A Method and System for Predicting the Performance of Water-Lubricated Bearings Based on Graph Neural Network Finite Element Method

By constructing a graph dataset and training a graph neural network surrogate model using the finite element method based on graph neural networks, the problems of long prediction time, grid topology ignoring, and difficulty in predicting complex structures in the performance prediction of water-lubricated bearings are solved, and fast and accurate prediction of pressure field and performance indicators is achieved.

CN122864194BActive Publication Date: 2026-11-10SANYA SCI & EDUCATION INNOVATION PARK WUHAN UNIV OF TECH
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611351457.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-09-02
Publication Date
2026-11-10
Estimated Expiration
2046-09-02

AI Technical Summary

Technical Problem

The performance prediction of water-lubricated bearings faces challenges such as long finite element solution time per working condition, neglecting mesh topology, difficulty in predicting complex grooves and boundaries, and separation of pressure field from engineering performance indicators.

Method used

By employing a finite element method based on graph neural networks, and constructing a graph dataset and a graph neural network surrogate model, the topological relationships and local coupling relationships of the finite element mesh are learned, enabling rapid prediction of the performance of water-lubricated bearings.

Benefits of technology

It improves the efficiency of predicting the performance of water-lubricated bearings, preserves the topological information of the finite element mesh, adapts to complex grooves and boundary structures, realizes unified prediction of pressure field and engineering performance, and reduces computational costs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122864194B_ABST
    Figure CN122864194B_ABST
Patent Text Reader

Abstract

This invention relates to the field of water-lubricated bearing performance prediction, and provides a method and system for predicting the performance of water-lubricated bearings based on graph neural network learning of finite element data. The method employs the finite element method to establish a finite element simulation sample library for water-lubricated bearings. This library covers finite element simulation samples corresponding to different journal surface running linear velocities, bearing external loads, journal eccentricities, groove structure parameters, and material parameters. The mesh, input parameters, and results of the finite element simulation samples are converted into graph data. Based on the graph data, a graph neural network surrogate model is constructed and trained to learn the nonlinear mapping relationship between finite element input features and finite element output results. In engineering applications, the trained graph neural network surrogate model is used to replace the finite element solution. This invention can quickly predict the water film pressure field, load-bearing capacity, friction force, friction coefficient, maximum water film pressure, minimum water film thickness, and lubrication state indicators of water-lubricated bearings.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of performance prediction for water-lubricated bearings, and more particularly to a method and system for predicting the performance of water-lubricated bearings based on graph neural network finite element method. Background Technology

[0002] Water-lubricated bearings are widely used in ship propulsion shafting, marine engineering equipment, and large rotating machinery, offering advantages such as environmental friendliness, compact structure, and ease of maintenance. For ultra-long shafting systems in ships or large vessels, stern bearings are affected by rotational speed, load, shaft misalignment, groove structure, liner deformation, and boundary conditions during operation, resulting in significant spatial non-uniformity in their water film pressure field and load-bearing capacity. Since water viscosity is much lower than oil, the water film load-bearing capacity is highly sensitive to film thickness, eccentricity, local grooves, material deformation, and boundary leakage. Therefore, the performance calculation of water-lubricated bearings exhibits significant nonlinearity and multi-parameter coupling characteristics.

[0003] The Finite Element Method (FEM) can handle complex geometries, groove regions, material deformation, and boundary coupling problems in water-lubricated bearings, making it suitable for obtaining high-precision performance results. However, in multi-condition scanning, parameter optimization, and online evaluation of water-lubricated bearings, the FEM typically requires repeated modeling, mesh processing, solution calculation, and post-processing analysis, resulting in a long computation cycle. This makes it difficult to meet the needs of large-scale, rapid evaluation in the engineering design phase of water-lubricated bearings.

[0004] Existing neural networks or conventional multilayer perceptrons typically map the operating parameters of water-lubricated bearings directly to performance indicators, or directly map the spatial coordinates of water-lubricated bearings to pressure values. While these methods offer rapid prediction capabilities, they struggle to fully represent the topological relationships and local spatial couplings between nodes in a finite element mesh, particularly prone to prediction biases at groove edges, end boundaries, local high-pressure zones, and areas of abrupt changes in film thickness.

[0005] Therefore, the field of water-lubricated bearing performance prediction urgently needs a graph neural network fast prediction method that does not rely on physical information loss function and directly learns finite element solution data, so as to realize the rapid prediction and analysis of multi-condition pressure field and performance index of water-lubricated bearings while maintaining the expressive power of finite element data.

[0006] Therefore, the following technical problems urgently need to be solved in the field of water-lubricated bearing performance prediction:

[0007] (1) Solve the problem of long time consumption of finite element solution for each working condition in the performance prediction of water-lubricated bearings.

[0008] (2) Solve the problem that ordinary data-driven models ignore mesh topology in the performance prediction of water-lubricated bearings.

[0009] (3) Solve the problem of difficulty in predicting complex grooves, boundaries and local high pressure areas in the performance prediction process of water-lubricated bearings.

[0010] (4) Solve the problem of separation between pressure field and engineering performance index in the performance prediction process of water-lubricated bearings. Summary of the Invention

[0011] The purpose of this invention is to overcome the shortcomings of existing technologies by proposing a method and system for predicting the performance of water-lubricated bearings based on graph neural network finite element method. This method can quickly predict the water film pressure field, bearing capacity, friction force, friction coefficient, maximum water film pressure, minimum water film thickness, and lubrication state indicators of water-lubricated bearings, thereby reducing the calculation cost of traditional finite element analysis and improving the efficiency of bearing performance evaluation.

[0012] To achieve the above objectives, the present invention adopts the following technical solution:

[0013] In a first aspect, the present invention provides a method for predicting the performance of water-lubricated bearings based on graph neural network finite element method, comprising the following steps:

[0014] S1. The parameter acquisition device collects the geometric parameters, groove structure parameters, material parameters, and operating condition parameters of the water-lubricated bearing. The parameter acquisition device includes: an input terminal, a data storage device, sensors, and a data acquisition unit. The sensors include: a speed sensor, a load sensor, a displacement sensor, a temperature sensor, a pressure sensor, and a flow sensor. The data acquisition unit receives the signals output by each sensor, performs sampling, analog-to-digital conversion, and filtering, and then sends them to the computing device. The computing device is a workstation or server equipped with a processor, memory, graphics processor, and finite element analysis program.

[0015] S2. The computing device establishes a water film fluid domain using the finite element method based on the geometric parameters, groove structure parameters, material parameters, and operating condition parameters of the water-lubricated bearing. It then discretizes the water film fluid domain into a mesh along the circumferential and axial directions of the water film to construct a parametric finite element analysis model of the water-lubricated bearing, resulting in a finite element mesh. The finite element mesh includes: finite element nodes, finite element elements formed by connecting multiple finite element nodes in a specific order, and the topological connection relationships between finite element nodes. The topological connection relationships between finite element nodes are determined by the node number and node arrangement order, as well as the relative position, connection direction, and mesh distance between adjacent finite element nodes.

[0016] S3. The computing device sets the range of values ​​for the geometric parameters, groove structure parameters, material parameters, and operating condition parameters of the water-lubricated bearing. It generates multiple parameter combinations according to preset sampling rules and inputs each parameter combination into the parametric finite element analysis model. Based on the operating condition parameters and parameter combinations, the computing device applies loads and boundary conditions to the parametric finite element analysis model, performs finite element solutions, obtains the water film pressure values ​​and corresponding performance index data at each finite element node, and constructs a finite element simulation sample library for the water-lubricated bearing. The finite element simulation sample library includes: the water film pressure values ​​of each finite element node, the bearing capacity, friction force, friction coefficient, maximum water film pressure, minimum water film thickness, lubrication state index obtained from the finite element simulation samples, and the circumferential position, axial position, local film thickness, groove area identification, boundary state and corresponding operating condition parameters of each finite element node, as well as the relative position, connection direction and mesh distance between adjacent finite element nodes.

[0017] S4. The computing device maps each finite element node in the finite element mesh of the finite element simulation sample library to a graph node, and maps finite element nodes with direct connections in the finite element mesh to graph edges, thus obtaining the graph topology. The circumferential position, axial position, local film thickness, trench region identification, boundary state, and corresponding working parameters of the finite element nodes are used as graph node features, and the relative position, connection direction, and mesh distance between adjacent finite element nodes are used as graph edge features. The graph topology, graph node features, and graph edge features together constitute the graph data.

[0018] S5. The computing device constructs a graph dataset based on graph data and its corresponding finite element simulation samples. Each graph sample in the graph dataset includes: graph topology, graph node features, graph edge features, global operating condition features, finite element node-level labels, and graph-level labels. Among them, graph node features include: circumferential position, axial position, local film thickness, groove region identification, and boundary state of finite element nodes; graph edge features include: relative position, connection direction, and mesh distance between adjacent graph nodes; global operating condition features include geometric parameters of water-lubricated bearings, groove structure parameters, material parameters, and operating condition parameters; finite element node-level labels are determined based on the water film pressure values ​​of each node obtained from the finite element solution; graph-level labels are further calculated based on the finite element solution results to obtain: bearing capacity, friction force, friction coefficient, maximum water film pressure, minimum water film thickness, and lubrication state indicators.

[0019] S6. The computing device constructs a graph neural network proxy model based on the graph dataset. The graph topology, graph node features, graph edge features, and global operating condition features in the graph dataset are input into the graph neural network proxy model. The graph neural network proxy model learns the local correlation between adjacent graph nodes through multi-layer graph message passing. The water film pressure values ​​of each finite element node obtained from the finite element solution are used as finite element node-level supervision labels. The bearing capacity, friction force, friction coefficient, maximum water film pressure, minimum water film thickness, and lubrication state indicators obtained from the finite element simulation samples are used as graph-level supervision labels. The graph neural network proxy model is trained under supervision to learn the nonlinear mapping relationship between the finite element input features and the finite element output results, and the trained graph neural network proxy model is obtained, which is used to predict the performance of water-lubricated bearings.

[0020] S7. In engineering applications, the parameter acquisition device collects new operating condition parameters of the water-lubricated bearing under test. The calculation device obtains the finite element mesh corresponding to the new operating condition parameters according to the water film fluid domain discretization rule adopted in step S2. The finite element nodes in the finite element mesh are mapped to new graph nodes. New graph edges are constructed based on the direct connection relationship between the finite element nodes. Based on the new operating condition parameters, geometric parameters, groove structure parameters, and material parameters of the water-lubricated bearing under test, new graph node features, graph edge features, and global operating condition features are generated as input graph data corresponding to the new operating condition parameters. The input graph data corresponding to the new operating condition parameters is input into the trained graph neural network surrogate model. After one forward propagation, the water film pressure field, bearing capacity, friction force, friction coefficient, maximum water film pressure, minimum water film thickness, and lubrication state index of the multi-condition water-lubricated bearing are predicted. The result output device outputs the prediction results. The prediction results are used for multi-condition performance evaluation of water-lubricated bearings, structural scheme selection, operating parameter adjustment, and abnormal lubrication state identification. In the actual engineering of water-lubricated bearings, it provides application analysis and theoretical guidance for specific performance indicators.

[0021] Secondly, the present invention provides a water-lubricated bearing performance prediction system based on graph neural network finite element method for performing the method described in the first aspect, comprising a parameter acquisition device, a calculation device, and a result output device connected in sequence, wherein...

[0022] The parameter acquisition equipment is used to collect geometric parameters, groove structure parameters, material parameters, and operating condition parameters of water-lubricated bearings.

[0023] The parameter acquisition device includes: an input terminal, a data storage device, a sensor, and a data acquisition device. The sensor is connected to the data acquisition device, and both the data acquisition device and the input terminal are connected to the data storage device.

[0024] The input terminal is used for manual input of design parameters and preset operating parameters of water-lubricated bearings: journal radius, journal length, radial clearance, number of grooves, groove width, material elastic modulus, lubricating medium viscosity, preset speed and preset load;

[0025] Data storage devices are used to read pre-stored parameter files or database data: geometric parameters of water-lubricated bearings, groove structure parameters, material parameters, and historical operating condition parameters; the data storage devices can be local hard drives, network storage devices, or database servers;

[0026] The sensors include: a speed sensor, a load sensor, a displacement sensor, a temperature sensor, a pressure sensor, and a flow sensor. These sensors are used to collect actual operating parameters of the water-lubricated bearing: the speed sensor collects the journal speed, the load sensor collects the external load on the bearing, the displacement sensor collects the journal displacement and uses this to determine the journal eccentricity and offset angle, the temperature sensor collects the lubricating medium temperature, the pressure sensor collects the supply pressure, and the flow sensor collects the supply flow rate. These operating parameters are used to determine the motion boundary conditions and load boundary conditions during the finite element analysis of the water-lubricated bearing. Specifically, the journal speed is used to determine the running speed on the journal surface, the external load on the bearing is used to apply load constraints, and the journal displacement is used to calculate the journal eccentricity and offset angle, thereby obtaining the water film pressure distribution and bearing performance indicators under the corresponding operating conditions.

[0027] The data acquisition device is used to: receive signals output from various sensors, perform sampling, analog-to-digital conversion, and filtering, and then send them to the computing device;

[0028] The computing device is a workstation or server equipped with a processor, memory, graphics processor, and finite element analysis program. The computing device includes, in sequence, a data reading module, a finite element simulation sample library construction module, a graph data conversion module, a graph feature encoding module, and a graph neural network proxy module, wherein:

[0029] The data reading module is used to read the geometric parameters, groove structure parameters, material parameters, operating condition parameters, finite element mesh files, and finite element result files of water-lubricated bearings.

[0030] The finite element simulation sample library construction module is used to: establish a water film fluid domain using the finite element method based on the geometric parameters, groove structure parameters, material parameters, and operating condition parameters of the water-lubricated bearing; discretize the water film fluid domain along the circumferential and axial directions of the water film; construct a parametric finite element analysis model of the water-lubricated bearing; and obtain a finite element mesh. The finite element mesh includes: finite element nodes, finite element elements formed by connecting multiple finite element nodes in a specific order, and the topological connection relationships between finite element nodes. The topological connection relationships between finite element nodes are determined by the node number and node arrangement order; the relative position, connection direction, and mesh distance between adjacent finite element nodes are also defined; and the value ranges of the geometric parameters, groove structure parameters, material parameters, and operating condition parameters of the water-lubricated bearing are set. Multiple parameter combinations are generated according to preset sampling rules, and each parameter combination is input into the parametric finite element analysis model. Based on the operating condition parameters and parameter combinations, loads and boundary conditions are applied to the parametric finite element analysis model, and finite element solutions are performed to obtain the water film pressure value and corresponding performance index data at each finite element node. A finite element simulation sample library for water-lubricated bearings is constructed. The finite element simulation sample library includes: water film pressure value of each finite element node, bearing capacity, friction force, friction coefficient, maximum water film pressure, minimum water film thickness, lubrication state index obtained from the finite element simulation samples, as well as the circumferential position, axial position, local film thickness, groove area identification, boundary state and corresponding operating condition parameters of each finite element node, and the relative position, connection direction and mesh distance between adjacent finite element nodes.

[0031] The graph data conversion module is used to: map finite element nodes in the finite element mesh of the finite element simulation sample library to graph nodes, and map finite element nodes with direct connections in the finite element mesh to graph edges, thus obtaining a graph topology; and use the circumferential position, axial position, local film thickness, trench region identification, boundary state, and corresponding working parameters of the finite element nodes as graph node features, and the relative position, connection direction, and mesh distance between adjacent finite element nodes as graph edge features. The graph topology, graph node features, and graph edge features together constitute the graph data.

[0032] The graph feature encoding module is used to: construct a graph dataset based on graph data and its corresponding finite element simulation samples. Each graph sample in the graph dataset includes: graph topology, graph node features, graph edge features, global operating condition features, finite element node-level labels, and graph-level labels. Among them, graph node features include: circumferential position, axial position, local film thickness, groove region identification, and boundary state of finite element nodes; graph edge features include: relative position, connection direction, and mesh distance between adjacent graph nodes; global operating condition features include geometric parameters of water-lubricated bearings, groove structure parameters, material parameters, and operating condition parameters; finite element node-level labels are determined based on the water film pressure values ​​of each node obtained from the finite element solution; graph-level labels are further calculated based on the finite element solution results to obtain: bearing capacity, friction force, friction coefficient, maximum water film pressure, minimum water film thickness, and lubrication state indicators.

[0033] The graph neural network proxy module is used to: construct a graph neural network proxy model based on a graph dataset. The graph topology, node features, edge features, and global operating condition features from the graph dataset are input into the graph neural network proxy model. The graph neural network proxy model learns the local relationships between adjacent graph nodes through multi-layer graph message passing. The water film pressure values ​​of each finite element node obtained from the finite element method are used as finite element node-level supervision labels, and the bearing capacity, friction force, friction coefficient, maximum water film pressure, minimum water film thickness, and lubrication state indicators obtained from finite element simulation samples are used as graph-level supervision labels. This supervises the training of the graph neural network proxy model, learning the nonlinear mapping relationship between the finite element input features and the finite element output results, resulting in a trained graph neural network proxy model. In engineering applications, the trained graph neural network proxy model performs a forward propagation on the input graph data corresponding to new operating condition parameters to predict the water film pressure field, bearing capacity, friction force, friction coefficient, maximum water film pressure, minimum water film thickness, and lubrication state indicators of multi-condition water-lubricated bearings.

[0034] The output device is a display or industrial control terminal, which is used to output the predicted results of water film pressure field and bearing performance indicators. The prediction results are used for multi-condition performance evaluation of water-lubricated bearings, structural scheme selection, operation parameter adjustment and abnormal lubrication state identification, and provide application analysis and theoretical guidance for specific performance indicators in the actual engineering of water-lubricated bearings.

[0035] Thirdly, the present invention provides an electronic device comprising: at least one processor, at least one memory, and a communication interface, wherein the processor, memory, and communication interface communicate with each other; the memory stores program instructions executable by the processor, and the processor invokes the program instructions to execute the method described in the first aspect.

[0036] Fourthly, the present invention provides a non-transitory computer-readable storage medium storing computer instructions that cause a computer to perform the method described in the first aspect.

[0037] Compared with existing technical solutions, the beneficial effects of the embodiments of the present invention are reflected in the following aspects:

[0038] (1) The embodiments of the present invention can improve the performance prediction efficiency of water-lubricated bearings under multiple working conditions. The embodiments of the present invention establish a finite element simulation sample library containing different journal surface running linear velocities, bearing external loads, journal eccentricity, groove structure and material parameters of water-lubricated bearings through the finite element method. Each finite element simulation sample in the finite element simulation sample library is converted into graph data, and a graph neural network surrogate model is trained to learn the nonlinear mapping relationship between finite element input features and finite element output results. After the graph neural network surrogate model is trained, it is used in engineering applications to replace repeated finite element solutions. The graph neural network surrogate model can quickly obtain the pressure field and performance indicators through one forward propagation, achieving fast... This invention provides rapid prediction and analysis of the pressure field, load-bearing capacity, friction force, friction coefficient, maximum pressure, minimum film thickness, and lubrication state indicators of water-lubricated bearings. New operating conditions do not require a complete finite element analysis (FEM) solution. This embodiment effectively reduces the reliance on condition-by-condition FEM solutions, solving the problem of time-consuming condition-by-condition FEM solutions in existing water-lubricated bearing performance prediction processes. Furthermore, since new operating conditions do not require repeated FEM modeling and iterative solutions, it also reduces online analysis costs. It is suitable for rapid design screening, digital twin generation, and operational state prediction of water-lubricated bearing performance, providing specific data application analysis and theoretical guidance in practical engineering applications of water-lubricated bearings.

[0039] (2) In this embodiment of the invention, the finite element nodes are converted into a graph structure. The connection relationship of the finite element mesh is expressed through graph nodes and graph edges, so that the graph neural network model can learn the spatial coupling relationship and pressure field propagation law between adjacent nodes, retain the finite element mesh topology information, and solve the problem that the ordinary data-driven model ignores the mesh topology in the existing water-lubricated bearing performance prediction process.

[0040] (3) In this embodiment of the invention, local structural information in the finite element mesh is preserved through node features, edge features, and graph message passing mechanism. Node features and edge features express the groove region, boundary nodes, material parameters, and local film thickness variations, which can improve the prediction capability under complex groove structures in the performance prediction process of water-lubricated bearings. Furthermore, in this embodiment of the invention, the connection relationship of the finite element mesh is expressed through graph nodes and graph edges, enabling the graph neural network surrogate model to learn the local spatial coupling and pressure field propagation laws, which can adapt to the complex groove and boundary structure of water-lubricated bearings and solve the problem of difficulty in predicting complex grooves, boundaries, and local high-pressure areas in the existing performance prediction process of water-lubricated bearings.

[0041] (4) The embodiments of the present invention simultaneously learn the pressure field and post-processing performance indicators of water-lubricated bearings, forming a rapid prediction result that can be directly used to design and evaluate the performance of water-lubricated bearings, thus solving the problem of separation between pressure field and engineering performance indicators in the existing water-lubricated bearing performance prediction process.

[0042] (5) The graph neural network proxy model in the embodiment of the present invention simultaneously outputs the finite element mesh node-level pressure and graph-level performance index of the water-lubricated bearing, which can directly serve the design evaluation and parameter optimization of the water-lubricated bearing, realize the unified prediction of the pressure field and engineering performance of the water-lubricated bearing, and provide application analysis and theoretical guidance of specific performance indexes in the actual engineering of water-lubricated bearings.

[0043] (6) The embodiments of the present invention take finite element label supervised learning as the core, use graph neural network to map the traditional finite element mesh structure, and use graph neural network model to realize the rapid and accurate prediction of pressure field and performance of water-lubricated bearings. It can solve the engineering problems of long solution time of traditional finite element and data-driven model ignoring mesh topology. Attached Figure Description

[0044] Figure 1 This is a flowchart of the water-lubricated bearing performance prediction method based on graph neural network finite element method in an embodiment of the present invention.

[0045] Figure 2 This is an offline training flowchart of constructing a graph neural network proxy model based on finite element simulation samples in an embodiment of the present invention.

[0046] Figure 3 This is a schematic diagram of the finite element analysis model and mesh generation of the water-lubricated bearing in an embodiment of the present invention.

[0047] Figure 4 This is a schematic diagram of the mapping from finite element mesh to graph structure in an embodiment of the present invention.

[0048] Figure 5 This is a schematic diagram of the topological relationship of the finite element structure of the water-lubricated bearing in an embodiment of the present invention.

[0049] Figure 6 This is a schematic diagram of the construction of graph node features and graph edge features of a water-lubricated bearing in an embodiment of the present invention.

[0050] Figure 7 This is a schematic diagram illustrating the construction of the global operating condition characteristics of the water-lubricated bearing in an embodiment of the present invention.

[0051] Figure 8 This is a schematic diagram of the encoding of the input features of the graph neural network in an embodiment of the present invention.

[0052] Figure 9 This is a flowchart of the neighborhood message passing in a graph neural network embodiment of the present invention.

[0053] Figure 10 This is a flowchart illustrating the prediction of water-lubricated bearing performance using a trained graph neural network surrogate model in an embodiment of the present invention.

[0054] Figure 11 This is a flowchart of online prediction of water-lubricated bearing performance based on a graph neural network surrogate model in an embodiment of the present invention.

[0055] Figure 12 This is a structural block diagram of the graph neural network proxy model in an embodiment of the present invention.

[0056] Figure 13 This is a structural block diagram of the water-lubricated bearing performance prediction system based on graph neural network finite element method in an embodiment of the present invention.

[0057] Figure 14 This is a structural block diagram of the parameter acquisition device in an embodiment of the present invention.

[0058] Figure 15 This is a structural block diagram of the computing device in an embodiment of the present invention.

[0059] Figure 16 This is a graph showing the prediction loss using a graph neural network in an embodiment of the present invention.

[0060] Figure 17 This is a water film pressure cloud map obtained by finite element solution in an embodiment of the present invention.

[0061] Figure 18 This is a water film pressure cloud map predicted by the graph neural network surrogate model in an embodiment of the present invention.

[0062] Figure 19 This is a pressure error cloud map obtained from finite element analysis and graph neural network prediction in an embodiment of the present invention.

[0063] Figure 20 This is a comparison diagram of the mid-section pressure in an embodiment of the present invention. Detailed Implementation

[0064] Example 1

[0065] See Figure 1 As shown, this embodiment of the invention provides a method for predicting the performance of water-lubricated bearings based on graph neural network finite element method, including the following steps:

[0066] S1. The parameter acquisition device collects the geometric parameters, groove structure parameters, material parameters, and operating condition parameters of the water-lubricated bearing. The parameter acquisition device includes: an input terminal, a data storage device, sensors, and a data acquisition unit. The sensors include: a speed sensor, a load sensor, a displacement sensor, a temperature sensor, a pressure sensor, and a flow sensor. The data acquisition unit receives the signals output by each sensor, performs sampling, analog-to-digital conversion, and filtering, and then sends them to the computing device. The computing device is a workstation or server equipped with a processor, memory, graphics processor, and finite element analysis program.

[0067] S2, see also Figure 2 As shown, the computing device establishes a water film fluid domain using the finite element method based on the geometric parameters, groove structure parameters, material parameters, and operating condition parameters of the water-lubricated bearing. The water film fluid domain is then discretized into a mesh along the circumferential and axial directions to construct a parametric finite element analysis model of the water-lubricated bearing. (See [link to relevant documentation]). Figure 3 As shown, a finite element mesh is obtained. The finite element mesh includes: finite element nodes, finite element elements formed by connecting multiple finite element nodes in a specific order, and the topological connection relationship between finite element nodes. The topological connection relationship between finite element nodes is determined by the number of each finite element node and the node arrangement order: the relative position, connection direction and mesh distance between adjacent finite element nodes.

[0068] S3. The computing device sets the range of values ​​for the geometric parameters, groove structure parameters, material parameters, and operating condition parameters of the water-lubricated bearing. It generates multiple parameter combinations according to preset sampling rules and inputs each parameter combination into the parametric finite element analysis model. Based on the operating condition parameters and parameter combinations, the computing device applies loads and boundary conditions to the parametric finite element analysis model, performs finite element solutions, obtains the water film pressure values ​​and corresponding performance index data at each finite element node, and constructs a finite element simulation sample library for the water-lubricated bearing. The finite element simulation sample library includes: the water film pressure values ​​of each finite element node, the bearing capacity, friction force, friction coefficient, maximum water film pressure, minimum water film thickness, lubrication state index obtained from the finite element simulation samples, and the circumferential position, axial position, local film thickness, groove area identification, boundary state and corresponding operating condition parameters of each finite element node, as well as the relative position, connection direction and mesh distance between adjacent finite element nodes.

[0069] S4, see also Figure 4As shown, the computing device maps each finite element node in the finite element mesh of the finite element simulation sample library to a graph node, and maps finite element nodes with direct connections in the finite element mesh to graph edges, thus obtaining a graph topology. The circumferential position, axial position, local film thickness, trench region identification, boundary state, and corresponding working parameters of the finite element nodes are used as graph node features, and the relative position, connection direction, and mesh distance between adjacent finite element nodes are used as graph edge features. The graph topology, graph node features, and graph edge features together constitute the graph data.

[0070] S5, see also Figure 5 As shown, the computing device constructs a graph dataset based on graph data and its corresponding finite element simulation samples. Each graph sample in the graph dataset includes: graph topology, graph node features, graph edge features, global load case features, finite element node-level labels, and graph-level labels. See [link to relevant documentation]. Figure 6 As shown, the graph node features include: circumferential position, axial position, local film thickness, trench region identification, and boundary status of the finite element node. Graph edge features include: relative position between adjacent graph nodes, connection direction, and mesh distance. (See [link to relevant documentation]). Figure 7 As shown, the global operating condition characteristics include the geometric parameters of the water-lubricated bearing, the groove structure parameters, the material parameters, and the operating condition parameters; the finite element node-level labels are determined based on the water film pressure values ​​of each node obtained from the finite element solution, and the graph-level labels are further calculated based on the finite element solution results to obtain: bearing capacity, friction force, friction coefficient, maximum water film pressure, minimum water film thickness, and lubrication status indicators;

[0071] S6, see also Figure 8 As shown, the computing device constructs a graph neural network proxy model based on a graph dataset. The graph topology, graph node features, graph edge features, and global operating condition features from the graph dataset are input into the graph neural network proxy model. (See [link to documentation]). Figure 9 As shown, the graph neural network surrogate model learns the local relationships between adjacent graph nodes through multi-layer graph message passing. The water film pressure values ​​of each finite element node obtained from the finite element method are used as finite element node-level supervision labels, while the bearing capacity, friction force, friction coefficient, maximum water film pressure, minimum water film thickness, and lubrication state indicators obtained from finite element simulation samples are used as graph-level supervision labels. This supervises the training of the graph neural network surrogate model, learning the nonlinear mapping relationship between finite element input features and finite element output results, thus obtaining the trained graph neural network surrogate model. (See [link to documentation]). Figure 10 As shown, this is used to predict the performance of water-lubricated bearings;

[0072] S7, see also Figure 11As shown, in engineering applications, the parameter acquisition device collects new operating condition parameters of the water-lubricated bearing under test. The calculation device obtains the finite element mesh corresponding to the new operating condition parameters according to the water film fluid domain discretization rule adopted in step S2. The finite element nodes in the finite element mesh are mapped to new graph nodes. New graph edges are constructed based on the direct connection relationship between the finite element nodes. Based on the new operating condition parameters, geometric parameters, groove structure parameters, and material parameters of the water-lubricated bearing under test, new graph node features, graph edge features, and global operating condition features are generated as input graph data corresponding to the new operating condition parameters. The input graph data corresponding to the new operating condition parameters is input into the trained graph neural network surrogate model. After one forward propagation, the water film pressure field, bearing capacity, friction force, friction coefficient, maximum water film pressure, minimum water film thickness, and lubrication state index of the multi-condition water-lubricated bearing are predicted. The result output device outputs the prediction results. The prediction results are used for multi-condition performance evaluation of water-lubricated bearings, structural scheme selection, operating parameter adjustment, and abnormal lubrication state identification. It provides application analysis and theoretical guidance for specific performance indicators in the actual engineering of water-lubricated bearings.

[0073] In step S1, the geometric parameters of the water-lubricated bearing include: journal radius, bearing inner radius, journal length, radial clearance, length-to-diameter ratio, and bearing wrap angle; the groove structure parameters include the number of grooves, groove type, groove circumferential position, groove axial position, groove width, groove depth, and groove length; the material parameters include the elastic modulus, Poisson's ratio, material density, surface roughness, and liner thickness of the bearing liner and journal materials; the operating condition parameters include: journal speed, journal surface running linear velocity calculated based on journal speed and journal radius, bearing external load, journal eccentricity, misalignment angle, fluid supply pressure, fluid supply flow rate, dynamic viscosity, density, temperature, and pressure parameters of the lubricating medium. The operating condition parameters are used to determine the motion boundary conditions and load boundary conditions in the finite element analysis process of the water-lubricated bearing. Among them, the journal speed is used to determine the journal surface running speed, the bearing external load is used to apply load constraints, and the journal displacement is used to calculate the journal eccentricity and misalignment angle, thereby obtaining the water film pressure distribution and bearing performance indicators under the corresponding operating conditions.

[0074] Input terminals or data storage devices collect geometric parameters, groove structure parameters, and material parameters of water-lubricated bearings.

[0075] Sensors collect actual operating parameters of water-lubricated bearings: speed sensor collects journal speed, load sensor collects bearing external load, displacement sensor collects journal displacement and determines journal eccentricity and offset angle accordingly, temperature sensor collects lubricating medium temperature, pressure sensor collects supply pressure, and flow sensor collects supply flow rate.

[0076] The input terminal is used for manual input of design parameters and preset operating parameters of water-lubricated bearings: journal radius, journal length, radial clearance, number of grooves, groove width, material elastic modulus, lubricating medium viscosity, preset speed and preset load.

[0077] Data storage devices are used to read pre-stored parameter files or database data: geometric parameters of water-lubricated bearings, groove structure parameters, material parameters, and historical operating condition parameters. The data storage devices can be local hard drives, network storage devices, or database servers.

[0078] Operating condition parameters are used to determine the motion boundary conditions and load boundary conditions in the finite element analysis process of water-lubricated bearings. Among them, the journal speed is used to determine the running speed of the journal surface, the bearing external load is used to apply load constraints, and the journal displacement is used to calculate the journal eccentricity and offset angle, thereby obtaining the water film pressure distribution and bearing performance indicators under the corresponding operating conditions.

[0079] See Figure 12 As shown, the graph neural network proxy model includes: node feature encoder, edge feature encoder, global working condition feature encoder, multi-layer graph message passing module, graph-level pooling module, graph node-level stress field decoder, and graph-level performance index decoder.

[0080] The node feature encoder is used to encode the features of finite element nodes, mapping the node attributes of finite element nodes—circumferential position, axial position, local film thickness, groove region identifier, and boundary state—to graph node implicit feature representations.

[0081] The edge feature encoder is used to encode the connection relationships of finite element nodes: the relative positions, connection directions and mesh distances between adjacent finite element nodes, and map them into graph edge latent feature representations.

[0082] The global operating condition feature encoder is used to encode the global operating condition features of water-lubricated bearings to obtain a global operating condition implicit feature representation that characterizes the overall operating condition state.

[0083] The multi-layer graph message passing module is used to: iteratively update the implicit feature representations of graph nodes, graph edges, and global working conditions based on the graph topology; aggregate neighborhood node information through a multi-layer message passing mechanism; learn the spatial topological relationships and local coupling relationships in the finite element mesh; and obtain the updated implicit feature representations of graph nodes.

[0084] The graph-level pooling module is used to aggregate the updated implicit feature representations of graph nodes to obtain a graph-level feature representation that characterizes the overall state of the water-lubricated bearing.

[0085] The graph node-level pressure field decoder is used to: predict the water film pressure value corresponding to each graph node based on the updated implicit features of the graph nodes. The graph nodes are obtained by mapping one-to-one from the finite element mesh nodes. The graph node pressure prediction result corresponds to the water film pressure value at the finite element node position, thereby realizing the prediction of the water film pressure field.

[0086] The graph-level performance index decoder is used to predict the overall performance index of water-lubricated bearings based on graph-level feature representations, including load capacity, friction force, friction coefficient, maximum water film pressure, minimum water film thickness, and lubrication status indexes.

[0087] The node feature encoder, edge feature encoder, and global working condition feature encoder all establish data flow connections with the multi-layer graph message passing module. The multi-layer graph message passing module establishes data flow connections with the graph-level pooling module. The graph-level pooling module establishes data flow connections with the graph node-level pressure field decoder and the graph-level performance index decoder, respectively.

[0088] In step S6, supervised training of the graph neural network surrogate model is performed, including the following steps:

[0089] The water film pressure values ​​of each finite element node obtained from the finite element method are used as finite element node-level supervision labels, and the bearing capacity, friction force, friction coefficient, maximum water film pressure, minimum water film thickness, and lubrication state indices obtained from the finite element simulation samples are used as graph-level supervision labels. The graph neural network surrogate model is trained under supervision. The loss function of the supervision training is composed of graph node-level pressure field supervision loss and graph-level performance index supervision loss. The graph node-level pressure field supervision loss is used to measure the error between the water film pressure prediction value output by the graph node-level pressure field decoder and the finite element node water film pressure value, and the graph-level performance index supervision loss is used to measure the error between the performance index prediction value output by the graph-level performance index decoder and the corresponding performance index calculated by the finite element method.

[0090] The graph edges are established based on the adjacency relationship of the finite element mesh, the shared edge relationship of the elements, the shared element relationship of the nodes, or the adjacency relationship of the regular mesh.

[0091] The relative positions and connections between finite element nodes are used to construct edge features, while operating condition parameters simultaneously constitute global operating condition features. The graph topology, node features, edge features, and global operating condition features together form the input graph of the graph neural network surrogate model, used to predict node-level water film pressure fields and graph-level bearing performance indicators. Finite element simulation samples are not used as model inputs. The water film pressure values ​​of each finite element node obtained from the finite element solution are used as finite element node-level supervision labels, while the bearing capacity, friction force, friction coefficient, maximum water film pressure, minimum water film thickness, and lubrication state indicators obtained from the finite element simulation samples are used as graph-level supervision labels to supervise the training of the graph neural network surrogate model, calculate the training loss, and update the model parameters.

[0092] For the new operating conditions, it is only necessary to establish a finite element mesh according to step S2, obtain geometric parameters, groove structure parameters, material parameters and operating condition parameters, and construct the input graph according to step S4. Without resolving the finite element problem, the trained graph neural network surrogate model can be used to obtain the prediction results of water film pressure field and bearing performance indicators. This can solve the problem that the traditional finite element method requires repeated solutions and has a long calculation time under multiple operating conditions, making it difficult to meet the requirements of rapid performance evaluation and online status judgment of water-lubricated bearings under multiple operating conditions.

[0093] This invention establishes a sample library containing the running linear velocity of different journal surfaces, external loads on the bearing, journal eccentricity, groove structure, and material parameters of water-lubricated bearings using the finite element method. Each finite element simulation sample is converted into graph data, and a graph neural network surrogate model is trained to learn the nonlinear mapping relationship between the finite element input features and the finite element output results. A GNN (Graph Neural Network) is a deep learning algorithm specifically designed to process graph-structured data, capable of extracting features from nodes, edges, and the overall graph to perform tasks such as classification and prediction. In engineering applications, the trained graph neural network surrogate model replaces repetitive finite element solutions, enabling rapid prediction and analysis of the pressure field, load-bearing capacity, friction force, friction coefficient, maximum pressure, minimum film thickness, and lubrication state indicators of water-lubricated bearings. Using finite element simulation samples as the training benchmark and the graph neural network surrogate model as the surrogate solver, after the graph neural network surrogate model is trained, for new operating parameters of the water-lubricated bearing, only graph data construction and one forward propagation are needed to output the pressure field and performance indicators.

[0094] See Figure 13 As shown, this embodiment of the invention also provides a water-lubricated bearing performance prediction system based on graph neural network finite element method for performing the above method, including a parameter acquisition device, a calculation device, and a result output device connected in sequence.

[0095] The parameter acquisition equipment is used to collect geometric parameters, groove structure parameters, material parameters, and operating condition parameters of water-lubricated bearings. (See also...) Figure 14 As shown, the parameter acquisition device includes: an input terminal, a data storage device, a sensor, and a data acquisition unit. The sensor is connected to the data acquisition unit, and both the data acquisition unit and the input terminal are connected to the data storage device.

[0096] The input terminal is used for manual input of design parameters and preset operating parameters of water-lubricated bearings: journal radius, journal length, radial clearance, number of grooves, groove width, material elastic modulus, lubricating medium viscosity, preset speed and preset load;

[0097] Data storage devices are used to read pre-stored parameter files or database data: geometric parameters of water-lubricated bearings, groove structure parameters, material parameters, and historical operating condition parameters; the data storage devices can be local hard drives, network storage devices, or database servers;

[0098] The sensors include: a speed sensor, a load sensor, a displacement sensor, a temperature sensor, a pressure sensor, and a flow sensor. These sensors are used to collect actual operating parameters of the water-lubricated bearing: the speed sensor collects the journal speed, the load sensor collects the external load on the bearing, the displacement sensor collects the journal displacement and uses this to determine the journal eccentricity and offset angle, the temperature sensor collects the lubricating medium temperature, the pressure sensor collects the supply pressure, and the flow sensor collects the supply flow rate. These operating parameters are used to determine the motion boundary conditions and load boundary conditions during the finite element analysis of the water-lubricated bearing. Specifically, the journal speed is used to determine the running speed on the journal surface, the external load on the bearing is used to apply load constraints, and the journal displacement is used to calculate the journal eccentricity and offset angle, thereby obtaining the water film pressure distribution and bearing performance indicators under the corresponding operating conditions.

[0099] The data acquisition device is used to: receive signals output from various sensors, perform sampling, analog-to-digital conversion, and filtering, and then send them to the computing device;

[0100] The computing device is a workstation or server equipped with a processor, memory, graphics processor, and finite element analysis programs. See [link to documentation]. Figure 15 As shown, the computing device includes a data reading module, a finite element simulation sample library construction module, a graph data conversion module, a graph feature encoding module, and a graph neural network proxy module connected in sequence, wherein:

[0101] The data reading module is used to read the geometric parameters, groove structure parameters, material parameters, operating condition parameters, finite element mesh files, and finite element result files of water-lubricated bearings.

[0102] The finite element simulation sample library construction module is used to: establish a water film fluid domain using the finite element method based on the geometric parameters, groove structure parameters, material parameters, and operating condition parameters of the water-lubricated bearing; discretize the water film fluid domain along the circumferential and axial directions of the water film; construct a parametric finite element analysis model of the water-lubricated bearing; and obtain a finite element mesh. The finite element mesh includes: finite element nodes, finite element elements formed by connecting multiple finite element nodes in a specific order, and the topological connection relationships between finite element nodes. The topological connection relationships between finite element nodes are determined by the node number and node arrangement order; the relative position, connection direction, and mesh distance between adjacent finite element nodes are also defined; and the value ranges of the geometric parameters, groove structure parameters, material parameters, and operating condition parameters of the water-lubricated bearing are set. Multiple parameter combinations are generated according to preset sampling rules, and each parameter combination is input into the parametric finite element analysis model. Based on the operating condition parameters and parameter combinations, loads and boundary conditions are applied to the parametric finite element analysis model, and finite element solutions are performed to obtain the water film pressure value and corresponding performance index data at each finite element node. A finite element simulation sample library for water-lubricated bearings is constructed. The finite element simulation sample library includes: water film pressure value of each finite element node, bearing capacity, friction force, friction coefficient, maximum water film pressure, minimum water film thickness, lubrication state index obtained from the finite element simulation samples, as well as the circumferential position, axial position, local film thickness, groove area identification, boundary state and corresponding operating condition parameters of each finite element node, and the relative position, connection direction and mesh distance between adjacent finite element nodes.

[0103] The graph data conversion module is used to: map finite element nodes in the finite element mesh of the finite element simulation sample library to graph nodes, and map finite element nodes with direct connections in the finite element mesh to graph edges, thus obtaining a graph topology; and use the circumferential position, axial position, local film thickness, trench region identification, boundary state, and corresponding working parameters of the finite element nodes as graph node features, and the relative position, connection direction, and mesh distance between adjacent finite element nodes as graph edge features. The graph topology, graph node features, and graph edge features together constitute the graph data.

[0104] The graph feature encoding module is used to: construct a graph dataset based on graph data and its corresponding finite element simulation samples. Each graph sample in the graph dataset includes: graph topology, graph node features, graph edge features, global operating condition features, finite element node-level labels, and graph-level labels. Among them, graph node features include: circumferential position, axial position, local film thickness, groove region identification, and boundary state of finite element nodes; graph edge features include: relative position, connection direction, and mesh distance between adjacent graph nodes; global operating condition features include geometric parameters of water-lubricated bearings, groove structure parameters, material parameters, and operating condition parameters; finite element node-level labels are determined based on the water film pressure values ​​of each node obtained from the finite element solution; graph-level labels are further calculated based on the finite element solution results to obtain: bearing capacity, friction force, friction coefficient, maximum water film pressure, minimum water film thickness, and lubrication state indicators.

[0105] The graph neural network proxy module is used to: construct a graph neural network proxy model based on a graph dataset. The graph topology, node features, edge features, and global operating condition features from the graph dataset are input into the graph neural network proxy model. The graph neural network proxy model learns the local relationships between adjacent graph nodes through multi-layer graph message passing. The water film pressure values ​​of each finite element node obtained from the finite element method are used as finite element node-level supervision labels, and the bearing capacity, friction force, friction coefficient, maximum water film pressure, minimum water film thickness, and lubrication state indicators obtained from finite element simulation samples are used as graph-level supervision labels. This supervises the training of the graph neural network proxy model, learning the nonlinear mapping relationship between the finite element input features and the finite element output results, resulting in a trained graph neural network proxy model. In engineering applications, the trained graph neural network proxy model performs a forward propagation on the input graph data corresponding to new operating condition parameters to predict the water film pressure field, bearing capacity, friction force, friction coefficient, maximum water film pressure, minimum water film thickness, and lubrication state indicators of multi-condition water-lubricated bearings.

[0106] The output device is a display or industrial control terminal, which is used to output the predicted results of water film pressure field and bearing performance indicators. The prediction results are used for multi-condition performance evaluation of water-lubricated bearings, structural scheme selection, operation parameter adjustment and abnormal lubrication state identification, and provide application analysis and theoretical guidance for specific performance indicators in the actual engineering of water-lubricated bearings.

[0107] When the graph neural network proxy module performs supervised training on the graph neural network proxy model, the loss function of supervised training is composed of graph node-level pressure field supervision loss and graph-level performance index supervision loss. The graph node-level pressure field supervision loss is used to measure the error between the water film pressure prediction value output by the graph node-level pressure field decoder and the water film pressure value of the finite element node. The graph-level performance index supervision loss is used to measure the error between the performance index prediction value output by the graph-level performance index decoder and the corresponding performance index calculated by the finite element.

[0108] Example 2

[0109] This invention provides a method for predicting the performance of water-lubricated bearings based on graph neural network finite element method, comprising the following steps:

[0110] See Figure 2 As shown, firstly, geometric parameters, groove structure parameters, material parameters, and operating condition parameters of the water-lubricated bearing are collected. The geometric and operating condition parameters are used to determine the spatial location, local film thickness, and boundary conditions of each node in the diagram. The groove structure and material parameters are used to determine the groove region identifier, local structural properties, and material properties. Based on these parameters, a finite element analysis model of the water-lubricated bearing is established. (See [reference needed]). Figure 3 As shown, the system inputs information such as bearing radius, axial length, radial clearance, groove shape, groove depth, groove width, liner material parameters, dynamic viscosity of lubricating medium, running linear velocity of journal surface, external load on bearing, journal eccentricity, and boundary conditions of water-lubricated bearings. The system then solves the problem using the finite element method to obtain results such as pressure field, nodal pressure value, element pressure value, maximum pressure, load-bearing capacity, friction force, friction coefficient, and minimum film thickness under different working conditions.

[0111] Finite element simulation sample dataset for multi-condition finite element simulation ,in, This is a finite element simulation sample dataset containing finite element simulation samples under multiple working conditions. The total number of working condition samples in the dataset. For sample number, It is a positive integer. For the first Each working condition is input as a graph structure derived from a finite element mesh, including: nodes, edges, node features, edge features, and working condition features. For the first The finite element solution labels corresponding to each working condition are used as the target values ​​for supervised learning of GNN (Graph Neural Network). This represents a complete set of "graph input-FEM label" training sample pairs; ,in, For the first Labels of nodal or element pressure fields obtained by the finite element method under various working conditions. These are the engineering performance indicators obtained through post-processing under the corresponding working conditions. Overall, each finite element working condition sample is organized into a graphical input. With tags The training pairs formed, all These samples together constitute the training dataset. It is used to train graph neural network agent models.

[0112] Finite element calculation data of water-lubricated bearings have natural graph structure characteristics, with clear topological connections between finite element nodes, element centers, and adjacent elements. The changes in the pressure field on the mesh are closely related to local adjacency relationships, groove transition relationships, and boundary regions. Introducing the topological relationships of the finite element mesh into the graph neural network is beneficial for learning the spatial propagation law in the finite element pressure field, the distribution characteristics of local high-pressure areas, and the influence of operating parameters on the overall performance indicators.

[0113] To improve data quality, the finite element calculation results for water-lubricated bearings undergo unified coordinate mapping, node sorting, outlier removal, and normalization. For samples with different mesh densities or different groove structures, a unified interpolation mesh, local subgraph partitioning, or variable-size graph input based on the finite element connection matrix can be used.

[0114] See Figure 4 As shown, the conversion of the finite element mesh of a water-lubricated bearing to a graphical structure includes the following steps:

[0115] In the finite element mesh of a water-lubricated bearing, nodes, control volume centers, or element centers are defined as graph nodes. If the k-th sample contains N finite element nodes, and these N finite element nodes correspond to a total of N graph nodes, then a node set V = {V1, V2, ..., V} is constructed. i}, where V represents the set of graph nodes, i is the number of the finite element node, and N is a positive integer representing the total number of graph nodes. Let represent the i-th graph node, where i = 1, 2, ..., N. In this embodiment, each graph node... Each node corresponds one-to-one with a finite element node in the fluid domain of a water-lubricated bearing. After the finite element node is transformed, each graphical node retains its spatial location, local film thickness, groove region, boundary state, material properties, and corresponding operating parameters.

[0116] See Figure 5 As shown, graph edges are established based on the finite element mesh topology of the water-lubricated bearing. If two finite element nodes share an element edge, two elements share a common edge, or they are adjacent in the finite element mesh, for regular circumferential-axial meshes, circumferential adjacency, axial adjacency, and necessary diagonal adjacency can be used. The graph structure is used to express the spatial adjacency relationships in the finite element mesh. For example, the graph structure expression is: ,in, For the first Each working condition is input as a graph structure derived from a finite element mesh, including: nodes, edges, node features, edge features, and working condition features. For a set of graph nodes, For the graph edge set, The node feature matrix, The edge feature matrix, For the first The first finite element simulation sample or the first The global working condition feature vector for each working condition.

[0117] Extracting the nodal features, edge features, and operating condition features of finite element nodes for water-lubricated bearings includes the following steps:

[0118] See Figure 6 As shown, the finite element node features of a water-lubricated bearing include: circumferential coordinate θ, axial coordinate z, sinθ, cosθ, local film thickness h, normalized film thickness, groove identifier, boundary identifier, material parameters, journal surface running linear velocity, bearing external load, journal eccentricity, dynamic viscosity of lubricating medium, and geometric parameters, which are used as input descriptions for the finite element node data of the water-lubricated bearing.

[0119] For example, ,in, For the first The original input feature vectors of the graph nodes corresponding to the finite element nodes, where graph node Vi corresponds to the i-th finite element node. For water-lubricated bearings, use the finite element node numbering. It is a positive integer; For the first The circumferential coordinates of each finite element node represent the position of the finite element node in the circumferential direction of the bearing. The axial coordinates of the finite element node represent its position along the bearing length direction; Encoding the circumferential coordinates periodically can eliminate the discontinuity of the circumferential periodic boundary and improve the model's ability to express the circumferential periodic pressure distribution. The local water film thickness at the finite element node is a core parameter affecting the pressure field, bearing capacity, and friction performance. To identify the trench area and distinguish the trench area, load-bearing area and trench inlet / outlet area where the finite element node is located; Use boundary markers to determine whether a finite element node is located at an axial end, a periodic boundary, an inlet / outlet boundary, or an internal region. These are characteristic parameters of the material and liner, characterizing local material properties and mechanical properties; The linear velocity of the journal surface. External load applied to the bearing For journal eccentricity, For the dynamic viscosity of the lubricating medium, For the bearing radius, axial length Radial clearance Together, they constitute the bearing's operating conditions and basic geometric parameters.

[0120] See Figure 6 As shown, It belongs to the local features of the finite element nodes and is used to describe the first... The spatial and structural properties of each finite element node.

[0121] See Figure 7 As shown, If it belongs to the global operating condition or geometric parameter, it will be broadcast to all finite element nodes of the water-lubricated bearing, so that each finite element node of the water-lubricated bearing can perceive the operating conditions corresponding to the current sample.

[0122] The edge features of finite element nodes are used to help graph neural networks distinguish different spatial connections. The edge features of finite element nodes include: relative circumferential distance, relative axial distance, Euclidean distance between adjacent finite element nodes, whether they cross the trench boundary, whether they are located at the end boundary, the membrane thickness difference between adjacent nodes, and the mesh connection type.

[0123] For example, the edge features of a finite element node are represented as follows: ,in, Finite element nodes With finite element nodes Edge features between nodes are used to describe finite element nodes. With finite element nodes The geometric relationships, membrane thickness variations, regional transitions, and connection types between these elements help Graph Surrogate (graph neural network proxy model) accurately learn the local variation patterns of the pressure field during message passing; , These are all finite element node numbers for water-lubricated bearings. , All are positive integers; Representing finite element nodes With finite element nodes The relative circumferential distance between them Indicates relative axial distance. This represents the physical spatial distance between two finite element nodes. Represents the local film thickness difference, This is a zone crossing identifier used to mark whether a section crosses a trench, material, or structural partition. This is an edge connection type used to distinguish finite element mesh adjacency relationships such as circumferential, axial, and diagonal connections.

[0124] See Figure 8 As shown, constructing a graph neural network surrogate model for a water-lubricated bearing includes the following steps:

[0125] This invention constructs a graph neural network surrogate model for learning finite element metadata of water-lubricated bearings, denoted as Graph Surrogate. Instead of directly inputting operating parameters into a conventional fully connected network, the graph neural network surrogate model uses the finite element mesh topology of the water-lubricated bearing as a basis, organizing the finite element nodes, element centers, or control volume centers of the water-lubricated bearing into a graph structure. It learns the spatial propagation law of the pressure field in the local mesh neighborhood through graph message passing. Since the pressure field of a water-lubricated bearing is influenced by the groove structure, local film thickness, boundary position, and operating parameters, its distribution exhibits significant local correlation and non-uniformity. Using a graph neural network can better preserve the spatial connectivity relationships in the finite element metadata than a conventional MLP (Multilayer Perceptron).

[0126] See Figure 8 As shown, the encoding of nodes, edges, and working conditions includes the following steps:

[0127] The graph neural network surrogate model first encodes node features, edge features, and operating condition features respectively. The node encoder is used to map the low-dimensional description of each finite element node or element to a high-dimensional latent space; the edge encoder is used to map the mesh connection relationship between adjacent finite element nodes to edge latent vectors; and the operating condition encoder is used to characterize the global operating conditions of the current sample.

[0128] The graph neural network surrogate model first completes the encoding processing of node features, edge features, and global operating condition features respectively. The specific mapping rule is as follows: each finite element node in the finite element mesh of the water-lubricated bearing is mapped one-to-one to a graph node, and the direct mesh connection relationship between finite element nodes is mapped to a graph edge.

[0129] For example, ,in, Represents graph nodes The initial latent space representation after node encoding, i.e., the node feature representation of layer 0 of the graph neural network. The node feature encoding function consists of a linear layer, a multilayer perceptron, or a normalized layer, and is used to map the original node features to a high-dimensional latent space. Represents graph nodes The original input feature vector, graph nodes Each element corresponds one-to-one with the i-th finite element node. Represents graph edges ( , The edge latent vectors obtained after edge encoding are used for subsequent message passing and neighborhood feature aggregation, helping the model distinguish different edge types such as circumferential adjacency, axial adjacency, cross-groove connection, and shared unit connection. This is the edge feature encoding function, used to map the original edge features into a high-dimensional implicit representation of the edge; Represents graph nodes With graph nodes The original edge feature vectors between them; the graph edges of the graph neural network are constructed from the mesh connection relationships between the corresponding finite element nodes. For the first The global latent vector corresponding to each working condition sample is broadcast and fused into all graph nodes and edges corresponding to that working condition sample, so that the graph neural network surrogate model can adapt to the water-lubricated bearing performance prediction task under different speeds, loads, eccentricities, groove structures and material parameters. This is a working condition feature encoding function used to map global working condition parameters to a unified working condition latent space.

[0130] See Figure 9 As shown, the message passing in the diagram includes the following steps:

[0131] The core of the graph neural network agent model is the multi-layer graph message passing module. In the l-th layer graph message passing process, the central graph node... Receive all adjacent nodes in its neighborhood set N(i) The transmitted information, while integrating edge hidden features and global working condition hidden features, completes the iterative update of the hidden state of its own node. Graph Node Each node corresponds one-to-one with the i-th finite element node in the finite element mesh, and the graph edges... It is constructed based on the direct mesh topology connection relationship between finite element nodes.

[0132] For example, in the k-th working condition sample, the mathematical expression for message generation at layer l is: ,in, Indicates the first The layer consists of neighbor graph nodes Transmitted to the central node The news, For message generation functions, For the central graph node No. Hidden layer state, For neighboring nodes No. Hidden layer state, It is the edge of the graph The implicit feature vectors are used to characterize the mesh topology, relative spatial coordinates, and local geometric associations between adjacent finite element nodes. In the graph structure corresponding to the k-th working case sample, the graph nodes are connected. With graph nodes The graph edges are derived from the finite element mesh topology. For the first The global latent vector corresponding to each working condition sample is broadcast and fused into all graph nodes and edges corresponding to that working condition sample, enabling the graph neural network surrogate model to adapt to the task of predicting the performance of water-lubricated bearings under different speeds, loads, eccentricities, groove structures and material parameters.

[0133] In the k-th set of work condition samples, the mathematical expression for solving the neighborhood attention weights of the l-th layer is: ,in, Represents graph nodes For graph nodes Learnable weights; This means normalizing the attention scores of all neighboring graph nodes Vj of the central graph node Vi, so that the sum of the weights of all neighbors is 1. This is the neighborhood weight learning function.

[0134] In the k-th working condition sample, the mathematical expression for the weighted message aggregation of the l-th layer is: ,in, The neighborhood composite message obtained by weighted aggregation of the central graph node Vi based on attention weights. Represents graph nodes In the The hidden representation after the layer update, This is the function for updating the state of graph nodes.

[0135] In the k-th set of working condition samples, the mathematical expression for the hidden state update of the node in the l-th layer is: .

[0136] To improve the training stability of deep networks, residual connections, normalization, and non-linear activation operations can be added to the message-passing layer. ,in, This represents a normalization function, including layer normalization functions and batch normalization functions, used to normalize the hidden states of nodes after residual connections.

[0137] This structure can gradually integrate neighborhood information while preserving the original node features, enabling the model to learn the spatial distribution law of the finite element pressure field in the local mesh. The graph message passing module can adopt one or more of GCN (Graph Convolutional Network), GAT (Graph Attention Network), GraphSAGE (Graph Sample and Aggregate), GIN (Graph Isomorphism Network), or continuous kernel graph convolution.

[0138] See Figure 9 As shown, the pressure field decoding and performance index output includes the following steps:

[0139] After multi-layer graph message passing, each node obtains a latent vector that integrates neighborhood information and operating condition information. The pressure field decoder maps the node's latent vector to the corresponding node's predicted pressure value. ,in, For the first The GNN predicted pressure value of each graph node corresponds to the finite element node pressure, element center pressure, or water film control volume pressure. The graph node-level stress decoding function is typically composed of a multilayer perceptron, a linear layer, or a combination thereof, and is used to map the hidden representation of nodes to stress values. For the first Each graph node passes through The final hidden representation after message passing in the layered graph includes the node's own characteristics, neighboring node information, edge connection information, and working condition information; For the first The global latent vector corresponding to each working condition sample is broadcast and fused into all graph nodes and edges corresponding to that working condition sample, enabling the graph neural network surrogate model to adapt to the task of predicting the performance of water-lubricated bearings under different speeds, loads, eccentricities, groove structures and material parameters.

[0140] For the entire graph, a graph-level representation can be obtained through global pooling: ,in, The global graph hidden representation corresponding to the k-th group of water-lubricated bearing working condition samples is a comprehensive feature vector that fully represents the overall lubrication state of the bearing, the topological coupling relationship of the finite element mesh, and the distribution of local features in the whole field under this working condition. This is a global pooling function used to aggregate the final hidden representations of all nodes into a graph-level vector. It can employ average pooling, max pooling, attention pooling, or weighted pooling. This represents the final hidden state of a graph node after L-layer graph message passing iterations.

[0141] ,in, This is the graph-level performance index vector obtained from GNN predictions. This is a graph-level performance index decoding function used to map global graph representation and working condition coding into engineering performance indicators; This is the global graph hidden representation corresponding to the k-th group of water-lubricated bearing operating condition samples. It is a comprehensive feature vector that fully represents the overall lubrication state of the bearing, the topological coupling relationship of the finite element mesh, and the distribution of local features in the entire field under this operating condition. This is the global working condition feature encoding vector corresponding to the kth working condition sample; To predict bearing capacity, To predict friction, To predict the coefficient of friction, To predict maximum pressure, To predict the minimum film thickness.

[0142] This design enables the graph neural network surrogate model to output both node-level pressure fields and graph-level engineering performance indicators, thereby achieving a unified approach to pressure distribution prediction and rapid performance evaluation.

[0143] See Figure 10 As shown, training a graph neural network agent model includes the following steps:

[0144] The supervised loss function for the pressure field at the graph nodes is: ,in, The pressure field supervision loss is used to measure the error between the pressure field predicted by the GNN and the pressure field predicted by the FEM. The total number of nodes in the graph can correspond to the number of FEM nodes, the number of element centers, or the number of discrete control volumes in the water film. For GNN in the first The predicted pressure values ​​at each node The first solution obtained by FEM Pressure labels at each finite element mesh node.

[0145] The supervised loss function for engineering performance indicators is: ,in, To monitor the loss of engineering performance indicators, This refers to the vector of engineering performance indicators predicted by GNN, such as load-bearing capacity, friction force, friction coefficient, maximum pressure, and minimum film thickness. Labels for engineering performance indicators obtained from FEM post-processing.

[0146] The total loss function is: ,in, The total training loss of the graph neural network surrogate model is... The weighting coefficients for the pressure field loss at the graph nodes. The weighting coefficient for the loss of graph-level performance indicators.

[0147] To avoid performance indicators with different dimensions dominating training, pressure, load-bearing capacity, friction, coefficient of friction, and maximum pressure can be normalized separately.

[0148] For high-pressure peak areas, trench edge areas, and boundary areas, data weighting items can also be set based on FEM labels:

[0149] ,in, The sample weights, determined by FEM labels or region categories, are used to enhance the model's learning ability in high-pressure areas, trench transition zones, or areas of drastic local change.

[0150] The graph neural network surrogate model is trained using FEM solution data as labels, and a mapping relationship between FEM input features and pressure field and performance indicators is established through supervised learning.

[0151] See Figure 11 As shown, the performance prediction and analysis of water-lubricated bearings under multiple operating conditions includes the following steps:

[0152] After the graph neural network surrogate model is trained, for a new water-lubricated bearing operating condition, the first step is to construct a graph based on its geometric parameters, groove parameters, operating condition parameters, and FEM mesh topology. Then, the graph is input into the trained graph neural network agent model, and the output graph node stress field and performance indicators are generated.

[0153] ,in, This is the graph node-level water film pressure field prediction vector output by the graph neural network surrogate model. This is the vector of engineering performance indicators predicted by GNN. The input graph sample constructed for any new water-lubricated bearing operating condition fully includes the graph node set, graph edge set, original node features, edge features, and global operating condition features. This indicates a graph neural network surrogate model whose parameters have been trained and determined.

[0154] The graph neural network surrogate model can output the nodal pressure field and performance indicators through a single forward propagation. Compared with the traditional finite element method, this process does not require repeated finite element iteration, nonlinear convergence calculation, and complete post-processing, making it suitable for rapid screening of multiple operating conditions, groove parameter optimization, preliminary design of bearing structures, and online evaluation using digital twins.

[0155] The system output includes: a comparison of the pressure cloud map predicted by the finite element label and the graph neural network, an error cloud map, the pressure curve of the mid-section, the bearing capacity error, the maximum pressure error, the friction coefficient error, and a comparison of the calculation time.

[0156] Example 3

[0157] This invention focuses on the rapid prediction of the water film pressure field in water-lubricated bearings. Finite element simulation samples are used as high-fidelity labels to train a graph neural network surrogate model. The bearing computational domain is discretized along the circumferential and axial directions. Finite element nodes, element centers, or water film control volume centers are mapped as graph nodes, and the connections between adjacent nodes or adjacent elements are mapped as graph edges.

[0158] The bearing radius is set to R = 31 mm, the bearing length to B = 80 mm, the radial clearance to c = 60 μm, the dynamic viscosity of the lubricating medium to μ = 1.002 × 10^-3 Pa·s, and the number of circumferential discrete nodes to N. θ =31, the number of discrete nodes along the axis is N. z =15. The sample operating conditions include different journal surface running linear velocities, bearing external loads, and journal eccentricities. The corresponding pressure fields are obtained by finite element analysis or equivalent high-fidelity numerical data.

[0159] The training dataset is represented as ,in, This is a finite element simulation sample dataset containing finite element simulation samples under multiple working conditions. Input the graph structure corresponding to the k-th working condition. The finite element solution label is the label corresponding to the k-th working condition. The finite element solution label can be the pressure field of the finite element mesh node, the pressure field of the element, or the engineering performance indicators such as bearing capacity, maximum pressure, and average pressure obtained from finite element post-processing.

[0160] For regular circumferential-axial water film meshes, each discrete location is defined as a graph node, and circumferential periodic adjacent edges and axial adjacent edges are established. Periodic connections are used in the circumferential direction to allow the first and last nodes to represent the circumferential closed characteristics of a bearing; only adjacent section nodes are connected in the axial direction. For unstructured finite element meshes, graph edges can also be constructed directly by reading the element connection matrix or shared edge relationships.

[0161] See Figure 6 As shown, the input features of each node include: sinθ, cosθ, z / B, h / c, U, F, and ε. Among them, sinθ and cosθ are used to express circumferential periodicity, z / B represents the normalized axial position, h / c represents the normalized film thickness, and U, F, and ε represent the current operating conditions. The above features are only used as input descriptions of finite metadata and are not used to construct Reynolds equation residuals, boundary residuals, or cavitation constraints.

[0162] In this embodiment of the invention, a graph neural network proxy model is used as a finite metadata proxy model. The graph neural network proxy model achieves end-to-end millisecond-level prediction from geometric / boundary features to node pressure fields through encoding, aggregation, and decoding.

[0163] See Figure 9 As shown, this embodiment of the invention employs self-node mapping, neighborhood message aggregation, layer normalization, GELU (Gaussian Error Linear Unit) activation function, and residual connections. This structure can fuse neighborhood node information while preserving the node's own information, thereby learning the spatial distribution law of the finite element pressure field in the water film mesh.

[0164] During training, the pressure field supervision loss is defined as To improve training stability, stress values ​​are normalized to the maximum stress of the samples. The optimizer used is AdamW, with a learning rate of 2×10^-3, a hidden dimension of 64, and 3 graph convolutional layers. This training process does not include Reynolds residuals, boundary condition residuals, cavitation non-negativity loss, and load integral balance loss from the physical information neural network.

[0165] After training, for a new water-lubricated bearing operating condition, it is only necessary to construct a graph input based on its geometric parameters, mesh connection relationship and operating condition parameters, and input the graph into the trained graph neural network surrogate model. The trained graph neural network surrogate model outputs the nodal pressure field through one forward propagation, without having to re-execute the complete finite element modeling, iterative solution and post-processing process.

[0166] After obtaining the pressure field, engineering performance indicators can be calculated in a manner consistent with finite element post-processing. Bearing capacity can be calculated using area integration. Obtain, maximum pressure average pressure This allows for a comparison of the pressure field and performance indicators between finite element method (FEM) labels and graph neural network (BNN) predictions.

[0167] Figure 16 This is a graph showing the prediction loss using a graph neural network in an embodiment of the present invention. See also... Figure 16 As shown, the supervised loss decreases rapidly in the early stages of training and then gradually tends to a lower level, indicating that the graph neural network surrogate model can learn the main mapping relationships from the finite element pressure field samples. Local fluctuations during training mainly arise from differences in pressure peak values ​​among samples under different working conditions and gradient changes caused by batch updates.

[0168] Figure 17 This is a water film pressure cloud map obtained by finite element analysis in an embodiment of the present invention. Figure 18 This is a water film pressure cloud map predicted by the graph neural network surrogate model in an embodiment of the present invention. Figure 17 , Figure 18 The study compares the pressure field predicted by the finite element method with that predicted by the graph neural network. The two methods show good consistency in the location of the high-pressure zone, the axial distribution range, and the circumferential pressure peak region, indicating that the graph neural network can capture the main spatial distribution characteristics of the water film pressure field.

[0169] Figure 19 This is a pressure error cloud map obtained from finite element analysis and graph neural network prediction in an embodiment of the present invention. Figure 19 The results show that the errors are mainly concentrated at the edge of the high-pressure peak and in the local transition region where the pressure gradient is large.

[0170] Figure 20 This is a comparison diagram of the mid-section pressure in an embodiment of the present invention. See also... Figure 20 As shown, the finite element curve and the graph neural network prediction curve maintain good overlap in the pressure rise, peak region and fall region, indicating that the trained graph neural network surrogate model can not only output pressure cloud map, but also accurately reproduce the pressure change law on the key section.

[0171] Optional alternatives to embodiments of the present invention include:

[0172] (1) The nodes in the graph can be finite element nodes, finite element element centers, control volume centers or unified interpolation grid points.

[0173] (2) The graph edges can be formed by finite element connection matrix, node shared element relationship, element shared edge relationship, four-neighborhood, eight-neighborhood or distance threshold neighborhood.

[0174] (3) Graph neural networks can be GCN (Graph Convolutional Network), GAT (Graph Attention Network), GraphSAGE (Graph Sample and Aggregate), GIN (Graph Isomorphism Network), continuous kernel graph convolution, or multi-layer residual graph network.

[0175] (4) Performance indicators can be directly output by the graph-level decoder or obtained by statistical or integral methods consistent with the predicted pressure field and finite element post-processing.

[0176] (5) The loss function can be the mean square error, the mean absolute error, the peak region weighted supervision loss, or the multi-task uncertainty weighted loss.

[0177] (6) The training data can come from finite element method, CFD (Computational Fluid Dynamics) or other high-fidelity numerical methods; when the labels come from finite element method, this method is preferably described as a fast prediction method for the performance of water-lubricated bearings based on graph neural network learning of finite element data.

[0178] This invention also provides an electronic device, including: at least one processor, at least one memory, and a communication interface, wherein the processor, memory, and communication interface communicate with each other; the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the method in this invention embodiment.

[0179] This invention also provides a non-transitory computer-readable storage medium storing computer instructions that cause a computer to execute the methods described in this invention.

[0180] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the performance of water-lubricated bearings based on graph neural network finite element method, characterized in that, Includes the following steps: S1. The parameter acquisition equipment collects the geometric parameters, groove structure parameters, material parameters, and operating condition parameters of the water-lubricated bearing. The parameter acquisition equipment includes: an input terminal, a data storage device, sensors, and a data acquisition device. The sensors include: a speed sensor, a load sensor, a displacement sensor, a temperature sensor, a pressure sensor, and a flow sensor. The data acquisition device receives the signals output by each sensor, performs sampling, analog-to-digital conversion, and filtering, and then sends them to the computing device. The computing device is a workstation or server equipped with a processor, memory, graphics processor, and finite element analysis program. S2. The computing device establishes a water film fluid domain using the finite element method based on the geometric parameters, groove structure parameters, material parameters, and operating condition parameters of the water-lubricated bearing. It then discretizes the water film fluid domain into a mesh along the circumferential and axial directions of the water film to construct a parametric finite element analysis model of the water-lubricated bearing, resulting in a finite element mesh. The finite element mesh includes: finite element nodes, finite element elements formed by connecting multiple finite element nodes in a specific order, and the topological connection relationships between finite element nodes. The topological connection relationships between finite element nodes are determined by the node number and node arrangement order, including the relative position, connection direction, and mesh distance between adjacent finite element nodes. S3. The computing device sets the range of values ​​for the geometric parameters, groove structure parameters, material parameters, and operating condition parameters of the water-lubricated bearing. It generates multiple parameter combinations according to preset sampling rules and inputs each parameter combination into the parametric finite element analysis model. Based on the operating condition parameters and parameter combinations, the computing device applies loads and boundary conditions to the parametric finite element analysis model, performs finite element solutions, obtains the water film pressure values ​​and corresponding performance index data at each finite element node, and constructs a finite element simulation sample library for the water-lubricated bearing. The finite element simulation sample library includes: the water film pressure values ​​of each finite element node, the bearing capacity, friction force, friction coefficient, maximum water film pressure, minimum water film thickness, lubrication state index obtained from the finite element simulation samples, and the circumferential position, axial position, local film thickness, groove area identification, boundary state and corresponding operating condition parameters of each finite element node, as well as the relative position, connection direction and mesh distance between adjacent finite element nodes. S4. The computing device maps each finite element node in the finite element mesh of the finite element simulation sample library to a graph node, and maps finite element nodes with direct connections in the finite element mesh to graph edges, thus obtaining the graph topology. The circumferential position, axial position, local film thickness, trench region identification, boundary state, and corresponding working parameters of the finite element nodes are used as graph node features, and the relative position, connection direction, and mesh distance between adjacent finite element nodes are used as graph edge features. The graph topology, graph node features, and graph edge features together constitute the graph data. S5. The computing device constructs a graph dataset based on graph data and its corresponding finite element simulation samples. Each graph sample in the graph dataset includes: graph topology, graph node features, graph edge features, global operating condition features, finite element node-level labels, and graph-level labels. Among them, graph node features include: circumferential position, axial position, local film thickness, groove region identification, and boundary state of finite element nodes; graph edge features include: relative position, connection direction, and mesh distance between adjacent graph nodes; global operating condition features include geometric parameters of water-lubricated bearings, groove structure parameters, material parameters, and operating condition parameters; finite element node-level labels are determined based on the water film pressure values ​​of each node obtained from the finite element solution; graph-level labels are further calculated based on the finite element solution results to obtain: bearing capacity, friction force, friction coefficient, maximum water film pressure, minimum water film thickness, and lubrication state indicators. S6. The computing device constructs a graph neural network proxy model based on the graph dataset. The graph topology, graph node features, graph edge features, and global operating condition features in the graph dataset are input into the graph neural network proxy model. The graph neural network proxy model learns the local correlation between adjacent graph nodes through multi-layer graph message passing. The water film pressure values ​​of each finite element node obtained from the finite element solution are used as finite element node-level supervision labels. The bearing capacity, friction force, friction coefficient, maximum water film pressure, minimum water film thickness, and lubrication state indicators obtained from the finite element simulation samples are used as graph-level supervision labels. The graph neural network proxy model is trained under supervision to learn the nonlinear mapping relationship between the finite element input features and the finite element output results, and the trained graph neural network proxy model is obtained, which is used to predict the performance of water-lubricated bearings. S7. In engineering applications, the parameter acquisition device collects new operating condition parameters of the water-lubricated bearing under test. The calculation device obtains the finite element mesh corresponding to the new operating condition parameters according to the water film fluid domain discretization rule adopted in step S2. The finite element nodes in the finite element mesh are mapped to new graph nodes. New graph edges are constructed based on the direct connection relationship between the finite element nodes. Based on the new operating condition parameters, geometric parameters, groove structure parameters, and material parameters of the water-lubricated bearing under test, new graph node features, graph edge features, and global operating condition features are generated as input graph data corresponding to the new operating condition parameters. The input graph data corresponding to the new operating condition parameters is input into the trained graph neural network surrogate model. After one forward propagation, the water film pressure field, bearing capacity, friction force, friction coefficient, maximum water film pressure, minimum water film thickness, and lubrication state index of the multi-condition water-lubricated bearing are predicted. The result output device outputs the prediction results. The prediction results are used for multi-condition performance evaluation of water-lubricated bearings, structural scheme selection, operating parameter adjustment, and abnormal lubrication state identification. In the actual engineering of water-lubricated bearings, it provides application analysis and theoretical guidance for specific performance indicators.

2. The method for predicting the performance of water-lubricated bearings based on graph neural network finite element method as described in claim 1, characterized in that: In step S1, the geometric parameters of the water-lubricated bearing include: journal radius, bearing inner radius, journal length, radial clearance, length-to-diameter ratio, and bearing wrap angle; the groove structure parameters include the number of grooves, groove type, groove circumferential position, groove axial position, groove width, groove depth, and groove length; the material parameters include the elastic modulus, Poisson's ratio, material density, surface roughness, and liner thickness of the bearing liner and journal materials; the operating condition parameters include: journal speed, journal surface running linear velocity calculated based on journal speed and journal radius, bearing external load, journal eccentricity, misalignment angle, fluid supply pressure, fluid supply flow rate, dynamic viscosity, density, temperature, and pressure parameters of the lubricating medium. These operating condition parameters are used to determine the motion boundary conditions and load boundary conditions in the finite element analysis process of the water-lubricated bearing. Specifically, the journal speed is used to determine the journal surface running speed, the bearing external load is used to apply load constraints, and the journal displacement is used to calculate the journal eccentricity and misalignment angle, thereby obtaining the water film pressure distribution and bearing performance indicators under the corresponding operating conditions.

3. The method for predicting the performance of water-lubricated bearings based on graph neural network finite element method as described in claim 2, characterized in that: The input terminal or data storage device collects the geometric parameters, groove structure parameters, and material parameters of the water-lubricated bearing. Sensors collect actual operating parameters of water-lubricated bearings: speed sensor collects journal speed, load sensor collects bearing external load, displacement sensor collects journal displacement and determines journal eccentricity and offset angle accordingly, temperature sensor collects lubricating medium temperature, pressure sensor collects supply pressure, and flow sensor collects supply flow. The input terminal is used for manual input of design parameters and preset operating parameters of water-lubricated bearings: journal radius, journal length, radial clearance, number of grooves, groove width, material elastic modulus, lubricating medium viscosity, preset speed and preset load; Data storage devices are used to read pre-stored parameter files or database data: geometric parameters of water-lubricated bearings, groove structure parameters, material parameters, and historical operating condition parameters. The data storage devices can be local hard drives, network storage devices, or database servers. Operating condition parameters are used to determine the motion boundary conditions and load boundary conditions in the finite element analysis process of water-lubricated bearings. Among them, the journal speed is used to determine the running speed of the journal surface, the bearing external load is used to apply load constraints, and the journal displacement is used to calculate the journal eccentricity and offset angle, thereby obtaining the water film pressure distribution and bearing performance indicators under the corresponding operating conditions.

4. The method for predicting the performance of water-lubricated bearings based on graph neural network finite element method as described in claim 3, characterized in that: The graph neural network proxy model includes: a node feature encoder, an edge feature encoder, a global working condition feature encoder, a multi-layer graph message passing module, a graph-level pooling module, a graph node-level stress field decoder, and a graph-level performance index decoder. The node feature encoder is used to encode the features of finite element nodes, mapping the node attributes of finite element nodes—circumferential position, axial position, local film thickness, groove region identifier, and boundary state—to graph node implicit feature representations. The edge feature encoder is used to encode the connection relationships of finite element nodes: the relative positions, connection directions and mesh distances between adjacent finite element nodes, and map them into graph edge latent feature representations. The global operating condition feature encoder is used to encode the global operating condition features of water-lubricated bearings to obtain a global operating condition implicit feature representation that characterizes the overall operating condition state. The multi-layer graph message passing module is used to: iteratively update the implicit feature representations of graph nodes, graph edges, and global working conditions based on the graph topology; aggregate neighborhood node information through a multi-layer message passing mechanism; learn the spatial topological relationships and local coupling relationships in the finite element mesh; and obtain the updated implicit feature representations of graph nodes. The graph-level pooling module is used to aggregate the updated implicit feature representations of graph nodes to obtain a graph-level feature representation that characterizes the overall state of the water-lubricated bearing. The graph node-level pressure field decoder is used to: predict the water film pressure value corresponding to each graph node based on the updated implicit features of the graph nodes. The graph nodes are obtained by mapping one-to-one from the finite element mesh nodes. The graph node pressure prediction result corresponds to the water film pressure value at the finite element node position, thereby realizing the prediction of the water film pressure field. The graph-level performance index decoder is used to predict the overall performance index of water-lubricated bearings based on graph-level feature representations, including load capacity, friction force, friction coefficient, maximum water film pressure, minimum water film thickness, and lubrication status indexes. The node feature encoder, edge feature encoder, and global working condition feature encoder all establish data flow connections with the multi-layer graph message passing module. The multi-layer graph message passing module establishes data flow connections with the graph-level pooling module. The graph-level pooling module establishes data flow connections with the graph node-level pressure field decoder and the graph-level performance index decoder, respectively.

5. The method for predicting the performance of water-lubricated bearings based on graph neural network finite element method as described in claim 4, characterized in that: In step S6, supervised training of the graph neural network surrogate model is performed, including the following steps: The water film pressure values ​​of each finite element node obtained from the finite element method are used as finite element node-level supervision labels, and the bearing capacity, friction force, friction coefficient, maximum water film pressure, minimum water film thickness, and lubrication state indices obtained from the finite element simulation samples are used as graph-level supervision labels. The graph neural network surrogate model is trained under supervision. The loss function of the supervision training is composed of graph node-level pressure field supervision loss and graph-level performance index supervision loss. The graph node-level pressure field supervision loss is used to measure the error between the water film pressure prediction value output by the graph node-level pressure field decoder and the finite element node water film pressure value, and the graph-level performance index supervision loss is used to measure the error between the performance index prediction value output by the graph-level performance index decoder and the corresponding performance index calculated by the finite element method.

6. The method for predicting the performance of water-lubricated bearings based on graph neural network finite element method as described in claim 5, characterized in that: The graph edges are established based on the finite element mesh adjacency relationship, element shared edge relationship, node shared element relationship, or regular mesh adjacency relationship.

7. A performance prediction system for water-lubricated bearings based on graph neural network finite element method for performing the method according to any one of claims 1-6, characterized in that, It includes parameter acquisition devices, computing devices, and result output devices connected in sequence, among which, The parameter acquisition equipment is used to collect geometric parameters, groove structure parameters, material parameters, and operating condition parameters of water-lubricated bearings. The parameter acquisition device includes: an input terminal, a data storage device, a sensor, and a data acquisition device. The sensor is connected to the data acquisition device, and both the data acquisition device and the input terminal are connected to the data storage device. The input terminal is used for manual input of design parameters and preset operating parameters of water-lubricated bearings: journal radius, journal length, radial clearance, number of grooves, groove width, material elastic modulus, lubricating medium viscosity, preset speed and preset load; Data storage devices are used to read pre-stored parameter files or database data: geometric parameters of water-lubricated bearings, groove structure parameters, material parameters, and historical operating condition parameters; the data storage devices can be local hard drives, network storage devices, or database servers; The sensors include: a speed sensor, a load sensor, a displacement sensor, a temperature sensor, a pressure sensor, and a flow sensor. These sensors are used to collect actual operating parameters of the water-lubricated bearing: the speed sensor collects the journal speed, the load sensor collects the external load on the bearing, the displacement sensor collects the journal displacement and uses this to determine the journal eccentricity and offset angle, the temperature sensor collects the lubricating medium temperature, the pressure sensor collects the supply pressure, and the flow sensor collects the supply flow rate. These operating parameters are used to determine the motion boundary conditions and load boundary conditions during the finite element analysis of the water-lubricated bearing. Specifically, the journal speed is used to determine the running speed on the journal surface, the external load on the bearing is used to apply load constraints, and the journal displacement is used to calculate the journal eccentricity and offset angle, thereby obtaining the water film pressure distribution and bearing performance indicators under the corresponding operating conditions. The data acquisition device is used to: receive signals output from various sensors, perform sampling, analog-to-digital conversion, and filtering, and then send them to the computing device; The computing device is a workstation or server equipped with a processor, memory, graphics processor, and finite element analysis program. The computing device includes, in sequence, a data reading module, a finite element simulation sample library construction module, a graph data conversion module, a graph feature encoding module, and a graph neural network proxy module, wherein: The data reading module is used to read the geometric parameters, groove structure parameters, material parameters, operating condition parameters, finite element mesh files, and finite element result files of water-lubricated bearings. The finite element simulation sample library construction module is used to: establish a water film fluid domain using the finite element method based on the geometric parameters, groove structure parameters, material parameters, and operating condition parameters of the water-lubricated bearing; discretize the water film fluid domain along the circumferential and axial directions of the water film; construct a parametric finite element analysis model of the water-lubricated bearing; and obtain a finite element mesh. The finite element mesh includes: finite element nodes, finite element elements formed by connecting multiple finite element nodes in a specific order, and the topological connections between finite element nodes. The topological connections between finite element nodes are determined by the node numbers and node arrangement order, including the relative positions, connection directions, and mesh distances between adjacent finite element nodes; and set the geometric parameters, groove structure parameters, material parameters, and operating condition parameters of the water-lubricated bearing. The range of values ​​is determined by generating multiple parameter combinations according to preset sampling rules. Each parameter combination is then input into the parametric finite element analysis model. Based on the operating condition parameters and parameter combinations, loads and boundary conditions are applied to the parametric finite element analysis model, and finite element solutions are performed to obtain the water film pressure values ​​and corresponding performance index data at each finite element node. A finite element simulation sample library for water-lubricated bearings is constructed. The finite element simulation sample library includes: water film pressure values ​​at each finite element node; bearing capacity, friction force, friction coefficient, maximum water film pressure, minimum water film thickness, lubrication state index obtained from the finite element simulation samples; circumferential position, axial position, local film thickness, groove area identification, boundary state and corresponding operating condition parameters of each finite element node; and relative position, connection direction and mesh distance between adjacent finite element nodes. The graph data conversion module is used to: map finite element nodes in the finite element mesh of the finite element simulation sample library to graph nodes, and map finite element nodes with direct connections in the finite element mesh to graph edges, thus obtaining a graph topology; and use the circumferential position, axial position, local film thickness, trench region identification, boundary state, and corresponding working parameters of the finite element nodes as graph node features, and the relative position, connection direction, and mesh distance between adjacent finite element nodes as graph edge features. The graph topology, graph node features, and graph edge features together constitute the graph data. The graph feature encoding module is used to: construct a graph dataset based on graph data and its corresponding finite element simulation samples. Each graph sample in the graph dataset includes: graph topology, graph node features, graph edge features, global operating condition features, finite element node-level labels, and graph-level labels. Among them, graph node features include: circumferential position, axial position, local film thickness, groove region identification, and boundary state of finite element nodes; graph edge features include: relative position, connection direction, and mesh distance between adjacent graph nodes; global operating condition features include geometric parameters of water-lubricated bearings, groove structure parameters, material parameters, and operating condition parameters; finite element node-level labels are determined based on the water film pressure values ​​of each node obtained from the finite element solution; graph-level labels are further calculated based on the finite element solution results to obtain: bearing capacity, friction force, friction coefficient, maximum water film pressure, minimum water film thickness, and lubrication state indicators. The graph neural network proxy module is used to: construct a graph neural network proxy model based on a graph dataset. The graph topology, node features, edge features, and global operating condition features from the graph dataset are input into the graph neural network proxy model. The graph neural network proxy model learns the local relationships between adjacent graph nodes through multi-layer graph message passing. The water film pressure values ​​of each finite element node obtained from the finite element method are used as finite element node-level supervision labels, and the bearing capacity, friction force, friction coefficient, maximum water film pressure, minimum water film thickness, and lubrication state indicators obtained from finite element simulation samples are used as graph-level supervision labels. This supervises the training of the graph neural network proxy model, learning the nonlinear mapping relationship between the finite element input features and the finite element output results, resulting in a trained graph neural network proxy model. In engineering applications, the trained graph neural network proxy model performs a forward propagation on the input graph data corresponding to new operating condition parameters to predict the water film pressure field, bearing capacity, friction force, friction coefficient, maximum water film pressure, minimum water film thickness, and lubrication state indicators of multi-condition water-lubricated bearings. The output device is a display or industrial control terminal, which is used to output the predicted results of water film pressure field and bearing performance indicators. The prediction results are used for multi-condition performance evaluation of water-lubricated bearings, structural scheme selection, operation parameter adjustment and abnormal lubrication state identification, and provide application analysis and theoretical guidance for specific performance indicators in the actual engineering of water-lubricated bearings.

8. The water-lubricated bearing performance prediction system based on graph neural network finite element method as described in claim 7, characterized in that: When the graph neural network proxy module performs supervised training on the graph neural network proxy model, the loss function of the supervised training is composed of graph node-level pressure field supervision loss and graph-level performance index supervision loss. The graph node-level pressure field supervision loss is used to measure the error between the water film pressure prediction value output by the graph node-level pressure field decoder and the water film pressure value of the finite element node. The graph-level performance index supervision loss is used to measure the error between the performance index prediction value output by the graph-level performance index decoder and the corresponding performance index calculated by the finite element.

9. An electronic device, characterized in that, include: At least one processor, at least one memory, and a communication interface, wherein the processor, memory, and communication interface communicate with each other; The memory stores program instructions that can be executed by a processor, which invokes the program instructions to perform the method described in any one of claims 1-6.

10. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions that cause the computer to perform the method according to any one of claims 1-6.

Citation Information

Patent Citations

  • Method, system and equipment for determining lubricating performance of sliding bearing and medium

    CN121351626A

  • Machine tool spindle angular contact ball bearing stress field digital twinning modeling method based on graph neural network

    CN121638048A