Intelligent prediction maintenance system and method for high-speed continuous rolling cold pilger mill

Abstract

The invention belongs to the technical field of equipment maintenance, and discloses an intelligent prediction maintenance system and method for a high-speed continuous rolling cold pilger mill. The method comprises the following steps: constructing a data acquisition network of the high-speed continuous rolling cold pilger mill, and acquiring operation data of each part of the high-speed continuous rolling cold pilger mill by using the data acquisition network; constructing a weighted undirected connected graph by taking each part of the high-speed continuous rolling cold pilger mill as a node; dividing the weighted undirected connected graph into a plurality of grids, and combining nodes in each grid into a super node to obtain a simple connected graph; presetting a maintenance rule, and performing node detection on the simple connected graph by using an improved scanning line algorithm to obtain super nodes in abnormal states; according to the method, the super nodes in the abnormal state are expanded to the component set, the maintenance suggestions are generated for the components in the component set, various possible abnormal conditions can be comprehensively considered, and diagnosis analysis is more comprehensive and accurate.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment maintenance, and more specifically, to an intelligent predictive maintenance system and method for a high-speed continuous rolling cold rolling tube mill. Background Art

[0002] The high-speed continuous rolling cold rolling tube mill is one of the key equipment in the steel production line, and its operating state directly affects the quality and production efficiency of steel products; the high-speed continuous rolling cold rolling tube mill consists of multiple complex functional systems, such as a feeding system, a rolling system, a cooling system, a coiling system, etc., and these systems contain a large number of key components, such as rolls, rolling stands, roll adjusting devices, roll bearings, motors, hydraulic cylinders, etc.; these key components bear huge loads and working stresses during the operation of the high-speed continuous rolling cold rolling tube mill and are extremely prone to failure; once a failure occurs, it will not only cause the equipment to stop running, resulting in economic losses, but may also trigger safety accidents, causing casualties and environmental pollution; therefore, timely detection and diagnosis of the operating faults of the high-speed continuous rolling cold rolling tube mill and preventive maintenance are of great significance for ensuring the reliable operation of the equipment, extending its service life, and reducing maintenance costs.

[0003] However, during the operation of the high-speed continuous rolling cold rolling tube mill, the prior art lacks the ability to comprehensively collect and monitor data of the equipment, making it difficult to accurately evaluate the actual operating state of the equipment; this results in the hidden dangers of equipment failures not being detected and warned in a timely manner, and sudden failures often occur, causing equipment downtime and production capacity losses, bringing significant economic losses to enterprises; secondly, the prior art cannot effectively model and analyze the mutual correlation relationships between the components of the equipment, ignoring the transmission and diffusion effects of faults within the system; this may lead to the simple localization of faults, the root problems cannot be solved, and the faults occur repeatedly, increasing the maintenance cost and economic losses; furthermore, the prior art is difficult to effectively simplify and abstract the large and complex continuous rolling cold rolling tube mill, and the computational complexity of state evaluation and fault diagnosis is too high, with low analysis efficiency; this makes it impossible to diagnose and handle faults in a timely manner, delaying the maintenance time and exacerbating the economic losses; in addition, the prior art lacks the intelligent analysis ability based on multiple maintenance rules, cannot comprehensively consider various abnormal situations, resulting in one-sided and inaccurate fault diagnosis and analysis, and poor maintenance measure effects.

[0004] In view of this, the present invention proposes an intelligent predictive maintenance system and method for a high-speed continuous rolling cold rolling tube mill to solve the above problems. Summary of the Invention

[0005] To overcome the above defects of the prior art and to achieve the above object, the present invention provides the following technical solution: An intelligent predictive maintenance system for a high-speed continuous rolling cold rolling tube mill, comprising: a network construction module for constructing a data acquisition network of the high-speed continuous rolling cold rolling tube mill and collecting the operation data of each component of the high-speed continuous rolling cold rolling tube mill by using the data acquisition network;

[0006] A graph construction module for constructing a weighted undirected connected graph with each component of the high-speed continuous rolling cold rolling tube mill as a node and taking the operation data as the attribute of the corresponding node in the weighted undirected connected graph;

[0007] A simplification module for dividing the weighted undirected connected graph into several grids and merging the nodes within each grid into a super node to obtain a simplified connected graph;

[0008] A maintenance module for presetting maintenance rules and using an improved scan line algorithm to detect nodes in the simplified connected graph to obtain super nodes with abnormal status; expanding the super nodes with abnormal status to a component set and generating maintenance suggestions for the components within the component set; the various modules are connected by wired and / or wireless means.

[0009] Furthermore, the construction method of the data acquisition network includes:

[0010] Identifying the main functional units of the high-speed continuous rolling cold rolling tube mill, identifying the key components of each main functional unit, determining the monitored parameters for each key component, and classifying the monitored parameters by type; dividing the monitored parameters of each type into corresponding data acquisition tasks, and decomposing all data acquisition tasks into several subtasks; sorting the subtasks in descending order according to the weights of the subtasks to form an ordered subtask execution sequence;

[0011] Analyzing the components involved in the subtasks and the data types to be collected, selecting the sensor type according to the structure and working environment of the components, and determining the installation positions of the sensors; defining the data acquisition process including data sampling, preprocessing and storage; taking the installation positions of the sensors as data acquisition points; using the K-means++ clustering algorithm to cluster all data acquisition points to obtain the clustering centers, and taking the clustering centers as the initial positions of the data aggregation nodes; constructing an undirected graph, where the nodes in the undirected graph are all data acquisition points and data aggregation nodes, and the edges of the undirected graph are the connections between them; defining the comprehensive weight of the edges as the distance from the data acquisition points to the data aggregation nodes;

[0012] Using the simulated annealing algorithm to find the minimum spanning tree on the undirected graph, and the nodes on the minimum spanning tree are the positions of the optimized data aggregation nodes; optimizing the positions of the optimized data aggregation nodes to obtain the data acquisition network.

[0013] Furthermore, the method for graph optimization of the positions of the optimized data aggregation nodes includes:

[0014] Taking all the optimized data aggregation nodes as seed points, for each seed point, finding another seed point closest to it, making a line segment on the perpendicular bisector of the two seed points, and extending this line segment to infinity; repeating the operation of making line segments for all seed points, and finally obtaining a series of line segments, which divide the plane where the undirected graph is located into several regions, corresponding each region to a data aggregation node to obtain a multi-region graph;

[0015] Projecting all data collection points onto the multi-region graph and judging which region they fall into respectively; for all data collection points falling into a certain region, the data aggregation node corresponding to this region is responsible for collecting their data; each data collection point samples the relevant parameters of the measured component in real time according to the preset sampling frequency and accuracy requirements to obtain data; the collected data is transmitted to the corresponding data aggregation node by wireless or wired means; configuring a data cache space at each data aggregation node, the data aggregation node collects the data transmitted by all data collection points in the corresponding region and caches it, and uses the data stream algorithm to perform real-time preprocessing on the cached data;

[0016] Presetting a load threshold, monitoring the load of each data aggregation node, when the load of a certain data aggregation node is greater than the load threshold, using the ant colony algorithm to dynamically adjust the corresponding region between adjacent data aggregation nodes, and setting a backup node for each data aggregation node, thus completing the construction of the data collection network.

[0017] Furthermore, the method for decomposing all data collection tasks into several subtasks includes:

[0018] Initializing a tree structure, regarding each data collection task as the root node of the tree structure, and storing the basic information of the data collection task in the root node, where the basic information includes the task name and time constraint;

[0019] Defining decomposition conditions, including component type, data type, and spatial location; storing the decomposition conditions as the attributes of the nodes of the tree structure;

[0020] For the root node, decomposing it according to the decomposition conditions to obtain the child nodes of the root node, and constructing the next layer of the tree structure based on the obtained child nodes; for each child node, continuing the recursive decomposition until the preset decomposition granularity is reached;

[0021] During the decomposition process, check whether adjacent child nodes meet the merging conditions; the merging conditions include the same component type and data type; merge adjacent child nodes that meet the merging conditions into a new node and replace it with the position of any one of these adjacent child nodes; define the weight \(W_T\) of each child node \(T\) as \(W_T = 1 - exp(-k\times(\alpha\times R(T)+\beta\times N(T)+\gamma\times M(T)))\); where \(R(T)\) represents the computing resources required to collect the data corresponding to the child node \(T\); \(N(T)\) represents the number of execution steps for collecting the data corresponding to the child node \(T\), and \(M(T)\) represents the data transmission volume for collecting the data corresponding to the child node \(T\) per unit time; \(k\) is an adjustment parameter, and \(\alpha\), \(\beta\), and \(\gamma\) are corresponding weight coefficients, and \(\alpha+\beta+\gamma = 1\).

[0022] Construct a two-dimensional coordinate system in the space where the tree structure is located, and construct a corresponding line segment for each child node in the tree structure. The length of the line segment is the weight of the corresponding child node; construct an index for each node; the index is the starting coordinate and the ending coordinate of the corresponding line segment in the two-dimensional coordinate system; and the child node corresponds to a subtask; traverse the child nodes of the tree structure to obtain the subtasks represented by each child node.

[0023] Furthermore, the operating data includes temperature data, pressure data, speed / displacement data, vibration data, current / power data, status data, and process parameters.

[0024] The temperature data includes the roll surface temperature, roll bearing temperature, motor winding temperature, hydraulic system temperature, and cooling water temperature; the pressure data includes the rolling pressure of the rolling mill, the hydraulic system pressure, and the cooling water pressure; the speed / displacement data includes the roll speed, the feeding vehicle speed, the coiler winding speed, and the hydraulic cylinder piston displacement; the vibration data includes roll vibration, rolling mill frame vibration, and motor vibration; the current / power data includes the main motor current, the auxiliary motor current, and the total power consumption; the status data includes the status of each actuator and the status of various alarm signals; the process parameters include the rolling torque, the rolling tension, the rolling deformation rate, and the cooling medium flow rate.

[0025] Furthermore, the construction method of the weighted undirected connected graph includes:

[0026] Analyze the roles of each component during operation. For components that are directly connected or have interactions, connect an undirected edge between them; use all components as the nodes of the weighted undirected connected graph; assign a corresponding weight value \(w(e)\) to each undirected edge \(e\); obtain the weighted undirected connected graph.

[0027] Weight value Among them, the importance levels of the two components connected by the undirected edge e are r1 and r2 respectively, s_e is the influence degree score between the two components connected by the undirected edge e; h is the normalization coefficient; RPN_e is the risk priority number caused by the breakage of the two components connected by the undirected edge e; RPNmax is the maximum value of the risk priority numbers in all connections; a1, a2, a3 are the weights of the corresponding items.

[0028] The ways to obtain the risk priority number include:

[0029] Define the breakage of the two components connected by the undirected edge e as the top event TE of the fault tree, analyze the direct causes leading to the occurrence of the top event TE, connect these direct causes to the top event with logic gates, and construct the first level of the fault tree; for each direct cause, continue to analyze its direct causes, use these direct causes as the next level of the first level of the fault tree, and connect them to the first level with logic gates.

[0030] Preset the basic event as a fault state or human error that does not need or cannot be further decomposed. Decompose step by step like this until it is decomposed to the basic event; for each basic event X, construct a Bayesian network model, use the basic event X as a network node of the Bayesian network model, and use various factors that may cause X to occur as the parent nodes of X; based on historical data, learn the conditional probability distribution between the network nodes in the Bayesian network model; use the probability propagation algorithm of the Bayesian network to calculate the prior probability P(X) of the basic event X.

[0031] Combined with the Bayesian network model, the fault tree calculates the probability of occurrence of each intermediate event from bottom to top until the probability of occurrence P(TE) of the top event TE is calculated; based on the probability of occurrence P(TE) of the top event TE, the risk priority number RPN = P(TE) × 10k1 is calculated; where, k1 is a range adjustment coefficient.

[0032] Furthermore, the way to divide the weighted undirected connected graph into several grids includes:

[0033] Project the weighted undirected connected graph onto a two-dimensional plane, the coordinates of the nodes are their coordinates on the two-dimensional plane, construct an initial rectangle R, and for each undirected edge e in the weighted undirected connected graph, calculate the vector between the two endpoints of the undirected edge e.

[0034] Construct a rectangular frame according to the vertical direction of the vector. The boundary lines of the rectangular frame are perpendicular to the vector, and the width of the rectangular frame is the width w of the preset grid; add the rectangular frame to a preset polygon list L; repeat until all undirected edges are processed; initialize an active edge table A and a vertex sorted list S. Take out a rectangular frame P from L, sort all the edges of P in ascending order of the ordinate of the vertices in the two-dimensional plane, and add them to S; traverse S. For each vertex, if the current vertex is the start vertex, find an edge in A such that the current vertex is on the right side of the edge, denoted as the right side edge, and insert a new edge above the right side edge. The start vertex of the new edge is the current vertex. If the current vertex is the end vertex, find an edge in A such that the current vertex is on the left side of the edge, denoted as the left side edge, delete the left side edge, and update the coordinates of the vertices of all the edges in A that intersect with the current vertex.

[0035] During the traversal process, for each pair of intersecting edges in A, calculate their intersection point p. Take p as the splitting point, divide the intersecting edges into two, forming four new edges; connect these four new edges in counterclockwise order to form a new polygon area r1; traverse the non-intersecting edges in A, connect their start vertices and end vertices to form a new polygon area r2; take all the new polygon areas as the union area, and repeat until all rectangular frames are processed; perform a difference operation on the union area and the initial rectangle R to obtain the final grid division result S'. For each grid in S', collect all the nodes that fall within the grid as the set of nodes of the grid.

[0036] Furthermore, the method for obtaining the reduced connected graph includes:

[0037] Regard all the nodes in each grid as objects, and connect the objects through edges, where the edges are denoted as object edges; calculate the importance weight w_i = b1×deg_i + b2×value_i of each object i; where deg_i is the node degree of object i, representing the number of edges connecting the node corresponding to object i to other nodes; value_i is the weighted value of the running data of the component corresponding to object i; b1 and b2 are relative adjustment parameters;

[0038] Select the node corresponding to the object with the largest importance weight in the corresponding grid as the living cell, and the remaining nodes as non-living cells; define the pulse intensity, and use the living cells to release pulses; the definition formula of the pulse intensity is:

[0039] where P0 is the preset initial pulse intensity, δ is the distance attenuation coefficient, d′ is the Euclidean distance to the nearest living cell; ε is the oscillation intensity coefficient, ω is the oscillation angular frequency, θ is the oscillation initial phase, and θ is the data value attenuation coefficient. is the decay rate of the data value; value is the weighted value of the operation data of the corresponding components of all non-living cells;

[0041] Non-living cells select the nearest living cell according to the received maximum pulse intensity and form an affinity connection with it; according to the affinity connection, all non-living cells are aggregated around the corresponding living cells to form several clusters, and each cluster is regarded as a new supernode;

[0042] After the above processing, the original weighted undirected connected graph is simplified into a new reduced connected graph G'(V', E'), where V' is the set of supernodes in all grids and E' is the set of new connection edges;

[0043] On the reduced connected graph G'(V', E'), the connection edges between supernodes retain the weights of the corresponding directed edges in the original weighted undirected connected graph.

[0044] Further, the method for detecting nodes in the reduced connected graph includes:

[0045] Create an empty status list and a priority queue; map the reduced connected graph to a two-dimensional plane, set a unique ID for each supernode in the reduced connected graph, and generate events for each supernode. The events include start events and end events; the start event contains information: supernode ID, event type, and the position coordinates in the two-dimensional plane; sort all events according to the abscissa in the two-dimensional plane and insert them into the priority queue;

[0046] Create a scanning line Lw perpendicular to the x-axis of the two-dimensional plane. Lw scans the entire reduced connected graph from left to right. When Lw moves to a new abscissa, process all events at this abscissa. For start events, add the corresponding supernode v' to the status list, check the connection relationship between v' and other supernodes in the status list, and update the adjacency information of v'; for end events, remove the corresponding supernode v' from the status list and update the adjacency information of other supernodes affected by the removal of v';

[0047] When processing each event, obtain the original node operation data contained in the current supernode v', evaluate the status of v' according to the preset maintenance rules, and when detecting a supernode v' with an abnormal status, calculate the influence coefficient between v' and its adjacent supernodes; the influence coefficient is the weight of the connection edge between v' and its adjacent supernodes;

[0048] For each adjacent supernode u of v', calculate its degree of influence where D(u) is the degree of u, and IC(v', u) is the influence coefficient between v' and its adjacent supernode u;

[0049] Set an impact threshold. If AI(u) is greater than the impact threshold, then u is significantly affected. For each adjacent supernode u that is significantly affected, update its risk level to Risk(u) = AI(u) × S(v'); where S(v') is the anomaly severity of v'.

[0050] If Risk(u) is greater than the preset risk threshold, then define the state of the corresponding adjacent supernode u as abnormal. After the scan is completed, output the state of each supernode.

[0051] An intelligent predictive maintenance method for a high-speed continuous rolling cold rolling tube mill, which is implemented based on the described intelligent predictive maintenance system for a high-speed continuous rolling cold rolling tube mill, includes: Step 1, construct a data acquisition network for the high-speed continuous rolling cold rolling tube mill, and use the data acquisition network to collect the operation data of each component of the high-speed continuous rolling cold rolling tube mill;

[0052] Step 2, take each component of the high-speed continuous rolling cold rolling tube mill as a node to construct a weighted undirected connected graph, and use the operation data as the attribute of the corresponding node in the weighted undirected connected graph;

[0053] Step 3, divide the weighted undirected connected graph into several grids, and merge the nodes within each grid into a supernode to obtain a simplified connected graph;

[0054] Step 4, preset maintenance rules, and use the improved scan line algorithm to detect the nodes of the simplified connected graph to obtain the supernodes with abnormal states; expand the supernodes with abnormal states to the component set, and generate maintenance suggestions for the components within the component set.

[0055] Technical effects and advantages of an intelligent predictive maintenance system and method for a high-speed continuous rolling cold rolling tube mill according to the present invention: The present invention realizes all-round monitoring and data collection of the entire life cycle of equipment, can accurately grasp the real-time operation status of the equipment, timely discover potential fault hazards, avoid the occurrence and expansion of faults, and thus minimize the economic losses caused by equipment failures; Secondly, a mathematical model of the mutual correlation relationship between equipment components is established, which can scientifically analyze the transmission and diffusion laws of faults in the system, contribute to the root cause solution of fault problems, and prevent the recurrence of faults; Furthermore, graph theory algorithms and artificial intelligence technologies are adopted to efficiently simplify and abstract large-scale complex systems, improve the computational efficiency of fault diagnosis, and shorten the analysis and decision-making time; In addition, a variety of maintenance rules are integrated, including threshold rules, trend rules, cycle rules, and association rules, which can comprehensively consider various possible abnormal situations, and the diagnostic analysis is more comprehensive and accurate; Finally, based on the intelligent analysis results, a targeted maintenance advice list can be generated for on-site maintenance personnel, clearly indicating abnormal components, fault types, and maintenance measures, greatly improving the quality and efficiency of maintenance work. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 FIG. is a schematic diagram of an intelligent predictive maintenance system for a high-speed continuous rolling cold rolling tube mill according to the present invention;

[0057] Figure 2 FIG. is a schematic diagram of an intelligent predictive maintenance method for a high-speed continuous rolling cold rolling tube mill according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0058] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0059] Embodiment 1

[0060] Please refer to Figure 1 shown. An intelligent predictive maintenance system for a high-speed continuous rolling cold rolling tube mill in this embodiment includes:

[0061] A network construction module for constructing a data collection network for the high-speed continuous rolling cold rolling tube mill and collecting the operation data of each component of the high-speed continuous rolling cold rolling tube mill by using the data collection network;

[0062] A graph construction module for constructing a weighted undirected connected graph with each component of the high-speed continuous rolling cold rolling tube mill as nodes and taking the operation data as the attributes of the corresponding nodes in the weighted undirected connected graph;

[0063] A simplification module is used to divide a weighted undirected connected graph into several grids, and merge the nodes within each grid into a supernode to obtain a simplified connected graph;

[0064] A maintenance module is used to preset maintenance rules, and use an improved scan line algorithm to detect nodes in the simplified connected graph to obtain supernodes with abnormal status; expand the supernodes with abnormal status to a component set, and generate maintenance suggestions for the components within the component set; each module is connected by wired and / or wireless means to achieve data transmission between modules.

[0065] Identify the main functional units of a high-speed continuous rolling cold rolling tube mill, such as a feeding system, a rolling system, a cooling system, a coiling system, etc.; identify the key components of each main functional unit. For example, the key components in the rolling system include rolls, rolling stands, roll adjusting devices, and roll bearings; for each key component, determine the monitored parameters; for example, for rolls, the parameters to be monitored include roll temperature, roll pressure, roll speed, and roll vibration; classify the monitored parameters by type, such as temperature data, pressure data, speed data, vibration data, current data, and displacement data, etc.

[0066] Divide the monitored parameters of each type into corresponding data acquisition tasks; decompose all data acquisition tasks into several subtasks, sort the subtasks in descending order according to the weights of the subtasks to form an ordered subtask execution sequence; each subtask is responsible for data acquisition of one or more related components.

[0067] The decomposition process is as follows:

[0068] Initialize a tree structure, such as an R-tree, a segment tree, etc.; regard each data acquisition task as the root node of the tree structure, and store the basic information of the data acquisition task in the root node. The basic information includes the task name and time constraints; it should be noted that the task name is a naming or identification of the data acquisition task, used to distinguish different data acquisition tasks. For example, roll temperature acquisition, roll pressure acquisition, etc.; the time constraint refers to the time requirement for the data acquisition task, and the time requirement includes sampling frequency, real-time performance, delay tolerance, and task execution cycle.

[0069] The sampling frequency refers to how often data needs to be sampled, such as once per second, once per minute, etc.; real-time performance refers to how long it takes to process and output the results after the data is collected for real-time monitoring; delay tolerance refers to how long to alarm if data is not received to determine that data acquisition is abnormal; if it is a periodically repeated task, what is the execution cycle.

[0070] These time constraints reflect the real-time requirements for the data acquisition task, such as time real-time performance and reliability, which are important conditions for task execution and need to be stored at the root node. By storing the task name and time constraints, the specific content and corresponding requirements of each data acquisition task can be clarified, providing a basis for subsequent task decomposition and scheduling.

[0071] Define the decomposition conditions, including component type, data type, and spatial location (the physical location of the monitored component in the device); store the decomposition conditions as attributes of the nodes in the tree structure.

[0072] For the root node, decompose it according to the decomposition conditions to obtain the child nodes of the root node, and build the next layer of the tree structure based on the obtained child nodes; for each child node, continue to recursively decompose until the preset decomposition granularity is reached; decompose a large data acquisition task into multiple small tasks according to the conditions, so that each small task has relatively single characteristics and is easier to allocate execution resources and arrange the execution plan.

[0073] During the decomposition process, check whether adjacent child nodes meet the merging conditions; the merging conditions include the same component type and data type; merge the adjacent child nodes that meet the merging conditions into a new node and replace it with the position of any one of these adjacent child nodes; define the weight \(W_T\) of each child node \(T = 1 - \exp(-k\times(\alpha\times R(T)+\beta\times N(T)+\gamma\times M(T)))\); where \(R(T)\) represents the computing resources required to collect the data corresponding to the child node \(T\), such as CPU time, memory occupancy, etc.; \(N(T)\) represents the number of execution steps for collecting the data corresponding to the child node \(T\), and \(M(T)\) represents the data transmission volume of collecting the data corresponding to the child node \(T\) per unit time; \(k\) is a tuning parameter used to control the scaling degree of the mapping, and \(\alpha\), \(\beta\), and \(\gamma\) are the corresponding weight coefficients representing the relative importance of each factor, and \(\alpha+\beta+\gamma = 1\).

[0074] Construct a two-dimensional coordinate system in the space where the tree structure is located, construct a corresponding line segment for each child node in the tree structure, and the length of the line segment is the weight of the corresponding child node; construct an index for each node for quick search and positioning; the index is the starting coordinate and ending coordinate of the corresponding line segment in the two-dimensional coordinate system; and the child node corresponds to the subtask; traverse the child nodes of the tree structure, obtain the subtasks represented by each child node, sort the subtasks in descending order according to the weight of the subtasks, and form an ordered subtask execution sequence as the final result of the decomposition.

[0075] Analyze the components involved in the subtask and the data types to be collected, and select the sensor type according to the structure and working environment of the component, such as temperature sensor, pressure sensor, vibration sensor, encoder, current transformer, etc.; determine the installation position of the sensor, which should be as close as possible to the component to be measured without affecting the normal operation of the component.

[0076] The process of defining data acquisition includes data sampling, preprocessing, and storage; the installation location of the sensor is used as the data acquisition point; the K-means++ clustering algorithm is used to cluster all data acquisition points to obtain the cluster centers, and the cluster centers are used as the initial positions of the data aggregation nodes; an undirected graph is constructed, where the nodes in the undirected graph are all data acquisition points and data aggregation nodes, and the edges of the undirected graph are the connections between them; the comprehensive weight of the edge is defined as the distance from the data acquisition point to the data aggregation node, and the closer the distance, the smaller the comprehensive weight.

[0077] The simulated annealing algorithm is used to find the minimum spanning tree on the undirected graph, and the nodes on the minimum spanning tree are the positions of the optimized data aggregation nodes; the positions of the optimized data aggregation nodes are graph-optimized.

[0078] Specifically, all optimized data aggregation nodes are used as seed points. For each seed point, find the other seed point closest to it, draw a line segment on the perpendicular bisector of the two seed points, and extend this line segment to infinity; repeat drawing line segments for all seed points. Finally, a series of line segments are obtained, and these line segments divide the plane where the undirected graph is located into several regions. The distance from any point in each region to the seed point corresponding to that region is closer than the distance to other seed points; each region corresponds to a data aggregation node to obtain a multi-region graph.

[0079] All data acquisition points are projected onto the multi-region graph to determine which regions they fall into; for all data acquisition points that fall into a certain region, the data aggregation node corresponding to that region is responsible for collecting their data.

[0080] Each data acquisition point samples the relevant parameters (such as temperature, pressure, etc.) of the measured component in real time according to the preset sampling frequency and accuracy requirements to obtain data; the collected data is transmitted to the corresponding data aggregation node wirelessly or wiredly; a certain amount of data cache space is configured at each data aggregation node, and the data aggregation node collects the data transmitted from all data acquisition points in the corresponding region and caches it, and uses data flow algorithms (such as sliding window, Bayesian estimation, etc.) to preprocess the cached data in real time; reduce data redundancy, extract valuable features, and reduce the data transmission volume.

[0081] A preset load threshold is set to monitor the load of each data aggregation node. When the load of a certain data aggregation node is greater than the load threshold, the ant colony algorithm is used to dynamically adjust the corresponding region between adjacent data aggregation nodes, and a backup node is set for each data aggregation node. Once the data aggregation node fails, the backup node immediately takes over, and thus the construction of the data acquisition network is completed.

[0082] The operating data includes temperature data, pressure data, speed / displacement data, vibration data, current / power data, status data, and process parameters; the temperature data includes the roll surface temperature, roll bearing temperature, motor winding temperature, hydraulic system temperature, and cooling water temperature; the pressure data includes the rolling pressure of the rolling mill, hydraulic system pressure, and cooling water pressure; the speed / displacement data includes the roll speed, feeding vehicle speed, coiler winding speed, and hydraulic cylinder piston displacement; the vibration data includes roll vibration, rolling mill frame vibration, and motor vibration; the current / power data includes the main motor current, auxiliary motor current, and total power consumption; the status data includes the status of each actuator (such as the door opening / closing status) and the status of various alarm signals; the process parameters include rolling torque, rolling tension, rolling deformation rate, and cooling medium flow rate; by comprehensively monitoring these key operating data, the actual operating status of the equipment can be reflected, providing data support for equipment status assessment, fault diagnosis, and predictive maintenance.

[0083] The construction method of the weighted undirected connected graph includes:

[0084] Analyze the functions of each component during operation. For components that are directly connected or have interactions, connect an undirected edge between them to indicate the association between them; take all components as the nodes of the weighted undirected connected graph; assign a corresponding weight value w(e) to each undirected edge e to represent the importance level of this connection; obtain the weighted undirected connected graph;

[0085] Weight value Among them, the importance levels of the two components connected by the undirected edge e are r1 and r2 respectively, s_e is the impact degree score between the two components connected by the undirected edge e; h is the normalization coefficient used to normalize the impact degree score; RPN_e is the risk priority number caused by the breakage of the two components connected by the undirected edge e; RPNmax is the maximum value of the risk priority numbers in all connections, used for normalization, and a1, a2, a3 are the weights of the corresponding items used to control the impact degree.

[0086] The impact degree score aims to quantify the impact degree of the connection breakage on the equipment operation, and evaluate the impact degree score from the aspects of production impact, safety impact, maintenance difficulty, and fault transmission impact;

[0087] The production impact is: evaluate the equipment downtime caused by the connection breakage, and evaluate the direct economic losses such as production capacity loss and product loss caused thereby. The scoring range can be set from 0 to 4 points, and the higher the score, the greater the production impact;

[0088] The safety impact is: evaluate the risk of personal injury that may be caused by the connection breakage, and evaluate the potential hazards to the environment (such as pollutant leakage, etc.); the scoring range can be set from 0 to 3 points, and the higher the score, the greater the safety impact;

[0089] The maintenance difficulty is as follows: evaluate the resources such as time, manpower, spare parts, etc. required for maintenance, and evaluate the complexity and technical difficulty of maintenance; the scoring range can be set from 0 to 2 points, and the higher the score, the greater the maintenance difficulty.

[0090] The fault transmission impact is as follows: evaluate the possible faults of other components caused by the connection breakage, and evaluate the transmission and diffusion degree of the fault in the system; the scoring range can be set from 0 to 2 points, and the higher the score, the greater the fault transmission impact.

[0091] In the actual scoring process, form a review group including multiple fields such as design, process, maintenance, safety, etc., and score each scoring item and form a consensus according to the specific situation and historical data of the equipment.

[0092] The methods for obtaining the risk priority number include:

[0093] Define the breakage of two components connected by the undirected edge e as the vertex event TE of the fault tree, analyze the direct causes leading to the occurrence of the vertex event TE, such as material defect (M), improper design (D), poor installation (I), wear (W), overload (O), etc., and connect these direct causes to the vertex event with logic gates to construct the first level of the fault tree; for example, if an AND gate is used for connection, then TE = M ∩ D ∩ I ∩ W ∩ O.

[0094] For each direct cause, continue to analyze its direct causes, and take these direct causes as the next level of the first level of the fault tree, and connect them to the first level with logic gates; for example, for the material defect M, it may be caused by the raw material quality problem (Q) and the process defect (P), then M = Q ∪ P.

[0095] Preset the basic event as a fault state or human error that does not need or cannot be further decomposed, and decompose it step by step until it is decomposed to the basic event; for each basic event X, estimate its occurrence probability value P(X). Specifically, construct a Bayesian network model, take the basic event X as a network node of the Bayesian network model, and take various factors that may cause X to occur as the parent nodes of X.

[0096] Based on historical data, learn the conditional probability distribution between the network nodes in the Bayesian network model; use the probability propagation algorithm of the Bayesian network (such as the enumeration method, variable elimination method, belief propagation algorithm, etc.) to calculate the prior probability P(X) of the basic event X.

[0097] Historical data includes the occurrence records of X, the status records of the parent nodes of X, and other relevant data; the occurrence records of X include the number of times X occurs under different times and conditions; the status records of the parent nodes of X record the status of various factors (i.e., the parent nodes of X) that lead to X for each occurrence or non-occurrence of X; other relevant data such as equipment operation parameters, environmental conditions, maintenance records, etc. are data that may affect the occurrence of X.

[0098] For the combined Bayesian network model, the fault tree calculates the occurrence probability of each intermediate event from bottom to top until the occurrence probability P(TE) of the top event TE is calculated; it should be noted that an intermediate event is composed of other events (including basic events or other intermediate events) combined through logic gates (AND gate or OR gate), reflecting the intermediate state of the propagation and development of faults in the system.

[0099] For example, for the event Y = X1 ∩ X2 ∩... ∩ Xn, then P(Y) = P(X1) × P(X2) ×... × P(Xn); X1, X2... Xn are all basic events or other intermediate events; based on the occurrence probability P(TE) of the top event TE, the risk priority number RPN = P(TE) × 10k1 is calculated; where k1 is a range adjustment coefficient used to expand the probability value to an appropriate numerical range, usually taking values from 5 to 8.

[0100] Through the above steps, we can construct the weighted undirected connected graph model G(V, E) of the high-speed continuous rolling cold rolling tube mill. This model not only describes the topological connection relationship between the components of the equipment but also contains the operation data information of the components, laying a foundation for subsequent minimum spanning tree solving, important connection extraction, and fault diagnosis analysis.

[0101] The ways to divide the weighted undirected connected graph into several grids include:

[0102] Project the weighted undirected connected graph onto a two-dimensional plane, the coordinates of the nodes are their coordinates on the two-dimensional plane, and an initial rectangle R is constructed. Its boundary is large enough to contain all the nodes; for each undirected edge e in the weighted undirected connected graph, calculate the vector between the two endpoints of the undirected edge e.

[0103] According to the vertical direction of the vector, construct a rectangular frame. The boundary line of the rectangular frame is perpendicular to the vector, and the width of the rectangular frame is the preset width w of the grid; add the rectangular frame to a preset polygon list L; repeat until all undirected edges are processed. At this time, the polygon list L contains the rectangular frames corresponding to all undirected edges.

[0104] Initialize an active edge table A and a vertex sorted list S. Take a rectangular box P from L, sort all the edges of P according to the ascending order of the vertical coordinates of the vertices in the two-dimensional plane, and add them to S. Traverse S. For each vertex, perform different operations according to its type (start vertex or end vertex) to update A. It should be explained that the start vertex refers to the left endpoint of the edge of the rectangular box P; when scanning from left to right, this endpoint is the start of the edge; the end vertex (right vertex) refers to the right endpoint of the edge of the rectangular box P; when scanning from left to right, this endpoint is the end of the edge.

[0105] Specifically, for the different operations, if the current vertex is a start vertex (left vertex), find an edge in A such that the current vertex is on the right side of the edge, denoted as the right side edge, and insert a new edge above the right side edge. The start vertex of the new edge is the current vertex. If the current vertex is an end vertex (right vertex), find an edge in A such that the current vertex is on the left side of the edge, denoted as the left side edge, delete the left side edge, and update the coordinates of the vertices of all the edges in A that intersect with the current vertex.

[0106] During the traversal, for each pair of intersecting edges in A, calculate their intersection point p. Take p as the splitting point to divide the intersecting edges into two parts each, forming four new edges; connect these four new edges in counterclockwise order to form a new polygon region r1; traverse the non-intersecting edges in A and connect their start vertices and end vertices to form a new polygon region r2; take all the new polygon regions as the union region and repeat until all the rectangular boxes are processed.

[0107] Perform a difference operation between the union region and the initial rectangle R to obtain the final grid division result S'. For each grid in S', collect all the nodes that fall within the grid as the set of nodes of the grid; the difference operation refers to finding the difference set between the union region and the initial rectangle R, that is, the remaining region after removing the part overlapping with the rectangle R from the union region.

[0108] Divide the weighted undirected connected graph into a series of irregular polygon grids, and each grid contains some nodes; this division method can flexibly divide the grids according to the actual trend of the edges, avoiding the problem of node dispersion caused by the fixed grid size in the grid division.

[0109] The ways to obtain the reduced connected graph include:

[0110] All nodes within each grid are regarded as objects, and the objects are connected by edges, which are denoted as object edges here; calculate the importance weight w_i of each object i = b1×deg_i + b2×value_i; where deg_i is the node degree of object i, representing the number of edges connecting the node corresponding to object i to other nodes; value_i is the weighted value of the operation data of the component corresponding to object i; b1 and b2 are relative adjustment parameters, controlling the relative importance of the degree and the data value in the weight;

[0111] Select the node corresponding to the object with the largest importance weight within the corresponding grid as the live cell, and the remaining nodes as non-live cells; define the pulse intensity, and use the live cells to release pulses. The intensity of the pulse decreases with the increase of the distance. The definition formula of the pulse intensity is:

[0112] where P0 is the preset initial pulse intensity, δ is the distance attenuation coefficient, which is a real number greater than 0; d′ is the Euclidean distance to the nearest live cell; ε is the oscillation intensity coefficient, representing the oscillation amplitude of the pulse intensity, used to simulate the oscillation phenomenon of ion concentration in the biofilm; ω is the oscillation angular frequency, representing the frequency of the pulse intensity oscillation, that is, the angular distance required for a complete oscillation period; θ is the oscillation initial phase, representing the initial phase of the pulse intensity oscillation, that is, the initial displacement at the start of the oscillation; different initial phases can cause the oscillation curve to shift; θ is the data value attenuation coefficient, representing the influence degree of the operation data value of the component corresponding to the non-live cell on the pulse intensity; is the data value attenuation rate, representing the attenuation rate of the operation data value of the component corresponding to the non-live cell with the distance; value is the weighted value of the operation data of the components corresponding to all non-live cells.

[0114] The non-live cells select the nearest live cell according to the maximum pulse intensity received and form an affinity connection with it; according to the affinity connection, all non-live cells are aggregated around the corresponding live cells to form several clusters, and each cluster is regarded as a new supernode; after the above processing, the original weighted undirected connected graph is simplified to a new reduced connected graph G'(V', E'), where V' is the set of all supernodes within the grids, and E' is the set of new connection edges; on the reduced connected graph G'(V', E'), the connection edges between supernodes retain the weights of the corresponding directed edges in the original weighted undirected connected graph.

[0115] The original large-scale weighted undirected connected graph is divided into many small grids. Inside each grid, the heuristic grid biofilm model algorithm is used to merge nodes into supernodes, which can greatly reduce the complexity of the problem and lay a foundation for subsequent analysis and calculation. At the same time, since the supernodes are generated in a data-driven manner, the characteristic information of the original data can be better retained.

[0116] The ways of presetting maintenance rules include:

[0117] Maintenance rules include threshold rules, trend rules, maintenance period rules, and association rules;

[0118] The threshold rule is: Set a threshold range for the operating data, such as the normal operating range of parameters like temperature, pressure, vibration, etc.; when the parameter exceeds the threshold range, trigger corresponding maintenance, and the status of the corresponding supernode is abnormal;

[0119] The trend rule is: Analyze the change trend of the operating data over time, such as the continuous rise in temperature, the intensification of vibration, etc., and the status of the corresponding supernode is abnormal;

[0120] The maintenance period rule is: According to the service life of the equipment, the workload, etc., set the time period for regular maintenance, such as a major overhaul every six months, and the status of the corresponding supernode is abnormal;

[0121] The association rule is: Consider the correlation between different components. The failure of one component may cause other components to also require maintenance, and the status of the corresponding supernode is abnormal.

[0122] Create an empty status list to store the status information of supernodes; create a priority queue to store the events to be processed; map the reduced connected graph to a two-dimensional plane, and set a unique ID for each supernode in the reduced connected graph. Generate events for each supernode, and the events include start events and end events; the start event contains information: supernode ID, event type (start / end), and the position coordinates in the two-dimensional plane; sort all the events according to the abscissa in the two-dimensional plane and insert them into the priority queue.

[0123] Create a scan line Lw perpendicular to the x-axis of the two-dimensional plane. Lw scans the entire reduced connected graph from left to right. When Lw moves to a new abscissa, process all the events at this abscissa. For the start event, add the corresponding supernode v' to the status list, check the connection relationship between v' and other supernodes in the status list, and update the adjacency information of v'; for the end event, remove the corresponding supernode v' from the status list and update the adjacency information of other supernodes affected by the removal of v'.

[0124] When processing each event, obtain the running data of the original nodes contained in the current supernode v', evaluate the status of v' according to the preset maintenance rules, and when detecting a supernode v' with an abnormal status, calculate the influence coefficient between v' and its adjacent supernodes; the influence coefficient is the weight of the connection edge between v' and its adjacent supernodes.

[0125] For each adjacent supernode u of v', calculate its degree of influence Where D(u) is the degree of u (the number of connection edges connected to u), and IC(v', u) is the influence coefficient between v' and the adjacent supernode u; set an influence threshold (e.g., 0.5), if AI(u) is greater than the influence threshold, then u is significantly affected; for each adjacent supernode u that is significantly affected, update its risk level to Risk(u) = AI(u) × S(v'); where S(v') is the severity of the abnormality of v' (set according to the type and degree of the abnormality, e.g., mild = 1, moderate = 2, severe = 3); if Risk(u) is greater than the preset risk threshold, then define the status of the corresponding adjacent supernode u as abnormal; after the scan is completed, output the status of each supernode.

[0126] For each supernode with an abnormal status, find all the original nodes (i.e., components) contained therein, take these nodes as a component set, analyze the running data of each component in the component set, and judge its type of abnormality. For example, if parameters such as the temperature and pressure of the component exceed the normal threshold range, the type of abnormality of the component is parameter overrun; if parameters such as the temperature and vibration of the component show an abnormal upward / downward trend, the type of abnormality of the component is trend abnormality; if the usage time of the component exceeds the preset maintenance cycle, the type of abnormality of the component is periodic maintenance; if the component has an associated relationship with other abnormal components (such as series, parallel, etc.), the type of abnormality of the component is associated abnormality.

[0127] For each type of abnormality, generate corresponding maintenance suggestions:

[0128] Parameter overrun: It is recommended to check the component, identify the cause (such as overload, wear, etc.), and perform repair or replacement if necessary; Trend abnormality: It is recommended to strengthen monitoring, make preparations for maintenance in advance, and stop the machine for maintenance in a timely manner if necessary; Periodic maintenance: Arrange regular inspections and replacements according to the preset maintenance plan; Associated abnormality: In addition to taking maintenance measures for the abnormal component, it is also necessary to check and maintain other associated components.

[0129] For each component in the component set, output its type of abnormality and the corresponding maintenance suggestions. For supernodes with a normal status, no maintenance suggestions are generated. Output the final maintenance suggestion list, including the list of abnormal components, types of abnormalities, maintenance suggestions, etc., to provide guidance for on-site maintenance personnel.

[0130] Expand the super nodes with abnormal status, analyze the specific conditions of the abnormal components, and generate targeted maintenance suggestions for each abnormal component, so as to provide decision-making support for the preventive maintenance of the equipment; at the same time, consider the correlation between components, avoid missing abnormalities, and improve the comprehensiveness and effectiveness of maintenance.

[0131] In this embodiment, the all-round monitoring and data collection of the whole life cycle of the equipment are realized, the real-time operation status of the equipment can be accurately grasped, potential fault hazards can be discovered in time, the occurrence and expansion of faults can be avoided, so as to minimize the economic losses caused by equipment faults; secondly, a mathematical model of the correlation between equipment components is established, which can scientifically analyze the transmission and diffusion laws of faults in the system, help to solve fault problems rootedly, and prevent the recurrence of faults; furthermore, graph theory algorithms and artificial intelligence technologies are adopted to simplify and abstract large-scale complex systems efficiently, improve the calculation efficiency of fault diagnosis, and shorten the time of analysis and decision-making; in addition, a variety of maintenance rules are integrated, including threshold rules, trend rules, cycle rules and association rules, which can comprehensively consider various possible abnormal situations, and the diagnostic analysis is more comprehensive and accurate; finally, based on the intelligent analysis results, a list of targeted maintenance suggestions can be generated for on-site maintenance personnel, clearly indicating abnormal components, fault types and maintenance measures, greatly improving the quality and efficiency of maintenance work.

[0132] Embodiment 2

[0133] Please refer to Figure 2 As shown, for the parts not described in detail in this embodiment, refer to the description content of Embodiment 1. A method for intelligent predictive maintenance of a high-speed continuous rolling cold rolling tube mill is provided, including: Step 1, construct a data acquisition network for the high-speed continuous rolling cold rolling tube mill, and use the data acquisition network to collect the operation data of each component of the high-speed continuous rolling cold rolling tube mill;

[0134] Step 2, take each component of the high-speed continuous rolling cold rolling tube mill as a node to construct a weighted undirected connected graph, and use the operation data as the attributes of the corresponding nodes in the weighted undirected connected graph;

[0135] Step 3, divide the weighted undirected connected graph into several grids, and merge the nodes in each grid into a super node to obtain a simplified connected graph;

[0136] Step 4, preset maintenance rules, and use an improved scan line algorithm to detect the nodes of the simplified connected graph to obtain super nodes with abnormal status; expand the super nodes with abnormal status to the component set, and generate maintenance suggestions for the components in the component set.

[0137] Embodiment 3

[0138] This embodiment discloses an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the operation mode of the intelligent predictive maintenance method for a high-speed continuous rolling cold rolling tube mill provided above.

[0139] Since the electronic device introduced in this embodiment is the electronic device adopted for implementing the intelligent predictive maintenance method for a high-speed continuous rolling cold rolling tube mill in the embodiments of the present application, based on the intelligent predictive maintenance method for a high-speed continuous rolling cold rolling tube mill introduced in the embodiments of the present application, those skilled in the art can understand the specific implementation manners and various variations of the electronic device in this embodiment. Therefore, the specific implementation of how this electronic device implements the method in the embodiments of the present application will not be described in detail here. As long as those skilled in the art implement the electronic device adopted for the intelligent predictive maintenance method for a high-speed continuous rolling cold rolling tube mill in the embodiments of the present application, it falls within the scope of protection of the present application.

[0140] The above formulas are all dimensionless and take their numerical values for calculation. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain a formula that is closest to the actual situation. The preset parameters and threshold selection in the formulas are set by those skilled in the art according to the actual situation.

[0141] The above are only the preferred embodiments of the present invention. The protection scope of the present invention is not limited to the above embodiments. All technical solutions falling within the concept of the present invention belong to the protection scope of the present invention. It should be noted that for ordinary technical users in the technical field, several improvements and refinements made without departing from the principle of the present invention should also be regarded as within the protection scope of the present invention.

Claims

1. An intelligent predictive maintenance system for a high-speed continuous cold rolling mill, characterized in that: include: A network construction module is used to construct a data acquisition network for a high-speed continuous cold rolling mill, and to collect the operating data of various components of the high-speed continuous cold rolling mill using the data acquisition network; A graph construction module is used to construct a weighted undirected connected graph by taking various components of the high-speed continuous cold rolling mill as nodes, and taking the operation data as the attributes of the corresponding nodes in the weighted undirected connected graph; The simplification module is used to divide the weighted undirected connected graph into several grids and merge the nodes in each grid into a super node to obtain a simplified connected graph; The maintenance module is used to preset maintenance rules and use the improved scanning line algorithm to detect nodes on the simplified connected graph to obtain super nodes with abnormal status; expand the super nodes with abnormal status to the component set, and generate maintenance suggestions for the components in the component set; each module is connected by wire and / or wireless means.

2. The intelligent predictive maintenance system for high-speed continuous cold rolling mill according to claim 1 is characterized in that: The data acquisition network is constructed in the following manner: Identify the main functional units of the high-speed continuous cold rolling mill, identify the key components of each main functional unit, determine the monitoring parameters for each key component, and classify the monitored parameters by type; divide each type of monitored parameters into corresponding data collection tasks, and decompose all data collection tasks into several subtasks; sort the subtasks in descending order according to their weights to form an orderly subtask execution sequence; Analyze the components involved in the subtask and the type of data that needs to be collected. According to the structure and working environment of the components, select the sensor type and determine the installation location of the sensor; define the data collection process including data sampling, preprocessing and storage; use the installation location of the sensor as the data collection point; use the K-means++ clustering algorithm to cluster all data collection points, obtain the cluster center, and use the cluster center as the initial location of the data aggregation node; construct an undirected graph, in which the nodes are all data collection points and data aggregation nodes, and the edges are the lines between them; define the comprehensive weight of the edge as the distance from the data collection point to the data aggregation node; The simulated annealing algorithm is used to find the minimum spanning tree on the undirected graph. The nodes on the minimum spanning tree are the locations of the optimized data aggregation nodes. The locations of the optimized data aggregation nodes are graph optimized to obtain the data collection network.

3. The intelligent predictive maintenance system for high-speed continuous cold rolling mill according to claim 2 is characterized in that: The method of performing graph optimization on the position of the optimized data aggregation node includes: All optimized data aggregation nodes are used as seed points. For each seed point, find another seed point closest to it, draw a line segment on the perpendicular bisector of the two seed points, and extend this line segment to infinity. Repeat the line segment drawing for all seed points to finally obtain a series of line segments. These line segments divide the plane where the undirected graph is located into several regions. Each region corresponds to a data aggregation node, and a multi-region graph is obtained. Project all data collection points on the multi-region map to determine which region they fall into; for all data collection points falling in a certain region, the data aggregation node corresponding to the region is responsible for collecting their data; each data collection point samples the relevant parameters of the measured component in real time according to the preset sampling frequency and accuracy requirements to obtain data; the collected data is transmitted to the corresponding data aggregation node via wireless or wired means; configure data cache space at each data aggregation node, the data aggregation node collects and caches the data from all data collection points in the corresponding region, and uses the data stream algorithm to pre-process the cached data in real time; A load threshold is preset to monitor the load of each data aggregation node. When the load of a data aggregation node is greater than the load threshold, the ant colony algorithm is used to dynamically adjust the corresponding area between adjacent data aggregation nodes, and a backup node is set for each data aggregation node to complete the construction of the data collection network.

4. The intelligent predictive maintenance system for high-speed continuous cold rolling mill according to claim 3 is characterized in that: The method of decomposing all data collection tasks into several subtasks includes: Initialize a tree structure, and regard each data collection task as the root node of the tree structure. The root node stores the basic information of the data collection task, including the task name and time constraint. Defining decomposition conditions, including component type, data type, and spatial position; storing the decomposition conditions as attributes of nodes in a tree structure; For the root node, decompose it according to the decomposition condition to obtain the child nodes of the root node, and build the next layer of the tree structure based on the obtained child nodes; for each child node, continue to recursively decompose until the preset decomposition granularity is reached; During the decomposition process, check whether the adjacent child nodes meet the merging conditions; the merging conditions include the same component type and data type; merge the adjacent child nodes that meet the merging conditions into a new node, and replace it with the position of any child node in the adjacent child nodes; define the weight of each child node T W_T = 1-exp(-k×(α×R(T)+β×N(T)+γ×M(T))); where R(T) represents the computing resources required to collect the data corresponding to the child node T; N(T) represents the number of execution steps for collecting the data corresponding to the child node T, and M(T) represents the data transmission volume of the data corresponding to the child node T per unit time; k is an adjustment parameter, α, β and γ are corresponding weight coefficients, and α+β+γ=1; Construct a two-dimensional coordinate system in the space where the tree structure is located, and construct a corresponding line segment for each child node in the tree structure. The length of the line segment is the weight of the corresponding child node; construct an index for each node; the index is the starting point coordinates and the end point coordinates of the corresponding line segment in the two-dimensional coordinate system; and the child node corresponds to the subtask; traverse the child nodes of the tree structure to obtain the subtask represented by each child node.

5. The intelligent predictive maintenance system for high-speed continuous cold rolling mill according to claim 4 is characterized in that: The operation data includes temperature data, pressure data, speed / displacement data, vibration data, current / power data, state data and process parameters; Temperature data include roller surface temperature, roller bearing temperature, motor winding temperature, hydraulic system temperature and cooling water temperature; pressure data include rolling pressure, hydraulic system pressure and cooling water pressure of the rolling mill; speed / displacement data include roller speed, feeder car speed, winder winding speed and hydraulic cylinder piston displacement; vibration data include roller vibration, rolling mill frame vibration and motor vibration; current / power data include main motor current, auxiliary motor current and total power consumption; status data include the status of each actuator and various alarm signal status; process parameters include rolling torque, rolling tension, rolling deformation rate and cooling medium flow rate.

6. The intelligent predictive maintenance system for high-speed continuous cold rolling mill according to claim 5 is characterized in that: The weighted undirected connected graph is constructed in the following manner: Analyze the role of each component in the operation process. For components that are directly connected or interact with each other, connect an undirected edge between them; treat all components as nodes of a weighted undirected connected graph; assign a corresponding weight w(e) to each undirected edge e; and obtain a weighted undirected connected graph; Weight Among them, the importance levels of the two components connected by the undirected edge e are r1 and r2 respectively, s_e is the impact score between the two components connected by the undirected edge e; h is the normalization coefficient; RPN_e is the risk priority number caused by the break of the two components connected by the undirected edge e; RPNmax is the maximum value of the risk priority number among all connections; a1, a2, and a3 are the weights of the corresponding items; The risk priority number can be obtained by: Define the breakage of two components connected by the undirected edge e as the vertex event TE of the fault tree, analyze the direct causes that lead to the vertex event TE, connect these direct causes with the vertex event with logic gates, and construct the first level of the fault tree; for each direct cause, continue to analyze the direct cause of its occurrence, and use these direct causes as the next level of the first level of the fault tree, and connect them with the first level with logic gates; The basic event is preset as a fault state or human error that does not need or cannot be further decomposed, and the decomposition is carried out step by step until the basic event is decomposed; for each basic event X, a Bayesian network model is constructed, and the basic event X is used as a network node of the Bayesian network model, and various factors that may cause X to occur are used as the parent node of X; based on historical data, the conditional probability distribution between each network node in the Bayesian network model is learned; the probability propagation algorithm of the Bayesian network is used to calculate the prior probability P(X) of the basic event X; Combined with the Bayesian network model, the fault tree calculates the probability of occurrence of each intermediate event from bottom to top until the probability of occurrence P(TE) of the vertex event TE is calculated; the risk priority number RPN=P(TE)×10k1 is calculated based on the probability of occurrence P(TE) of the vertex event TE; where k1 is a range adjustment coefficient.

7. The intelligent predictive maintenance system for high-speed continuous cold rolling mill according to claim 6 is characterized in that: The method of dividing the weighted undirected connected graph into a plurality of grids includes: Project the weighted undirected connected graph onto a two-dimensional plane. The coordinates of the nodes are their coordinates on the two-dimensional plane. Construct an initial rectangle R. For each undirected edge e in the weighted undirected connected graph, calculate the vector between the two endpoints of the undirected edge e. According to the vertical direction of the vector, construct a rectangular box, the boundary line of which is perpendicular to the vector, and the width of which is the preset grid width w; add the rectangular box to a preset polygon list L; repeat until all undirected edges have been processed; initialize an active edge table A and a vertex sorting list S, take out a rectangular box P from L, sort all the edges of P from small to large according to the vertical coordinates of the vertices on the two-dimensional plane, and add them to S; traverse S, for each vertex, if the current vertex is the starting vertex, find an edge from A so that the current vertex is on the right side of the edge, recorded as the right side edge, insert a new edge above the right side edge, the starting vertex of the new edge is the current vertex, if the current vertex is the ending vertex, find an edge from A so that the current vertex is on the left side of the edge, recorded as the left side edge, delete the left side edge, and update the coordinates of all vertices of the edges in A that intersect with the current vertex; During the traversal process, for each pair of intersecting edges in A, calculate their intersection point p, and use p as the dividing point to split the intersecting edges into two to form four new edges; connect these four new edges in counterclockwise order to form a new polygonal area r1; traverse the non-intersecting edges in A, connect its starting vertex and ending vertex to form a new polygonal area r2; take all new polygonal areas as union areas, and repeat until all rectangular boxes have been processed; perform a difference operation on the union area and the initial rectangle R to obtain the final grid division result S'. For each grid in S', collect all the nodes that fall within the grid as the set of nodes of the grid.

8. The intelligent predictive maintenance system for high-speed continuous cold rolling mill according to claim 7 is characterized in that: The method for obtaining the simplified connectivity graph includes: All nodes in each grid are regarded as objects, and objects are connected by edges, where the edges are recorded as object edges; calculate the importance weight of each object i w_i = b1×deg_i+b2×value_i; where deg_i is the node degree of object i, indicating the number of edges connecting the node corresponding to object i with other nodes; value_i is the weighted value of the operating data of the component corresponding to object i; b1 and b2 are relative adjustment parameters; Select the node corresponding to the object with the largest importance weight in the corresponding grid as the live cell, and the remaining nodes as the non-live cells; define the pulse intensity, and use the live cells to release the pulse; the definition formula of the pulse intensity is: Where P0 is the preset initial pulse intensity, δ is the distance attenuation coefficient, d′ is the Euclidean distance to the nearest living cell; ε is the oscillation intensity coefficient, ω is the oscillation angular frequency, θ is the initial oscillation phase, and θ is the data value attenuation coefficient. is the data value decay rate; value is the weighted value of the operating data of the corresponding components of all non-living cells; The non-living cells select the nearest living cells according to the maximum pulse intensity received and form affinity connections with them. Based on the affinity connections, all non-living cells are gathered around the corresponding living cells to form several clusters, and each cluster is regarded as a new super node. After the above processing, the original weighted undirected connected graph is simplified into a new simplified connected graph G'(V', E'), where V' is the set of super nodes in all grids, and E' is the set of new connecting edges; In the simplified connected graph G'(V', E'), the connecting edges between super nodes retain the weights of their corresponding directed edges in the original weighted undirected connected graph.

9. The intelligent predictive maintenance system for high-speed continuous cold rolling mill according to claim 8 is characterized in that: The method of performing node detection on the simplified connected graph includes: Create an empty state list and a priority queue; map the minimalist connected graph to a two-dimensional plane, set a unique ID for each supernode in the minimalist connected graph, generate events for each supernode, and the events include start events and end events; the start event contains information: supernode ID, event type, and location coordinates on the two-dimensional plane; sort all events by the horizontal coordinates of the two-dimensional plane, and insert them into the priority queue; Create a scan line Lw perpendicular to the x-axis of the two-dimensional plane. Lw scans the entire simplified connected graph from left to right. When Lw moves to a new horizontal coordinate, process all events on the horizontal coordinate. For the start event, add the corresponding super node v' to the state list, check the connection relationship between v' and other super nodes in the state list, and update the adjacency information of v'; for the end event, remove the corresponding super node v' from the state list, and update the adjacency information of other super nodes affected by the removal of v'; When processing each event, the operation data of the original node contained in the current super node v' is obtained, and the status of v' is evaluated according to the preset maintenance rules. When a super node v' with an abnormal status is detected, the influence coefficient between v' and the adjacent super nodes is calculated; the influence coefficient is the weight of the connection edge between v' and the adjacent super nodes; For each adjacent super node u of v', calculate its degree of influence Where D(u) is the degree of u, IC(v',u) is the influence coefficient between v' and the adjacent supernode u; Set an impact threshold. If AI(u) is greater than the impact threshold, u is significantly affected. For each adjacent super node u that is significantly affected, update its risk level to Risk(u) = AI(u) × S(v'); where S(v') is the severity of the abnormality of v'. If Risk(u) is greater than the preset risk threshold, the state of the corresponding adjacent supernode u is defined as abnormal; after the scan is completed, the state of each supernode is output.

10. A high-speed continuous rolling cold rolling mill intelligent predictive maintenance method, which is implemented based on the high-speed continuous rolling cold rolling mill intelligent predictive maintenance system according to any one of claims 1 to 9, characterized in that: include: Step 1: construct a data acquisition network for a high-speed continuous cold rolling mill, and use the data acquisition network to collect operating data of various components of the high-speed continuous cold rolling mill; Step 2: construct a weighted undirected connected graph by taking the components of the high-speed continuous cold rolling mill as nodes, and taking the operation data as the attributes of the corresponding nodes in the weighted undirected connected graph; Step 3: Divide the weighted undirected connected graph into several grids, and merge the nodes in each grid into a super node to obtain a simplified connected graph; Step 4: preset maintenance rules, and use the improved scanning line algorithm to detect nodes on the simplified connected graph to obtain super nodes with abnormal status; expand the super nodes with abnormal status to the component set, and generate maintenance suggestions for the components in the component set.

Images