenuation curve can be fitted. The attenuation ratio of the signal strength can then be calculated based on this curve.

[0065] Step S235: Convert the attenuation ratio into a transmission attenuation coefficient. The transmission attenuation coefficient is positively correlated with the attenuation ratio and is used to describe the degree of energy loss during cross-level signal transmission.

[0066] The attenuation ratio is the proportion of signal strength attenuation during cross-level propagation, while the transmission attenuation coefficient is a more intuitive and easier-to-use indicator used to describe the degree of energy loss during cross-level signal transmission.

[0067] The transmission attenuation coefficient is positively correlated with the attenuation ratio; that is, the larger the attenuation ratio, the larger the transmission attenuation coefficient. The attenuation ratio can be converted to the transmission attenuation coefficient using a simple linear transformation. For example, if the attenuation ratio is r and the transmission attenuation coefficient is k, then k = ar + b, where a and b are constants, and a > 0.

[0068] In practical applications, the values ​​of a and b can be determined based on specific needs and experimental data. Generally, when the attenuation ratio is 0, the transmission attenuation coefficient is also 0; when the attenuation ratio is 1, the transmission attenuation coefficient reaches its maximum value. Through this conversion, the attenuation ratio can be transformed into a transmission attenuation coefficient that better meets the needs of practical applications.

[0069] The transmission attenuation coefficient can be used to evaluate the quality of signal transmission across multiple levels. A large transmission attenuation coefficient indicates significant energy loss during transmission, which may affect the accuracy and reliability of signals at lower levels. Therefore, when designing signal acquisition and transmission systems, it is necessary to minimize the transmission attenuation coefficient and improve signal transmission efficiency.

[0070] Step S236: Combine the interdependence of the same level and the transmission attenuation coefficient of the cross level in hierarchical order to generate a multi-dimensional correlation strength index that includes intra-level and inter-level correlation characteristics. The dimension of the index matches the number of levels and the number of physical field types in the signal subset.

[0071] The multidimensional correlation strength index is a comprehensive indicator that takes into account both the interdependence of signals within the same level and the transmission attenuation coefficient of signals across levels. The interdependence within a level describes the coordinated change pattern between signals at the same level, while the transmission attenuation coefficient between levels describes the energy loss of signals propagating between different levels.

[0072] Combining interdependency and propagation attenuation coefficients in hierarchical order is to construct an ordered multi-dimensional index. The dimension of the index matches the number of levels and physical field types in the signal subset. By generating a multi-dimensional correlation strength index, the correlation between signals can be described more comprehensively and accurately. This index can be used to analyze the interaction between signals at different levels and with different physical field types, providing a basis for spectrum construction and fault early warning.

[0073] Step S240: Construct the initial topology of the dynamic correlation graph based on the multi-dimensional correlation strength index, where nodes correspond to signal types at each level, directed edges represent the signal transmission direction, and edge weights combine mutual dependence and transmission attenuation coefficient.

[0074] The initial topology of the dynamic correlation graph is a graphical representation built upon multi-dimensional correlation strength indices. Nodes correspond to signal types at each level, such as vibration and temperature signals in the equipment body monitoring layer, and surface stress and rolling force signals on the copper strip in the rolling process monitoring layer. Each node represents a signal type and has a clear physical meaning.

[0075] Directed edges represent the direction of signal transmission. Signals of different types exhibit certain transmission relationships, which can be represented by directed edges. For example, between the equipment body monitoring layer and the rolling process monitoring layer, the vibration signal of the roll bearing may affect the rolling force signal; therefore, a directed edge can be drawn from the vibration signal node to the rolling force signal node. Edge weights combine mutual dependence and transmission attenuation coefficients. Mutual dependence describes the degree of coordinated change between signals within the same layer, while the transmission attenuation coefficient describes the energy loss of the signal propagating between layers. By comprehensively considering these two factors, the correlation strength between signals can be represented more accurately. For example, if the mutual dependence between two signals is high and the transmission attenuation coefficient is low, the edge weight value will be large, indicating a strong correlation between the two signals; conversely, if the mutual dependence is low and the transmission attenuation coefficient is large, the edge weight value will be small, indicating a weak correlation between the two signals.

[0076] Step S250: Perform spatiotemporal dynamic optimization on the initial topology. By introducing rolling speed and copper strip thickness variation parameters, adjust the temporal evolution characteristics of edge weight values ​​to generate a rolling state correlation map that is dynamically updated with the rolling conditions. The attributes of nodes in the rolling state correlation map include the location of the signal field source and the characteristics of the propagation medium. The edge weight values ​​are obtained by weighted averaging of correlation strength indices within a sliding time window.

[0077] While the initial topology can demonstrate the basic relationships between signals, the actual rolling process involves constantly changing rolling conditions, which in turn alter the relationships between signals. Therefore, spatiotemporal dynamic optimization of the initial topology is necessary.

[0078] Rolling speed and copper strip thickness are two important rolling parameters. Changes in rolling speed affect the operating status of the equipment and the rolling process of the copper strip, thereby altering the correlation between signals. For example, when the rolling speed increases, the vibration of the rolls intensifies, which may cause changes in the correlation strength between vibration signals and other signals. Changes in copper strip thickness also affect the rolling process; copper strips of different thicknesses may require different rolling forces and stress distributions during rolling, thus affecting the correlation between signals.

[0079] By introducing parameters related to rolling speed and copper strip thickness variation, the time evolution characteristics of the edge weight values ​​can be adjusted. Specifically, the edge weight values ​​can be dynamically adjusted based on changes in rolling speed and copper strip thickness. For example, when the rolling speed increases, the edge weight values ​​related to the vibration signal are increased; when the copper strip thickness varies significantly, the edge weight values ​​related to the rolling force and strip shape signals are adjusted.

[0080] The edge weights are obtained by weighted averaging of correlation strength indices within a sliding time window. The sliding time window is a fixed-length time interval that slides along the time axis over time. Within each time window, correlation strength indices are calculated, and then these indices are weighted and averaged to obtain the edge weights. This method allows the edge weights to reflect the real-time correlation strength between signals and dynamically update as rolling conditions change.

[0081] The attributes of nodes in the rolling state correlation diagram include the location of the signal source and the characteristics of the propagation medium. The location of the signal source clarifies the origin of the signal, while the characteristics of the propagation medium describe the properties of the medium through which the signal passes during propagation.

[0082] As one implementation method, step S250 can be implemented as the following steps S251~S256: Step S251: Collect the copper strip thickness setting value and the actual rolling speed curve of the current rolling batch, divide the rolling speed curve into multiple characteristic time periods including acceleration section, uniform speed section and deceleration section, and extract the rolling speed change rate of each time period.

[0083] Collecting the setpoint for copper strip thickness and the actual rolling speed curve for the current rolling batch is the foundation for spatiotemporal dynamic optimization. The setpoint for copper strip thickness is the thickness specified in the production plan and is one of the important parameters for controlling the rolling process. The actual rolling speed curve records the change in rolling speed over time during the rolling process.

[0084] At the beginning of the rolling process, the rolling speed needs to be accelerated to reach the set value; this stage is called the acceleration stage. During the rolling process, in order to ensure the quality of the copper strip and production efficiency, the rolling speed is kept at a relatively stable level; this stage is called the uniform speed stage. At the end of the rolling process, the rolling speed needs to be decelerated to bring the rolling process to a smooth stop; this stage is called the deceleration stage.

[0085] Extracting the rolling speed change rate for each time period is to more accurately describe the variation in rolling speed. The rolling speed change rate is the rate of change of rolling speed over time, which can be obtained by differentiating the rolling speed curve. In the acceleration phase, the rolling speed change rate is positive and relatively large; in the constant speed phase, the rolling speed change rate is close to 0; and in the deceleration phase, the rolling speed change rate is negative.

[0086] Step S252: Construct a time-dynamic adjustment factor based on the rolling speed change rate, wherein the adjustment factor for the acceleration and deceleration sections is greater than that for the constant speed section, in order to enhance the sensitivity of signal correlation during unsteady rolling.

[0087] The time-dynamic adjustment factor is an adjustment parameter constructed based on the rolling speed change rate, used to adjust the time evolution characteristics of the edge weight values. In the unsteady rolling process (acceleration and deceleration sections), the rolling speed varies significantly, and the operating state of the equipment and the rolling process of the copper strip also change considerably. At this time, the correlation between signals may become more complex and sensitive. To enhance sensitivity to these changes, the adjustment factors for the acceleration and deceleration sections are set higher than those for the constant speed section. By constructing the time-dynamic adjustment factor, the sensitivity of signal correlation can be enhanced during unsteady rolling, allowing the edge weight values ​​to reflect the changing relationships between signals more promptly.

[0088] Step S253: Construct a spatial dynamic adjustment factor based on the copper strip thickness variation parameter. When the copper strip thickness variation exceeds the preset fluctuation range, increase the correlation weight adjustment amplitude between the corresponding rolling force signal and the strip shape signal.

[0089] The variation in copper strip thickness is another factor affecting the rolling process. Changes in copper strip thickness directly impact rolling force and strip shape. When the variation in copper strip thickness exceeds the preset fluctuation range, it indicates a potential anomaly in the rolling process, requiring an increase in the correlation weight adjustment between the corresponding rolling force signal and strip shape signal.

[0090] The preset fluctuation range is a range determined based on production experience and quality requirements. If the change in copper strip thickness is within this range, it indicates that the rolling process is relatively stable; if it exceeds this range, it may affect the quality of the copper strip.

[0091] The spatial dynamic adjustment factor is an adjustment parameter constructed based on the copper strip thickness variation parameter. When the copper strip thickness variation exceeds the preset fluctuation range, the value of the spatial dynamic adjustment factor is increased, thereby increasing the correlation weight adjustment amplitude between the corresponding rolling force signal and the strip shape signal.

[0092] By constructing a spatial dynamic adjustment factor, the correlation weight between the rolling force signal and the strip shape signal can be dynamically adjusted according to the change in copper strip thickness, thereby improving the sensitivity to changes in copper strip quality.

[0093] Step S254: Couple the time dynamic adjustment factor and the space dynamic adjustment factor to generate a time-space coupling adjustment coefficient. The coefficient is obtained by multiplying the rolling speed change rate and the copper strip thickness change.

[0094] The spatiotemporal coupling adjustment coefficient is an adjustment coefficient generated by comprehensively considering both time-based and space-based dynamic adjustment factors. By coupling these two adjustment factors, the influence of changes in rolling conditions on signal correlation can be reflected more comprehensively.

[0095] The rate of change of rolling speed reflects the temporal variation characteristics of the rolling process, while the change in copper strip thickness reflects the spatial variation characteristics of the rolling process. By multiplying these two parameters, a comprehensive spatiotemporal coupling adjustment coefficient can be obtained.

[0096] Step S255: Correct the edge weight values ​​in the initial topology by using the spatiotemporal coupling adjustment coefficient to obtain the corrected edge weight values; wherein, the correction formula is achieved by multiplying the original weight values ​​by the spatiotemporal coupling adjustment coefficient, and the original weight values ​​are the weighted average results of the correlation strength index within the sliding time window.

[0097] Adjusting the edge weights in the initial topology using spatiotemporal coupling coefficients is a key step in spatiotemporal dynamic optimization.

[0098] The original weight values ​​are the weighted average of the correlation strength indices within the sliding time window, reflecting the basic correlation strength between signals. The spatiotemporal coupling adjustment coefficient takes into account the influence of the rolling speed variation rate and the copper strip thickness variation on the signal correlation.

[0099] The correction formula is, for example,: Corrected edge weight value = Original weight value × Spatiotemporal coupling adjustment coefficient. Based on this, the edge weight value can be dynamically adjusted according to changes in rolling conditions. For example, when the rolling speed change rate is large and the copper strip thickness change is also large, the spatiotemporal coupling adjustment coefficient will increase, thereby increasing the corrected edge weight value, reflecting the enhanced correlation strength between signals. By correcting the edge weight value, the dynamic correlation graph can more accurately reflect the real-time correlation between signals, improving the accuracy of fault early warning and quality control.

[0100] Step S256: Smooth the corrected edge weight values, eliminate high-frequency fluctuations by moving average filtering, and generate an edge weight time series that changes continuously with the rolling conditions, which is used to update the dynamic topology of the rolling state correlation graph.

[0101] Moving average filtering calculates a smoothed edge weight value by averaging the edge weight values ​​over a certain time window. This filtering eliminates high-frequency fluctuations caused by measurement errors, noise, and other factors, allowing the edge weight values ​​to change continuously with the rolling conditions. The generated edge weight time series can be used to update the dynamic topology of the rolling state correlation graph. As the rolling conditions change, the edge weight values ​​are continuously updated, and the dynamic topology of the rolling state correlation graph adjusts accordingly, thus more accurately reflecting the correlation between signals.

[0102] Step S300: Generate deviation evolution sequences based on the rolling state correlation map to obtain a set of deviation evolution sequences that characterize the changes of abnormal features with rolling time. Each sequence element in the deviation evolution sequence set corresponds to an abnormal feature description of a rolling period.

[0103] The rolling state correlation map shows the correlation between signals. Based on this, the generation of deviation evolution sequence can be used to analyze the changes of abnormal features with rolling time in depth.

[0104] Abnormal characteristics refer to features that differ from the signal characteristics under normal rolling conditions. For example, under normal circumstances, the vibration signal of a roll bearing has a certain frequency and amplitude range. If the vibration signal at a certain moment exceeds this range, the signal can be considered to have abnormal characteristics.

[0105] The deviation evolution sequence set is a series of sequences characterizing the changes of abnormal features over rolling time. Each sequence element corresponds to an abnormal feature description during a specific rolling period. By analyzing the deviation evolution sequence set, we can understand the occurrence, development, and evolution of abnormal features, providing important basis for fault early warning and diagnosis.

[0106] As one implementation method, step S300 can be implemented as the following steps S310~S360: Step S310: Construct a normal state reference map library of the physical field linkage sensing network. The reference map library contains rolling state correlation map samples of copper strips with different rolling specifications and different material grades under normal production conditions. Each sample is labeled with the corresponding rolling speed, copper strip thickness, rolling force and other process parameters.

[0107] The normal state baseline map library of the physical field linkage sensing network is the basis for generating deviation evolution sequences. Copper strips with different rolling specifications (such as copper strip thickness and width) and material grades have different rolling state correlation maps under normal production conditions.

[0108] By collecting a large number of rolling state correlation graph samples under normal production conditions and labeling the corresponding process parameters such as rolling speed, copper strip thickness, and rolling force, a comprehensive benchmark graph library can be constructed. For example, for copper strips of different thicknesses, the rolling state correlation graphs may differ under the same rolling speed and rolling force. By collecting these graph samples under different conditions, accurate references can be provided for subsequent comparative analysis.

[0109] Step S320: Select the reference map sample that is most similar to the current rolling condition from the normal state reference map library, and determine the reference reference map with the highest similarity by calculating the process parameter matching degree. The process parameter matching degree is the normalized deviation value of parameters such as rolling speed, copper strip thickness, and material hardness.

[0110] The matching degree of process parameters is calculated by comprehensively considering the normalized deviation values ​​of multiple process parameters. Parameters such as rolling speed, copper strip thickness, and material hardness have a significant impact on the correlation between the rolling process and the rolling state.

[0111] First, each process parameter is normalized to map the value ranges of different parameters to the same interval, eliminating the influence of different parameter magnitudes. For example, a minimum-maximum normalization method can be used to map the value range of each process parameter to between 0 and 1. Then, the deviation between the current process parameter and the process parameters of the benchmark sample is calculated. Finally, the normalized deviations of each process parameter are considered comprehensively to calculate the process parameter matching degree. A weighted average method can be used, assigning weights based on the degree of influence of different process parameters on the rolling state. For example, the weights of rolling force and rolling speed can be set higher because they have a greater impact on the rolling process; the weights of width and thickness can be relatively lower. Based on the process parameter matching degree calculation, the benchmark sample with the highest similarity is selected as the reference benchmark sample. If multiple samples with similar scores exist, further screening can be performed by comparing the surface quality grades of the copper strips corresponding to the samples, selecting the sample with the higher surface quality grade as the reference.

[0112] As one implementation method, step S320 can be implemented as the following steps S321~S326: Step S321: Extract the set of key process parameters for the current rolling condition, including parameters such as copper strip material grade, target thickness, target width, rolling speed setting value, and rolling force setting value.

[0113] Extracting the key process parameters for the current rolling condition is crucial for accurately describing that condition. The copper strip material grade determines its physical and chemical properties, and different material grades may require different process parameters during rolling. Target thickness and target width are the copper strip dimensions specified in the production plan and are important parameters for controlling the rolling process. The rolling speed setpoint and rolling force setpoint directly affect the rolling quality and production efficiency of the copper strip.

[0114] Step S322: Read the process parameter annotation information of all samples from the normal state reference spectrum library, and construct a process parameter sample matrix, where each row corresponds to a reference spectrum sample and each column corresponds to a process parameter.

[0115] The process parameter sample matrix is ​​a matrix obtained by organizing the process parameter annotation information of all samples in the normal state benchmark spectrum library. Each row corresponds to a benchmark spectrum sample, and each column corresponds to a process parameter.

[0116] Step S323: Normalize the current set of process parameters and the sample matrix of process parameters to map each parameter value to the same dimension range, so as to eliminate the influence of differences in the magnitude of different parameters.

[0117] Normalization is performed to eliminate the impact of differences in the magnitude of different parameters on the calculation of process parameter matching degree. Different process parameters may have different dimensions and value ranges. Normalization methods such as minimum-maximum normalization and Z-score normalization can be used to compare different process parameters within the same dimension range, thereby improving the accuracy of process parameter matching degree calculation.

[0118] Step S324: Calculate the similarity between the current process parameter vector and each sample vector in the process parameter sample matrix, and generate a similarity score set.

[0119] The current process parameter vector is a vector composed of the set of key process parameters for the current rolling condition. Each sample vector in the process parameter sample matrix corresponds to the process parameters of a baseline spectrum sample.

[0120] There are several methods for calculating similarity, such as Euclidean distance and cosine similarity. Euclidean distance calculates the distance between two vectors; the smaller the distance, the higher the similarity. Cosine similarity calculates the cosine of the angle between two vectors; the closer the cosine value is to 1, the higher the similarity.

[0121] Step S325: Perform weighted optimization on the similarity score set, and assign weights according to the degree of influence of process parameters on the rolling state, wherein the weights of rolling force and rolling speed are higher than the weights of width and thickness.

[0122] Weighted optimization of the similarity score set is performed to more accurately reflect the influence of different process parameters on the rolling state. Different process parameters have varying degrees of influence on the correlation map between the rolling process and the rolling state; therefore, different weights need to be assigned to each process parameter.

[0123] Rolling force and rolling speed have a significant impact on the rolling process, directly affecting the thickness, shape, and quality of the copper strip. Therefore, the weights of rolling force and rolling speed are set higher than those of width and thickness. Through weighted optimization, the similarity score can more accurately reflect the degree of similarity between the current rolling condition and the benchmark sample.

[0124] Step S326: Select the benchmark map sample with the highest similarity score after weighted optimization as the reference benchmark map. When there are multiple samples with similar scores, further filter by comparing the copper strip surface quality level of the samples and select the sample with higher surface quality level as the reference.

[0125] The purpose of selecting the benchmark map sample with the highest weighted similarity score as the reference benchmark map is to find the normal state reference standard that is most similar to the current rolling conditions.

[0126] When multiple samples with similar scores exist, further screening is performed by comparing the surface quality grades of the copper strip corresponding to the samples. The surface quality grade of the copper strip is an important indicator for measuring the quality of the copper strip, and samples with higher surface quality grades are more representative of the ideal rolling state.

[0127] Step S330: Compare the nodes and edges of the rolling state association map with the reference map one by one, and calculate the node degree deviation vector, edge weight deviation matrix and map structure difference degree. Each element of the node degree deviation vector is the difference between the current node degree and the reference node degree, the elements of the edge weight deviation matrix are the relative deviations between the current edge weight and the reference edge weight, and the map structure difference degree is calculated by the graph editing distance algorithm.

[0128] The rolling state correlation graph is compared node-by-node and edge-by-edge with the reference graph to identify differences between the two. Node degree refers to the number of edges connecting a node in the graph, reflecting the importance of that node. Edge weight refers to the weight value of an edge in the graph, representing the strength of the correlation between nodes.

[0129] Each element of the node degree deviation vector is the difference between the current node degree and the reference node degree. The elements of the edge weight deviation matrix are the relative deviations between the current edge weight and the reference edge weight. The relative deviation can be calculated using the following formula: Relative deviation = (Current edge weight - Reference edge weight) / Reference edge weight.

[0130] Graph structural dissimilarity is calculated using the graph edit distance algorithm. Graph edit distance refers to the minimum number of editing operations (insertion, deletion, replacement of nodes or edges) required to transform one graph into another. By calculating graph structural dissimilarity, the degree of structural difference between two graphs can be measured.

[0131] By calculating the node degree deviation vector, edge weight deviation matrix, and graph structure difference degree, the differences between the rolling state correlation graph and the reference graph can be comprehensively described.

[0132] Step S340: Multi-dimensional fusion of node degree deviation vector, edge weight deviation matrix and graph structure difference degree, extract the dominant deviation direction through principal component analysis, and generate a comprehensive deviation index containing spatiotemporal characteristics. The time dimension of the comprehensive deviation index reflects the trend of deviation change with rolling time, and the spatial dimension reflects the distribution characteristics of deviation among different level signals.

[0133] The purpose of multi-dimensional fusion of node degree deviation vector, edge weight deviation matrix and graph structure difference is to integrate these different types of deviation information to obtain a more comprehensive and accurate deviation index.

[0134] In this step, principal component analysis (PCA) is used to extract the dominant deviation direction. First, the node degree deviation vector, edge weight deviation matrix, and graph structure difference are combined into a high-dimensional deviation eigenvector. Then, this high-dimensional deviation eigenvector is zero-mean processed to eliminate the mean differences in deviation values ​​across dimensions. Next, the covariance matrix of the centered deviation eigenvector is calculated; the elements of this matrix represent the degree of linear correlation between different deviation dimensions. Eigenvalue decomposition is performed on the covariance matrix to obtain eigenvalues ​​and eigenvectors. The eigenvalues ​​represent the variance contribution of the corresponding eigenvector, and the eigenvectors represent the dominant direction of deviation change. The top few eigenvectors whose cumulative variance contribution exceeds a preset proportion are selected as principal components to generate the principal component matrix. The centered deviation eigenvector is multiplied by the principal component matrix to obtain the dimensionality-reduced low-dimensional deviation vector. Each element of the low-dimensional deviation vector corresponds to the score of a principal component, reflecting the degree of deviation in that principal component direction. The low-dimensional deviation vector is then extended in terms of spatiotemporal dimension, associating it with the corresponding rolling timestamp and signal level information to generate a comprehensive deviation index containing both time series and spatial distribution characteristics. The time dimension of the comprehensive deviation index reflects the trend of deviation change with rolling time, while the spatial dimension reflects the distribution characteristics of deviation among different levels of signals.

[0135] As one implementation method, step S340 can be implemented as the following steps S341~S347: Step S341: Convert the node degree deviation vector into a column vector, expand the edge weight deviation matrix into a column vector in row priority order, and use the graph structure difference degree as a scalar value. The three are combined to obtain a high-dimensional deviation feature vector.

[0136] The node degree deviation vector represents the deviation of the degree of each node in the current rolling state correlation diagram from the reference baseline diagram. Converting it to a column vector is to standardize the data representation and facilitate subsequent matrix operations.

[0137] The edge weight deviation matrix records the relative deviation of edge weights between the current graph and the reference graph. Expanding it into a column vector in row-major order arranges the matrix elements into a one-dimensional vector. This expansion method fully preserves the edge weight deviation information and facilitates combination with other data. The graph structure dissimilarity is a scalar value calculated using the graph edit distance algorithm, reflecting the overall structural differences between the two graphs. Combining these three components forms a high-dimensional deviation feature vector, which integrates deviation information from node degree, edge weights, and graph structure.

[0138] Step S342: Perform zero-mean processing on the high-dimensional deviation feature vector to eliminate the mean difference of deviation values ​​in each dimension and generate a centered deviation feature vector.

[0139] The dimensions of a high-dimensional biased eigenvector may have different means, which can affect the performance of subsequent principal component analysis. Zero-mean processing aims to make the mean of each dimension zero. Specifically, this involves first calculating the mean of each dimension in the high-dimensional biased eigenvector, and then subtracting that mean from each data point in that dimension.

[0140] By applying this processing to all dimensions of the high-dimensional bias eigenvector, a centered bias eigenvector is obtained. This processing method ensures consistency in the mean across all dimensions, avoiding analytical errors caused by mean differences and enabling principal component analysis to extract the main information from the data more accurately.

[0141] Step S343: Calculate the covariance matrix of the centered bias eigenvectors. The elements of the covariance matrix represent the degree of linear correlation between different bias dimensions.

[0142] Covariance is a statistic that measures the linear relationship between two variables. For a centered bias eigenvector, its covariance matrix is ​​used to describe the linear association between its various dimensions.

[0143] By calculating the covariance matrix, we can understand the linear dependence between different bias dimensions, providing important information for subsequent eigenvalue decomposition.

[0144] Step S344: Perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and eigenvectors, where the eigenvalues ​​represent the variance contribution of the corresponding eigenvectors, and the eigenvectors represent the main direction of the deviation change.

[0145] Eigenvalues ​​reflect the variance of the data in the direction represented by the corresponding eigenvector, i.e., the degree of dispersion of the data in that direction. The larger the variance, the more information is contained in that direction, and the greater its contribution to the overall variation of the data. Eigenvectors represent the trend of data variation in that direction, i.e., the main direction of deviation change. Eigenvalue decomposition transforms the information of the covariance matrix into eigenvalues ​​and eigenvectors, facilitating subsequent principal component selection.

[0146] Step S345: Select the top few eigenvectors whose cumulative variance contribution exceeds a preset ratio as principal components, generate a principal component matrix, and the column vectors of the principal component matrix are the selected eigenvectors.

[0147] The preset proportion is usually determined based on actual needs. The variance contribution of each eigenvalue is calculated, and these variance contributions are then accumulated sequentially until the cumulative proportion exceeds the preset proportion. Assuming the condition is met when accumulating to the q-th eigenvalue, the first q eigenvectors are selected as principal components. Arranging these eigenvectors column-wise yields the principal component matrix, and the column vectors of this matrix are the selected principal components. In this way, the principal components that best represent the data changes can be selected from numerous eigenvectors, achieving data dimensionality reduction.

[0148] Step S346: Multiply the centered bias eigenvector with the principal component matrix to obtain the dimensionality-reduced low-dimensional bias vector. Each element of the low-dimensional bias vector corresponds to the score of a principal component, reflecting the degree of bias in the direction of that principal component.

[0149] Let the centered bias eigenvector be X and the principal component matrix be P, then the dimensionality-reduced low-dimensional bias vector is Y = XP.

[0150] Since the principal component matrix P is a p × q matrix (p is the original dimension, and q is the number of principal components), and the centered bias eigenvector X is an n × p matrix (n is the number of samples), the resulting low-dimensional bias vector Y is an n × q matrix. Each row of the low-dimensional bias vector corresponds to the score of a sample along the principal component direction, and each element reflects the degree of bias of that sample along the corresponding principal component direction. In this way, high-dimensional bias information is projected onto a few principal components, achieving effective dimensionality reduction of the data while preserving its main features.

[0151] Step S347: Extend the spatiotemporal dimension of the low-dimensional deviation vector, associate it with the corresponding rolling timestamp and signal level information, and generate a comprehensive deviation index that includes time series and spatial distribution characteristics.

[0152] While low-dimensional deviation vectors have reduced the dimensionality of deviation information, they lack temporal and spatial information. To more comprehensively describe the deviation situation, spatiotemporal dimensional expansion is needed.

[0153] The rolling timestamp records the rolling time for each sample, while the signal hierarchy information indicates the level to which the signal belongs (e.g., equipment body monitoring layer, rolling process monitoring layer, environmental correlation layer). The low-dimensional deviation vector is associated with the rolling timestamp and signal hierarchy information, assigning temporal and spatial attributes to each element of the low-dimensional deviation vector.

[0154] For example, low-dimensional deviation vectors can be arranged in chronological order to form a time series, while labeling each element with its corresponding signal level. The resulting comprehensive deviation index reflects both the trend of deviation change with rolling time (time dimension) and the distribution characteristics of deviation among different signal levels (spatial dimension).

[0155] Step S350: Construct a deviation evolution dynamic model based on the comprehensive deviation index, where the state variables of the model are the deviation values ​​of signals at each level.

[0156] The model parameters are determined by the physical laws of the rolling process, including the correlation between rolling force and copper strip deformation, and the mapping relationship between vibration signals and equipment wear. The deviation evolution dynamics model describes the change of the comprehensive deviation index over time. This model can predict the development trend of deviations, providing a basis for fault early warning.

[0157] The model's state variables are selected from the deviation values ​​of signals at each level. These deviation values ​​are an important component of the comprehensive deviation index, reflecting the degree of deviation of different level signals from the normal state. The model parameters are determined based on the physical laws of the rolling process. For example, there is a close correlation between rolling force and copper strip deformation. Based on the theories of materials mechanics and plastic deformation, a mathematical model between the two can be established. During the rolling process, the magnitude of the rolling force directly affects the thickness and shape changes of the copper strip. Through experimental data and theoretical analysis, the specific parameters of this correlation can be determined.

[0158] There is a certain mapping relationship between vibration signals and equipment wear. During normal operation, the vibration signal has a specific characteristic frequency and amplitude range. When wear occurs, the vibration signal changes. By analyzing a large amount of experimental data and applying machine learning algorithms, a mapping model between vibration signal characteristics and the degree of equipment wear can be established, thereby determining the model parameters. Based on these physical laws, the model parameters can enable the deviation evolution dynamics model to more accurately reflect the actual rolling process and improve the accuracy of fault prediction.

[0159] Step S360: Solve the deviation evolution dynamic model to obtain the trajectory curve of the signal deviation value of each level changing with rolling time. Sample the trajectory curve at a preset time interval to generate a set of deviation evolution sequences containing descriptions of abnormal features of each rolling period.

[0160] Numerical methods, such as the Euler method and the Runge-Kutta method, can be used to solve the dynamic model of deviation evolution. These methods can calculate the deviation values ​​of each level signal at different time points based on the initial conditions and parameters of the model, thereby obtaining the trajectory curve of the deviation value changing with rolling time.

[0161] The trajectory curve visually illustrates the dynamic changes in signal deviation values ​​at each level. For ease of analysis and processing, the trajectory curve needs to be sampled at preset time intervals. These preset time intervals can be determined based on actual needs and data processing capabilities; for example, they can be set to every minute, every five minutes, etc. After sampling, a series of discrete deviation value data points are obtained, each corresponding to a rolling period.

[0162] By analyzing the deviation value at each sampling point and combining the temporal and spatial dimensions of the comprehensive deviation index, an anomaly description for that rolling period can be obtained. For example, if the deviation value of a certain layer signal exceeds a preset threshold within a certain period and shows a continuous upward trend over time, then an anomaly can be described as existing in that period, characterized by excessive and continuously deteriorating deviation of the signal at that layer. Arranging the anomaly descriptions for each rolling period in chronological order generates a set of deviation evolution sequences containing the anomaly descriptions for each rolling period.

[0163] Step S400: Model the risk transmission path of the deviation evolution sequence set to generate a risk transmission network model that includes potential failure propagation paths and impact intensity.

[0164] The set of deviation evolution sequences records the abnormal characteristics of each rolling stage. By analyzing these sequences, possible propagation paths and impact ranges of faults can be discovered. Risk transmission path modeling aims to extract potential fault propagation patterns from these sequences and construct a network model that can describe the fault propagation process and its impact intensity.

[0165] This process requires considering the correlation between signals at different levels and the evolution trend of anomalous features. For example, if an anomaly occurs in a signal at a certain level, and anomalies also occur in signals at adjacent levels in subsequent time periods, it can be inferred that a fault propagation path exists from that level to adjacent levels. Simultaneously, it is also necessary to analyze the severity of each anomalous feature and its impact on other signals to determine the intensity of the fault propagation.

[0166] By modeling risk transmission paths, we can predict in advance the direction of fault propagation and the potential scope of its impact, providing a basis for taking effective fault prevention and control measures.

[0167] As one implementation method, step S400 can be implemented as the following steps S410~S470: Step S410: Construct a graph structure learning framework for a physical field linkage sensing network. The graph structure learning framework includes a feature input layer, a graph topology construction layer, a path mining layer, and an intensity evaluation layer. Each layer transmits feature information through a fully connected method.

[0168] The feature input layer is the framework's entry point, responsible for receiving the set of deviation evolution sequences as input data. This layer preprocesses the input data, performing operations such as normalization and feature extraction to improve data quality and usability. The graph topology construction layer constructs the graph's topology based on the input feature information. In this layer, anomalous features in the deviation evolution sequences are used as nodes in the graph, and the connections between nodes are determined based on the correlations between signals at different levels, thus forming a graph structure.

[0169] The path mining layer, based on the graph structure, uncovers potential fault propagation paths. These algorithms can identify possible fault propagation paths by combining graph traversal and search operations with the temporal sequence and correlation of anomaly features. The intensity assessment layer then evaluates the impact intensity of the discovered fault propagation paths. It considers factors such as the severity of node anomalies and the strength of edge correlations to calculate the impact intensity of each path.

[0170] Feature information is transmitted between layers through full connectivity, ensuring that each layer can make full use of the output of the previous layer, thereby enabling comprehensive analysis and processing of the set of deviation evolution sequences.

[0171] Step S420: Input the set of deviation evolution sequences into the feature input layer, divide the continuous evolution sequence into multiple time window segments through time series segmentation processing, each segment corresponds to a rolling process sub-stage, and generate a time window feature matrix. The row vectors of the matrix correspond to different time windows, and the column vectors correspond to the dimensions of the comprehensive deviation index.

[0172] Rolling processes can typically be divided into multiple sub-stages, such as the feeding stage, roughing stage, finishing stage, and unloading stage. Each sub-stage has different process characteristics and signal features. By dividing the deviation evolution sequence into time windows, each time window segment corresponds to a sub-stage of the rolling process.

[0173] The length of the time window can be determined based on the actual rolling process and data characteristics. The generated time window feature matrix is ​​a two-dimensional matrix, where the row vectors correspond to different time windows and the column vectors correspond to the dimensions of the comprehensive deviation index. Each row records the value of the comprehensive deviation index in each dimension within a time window. This matrix allows for a direct observation of the changes in deviation characteristics across different time windows and dimensions.

[0174] Step S430: Initialize the graph structure of the feature matrix of each time window through the graph topology construction layer, take the dimension of the comprehensive deviation index that exceeds the preset threshold as the graph node, construct the initial connection between the nodes based on the topological relationship of the rolling state association graph, and generate the time window association graph.

[0175] The main task of the graph topology construction layer is to convert the time window feature matrix into a graph structure. Dimensions with a comprehensive deviation index exceeding a preset threshold indicate that the dimension has abnormal characteristics, and these dimensions are used as nodes in the graph.

[0176] The preset threshold can be determined based on historical data and experience to judge whether the deviation in a certain dimension has reached an abnormal level. Initial connections between nodes are constructed based on the topological relationships of the rolling state correlation graph. The rolling state correlation graph records the correlations between signals at each level. Based on this graph, it can be determined which nodes have potential connections. For example, if the signal from a certain equipment body monitoring layer is correlated with the signal from the rolling process monitoring layer in the rolling state correlation graph, then during graph structure initialization, if the dimensions corresponding to these two signals both exceed the preset threshold, a connection edge can be established between their corresponding nodes. The generated time window correlation graph is a local graph structure that reflects the correlations between abnormal features within each time window.

[0177] As one implementation method, step S430 can be implemented as the following steps S431~S435: Step S431: Traverse each element of the time window feature matrix. When the value of a certain dimension of the comprehensive deviation index exceeds the preset threshold, mark the physical field signal node corresponding to that dimension as an abnormal node and record the signal type, level and time window number of the node.

[0178] Traversing the feature matrix within the time window is to comprehensively examine the overall deviation index value of each dimension. A preset threshold is the standard for judging anomalies; when the value of a certain dimension exceeds this threshold, it indicates that the physical field signal corresponding to that dimension has an anomaly.

[0179] The physical field signal nodes corresponding to this dimension are marked as anomalous nodes, and their signal type, level, and time window number are recorded. The signal type clarifies the nature of the physical field signal represented by the node, such as vibration signal, temperature signal, etc. The level indicates the node's hierarchical position in the rolling state correlation diagram, such as the equipment body monitoring level, rolling process monitoring level, etc. The time window number records the time information of the anomalous node's occurrence, facilitating subsequent time series analysis.

[0180] Step S432: Extract the neighboring node information of the abnormal node from the rolling state association graph, including the neighboring nodes at the same level and the upstream and downstream neighboring nodes across levels, and construct the potential connection candidate set of the abnormal node.

[0181] The rolling state correlation map describes the correlation between signals at each level. Based on this map, the adjacent nodes of abnormal nodes can be found.

[0182] Adjacent nodes include intra-level adjacent nodes and cross-level upstream / downstream adjacent nodes. Intra-level adjacent nodes are nodes at the same level as the abnormal node and are related to it. For example, in the equipment body monitoring layer, the intra-level adjacent nodes of a vibration signal node might be vibration signal nodes or temperature signal nodes in other parts of the equipment. Cross-level upstream / downstream adjacent nodes involve relationships between different levels. For instance, an abnormal node in the equipment body monitoring layer might affect certain nodes in the rolling process monitoring layer; these affected nodes are its cross-level downstream adjacent nodes. This adjacent node information is extracted to construct a potential connection candidate set for the abnormal node. This candidate set contains information on all nodes that might establish connections with the abnormal node.

[0183] Step S433: Perform physical correlation screening on the potential connection candidate set and retain candidate nodes that have a direct physical field coupling relationship with the abnormal nodes.

[0184] For example, candidate nodes include vibration signal nodes whose adjacent nodes include temperature signal nodes from the same equipment, and rolling force signal nodes whose adjacent nodes include plate shape signal nodes. Physical correlation screening ensures that the connections between nodes have actual physical meaning. In the potential connection candidate set, some nodes may have indirect or no actual physical connection with anomalous nodes, and these nodes need to be eliminated.

[0185] Candidate nodes with direct physical field coupling to anomalous nodes are retained. For example, during rolling, adjacent nodes of vibration signal nodes include temperature signal nodes from the same equipment, because equipment vibration may cause frictional heat, thus affecting the temperature signal. Adjacent nodes of rolling force signal nodes include strip shape signal nodes, because the magnitude of the rolling force directly affects the strip shape of the copper strip. This screening process makes the graph structure more reasonable and reflects the actual physical relationships.

[0186] Step S434: Construct a local connection graph of abnormal nodes based on the selected candidate nodes, where nodes are abnormal nodes and selected candidate nodes, edges represent the physical field coupling relationship between nodes, and the direction of the edges is determined according to the signal propagation direction.

[0187] A local connectivity graph of anomalous nodes is constructed based on the selected candidate nodes. The graph includes both anomalous nodes and selected candidate nodes, all of which have direct physical field coupling relationships with the anomalous nodes. Edges represent the physical field coupling relationships between nodes, and the direction of the edges is determined by the signal propagation direction. For example, if a vibration signal affects a temperature signal, then there is a directed edge from the vibration signal node to the temperature signal node, pointing from the vibration signal node to the temperature signal node.

[0188] Step S435: Perform weight assignment on the local connection graph. The edge weight value is calculated by weighting the edge weight value of the corresponding node pair in the rolling state association graph with the comprehensive deviation index of the current time window to generate the initial graph structure for each time window.

[0189] Edge weights reflect the strength of the association between nodes. The edge weights are calculated by weighting the edge weights of the corresponding node pairs in the rolling state association graph with the comprehensive deviation index of the current time window.

[0190] The edge weights in the rolling state correlation graph record the basic correlation strength between nodes, while the comprehensive deviation index for the current time window reflects the severity of node anomalies within that time window. Weighting both yields edge weights that better reflect the current situation. For example, the edge weights in the rolling state correlation graph can be multiplied by a coefficient related to the comprehensive deviation index to obtain new edge weights.

[0191] After applying these weights to the local connectivity graph, an initial graph structure for each time window is generated. This graph structure considers both the basic relationships between nodes and the anomalies of the current time window.

[0192] Step S440: Perform time-series correlation analysis on all time window association graphs through the path mining layer, and identify node connection paths across time windows through dynamic programming algorithm. The node connection paths are formed by connecting abnormal nodes in different time windows in chronological order, which conforms to the adjacency relationship of nodes in the physical field propagation path between the preceding and following windows.

[0193] The main task of the path mining layer is to mine node connection paths across time windows from all time window association graphs. These paths represent potential fault propagation paths.

[0194] Temporal correlation analysis is a correlation analysis method that considers the time factor, focusing on the correlation between anomalous nodes in different time windows. The node connection path is formed by connecting anomalous nodes from different time windows in chronological order, and must conform to the adjacency relationship of nodes in preceding and following windows along the physical field propagation path. For example, if a vibration signal node in a certain equipment body monitoring layer shows an anomaly in the first time window, and a rolling force signal node in the rolling process monitoring layer, which has a physical field propagation path adjacent to that vibration signal node, also shows an anomaly in the second time window, then these two nodes can be connected in chronological order to form a potential fault propagation path.

[0195] As one implementation, step S440 can be implemented as the following steps S441~S446: Step S441: Arrange all initial graph structures in the order of time windows to construct a time series graph sequence, wherein each time window corresponds to a graph structure in the sequence.

[0196] The initial graph structures of all time windows are arranged chronologically to form a time-series graph sequence. This sequence records the graph structure information of anomalous nodes within different time windows, providing a foundation for subsequent time-series correlation analysis. Each time window corresponds to a graph structure in the sequence, and the nodes and edges in the graph structure reflect the anomalous nodes and their relationships within that time window. By constructing the time-series graph sequence, the evolution of anomalous features over time can be clearly observed, facilitating the search for node connection paths across time windows.

[0197] Step S442: Define path state variables, which include information such as abnormal nodes in the current time window, cumulative path deviation value, and path length. The initial state is set to all abnormal nodes in the first time window, the cumulative deviation value is the deviation value of the node, and the path length is 1.

[0198] Path state variables are used to record relevant information during the path mining process. The abnormal node in the current time window represents the node where the path is currently located. The cumulative deviation value of the path is the sum of the deviation values ​​of all nodes on the path, reflecting the severity of the path's anomaly. The path length represents the number of nodes the path passes through.

[0199] The initial state is set to all anomalous nodes in the first time window, with the cumulative deviation value of each anomalous node being its deviation value, and the path length being 1. This is the starting point for path mining; from these initial states, possible path extensions are gradually sought.

[0200] Step S443: Iterate through the time series graph sequence. Starting from the second time window, for each time window's abnormal node, find the abnormal node in the previous time window that has an adjacent relationship with it, and construct possible path extension relationships.

[0201] The time series graph sequence is iteratively processed, starting from the second time window, processing the anomalous nodes of each time window sequentially. For each anomalous node in a time window, an anomalous node with an adjacent relationship to it in the previous time window is found.

[0202] Adjacency relationships can be determined based on the edges in the time window association graph. If an abnormal node in the previous time window is adjacent to the current abnormal node, then a possible path extension relationship can be constructed.

[0203] Step S444: Calculate the cumulative deviation value and path continuity index for each extended path, where the cumulative deviation value is the weighted sum of the current node deviation value and the cumulative deviation value of the preceding path, and the path continuity index is evaluated by the node type matching degree and the consistency of the physical field propagation path.

[0204] Calculating the cumulative deviation value and path continuity index for each extended path is to evaluate the quality of the path. The cumulative deviation value is a weighted sum of the deviation value of the current node and the cumulative deviation value of the preceding path. The weighting of the sum can be determined based on the importance of different nodes and the degree of influence of the deviation.

[0205] Path continuity is assessed through node type matching and consistency with the physical field propagation path. Node type matching refers to whether the signal types of consecutive nodes have a reasonable correlation. For example, the correlation between vibration signal nodes and temperature signal nodes is reasonable, while the correlation with some unrelated signal nodes may be unreasonable. Physical field propagation path consistency refers to whether the connection between consecutive nodes conforms to the propagation laws of the physical field.

[0206] Step S445: Perform pruning optimization on all possible extension paths, retain the path with the highest cumulative deviation value exceeding the preset threshold and the highest path continuity index, and remove redundant paths with low cumulative deviation value or poor continuity.

[0207] Pruning optimization aims to reduce the number of paths and improve the efficiency and accuracy of path discovery. All possible extension paths are evaluated, and the path with the highest cumulative deviation value exceeding a preset threshold and the highest path continuity index is retained.

[0208] The preset threshold can be determined based on the actual situation to judge whether the severity of an anomaly in a certain path has reached a level that requires attention. Redundant paths with low cumulative deviation values ​​or poor continuity are removed, as these paths may be caused by noise or random factors and have no practical significance for fault propagation analysis. Through pruning optimization, the path mining results can be made more concise and effective.

[0209] Step S446: After processing all time windows, output paths whose length exceeds the preset minimum length as potential fault propagation paths. The paths include a sequence of abnormal nodes arranged in chronological order and the connection relationships between the nodes.

[0210] After processing all time windows, the resulting paths are further filtered. Paths exceeding a preset minimum length are identified as potential fault propagation paths. The preset minimum length can be determined based on actual needs to exclude shorter, potentially unrepresentative paths. Potential fault propagation paths include a chronological sequence of anomalous nodes and the connections between them. These paths reflect the possible propagation process of the fault. By analyzing these paths, the development trend of the fault can be predicted in advance, allowing for corresponding prevention and control measures.

[0211] Step S450: Verify the validity of the identified cross-time window paths. By comparing the deviation evolution trend of nodes on the path with the physical field propagation law, eliminate false paths that do not conform to the propagation law and retain potential fault propagation paths.

[0212] Verifying the validity of identified cross-time-window paths ensures their authenticity and reliability. This involves comparing the deviation evolution trends of nodes along the path with the laws governing the propagation of physical fields. These laws describe the propagation mode and trends of signals within a physical field. For example, during rolling, the vibration signal from the equipment may be transmitted to the rolling force signal through the mill stand, and this transmission should exhibit a certain time delay and amplitude attenuation. If the deviation evolution trend of nodes along a cross-time-window path does not conform to these laws of physical field propagation, then the path may be a false path and needs to be discarded.

[0213] This verification process preserves the true potential failure propagation paths, improving the accuracy of the risk transmission network model.

[0214] Step S460: Calculate the cumulative impact strength of each potential fault propagation path. The cumulative impact strength is calculated by considering each path.

[0215] Specifically, it can be implemented as follows: Step S461~S466: Step S461: Extract the comprehensive deviation index value of each abnormal node on the potential fault propagation path, and the propagation coefficient of the connecting edge between nodes. The propagation coefficient is obtained from the edge weight value of the rolling state association graph, which reflects the efficiency of signal transmission between nodes.

[0216] Each abnormal node on the potential fault propagation path has a corresponding comprehensive deviation index value, which reflects the severity of the abnormality at that node. The propagation coefficient of the connecting edges between nodes is obtained from the edge weight values ​​of the rolling state correlation graph. The edge weight values ​​have already taken into account factors such as mutual dependence and propagation attenuation coefficient in previous steps, and can well reflect the efficiency of signal transmission between nodes.

[0217] Step S462: Perform time decay processing on the comprehensive deviation index value of each abnormal node. Adjust the deviation value according to the interval between the time window where the node is located and the current time window. The larger the interval, the larger the decay coefficient. Simulate the natural decay of the deviation effect over time.

[0218] As time progresses, the impact of deviations from abnormal nodes gradually decreases. Therefore, it is necessary to perform time decay processing on the comprehensive deviation index value of each abnormal node. The deviation value is adjusted based on the interval between the node's time window and the current time window.

[0219] A decay coefficient function can be set, which increases with the time interval. For example, suppose the interval between the time window of the node and the current time window is t, and the decay coefficient is f(t), where f(t) is a monotonically increasing function. Then the deviation value after time decay processing is the original deviation value multiplied by the decay coefficient f(t). This can simulate the natural decay process of the deviation's influence over time, making the calculation of the cumulative influence intensity more consistent with reality.

[0220] Step S463: Calculate the product of the attenuated node deviation value and the corresponding propagation coefficient to obtain the node influence contribution value, where the larger the propagation coefficient, the larger the node influence contribution value.

[0221] The node impact contribution value reflects the degree of influence of each abnormal node on the entire potential fault propagation path. The node impact contribution value is obtained by calculating the product of the attenuated node deviation value and the corresponding propagation coefficient.

[0222] If a node has a large attenuation deviation and a large propagation coefficient with its neighboring nodes, then the node's influence contribution will be large.

[0223] Step S464: Calculate the cumulative impact contribution of nodes according to their time sequence on the path to generate the preliminary cumulative impact intensity of the path.

[0224] By summing the impact contribution values ​​of each anomalous node on the path in chronological order, we obtain the preliminary cumulative impact strength of the path. This preliminary cumulative impact strength reflects the combined impact of all anomalous nodes on the path.

[0225] Step S465: Perform path length correction processing on the initial cumulative influence intensity. The longer the path length, the smaller the correction coefficient, to avoid the intensity being artificially high due to excessively long paths.

[0226] Path length has a certain impact on the cumulative impact strength. An excessively long path may lead to an inflated cumulative impact strength due to the inclusion of too many nodes. Therefore, path length correction is required for the initial cumulative impact strength.

[0227] A path length correction factor can be set, which decreases as the path length increases. For example, let the path length be L, the correction factor be g(L), and g(L) be a monotonically decreasing function. Then the corrected cumulative impact intensity is the initial cumulative impact intensity multiplied by the correction factor g(L). Through this correction, the cumulative impact intensity can more reasonably reflect the actual impact of the path.

[0228] Step S466: Combining the number of time windows in which the path appears with the process stage of copper strip rolling, make a final adjustment to the corrected cumulative influence intensity to generate a cumulative influence intensity that reflects the actual degree of influence of the path.

[0229] The number of time windows in which a path appears and the process stage of copper strip rolling also affect the actual impact of the path. If a path appears in multiple time windows, it indicates that the path has high stability and reliability, and its impact may be greater.

[0230] Different stages of the copper strip rolling process have varying sensitivities to faults. For example, in the finishing rolling stage, a small anomaly can have a significant impact on the quality of the copper strip. Therefore, the cumulative impact intensity is ultimately adjusted by considering the number of time windows in which the fault occurs and the specific stages of the copper strip rolling process. Adjustment coefficients corresponding to different time windows and process stages can be determined based on historical data and experience. Multiplying the adjusted cumulative impact intensity by these adjustment coefficients yields the cumulative impact intensity that reflects the actual degree of the fault's influence.

[0231] Step S470: Input the potential fault propagation path and the corresponding cumulative impact intensity into the intensity assessment layer. Adjust the assessment results of the impact intensity by introducing the path occurrence frequency and duration parameters to generate a risk transmission network model that includes path topology, propagation direction, cumulative impact intensity and time sequence characteristics.

[0232] The main task of the intensity assessment layer is to further evaluate and adjust the cumulative impact intensity of potential fault propagation paths. This involves introducing parameters related to path occurrence frequency and duration.

[0233] Path occurrence frequency refers to the number of times a path appears throughout the rolling process. A higher frequency indicates greater stability and reliability of the path, and potentially a stronger impact. Duration refers to the length of time the path persists; a longer duration indicates a more lasting impact. These two parameters are used to adjust the cumulative impact intensity.

[0234] Step S500: Output a fault warning instruction for copper strip rolling production based on the risk transmission network model. The fault warning instruction includes a fault type identifier and risk diffusion control parameters.

[0235] The risk transmission network model provides information on the propagation path and impact intensity of potential faults, which can be used to generate fault warning instructions. Fault type identifiers uniquely identify the nature and location of a fault and consist of an equipment category code, a signal field type code, and a fault mode code. The equipment category code indicates the type of equipment where the fault occurred, such as a rolling mill or tension roll; the signal field type code indicates the type of signal associated with the fault, such as vibration or temperature signals; and the fault mode code indicates the specific mode of the fault, such as wear or loosening. Risk diffusion control parameters are used to control the spread of faults, including information on the spatial range and temporal speed of risk diffusion. These parameters allow for the development of corresponding control measures, such as adjusting rolling speed or replacing equipment components, to reduce the impact of faults on production.

[0236] As one implementation method, step S500 can be implemented as the following steps S510~S570: Step S510: Analyze all potential fault propagation paths and their corresponding cumulative impact intensity in the risk transmission network model, sort them from high to low according to the cumulative impact intensity, and select the top-ranked paths as key risk paths.

[0237] Analyze the risk transmission network model to extract all potential failure propagation paths and their corresponding cumulative impact strengths. Sort these paths from highest to lowest cumulative impact strength. Higher cumulative impact strength indicates a greater impact of the failure along that path on production. Select the top-ranked paths as critical risk paths. The number of paths selected can be determined based on actual circumstances; for example, the top 5 paths with the highest cumulative impact strength can be chosen. These critical risk paths are the most critical failure propagation paths requiring the most attention; focusing on their analysis and handling can effectively reduce production risks.

[0238] Step S520: Perform path source analysis for each critical risk path. By traversing the abnormal node sequence on the path in reverse, determine the abnormal node at the beginning of the path as the initial fault source node, and record the signal type, the device to which the initial fault source node belongs, and the time window in which it is located.

[0239] Path source analysis aims to identify the root cause of a failure. Each critical risk path is traversed in reverse, starting from the last anomalous node and proceeding backward until the anomalous node at the beginning of the path is found, which is then used as the initial source node of the failure.

[0240] Record the signal type, associated device, and time window of the initial fault source node. The signal type clarifies the manifestation of the fault, the associated device determines the location of the fault, and the time window records the time when the fault occurred.

[0241] Step S530: Based on the signal type of the initial fault source node and the preset fault type mapping rule of the device, determine the corresponding fault type identifier. The fault type identifier consists of the device category code, signal field type code and fault mode code, which are used to uniquely identify the nature and location of the fault.

[0242] The preset fault type mapping rules are based on a large amount of historical data and experience, and record the fault types corresponding to different signal types and their respective devices.

[0243] Based on the mapping rule between the signal type of the initial fault source node and its associated equipment, the corresponding fault type identifier is determined. For example, if the signal type of the initial fault source node is a vibration signal and its associated equipment is a rolling mill bearing, according to the fault type mapping rule, the corresponding equipment category code might be 01 (representing rolling mill equipment), the signal field type code might be 02 (representing vibration signal), and the fault mode code might be 03 (representing bearing wear). Therefore, the fault type identifier would be 010203. This identifier uniquely identifies the nature and location of the fault, facilitating subsequent fault handling and management.

[0244] Step S540: Analyze the hierarchical and temporal distribution characteristics of abnormal nodes on the critical risk path, and calculate the spatial range and temporal speed of risk spread. The spatial range is evaluated by the number of levels and nodes involved in the path, and the temporal speed is calculated by the reciprocal of the time interval between the occurrence of adjacent abnormal nodes.

[0245] Analyze the hierarchical and temporal distribution characteristics of abnormal nodes on critical risk paths. The hierarchical distribution reflects the spread of fault signals across different levels, while the temporal distribution records the development process of the fault over time.

[0246] Calculate the spatial extent and temporal velocity of risk diffusion. Spatial extent is assessed by the number of levels and nodes involved in the path. If a critical risk path involves signal nodes across multiple levels and has a large number of nodes, it indicates a large spatial extent of risk diffusion. Temporal velocity is calculated by counting the reciprocal of the time interval between the occurrence of adjacent anomalous nodes.

[0247] Step S550: Assess the risk level comprehensively based on the cumulative impact intensity, the spatial range of risk diffusion, and the temporal speed. The higher the cumulative impact intensity, the larger the spatial range, and the faster the temporal speed, the higher the risk level.

[0248] The risk level is assessed by comprehensively considering the cumulative impact intensity, the spatial range of risk diffusion, and the temporal velocity. The cumulative impact intensity reflects the severity of the failure, the spatial range of risk diffusion indicates the scope of the failure's impact, and the temporal velocity reflects how quickly the failure spreads.

[0249] A risk assessment function can be set up, taking the cumulative impact intensity, spatial range of risk diffusion, and temporal velocity as input parameters. The function calculates the risk level. For example, the higher the cumulative impact intensity, the larger the spatial range, and the faster the temporal velocity, the higher the output value of the risk assessment function, and the higher the corresponding risk level. Risk levels can be divided into multiple levels, such as low risk, medium risk, and high risk, and different risk control measures can be taken according to different risk levels.

[0250] Step S560: Integrate the fault type identifier, initial fault source node location information, risk level and risk propagation parameters to generate a fault warning instruction containing a timestamp. The instruction includes information such as fault location code, risk level code, and recommended handling measure code.

[0251] The fault type identifier, initial fault source node location information, risk level, and risk propagation parameters are fused to generate a fault warning command. A timestamp is added to the command to accurately record the time of the fault occurrence.

[0252] The fault warning instruction includes information such as fault location code, risk level code, and recommended handling measure code. The fault location code is generated based on the fault type identifier and the initial fault source node location information, used to accurately indicate the location and nature of the fault. The risk level code is generated based on the risk level assessment results, facilitating rapid identification of the fault's severity. The recommended handling measure code is a code for the corresponding handling measures formulated according to different fault types and risk levels. For example, for a high-risk bearing wear fault, the recommended handling measure code might indicate that the bearing needs to be replaced immediately.

[0253] Step S570: Standardize the format of the fault warning command to make it conform to the communication protocol requirements of the copper strip rolling control system, including the command header, data field, check code and other structures, to ensure that the control system can correctly parse and execute it.

[0254] Standardizing the format of fault warning commands ensures their proper transmission and processing within the copper strip rolling control system. Based on the communication protocol requirements of the copper strip rolling control system, the fault warning commands are designed with a format including a command header, data field, and checksum. The command header identifies the command type and starting position; the data field contains the specific content of the fault warning command, such as fault location code, risk level code, and suggested handling measure code; and the checksum verifies the completeness and accuracy of the command. This format standardization ensures that the control system can correctly parse and execute fault warning commands, taking timely and appropriate measures to address faults and guaranteeing the safety and stability of copper strip rolling production.

[0255] It is understood that the various algorithms involved in the above descriptions of the embodiments of the present invention can all be obtained from relevant content in the prior art. In order to save space, they will not be elaborated on in the embodiments of the present invention.

[0256] Furthermore, those skilled in the art can supplement the details when implementing the solution of this invention based on common knowledge in the field. For example, they can use normalization to eliminate dimensional conflicts before feature fusion, use interpolation to eliminate dimensional differences, combine historical data, experience or business scenario requirements to reasonably set the threshold, train the model based on a general model training method, set the number of layers in the model structure based on actual needs, select the activation function, etc. This invention will not provide redundant descriptions of overly detailed implementation processes here.

[0257] For example, in step S236, the interdependence and transmission attenuation coefficients can be normalized to eliminate dimensional differences and generate dimensionless same-level and cross-level correlation strength indices. The normalized same-level and cross-level correlation strength indices are then combined in hierarchical order to generate a multi-dimensional correlation strength index containing intra- and inter-level correlation characteristics. Similarly, in step S254, the temporal and spatial dynamic adjustment factors can be normalized to convert them into dimensionless adjustment coefficients. The normalized temporal and spatial dynamic adjustment coefficients are then weighted and fused to generate a spatiotemporal coupling adjustment coefficient. In step S340, the node degree deviation vector, edge weight deviation matrix, and graph structure difference can be normalized to eliminate dimensional differences. The normalized node degree deviation vector, edge weight deviation matrix, and graph structure difference are then multi-dimensionally fused, and principal component analysis is used to extract the dominant deviation direction, generating a comprehensive bias index containing spatiotemporal characteristics. The difference index; during step S341, the node degree deviation vector, edge weight deviation matrix, and graph structure difference are normalized to eliminate dimensional differences; the normalized node degree deviation vector is converted into a column vector, the normalized edge weight deviation matrix is ​​expanded into a column vector in row-major order, and the normalized graph structure difference is used as a scalar value. The combination of these three results in a high-dimensional deviation feature vector; during step S463, the attenuated node deviation value and the corresponding propagation coefficient can be normalized to eliminate dimensional differences; the product of the normalized node deviation value and the normalized propagation coefficient is calculated to obtain the node influence contribution value. Other similar issues will not be exemplified.

[0258] Based on the same inventive concept, this application also provides a fault warning device for implementing the fault warning method for copper strip rolling production as described above. The solution provided by this device is similar to the implementation described in the above method; therefore, the specific limitations in one or more fault warning device embodiments provided below can be found in the limitations of the fault warning method for copper strip rolling production described above, and will not be repeated here.

[0259] In one embodiment, such as Figure 3As shown, a fault early warning device 400 is provided, including: a signal acquisition module 410, used to collect a set of rolling state signals from multiple monitoring locations during the copper strip rolling process; a graph construction module 420, used to construct a multimodal correlation graph of the rolling state signal set, generating a rolling state correlation graph containing correlation parameters between signals, where nodes in the rolling state correlation graph represent signal types and edges represent the dynamic correlation strength between different types of signals; a sequence generation module 430, used to generate a deviation evolution sequence based on the rolling state correlation graph, obtaining a set of deviation evolution sequences characterizing the changes in abnormal features with rolling time, where each sequence element in the deviation evolution sequence set corresponds to an abnormal feature description of a rolling period; a path modeling module 440, used to model the risk transmission path of the deviation evolution sequence set, generating a risk transmission network model containing potential fault propagation paths and influence strength; and an instruction generation module 450, used to output a fault early warning instruction for copper strip rolling production based on the risk transmission network model, where the fault early warning instruction includes a fault type identifier and risk diffusion control parameters.

[0260] Each module in the aforementioned fault warning device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of the computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0261] In this embodiment of the invention, a fault early warning system is also provided. This fault early warning system can be a server, and its internal structure diagram can be as follows: Figure 4 As shown, the fault early warning system includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs in the non-volatile storage media. The database stores a set of rolling status signals. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a fault early warning method for copper strip rolling production.

[0262] Those skilled in the art will understand that Figure 4The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the fault warning system applied thereto. A specific fault warning system may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0263] In one embodiment, a fault warning system is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0264] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0265] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0266] It should be noted that the object information (including but not limited to the object's device information, corresponding personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the object or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0267] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A fault early warning method for copper strip rolling production, characterized in that, The method includes: collecting a set of rolling status signals from multiple monitoring locations during the copper strip rolling process; constructing a multimodal correlation graph from the set of rolling status signals to generate a rolling status correlation graph containing correlation parameters between signals, wherein nodes in the rolling status correlation graph represent signal types and edges represent the dynamic correlation strength between different types of signals; generating a deviation evolution sequence based on the rolling status correlation graph to obtain a set of deviation evolution sequences characterizing the changes in abnormal features with rolling time, wherein each sequence element in the deviation evolution sequence set corresponds to an abnormal feature description of a rolling period; modeling a risk transmission path from the set of deviation evolution sequences to generate a risk transmission network model containing potential fault propagation paths and influence strengths; and outputting a fault warning instruction for copper strip rolling production based on the risk transmission network model, wherein the fault warning instruction contains a fault type identifier and risk diffusion control parameters.

2. The method according to claim 1, characterized in that, The construction of a multimodal correlation map of the rolling state signal set, generating a rolling state correlation map containing inter-signal correlation parameters, includes: dividing the rolling state signal set into hierarchical subsets according to monitoring location and physical field type, obtaining multi-level signal subsets, wherein the hierarchical division of the multi-level signal subsets is based on the structural topology of the rolling equipment and the signal propagation path; performing cross-level feature correlation processing on the multi-level signal subsets, constructing causal correlation relationships between upper and lower level signals through inter-level signal transfer functions, generating a hierarchical correlation feature matrix, wherein the row vectors of the hierarchical correlation feature matrix correspond to upper-level signal features, the column vectors correspond to lower-level signal features, and the matrix elements represent inter-level signal transfer efficiency; based on the... The hierarchical correlation feature matrix calculates the mutual dependence of signal types within the same level and the transmission attenuation coefficient of signals across levels, generating a multi-dimensional correlation strength index. The mutual dependence is used to describe the cooperative change law of signals within the same level, and the transmission attenuation coefficient is used to describe the energy loss characteristics of signals propagating between levels. Based on the multi-dimensional correlation strength index, an initial topology of the dynamic correlation graph is constructed, where nodes correspond to signal types at each level, directed edges represent the signal transmission direction, and edge weights combine mutual dependence and transmission attenuation coefficients. The initial topology is then subjected to spatiotemporal dynamic optimization processing. By introducing rolling speed and copper strip thickness variation parameters, the temporal evolution characteristics of edge weights are adjusted to generate a rolling state correlation graph that is dynamically updated with rolling conditions.

3. The method according to claim 2, characterized in that, The process of dividing the rolling state signal set into multi-level signal subsets based on monitoring location and physical field type includes: determining the spatial relationship of key equipment and the distribution of signal monitoring points by analyzing the equipment structure layout of the copper strip rolling production line; constructing a three-level signal acquisition architecture based on the equipment structure layout, wherein the first level is the equipment body monitoring layer, the second level is the rolling process monitoring layer, and the third level is the environmental correlation layer; classifying and aggregating the rolling state signal set according to the three-level signal acquisition architecture, merging signals of the same level and type of physical field into signal subgroups to generate equipment body monitoring signal groups, rolling process monitoring signal groups, and environmental correlation signal groups; performing internal signal consistency verification processing on each signal group, by comparing the time series trends of signals of the same physical field within the same level, eliminating abnormal signal samples whose trend deviation exceeds the preset allowable range, and retaining the valid signal units within the signal group; arranging the verified equipment body monitoring signal group, rolling process monitoring signal group, and environmental correlation signal group in hierarchical order to generate multi-level signal subsets containing hierarchical identifiers and physical field attributes.

4. The method according to claim 3, characterized in that, The process of calculating the cross-dependency of signal types within the same level and the transmission attenuation coefficient of signals across levels based on the hierarchical correlation feature matrix to generate a multi-dimensional correlation strength index includes: performing row normalization on the hierarchical correlation feature matrix so that the sum of the elements in each row is 1, obtaining a probabilistic hierarchical correlation matrix, where the matrix elements represent the probability of the influence of upper-level signal features on lower-level signal features; within the same level, treating signal type as a random variable, calculating mutual information values ​​by analyzing the joint probability distribution of the signal time series, where the mutual information values ​​are described by the difference between joint entropy and marginal entropy, used to quantify the nonlinear dependence between signals within the same level. The relationship is as follows: the mutual information values ​​are normalized to obtain the interdependence of signals at the same level; between levels, an attenuation model is constructed based on the physical characteristics of the signal propagation path, and the attenuation ratio of the signal strength is calculated by simulating the energy change process of the signal propagating from the upper node to the lower node; the attenuation ratio is converted into a transmission attenuation coefficient, which is positively correlated with the attenuation ratio and is used to describe the degree of energy loss during the signal transmission process across levels; the interdependence of signals at the same level and the transmission attenuation coefficient between levels are combined in hierarchical order to generate a multi-dimensional correlation strength index that includes intra-level and inter-level correlation characteristics.

5. The method according to claim 4, characterized in that, The spatiotemporal dynamic optimization of the initial topology involves adjusting the temporal evolution characteristics of edge weights by introducing rolling speed and copper strip thickness variation parameters. This includes: collecting the copper strip thickness setpoint and actual rolling speed curve for the current rolling batch; dividing the rolling speed curve into multiple characteristic time periods including acceleration, constant speed, and deceleration segments; extracting the rolling speed change rate for each time period; constructing a time-dynamic adjustment factor based on the rolling speed change rate, where the adjustment factors for acceleration and deceleration segments are greater than those for constant speed segments to enhance the sensitivity of signal correlation during unsteady rolling; and constructing a spatial dynamic adjustment factor based on the copper strip thickness variation parameters. When the change in copper strip thickness exceeds a preset fluctuation range, the correlation weight adjustment amplitude between the corresponding rolling force signal and the strip shape signal is increased; the time dynamic adjustment factor and the spatial dynamic adjustment factor are coupled to generate a spatiotemporal coupling adjustment coefficient, which is obtained by multiplying the rolling speed change rate and the change in copper strip thickness; the edge weight values ​​in the initial topology are corrected using the spatiotemporal coupling adjustment coefficient to obtain corrected edge weight values; the corrected edge weight values ​​are smoothed to generate a continuous edge weight time series that changes with the rolling conditions, which is used to update the dynamic topology of the rolling state correlation graph.

6. The method according to claim 1, characterized in that, The process of generating deviation evolution sequences based on the rolling state correlation map to obtain a set of deviation evolution sequences representing the changes of abnormal features with rolling time includes: constructing a normal state benchmark map library of a physical field linkage sensing network, wherein the benchmark map library contains rolling state correlation map samples of copper strips with different rolling specifications and different material grades under normal production conditions, and each sample is labeled with corresponding process parameters such as rolling speed, copper strip thickness, and rolling force; selecting the benchmark map sample most similar to the current rolling condition from the normal state benchmark map library, and determining the reference benchmark map with the highest similarity through process parameter matching degree calculation; comparing the nodes and edges of the rolling state correlation map and the reference benchmark map one by one, and calculating the node degree deviation vector, edge weight deviation matrix, and graph. The graph structure difference is calculated by multi-dimensionally fusing the node degree deviation vector, edge weight deviation matrix, and graph structure difference. Principal component analysis is used to extract the dominant deviation direction, generating a comprehensive deviation index that includes spatiotemporal characteristics. The time dimension of the comprehensive deviation index reflects the trend of deviation change with rolling time, and the spatial dimension reflects the distribution characteristics of deviation among different levels of signals. A deviation evolution dynamic model is constructed based on the comprehensive deviation index. The state variables of the model are the deviation values ​​of each level of signal, and the model parameters are determined by the physical laws of the rolling process. The deviation evolution dynamic model is solved to obtain the trajectory curves of the deviation values ​​of each level of signal change with rolling time. The trajectory curves are sampled at preset time intervals to generate a set of deviation evolution sequences containing descriptions of abnormal characteristics of each rolling period.

7. The method according to claim 6, characterized in that, The step of selecting the benchmark map sample most similar to the current rolling condition from the normal state benchmark map library and determining the reference benchmark map with the highest similarity through process parameter matching degree calculation includes: extracting the key process parameter set of the current rolling condition; reading the process parameter annotation information of all samples from the normal state benchmark map library and constructing a process parameter sample matrix, where each row corresponds to a benchmark map sample and each column corresponds to a process parameter; normalizing the current process parameter set and the process parameter sample matrix to map each parameter value to the same dimension range to eliminate the influence of differences in the magnitude of different parameters; calculating the similarity between the current process parameter vector and each sample vector in the process parameter sample matrix to generate a similarity score set; performing weighted optimization on the similarity score set, allocating weights according to the degree of influence of process parameters on the rolling state; selecting the benchmark map sample with the highest weighted optimization similarity score as the reference benchmark map, and when there are multiple samples with similar scores, further filtering is performed by comparing the copper strip surface quality grades corresponding to the samples, and selecting the sample with the higher surface quality grade as the reference.

8. The method according to claim 7, characterized in that, The process of multi-dimensionally fusing the node degree deviation vector, edge weight deviation matrix, and graph structure difference, and extracting the dominant deviation direction through principal component analysis to generate a comprehensive deviation index containing spatiotemporal characteristics, includes: converting the node degree deviation vector into a column vector; expanding the edge weight deviation matrix into a column vector in row-major order; and using the graph structure difference as a scalar value, combining the three to obtain a high-dimensional deviation feature vector; performing zero-mean processing on the high-dimensional deviation feature vector to eliminate the mean differences of deviation values ​​in each dimension, generating a centered deviation feature vector; calculating the covariance matrix of the centered deviation feature vector, where the elements of the covariance matrix represent the degree of linear correlation between different deviation dimensions; and performing special processing on the covariance matrix. Eigenvalue decomposition yields eigenvalues ​​and eigenvectors, where eigenvalues ​​represent the variance contribution of the corresponding eigenvectors, and eigenvectors represent the principal direction of deviation change. The top few eigenvectors whose cumulative variance contribution exceeds a preset proportion are selected as principal components, generating a principal component matrix. The column vectors of the principal component matrix are the selected eigenvectors. The centered deviation eigenvectors are multiplied by the principal component matrix to obtain a dimensionality-reduced low-dimensional deviation vector. Each element of the low-dimensional deviation vector corresponds to the score of a principal component, reflecting the degree of deviation in that principal component direction. The low-dimensional deviation vector is then extended in terms of spatiotemporal dimension and associated with the corresponding rolling timestamp and signal level information to generate a comprehensive deviation index containing both time series and spatial distribution characteristics.

9. The method according to claim 1, characterized in that, The step of modeling the risk transmission path of the deviation evolution sequence set to generate a risk transmission network model containing potential fault propagation paths and impact intensity includes: constructing a graph structure learning framework for a physical field linkage perception network, wherein the graph structure learning framework includes a feature input layer, a graph topology construction layer, a path mining layer, and an intensity evaluation layer, and each layer transmits feature information through a fully connected manner; inputting the deviation evolution sequence set into the feature input layer, dividing the continuous evolution sequence into multiple time window segments through time series segmentation processing, each segment corresponding to a rolling process sub-stage, generating a time window feature matrix, wherein the row vectors of the matrix correspond to different time windows, and the column vectors correspond to the dimensions of the comprehensive deviation index; initializing the graph structure of each time window feature matrix through the graph topology construction layer, taking the dimensions of the comprehensive deviation index exceeding a preset threshold as graph nodes, constructing initial connections between nodes based on the topological relationship of the rolling state association graph, and generating a time window association graph; and using the path mining layer to process all time windows... A temporal correlation analysis is performed on the inter-window association graph to identify node connection paths across time windows. These paths are formed by connecting abnormal nodes from different time windows in chronological order, conforming to the adjacency relationship of nodes in the physical field propagation path between windows. The effectiveness of the identified inter-time window paths is verified by comparing the deviation evolution trend of nodes on the path with the physical field propagation law. False paths that do not conform to the propagation law are eliminated, while potential fault propagation paths are retained. The cumulative impact intensity of each potential fault propagation path is calculated. The cumulative impact intensity is calculated by comprehensively considering the deviation value of each node on the path, the propagation coefficient between nodes, and the path length. The propagation coefficient is determined based on the edge weight value of the rolling state association graph. The potential fault propagation paths and their corresponding cumulative impact intensities are input into the intensity evaluation layer. The evaluation results of the impact intensity are adjusted by introducing path occurrence frequency and duration parameters, generating a risk transmission network model that includes path topology, propagation direction, cumulative impact intensity, and temporal characteristics.

10. A fault early warning system, characterized in that, include: processor; And a memory, wherein the memory stores a computer program that, when run by the processor, causes the processor to perform the method as described in any one of claims 1 to 9.

Images