A railway bogie structure optimization method and system based on deep learning

Abstract

The present application relates to the technical field of train bogie, and discloses a railway bogie structure optimization method and system based on deep learning, which comprises the following steps: collecting a three-dimensional geometric model of a bogie swing bolster to construct a finite element simulation model; applying combined loads to the swing bolster structure based on the finite element simulation to obtain a stress atlas of the swing bolster structure; extracting mechanical indexes to determine whether to perform structure optimization; when it is determined to perform optimization, training a deep learning model based on historical structure samples; inputting swing bolster structure parameters to be optimized into the deep learning model to obtain parameter adjustment suggestions for optimization; reconstructing a swing bolster structure model according to the parameter adjustment suggestions and performing finite element verification to determine whether the optimized structure meets target performance index requirements. The present application improves design efficiency, shortens development cycles, and realizes efficient collaborative optimization of railway bogie swing bolster structures in terms of strength, stiffness, modal and fatigue life and other multi-target performance.

Description

Technical Field

[0001] The present invention relates to the technical field of train bogies, and in particular to a railway bogie structure optimization method and system based on deep learning. Background Art

[0002] With the increase in railway vehicle operating speeds and train loads, the structural performance of railway bogies, as key components for train safety and comfort, plays a decisive role in the overall vehicle operating status. The bolster, as the core load-bearing structure in the bogie frame system, primarily bears the vertical loads between the car body and sideframes, lateral inertia forces, and longitudinal impact loads during braking and traction. In actual operation, it bears complex, multi-dimensional, and high-intensity dynamic conditions.

[0003] To improve the mechanical performance of bolsters, traditional structural optimization design primarily relies on finite element analysis, analyzing stress distribution and deformation patterns under different loading conditions to enhance the structure. For example, patent document CN112356865A achieves performance enhancements for a traditional K1-type bolster by reconfiguring the bolster's internal reinforcement ribs, sidewall support plates, and wall thickness distribution, combined with multi-condition static strength analysis. However, engineering applications still face several challenges. Traditional bolster structural design relies primarily on engineers setting and modifying parameters based on empirical rules. The structural optimization process requires multiple manual iterations, resulting in low development efficiency, long lead times, and uncertain results. Furthermore, stress analysis or local load simulation based on a single operating condition fails to fully capture the dynamic response of the structure under different operating scenarios, potentially invalidating the optimization results under other conditions. In most current optimization processes, simulation results cannot be directly fed back into the structure generation or optimization module. This lacks an end-to-end structure-performance mapping model and iterative correction capabilities, and the structural design process remains fragmented and static.

[0004] Therefore, it is necessary to design a railway bogie structure optimization method and system based on deep learning to solve the problems existing in current technology. Summary of the Invention

[0005] In view of this, the present invention proposes a railway bogie structure optimization method and system based on deep learning, aiming to solve the current problems of long design iteration cycle, low optimization space utilization, and lack of prediction and optimization linkage mechanism.

[0006] In one aspect, the present invention proposes a railway bogie structure optimization method based on deep learning, comprising:

[0007] Collect the 3D geometric model of the railway bogie bolster to be optimized and build a finite element simulation model;

[0008] Apply combined loads to the bolster structure based on finite element simulation to obtain the stress spectrum, displacement response spectrum, and modal characteristic spectrum of the bolster structure;

[0009] Extracting mechanical indicators from the atlas to determine whether to perform structural optimization, the mechanical indicators include maximum equivalent stress, maximum displacement, first-order modal frequency, and structural stiffness value;

[0010] When it is determined that structural optimization is to be performed, a deep learning model is trained based on historical structural samples, wherein the deep learning model includes a neural network model that introduces a multi-attention mechanism, the input of the deep learning model is a structural parameter tensor, and the output of the deep learning model is a structural performance prediction index;

[0011] Inputting the bolster structural parameters to be optimized into the deep learning model to obtain parameter adjustment suggestions for optimization, the parameter adjustment suggestions including wall thickness changes, rib layout adjustments, and support plate position corrections;

[0012] The bolster structure model is rebuilt based on the parameter adjustment suggestions, and finite element verification is performed to determine whether the optimized structure meets the target performance index requirements.

[0013] Furthermore, when applying a combined load to the bolster structure based on finite element simulation to obtain a stress spectrum, a displacement response spectrum, and a modal characteristic spectrum of the bolster structure, the following steps are included:

[0014] The 3D geometric model of the railway bogie bolster was meshed in the finite element analysis platform, and local mesh refinement was performed on the ear connection area, stiffener transition area, and hollow cavity.

[0015] Setting material properties of the bolster structure, wherein the material properties include elastic modulus, Poisson's ratio, and yield strength;

[0016] applying a combined load condition including a longitudinal load, a transverse load, and a vertical load to the bolster structure;

[0017] Solve the finite element model and obtain the stress spectrum, displacement response spectrum and modal characteristic spectrum.

[0018] Furthermore, the mechanical indicators in the atlas are extracted to determine whether to perform structural optimization, including:

[0019] Extract the maximum displacement amplitude and compare it with the allowable displacement tolerance of the structure installation;

[0020] Extract the first-order modal frequency value to determine whether it falls within the operating excitation frequency range;

[0021] Obtain the overall stiffness value of the structure based on the load-deformation relationship under loading conditions;

[0022] If any mechanical index exceeds the safety threshold, it is determined that the railway bogie bolster structure needs to be optimized.

[0023] Furthermore, when training a deep learning model based on historical structure samples, it includes:

[0024] Collect several bolster structure samples with different structural parameters and simulation results to construct a training dataset containing structural parameters, load conditions, and simulation performance indicators;

[0025] The structural parameters include wall thickness distribution, rib positions, and the number and size of support plates, and the structural parameters are encoded in tensor form as model input.

[0026] Furthermore, when training a deep learning model based on historical structure samples, it also includes:

[0027] The deep learning model includes a backbone network and a multi-attention module. The backbone network is a residual convolutional neural network, and the multi-attention module includes a spatial attention mechanism and a channel attention mechanism, which are used to enhance the feature expression ability of high stress areas.

[0028] Furthermore, when training a deep learning model based on historical structure samples, it also includes:

[0029] The deep learning model is trained through supervised learning, and the optimization goal is to minimize the error between the structural performance prediction value and the finite element simulation result. The structural performance prediction value includes the maximum equivalent stress, maximum displacement, first-order modal frequency and overall structural stiffness value.

[0030] Furthermore, when inputting the bolster structural parameters to be optimized into the deep learning model and obtaining parameter adjustment suggestions for optimization, the following steps are included:

[0031] Encode the parameters of the bolster structure's wall thickness distribution, reinforcement configuration, and support plate size into the current structural parameter tensor;

[0032] Inputting the current structural parameter tensor into the trained deep learning model to obtain the corresponding structural performance prediction index;

[0033] The parameter adjustment suggestion is obtained based on the difference between the structural performance prediction index and the preset target performance value and in combination with historical structural samples.

[0034] Furthermore, the bolster structure model is rebuilt based on the parameter adjustment suggestions, and finite element verification is performed to determine whether the optimized structure meets the target performance index requirements, including:

[0035] When all indicators of the optimized structure meet the target performance indicator requirements, the structural optimization is considered successful;

[0036] When any mechanical index of the optimized structure still exceeds the safety threshold, the model is optimized and the structure is optimized again.

[0037] Furthermore, when optimizing the model and re-optimizing the structure, the following steps are included:

[0038] Compare the error between the finite element simulation results of the optimized structure and the predicted value of the deep learning model. If the error exceeds the set threshold, the structure sample is added to the training dataset and the deep learning model is incrementally trained.

[0039] During incremental training, adjusting the channel weights of the multi-attention mechanism according to the error location;

[0040] Based on the updated deep learning model, parameter adjustment suggestions are re-obtained and a new bolster structure model is constructed;

[0041] The new model is imported into the finite element simulation platform for re-verification to determine whether it meets the performance target. If it still does not meet the target, the model optimization and structural optimization processes are continued until the preset number of times is reached or the target performance index requirements are met.

[0042] Compared with the existing technology, the present invention has the following advantages: through the intelligent optimization framework combining deep learning and finite element simulation, it solves the problems of reliance on manual experience, design-simulation separation, and insufficient adaptability to multiple working conditions in the optimization process of traditional railway bogie bolster structures. By collecting three-dimensional geometric models and applying combined loads to perform multi-working condition finite element simulation, the stress, deformation, and modal response of the bolster structure in actual operation are comprehensively extracted, achieving accurate evaluation of structural performance. With the help of a deep neural network model that introduces a multi-attention mechanism, a high-dimensional mapping relationship between structural parameters and performance responses is established, making performance prediction more refined, especially with stronger identification capabilities for high-stress areas and structural weaknesses. By training the model to generate wall thickness adjustment, rib optimization, and support layout suggestions for the current structural state, it realizes automation and personalization of structural optimization. A closed-loop iterative optimization mechanism based on the prediction results is constructed, which improves design efficiency, shortens the development cycle, and ensures that the optimized structure meets performance index requirements under multiple working conditions, thereby achieving efficient coordinated optimization of railway bogie bolster structures among multiple objective performance indicators such as strength, stiffness, modality, and fatigue life.

[0043] On the other hand, the present application also provides a railway bogie structure optimization system based on deep learning, which is used to apply the railway bogie structure optimization method based on deep learning, including:

[0044] A geometric modeling unit, used to acquire a three-dimensional geometric model of the railway bogie bolster to be optimized and to construct a parametric model for finite element analysis;

[0045] a simulation analysis unit, configured to apply a combined load condition including a longitudinal load, a transverse load, and a vertical load to the bolster structure in a finite element analysis platform, and generate a stress spectrum, a displacement response spectrum, and a modal characteristic spectrum;

[0046] a structural evaluation unit, configured to extract the maximum equivalent stress, maximum displacement, first-order modal frequency, and structural stiffness value from the atlas, and determine whether to perform structural optimization based on a preset safety threshold;

[0047] A model training unit is used to construct a training data set based on historical structural samples and train a deep learning neural network model that introduces a multi-attention mechanism. The input of the deep learning neural network model is a structural parameter tensor, and the deep learning neural network model outputs a structural performance prediction indicator;

[0048] The parameter optimization unit is used to input the parameter tensor of the structure to be optimized into the trained neural network model to obtain the structural performance prediction results, and output parameter adjustment suggestions for structural improvement based on the historical sample library. The parameter adjustment suggestions include changing the wall thickness, adjusting the rib layout, and correcting the support plate position;

[0049] A verification unit is used to generate a new rocker structure model according to the parameter adjustment suggestion and perform finite element simulation to verify whether the optimized structure meets the target performance index requirements.

[0050] It is understandable that the above-mentioned railway bogie structure optimization method and system based on deep learning have the same beneficial effects and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Various other advantages and benefits will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiment below. The accompanying drawings are for illustration purposes only and are not to be considered as limiting the present invention. The same reference symbols are used throughout the drawings to represent the same components. In the drawings:

[0052] Figure 1 A flowchart of a railway bogie structure optimization method based on deep learning provided by an embodiment of the present invention;

[0053] Figure 2 Functional block diagram of a railway bogie structure optimization system based on deep learning provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0054] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art. It should be noted that, unless there is a conflict, the embodiments of the present disclosure and the features in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

[0055] In the process of optimizing the bolster structure of traditional railway bogies, structural parameters are set manually based on experience, and performance indicators are verified through finite element simulation. Engineers need to manually adjust the wall thickness distribution, rib layout, and support plate position. After each parameter change, the finite element model needs to be rebuilt and multi-condition static and modal analysis needs to be performed. Due to the nonlinear coupling relationship between structural parameters and performance indicators, there is a lack of quantitative basis for the direction of parameter adjustment, resulting in the optimization process requiring multiple trial and error iterations. At the same time, the simulation results of a single condition cannot reflect the superposition effect of complex dynamic loads. For example, the stress concentration area under the combined action of lateral inertia force and longitudinal impact load may be missed, which makes the optimized structure at risk of local failure in the real operating environment.

[0056] For example, during the development of the bolster for the bogie of a railway freight car with a speed of 100 kilometers per hour, engineers used traditional methods to gradiently adjust the wall thickness of the hollow cavity to reduce the maximum equivalent stress. Each modification of the wall thickness requires re-dividing the mesh in the finite element platform and applying longitudinal, transverse, and vertical combined loads. A single simulation takes more than 8 hours. After three iterations, it was found that although the maximum displacement under the vertical load met the requirements, a new stress peak appeared in the ear connection area under the transverse load condition. At this time, the structural parameters of the stiffener transition zone need to be readjusted, resulting in the overall development cycle being extended to three weeks. What's more serious is that modal analysis shows that the first-order natural frequency of the optimized structure is close to the wheel-rail excitation frequency band, and there is a risk of resonance, forcing the design process to be reworked.

[0057] Without addressing these issues, structural optimization efficiency will be limited by manual experience and computing resources, making it difficult to meet the demands of high-frequency train development iterations. The lack of multi-condition dynamic response characteristics can lead to performance deviations in optimized solutions under complex load conditions, potentially triggering structural fatigue crack propagation or abnormal vibration, directly impacting bogie service life and operational safety. Furthermore, the disconnected process between parameter adjustment and performance verification will drive up engineering development costs and hinder the advancement of lightweight and high-performance railway equipment.

[0058] For this, see Figure 1 As shown, this application proposes a railway bogie structure optimization method based on deep learning, including:

[0059] S100: Collect the three-dimensional geometric model of the railway bogie bolster to be optimized and build a finite element simulation model.

[0060] S200: Applying a combined load to the bolster structure based on finite element simulation to obtain a stress spectrum, a displacement response spectrum, and a modal characteristic spectrum of the bolster structure.

[0061] S300: Extract mechanical indicators from the graph to determine whether to perform structural optimization. The mechanical indicators include maximum equivalent stress, maximum displacement, first-order modal frequency, and structural stiffness.

[0062] S400: When it is determined to perform structural optimization, a deep learning model is trained based on historical structural samples. The deep learning model includes a neural network model that introduces a multi-attention mechanism. The input of the deep learning model is a structural parameter tensor, and the output of the deep learning model is a structural performance prediction indicator.

[0063] S500: Input the bolster structure parameters to be optimized into the deep learning model to obtain parameter adjustment suggestions for optimization. The parameter adjustment suggestions include changing the wall thickness, adjusting the rib layout, and correcting the support plate position.

[0064] S600: Reconstruct the bolster structure model based on the parameter adjustment suggestions and perform finite element verification to determine whether the optimized structure meets the target performance index requirements.

[0065] The 3D geometric model of the railway bogie bolster is a three-dimensional digital representation of the bolster. This can be obtained using CAD software or 3D scanning technology. It accurately describes the bolster's geometry and spatial structure, providing a foundational model for subsequent finite element analysis. A finite element simulation model is a numerical calculation model built based on the 3D geometric model. ANSYS or ABAQUS software can be used for meshing and boundary condition setting to simulate the mechanical behavior of the bolster under load. Combined loads refer to the simultaneous longitudinal, transverse, and vertical loads acting on the bolster structure. Combined loads can be applied based on the vehicle limit loads specified in the railway industry standard TB / T 1335-1996 to simulate the complex stress conditions experienced during actual operation. Stress maps, displacement response maps, and modal characteristic maps are visualizations of the mechanical responses generated through finite element simulation. Post-processing modules can be used to extract maximum stress, displacement distribution, and natural frequency data for quantitative evaluation of structural performance. Mechanical indicators, including maximum equivalent stress, maximum displacement, first-order modal frequency, and structural stiffness, can be obtained by analyzing key data points in the simulation maps and used to determine whether the structure meets safety threshold requirements. A deep learning model refers to a neural network model trained based on historical structural samples. Specifically, it can be constructed using a residual convolutional network combined with a spatial attention mechanism and a channel attention mechanism to establish a mapping relationship between structural parameters and performance indicators. A structural parameter tensor refers to encoding parameters such as wall thickness distribution and rib position into a multidimensional array format. Specifically, it can be represented by a numerical matrix or eigenvector and used as input data for a deep learning model. Parameter adjustment suggestions refer to structural improvement plans output by the model, which may include operations such as increasing or decreasing local wall thickness and adjusting rib topology, to guide designers to quickly optimize weak areas of the structure. Finite element verification refers to re-simulating and analyzing the optimized model. Specifically, the stress, displacement, and modal parameters before and after optimization can be compared to confirm whether the structural performance meets the target requirements.

[0066] The core innovation of this application lies in the combination of deep learning technology and finite element simulation. By constructing a neural network model with a multi-attention mechanism, it achieves end-to-end mapping from structural parameters to performance indicators, and automatically generates structural optimization suggestions based on the model prediction results, forming a closed-loop iterative optimization process to improve the efficiency and reliability of railway bogie bolster structure design.

[0067] The working process and principles of this application involve collecting a three-dimensional geometric model of the railway bogie bolster to be optimized and constructing a finite element simulation model. Based on finite element simulation, a combined load is applied to the bolster structure to obtain stress maps, displacement response maps, and modal characteristic maps. By applying the combined loads to simulate actual operating conditions, multi-dimensional structural response data is obtained. Mechanical indicators from the maps are extracted to determine whether structural optimization is necessary. Mechanical indicators include maximum equivalent stress, maximum displacement, first-order modal frequency, and structural stiffness. These indicators comprehensively reflect the static strength, deformation characteristics, and dynamic response of the structure. When structural optimization is determined to be necessary, a deep learning model is trained based on historical structural samples. The deep learning model includes a neural network model that incorporates a multi-attention mechanism. Its input is a structural parameter tensor, and its output is a structural performance prediction indicator. The multi-attention mechanism enhances feature learning in high-stress areas. The structural parameters of the bolster to be optimized are input into the deep learning model to obtain parameter adjustment recommendations for optimization. Parameter adjustment recommendations include wall thickness changes, rib layout adjustments, and support plate position corrections. This data-driven approach provides quantitative optimization guidance. The bolster structural model was rebuilt based on the parameter adjustment recommendations, and finite element verification was performed to determine whether the optimized structure met the target performance requirements. This formed a closed-loop optimization process to ensure the effectiveness of the optimization results.

[0068] By combining finite element analysis and deep learning methods, automated optimization of the bolster structure was achieved. The deep learning model established a mapping relationship between structural parameters and performance indicators, reducing manual trial and error. Multi-case combined load analysis ensured the reliability of the optimization solution in complex dynamic environments. This end-to-end optimization process improved structural design efficiency.

[0069] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0070] A 3D CAD model of the bogie bolster of a certain high-speed EMU was collected and imported into the finite element analysis software ANSYS to construct a simulation model. The model was meshed, and local mesh refinement was performed in the lug connection area, the stiffener transition area, and the hollow cavity.

[0071] Set the bolster material properties, including an elastic modulus of 210 GPa, a Poisson's ratio of 0.3, and a yield strength of 355 MPa. Calculate the ultimate load case according to the TB / T 1335-1996 standard, applying a combined load of 150 kN longitudinally, 100 kN transversely, and 600 kN vertically to the bolster structure.

[0072] Solve the finite element model to obtain stress maps, displacement response maps, and modal characteristic maps. Extract mechanical indicators such as maximum equivalent stress, maximum displacement, first-order modal frequency, and structural stiffness. Compare these indicators to pre-set safety thresholds to determine whether structural optimization is necessary.

[0073] A historical sample dataset containing 1,000 sets of different structural parameters and simulation results was constructed. Structural parameters, including wall thickness distribution, rib location, and support plate quantity and size, were encoded in tensor form as model input.

[0074] A deep learning model was built, using the ResNet50 residual convolutional neural network as the backbone, and incorporating spatial and channel attention mechanisms. The model was trained using supervised learning, with the optimization objective being to minimize the error between the predicted structural performance and the finite element simulation results.

[0075] The bolster structural parameters to be optimized are encoded as tensors and fed into a trained deep learning model. The model outputs predicted structural performance metrics, which are then compared with target performance values ​​to generate parameter adjustment recommendations.

[0076] Modify the bolster CAD model based on the recommendations and re-import it into ANSYS for finite element verification. If the optimized structure meets the target performance requirements, the optimization is considered successful. Otherwise, the new structure sample is added to the training set, the deep learning model is incrementally trained, and the optimization process is repeated.

[0077] Through the above scheme, the present application realizes the automated optimization of the bolster structure of railway bogies. The deep learning model establishes a nonlinear mapping relationship between structural parameters and performance indicators, provides a quantitative optimization direction, reduces the number of manual trial and error, and shortens the structural design cycle. The multi-condition combined load analysis comprehensively evaluates the response characteristics of the structure in a complex dynamic environment, avoiding the risk of local failure that may be caused by single-condition optimization. The end-to-end optimization process realizes the automatic closed loop of parameter adjustment and performance verification, and improves the efficiency of structural design. The multi-attention mechanism enhances the feature learning ability of high-stress areas and improves the optimization accuracy of the model for key areas. Through a method that combines historical sample data drive and simulation verification, the present application improves the efficiency and reliability of bolster structure optimization, and provides an effective tool for the lightweight and high-performance design of railway bogies.

[0078] In some of the above-mentioned solutions of this application, insufficient meshing accuracy exists in the finite element simulation process, resulting in deviations in stress analysis of key areas, and the load combination conditions fail to fully cover the actual operating limit conditions, affecting the accuracy of subsequent structural optimization decisions.

[0079] The present application further proposes meshing the three-dimensional geometric model of the railway bogie bolster in a finite element analysis platform, and performing local mesh encryption on the ear connection area, the stiffener transition area, and the hollow cavity. Material properties of the bolster structure are set, including elastic modulus, Poisson's ratio, and yield limit. Combined load conditions including longitudinal load, transverse load, and vertical load are applied to the bolster structure. The finite element model is solved and stress maps, displacement response maps, and modal characteristic maps are obtained. The load combination is calculated based on the vehicle limit load in the railway industry standard TB / T 1335-1996.

[0080] Among them, the ear connection area is encrypted due to the dynamic load between the bogie and the car body, and the mesh size is reduced to 30% of the basic unit. The reinforcement transition zone adopts a gradual mesh transition strategy, and the adjacent unit size ratio is controlled within 1:1.5. A hexahedron-dominated hybrid mesh is used inside the hollow cavity, and a three-layer prismatic unit is set in the wall boundary layer. The material properties are calibrated based on experimental data, with the elastic modulus set to 206GPa, the Poisson's ratio to 0.3, and the yield limit to 345MPa. The longitudinal load is calculated according to Article 4.2.3 of the standard TB / T 1335-1996, the lateral load is taken as the equivalent load of the car body rolling moment, and the vertical load includes static and dynamic load components.

[0081] Specifically, mesh refinement in the lug connection area reduces the calculated error in the stress concentration factor around the bolt hole to less than 5%, while the gradient mesh in the transition zone eliminates sudden stress changes. A hexahedron-dominated hollow cavity mesh improves the efficiency of cavity buckling mode calculations by 40%. Precise settings for the elastic modulus and yield limit keep the deviation of the predicted stress within ±3% of the measured material value. Combined load conditions calculated in accordance with standard TB / T 1335-1996 cover three limit states: vehicle emergency braking, curve negotiation, and track irregularity excitation. Longitudinal loads are applied using a quasi-static superposition method, transverse loads are converted into distributed forces using a moment equivalence formula, and vertical loads are dynamically corrected based on the axle load distribution coefficient. During the finite element solution, an explicit dynamics algorithm is used to address the time-varying load characteristics. Modal analysis uses the Lanczos method to extract the first six vibration modes, resulting in a stress map with a resolution of 0.1 mm / pixel.

[0082] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0083] The 3D geometric model of the railway bogie bolster was meshed using a finite element analysis platform. Tetrahedral elements were used for meshing, with a cell size of 10 mm. Local mesh refinement was performed in the lug connection area, the stiffener transition zone, and the hollow cavity, with a cell size of 5 mm in the refined areas.

[0084] Set the material properties for the bolster structure. Select Q345 steel as the material, set the elastic modulus to 210 GPa, the Poisson's ratio to 0.3, and the yield strength to 345 MPa.

[0085] A combined load condition was applied to the bolster structure. The longitudinal load was set to 300kN, the transverse load to 150kN, and the vertical load to 600kN. This load combination was determined based on the vehicle limit load calculation method outlined in railway industry standard TB / T 1335-1996.

[0086] Solve the finite element model and obtain stress maps, displacement response maps, and modal characteristic maps. The stress map is expressed as von Mises equivalent stress, the displacement response map shows the displacement vectors at each node, and the modal characteristic map includes the first six modal frequencies and vibration shapes.

[0087] Through the above-mentioned technical solution, this application achieves accurate finite element analysis of railway bogie bolster structures. Local mesh refinement improves calculation accuracy in high-stress areas. Combined load cases are used to simulate the complex load conditions encountered in actual operation. The various maps obtained comprehensively reflect the mechanical response characteristics of the bolster structure under extreme operating conditions, providing a reliable data foundation for subsequent structural optimization.

[0088] In some of the above-mentioned solutions of this application, there is a problem of a single dimension of mechanical performance evaluation in the traditional rocker structure optimization process. Relying solely on a single indicator for judgment can easily lead to optimization decision deviations and cannot fully reflect the comprehensive performance of the structure under complex working conditions.

[0089] This application further proposes extracting mechanical indicators from the atlas to determine whether structural optimization is necessary. This includes extracting the maximum displacement amplitude and comparing it with the allowable displacement tolerance for structural installation. Extracting the first-order modal frequency value and determining whether it falls within the operating excitation frequency range. Determining the overall structural stiffness based on the load-deformation relationship under loading conditions. If any mechanical indicator exceeds a safety threshold, the railway bogie bolster structure is deemed to require structural optimization.

[0090] The maximum displacement amplitude is extracted by extracting the coordinate extreme points in the displacement response spectrum, and the allowable displacement tolerance of the structural installation is set according to the assembly tolerance of the bogie and car body connectors. The operating excitation frequency range is determined by collecting the bogie vibration spectrum data during actual train operation, covering the 0.5-15Hz frequency band corresponding to the vehicle operating speed. The overall structural stiffness value is obtained by calculating the displacement ratio of the load application point and the deformation monitoring point in the finite element model. The safety thresholds include the maximum equivalent stress not exceeding 80% of the material yield strength, the displacement tolerance not exceeding ±2mm, the interval between the modal frequency and the excitation frequency exceeding 3Hz, and the stiffness value not less than 1.2 times the design specification value.

[0091] Specifically, after generating a displacement response spectrum through finite element simulation, an extreme value search algorithm is used to locate the coordinates of the maximum displacement point. The displacement value is then compared with the installation clearance tolerance specified in the assembly process documentation. If the maximum displacement exceeds the upper tolerance limit, the structure is determined to be potentially causing component interference under load. In modal analysis, the first-order natural frequency is extracted and compared with the vibration frequency range that accounts for more than 5% of the measured track spectrum. If the natural frequency falls within this range, there is a risk of resonance. Using the force-displacement curve recorded during load application, a linear regression method is used to calculate the stiffness coefficient, which reflects the structure's comprehensive ability to resist deformation. If any of these three indicators exceeds a preset threshold, the structural optimization process is triggered. For example, in a measured case, the maximum displacement of a certain type of bolster under lateral load reached 2.5mm, exceeding the allowable tolerance of 2mm. At the same time, the first-order modal frequency of 12Hz falls within the high-incidence vibration frequency range of 10-14Hz during vehicle operation. In this case, the system automatically determines that structural optimization is required.

[0092] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0093] The maximum displacement amplitude is extracted and compared with the allowable displacement tolerance for structural installation. Specifically, the maximum displacement value is extracted from the displacement response spectrum obtained by finite element analysis and compared with the pre-set installation displacement tolerance. For example, for a certain model of railway bogie bolster, the installation displacement tolerance is set at 5mm. If the extracted maximum displacement value exceeds 5mm, structural optimization is determined to be necessary.

[0094] The first-order modal frequency is extracted to determine whether it falls within the operating excitation frequency range. Furthermore, the first-order modal frequency is extracted from the modal signature and compared with the typical operating excitation frequency range for railway vehicles. If the first-order modal frequency falls within the excitation frequency range of 20-40 Hz, a resonance risk is identified, requiring structural optimization.

[0095] The overall structural stiffness is derived based on the load-deformation relationship under the loading conditions. Specifically, the overall stiffness of the bolster structure is calculated by analyzing the combined loads applied in the finite element simulation and the corresponding structural deformations. For example, for vertical load conditions, the vertical stiffness can be estimated by the ratio of the vertical load to the maximum vertical deformation.

[0096] If any mechanical indicator exceeds a safety threshold, the railway bogie bolster structure is deemed to require structural optimization. As a preferred implementation, safety thresholds can be set, such as maximum equivalent stress not exceeding 80% of the material's yield strength, maximum displacement not exceeding 5mm, first-order modal frequency not falling within the 20-40Hz range, and overall structural stiffness not falling below a preset value. If any of these indicators exceeds the corresponding threshold, the structural optimization process is triggered.

[0097] Through the above technical solution, the present application realizes a comprehensive evaluation of the structural performance of railway bogie bolsters. As a result, potential structural defects can be discovered in a timely manner, providing a clear direction for subsequent optimization. Furthermore, by introducing a comprehensive judgment of multiple mechanical indicators, the one-sidedness that may exist in the evaluation of a single indicator is avoided, and the accuracy and reliability of structural optimization decisions are improved. Specifically, the maximum displacement evaluation ensures the stability of the structure during actual installation and operation. Modal frequency analysis effectively prevents the risk of resonance. The overall stiffness evaluation ensures the bearing capacity of the structure under complex loads. Through this multi-dimensional performance evaluation method, the present application improves the scientificity and rationality of the structural design of railway bogie bolsters.

[0098] In some of the above-mentioned solutions of this application, traditional rocker structure optimization relies on finite element simulation and manual experience iteration, which has the problem of insufficient sample data leading to insufficient training of deep learning models, making it difficult to accurately predict the complex nonlinear relationship between structural parameters and performance indicators.

[0099] This application further proposes training a deep learning model based on historical structural samples by collecting several bolster structure samples with different structural parameters and simulation results to construct a training dataset containing structural parameters, load conditions, and simulation performance indicators. Structural parameters include wall thickness distribution, rib location, and the number and size of support plates. These parameters are encoded in tensor form as model input.

[0100] Among them, the wall thickness distribution is expressed by the thickness value matrix of the three-dimensional grid node, and each grid unit corresponds to the thickness value to form a two-dimensional tensor. The position of the reinforcement rib is composed of the rib centerline coordinates and the cross-sectional dimensions to form a three-dimensional tensor. The number and size of the support plates are combined into a four-dimensional tensor through the position encoding matrix and the geometric parameter vector. The load condition data contains the amplitude vectors of the longitudinal, transverse and vertical loads, and the simulation performance indicators include the maximum equivalent stress scalar, the maximum displacement vector, the modal frequency sequence and the stiffness tensor. The training data set is normalized to map the parameters of different dimensions to a unified numerical range, and a sliding window sampling strategy is used to enhance data diversity.

[0101] Specifically, the wall thickness distribution tensor is generated through finite element mesh node mapping, and each node records the local thickness value to form a two-dimensional matrix that is topologically consistent with the geometric model. The rib position tensor is encoded in a polar coordinate system, with the center of the bolster as the origin, and records the radial distance, angle, and cross-sectional height and width of the ribs to form spatial distribution characteristics. The support plate parameters are marked with a Boolean matrix to mark the installation position, and a composite code is formed by combining the plate thickness and length parameters. During the training process, the input tensor extracts spatial features through the convolution layer, the fully connected layer fuses global parameters, and the output layer predicts performance indicators. By optimizing the network weights through batch training, the model can learn the mapping rules between structural parameters and stress, displacement, frequency and stiffness, providing a data-driven basis for subsequent parameter adjustments.

[0102] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0103] Several bolster structure samples with different structural parameters and simulation results were collected to construct a training dataset containing structural parameters, load conditions, and simulation performance indicators. Structural parameters include wall thickness distribution, rib location, and the number and size of support plates. Structural parameters are encoded in tensor form as model input.

[0104] Specifically, multiple bolster structural models were first constructed using 3D modeling software. Each model group differed in terms of wall thickness, rib layout, and support plate configuration. Next, finite element analysis was performed on each model group, applying a combination of longitudinal, transverse, and vertical load conditions. Furthermore, performance indicators such as maximum equivalent stress, maximum displacement, first-order modal frequency, and structural stiffness were extracted from the finite element analysis results. This generated a sample dataset containing structural parameters and corresponding performance indicators.

[0105] For example, 100 different bolster structure samples can be constructed. For each sample, the wall thickness distribution can be represented by a 10×10 matrix, the rib locations can be encoded by a 20×3 matrix, and the number and size of support plates can be represented by a 1×10 vector. These parameters are integrated into a three-dimensional tensor, which serves as the input to the deep learning model. Simultaneously, the performance indicators of each sample under different load conditions, such as maximum equivalent stress and maximum displacement, are recorded as the model's output labels.

[0106] Through the above technical solution, this application can construct a training dataset containing rich structural variants, providing sufficient learning samples for deep learning models. By encoding complex three-dimensional structural information into a standardized tensor form, it facilitates feature extraction and pattern recognition by neural networks. This data preprocessing method can improve the generalization ability of the model, enabling it to accurately predict performance indicators under different structural parameter combinations, thereby providing a reliable decision-making basis for subsequent structural optimization.

[0107] In some of the above-mentioned solutions of this application, when training a deep learning model based on historical structural samples, the backbone network may not be able to effectively capture the local characteristics of high-stress areas, resulting in insufficient prediction accuracy of the model for structural performance.

[0108] This application further proposes a deep learning model including a backbone network and multi-attention modules. The backbone network is a residual convolutional neural network, and the multi-attention module includes a spatial attention mechanism and a channel attention mechanism, which are used to enhance the feature expression ability of high stress areas.

[0109] The residual convolutional neural network utilizes a skip connection structure to achieve deep feature extraction, avoiding the vanishing gradient problem. The spatial attention mechanism enhances activation responses in high-stress areas by calculating the spatial weight matrix of the feature map. The channel attention mechanism selects channel information that contributes most to structural performance prediction by assigning weights to feature channels. Both attention modules operate in parallel on the feature map output by the backbone network, forming an optimized feature representation through weighted fusion.

[0110] Specifically, the residual convolutional neural network performs multi-layer convolution processing on the input wall thickness distribution and rib position parameters to extract structural features at different scales. The spatial attention mechanism generates a weight distribution based on the spatial position of the feature map. For example, the feature weights of the ear connection area and the rib transition area are increased to 0.8-0.9, and the weights of the remaining areas are reduced to 0.1-0.2. The channel attention mechanism performs global average pooling on the convolution channel to generate a channel weight vector, setting the weight coefficient of the maximum equivalent stress-related channel to 0.7-0.8 and the weight coefficient of the secondary feature channel to 0.3-0.4. After processing by the attention module, the local features of the high-stress area are enhanced in both the spatial dimension and the channel dimension, enabling the subsequent fully connected layer to more accurately predict the maximum displacement and modal frequency indicators.

[0111] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0112] The deep learning model consists of a backbone network and a multi-attention module. The backbone network adopts a residual convolutional neural network architecture, consisting of multiple stacked residual blocks. Each residual block contains two 3x3 convolutional layers and a skip connection. The multi-attention module includes spatial and channel attention mechanisms to enhance the feature representation of high-stress areas.

[0113] The spatial attention mechanism weights the spatial positions of feature maps by generating a two-dimensional attention map. In its implementation, the input feature map is first subjected to max-pooling and average-pooling operations. The pooled results are then concatenated and passed through a 7x7 convolutional layer to generate the spatial attention map. The channel attention mechanism adjusts the weights of different channels by generating a one-dimensional vector. This is achieved by performing global average-pooling and max-pooling on the input feature map, followed by a shared multi-layer perceptron to generate the channel attention vector.

[0114] During model training, the structural parameter tensor is first fed into the backbone network for feature extraction. The feature map is then weighted by the spatial attention module and the channel attention module. Finally, a fully connected layer outputs the structural performance prediction metrics. The model is trained using supervised learning, with the optimization objective being to minimize the mean squared error between the predicted values ​​and the finite element simulation results.

[0115] Through the above technical solution, this application can effectively capture the characteristics of high-stress areas in the bolster structure and improve the model's sensitivity to key areas. The introduction of the multi-attention mechanism enables the model to adaptively adjust the importance of different spatial positions and feature channels, thereby improving the prediction accuracy of complex structural responses. At the same time, the residual network structure helps alleviate the gradient vanishing problem in deep network training, enabling the model to learn deeper feature representations. This end-to-end structure-performance mapping model provides an efficient and reliable performance evaluation tool for bolster structure optimization, improving the efficiency of structural design and optimization.

[0116] In some of the above-mentioned schemes in this application, when training the deep learning model based on historical structural samples, a combination of a multi-attention mechanism and a residual convolutional neural network backbone network is adopted. Although it can enhance the feature expression ability of high stress areas, in the actual model training process, due to the complex nonlinear mapping relationship between structural parameters and performance indicators, there may still be a large deviation between the model prediction value and the finite element simulation result, resulting in insufficient reliability of subsequent parameter adjustment suggestions, and multiple iterative verifications are required to achieve the optimization goal, affecting the overall optimization efficiency.

[0117] This application further proposes that when training a deep learning model based on historical structural samples, it also includes: the deep learning model is trained through supervised learning, and the optimization goal is to minimize the error between the structural performance prediction value and the finite element simulation result. The structural performance prediction value includes the maximum equivalent stress, maximum displacement, first-order modal frequency and overall structural stiffness value.

[0118] Among them, the supervised learning method uses the structural parameter tensor in the training dataset and the corresponding finite element simulation results as input and output pairs, and adjusts the network weights through the back-propagation algorithm. The optimization goal is achieved by constructing a multi-index joint loss function, which includes the weighted sum of the maximum equivalent stress error term, the maximum displacement error term, the first-order modal frequency error term, and the structural stiffness error term. The spatial attention mechanism of the multi-attention module is used to capture the geometric features of the local stress concentration area, the channel attention mechanism is used to screen the key mechanical response channels, and the residual convolutional network retains the feature information of different levels through jump connections. The error minimization process uses the gradient descent algorithm, combined with the adaptive learning rate adjustment strategy to ensure the balanced convergence speed of each indicator.

[0119] Specifically, during the model training phase, the structural parameter tensors of historical structural samples are input into a neural network consisting of residual convolutional layers and multi-attention modules. The network output layer generates predicted values ​​for four performance metrics. The difference between the predicted values ​​and the finite element simulation results is calculated using a multi-metric joint loss function. Each metric error term in the loss function is weighted according to engineering requirements. During backpropagation, a spatial attention mechanism dynamically adjusts the spatial weight distribution of the feature map based on the error gradient, enhancing local feature learning in high-stress areas. A channel attention mechanism adjusts channel weights based on their contribution to the final error to suppress noise interference. The residual convolutional network uses skip connections to fuse shallow geometric features with deep abstract features, avoiding the vanishing gradient problem. By iteratively optimizing network parameters, the model gradually establishes a high-precision mapping between structural parameters and multiple performance metrics, converging the prediction errors within a preset threshold. The trained model accurately predicts performance metrics for different structural parameter combinations, providing a reliable basis for subsequent parameter adjustment recommendations and reducing the number of finite element verification iterations.

[0120] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0121] The deep learning model is trained using supervised learning. First, a large number of historical bolster structure samples and their corresponding finite element simulation results are collected to construct a training dataset. This dataset contains structural parameters (such as wall thickness distribution and rib location) and corresponding performance indicators (maximum equivalent stress, maximum displacement, first-order modal frequency, and overall structural stiffness).

[0122] Next, we designed the network architecture of the deep learning model, including convolutional, pooling, and fully connected layers. The input layer receives a tensor of structural parameters, and the output layer predicts four performance metrics. During training, we use backpropagation and gradient descent to update the network parameters. The loss function is defined as the mean squared error between the predicted values ​​and the actual simulation results.

[0123] Specifically, the training process includes the following steps:

[0124] Data preprocessing: Normalize the structural parameters and convert the performance indicators to the same scale.

[0125] Batch training: A batch of samples are randomly selected and input into the network each time.

[0126] Forward propagation: Calculate the predicted value of the network output.

[0127] Calculate the loss: Compare the predicted value with the true value to get the loss function value.

[0128] Backpropagation: Calculate the gradient of the loss function with respect to the parameters of each layer.

[0129] Parameter update: Update the network weights using an optimization algorithm such as Adam.

[0130] Repeat steps 2-6 until the preset number of training rounds is reached or the loss function converges.

[0131] After training is complete, the validation set is used to evaluate model performance. If the prediction error is large, optimization can be performed by adjusting the network structure, increasing the training data, or using regularization methods. Ultimately, a deep learning model is obtained that can accurately predict the performance of the bolster structure.

[0132] Through the above-mentioned technical solution, this application achieves rapid prediction of bolster structural performance. The deep learning model uses extensive historical data to learn the complex mapping relationship between structural parameters and performance indicators, avoiding the tedious finite element analysis process. After model training is complete, performance prediction results can be quickly obtained by simply inputting structural parameters, improving the efficiency of structural optimization. Furthermore, the supervised learning method ensures consistency between the model's predictions and actual simulation results, providing a reliable basis for performance evaluation for subsequent structural optimization.

[0133] In some of the above-mentioned solutions of this application, the deep learning model trained based on historical structural samples can predict the performance indicators of the bolster structure. However, in practical applications, there is still a challenge in how to effectively encode the parameters of the structure to be optimized into an input form that can be processed by the model and accurately generate adjustment suggestions that match the preset target performance values. In the parameter encoding process, the existing methods may cause large deviations between the model prediction results and the actual performance indicators due to inconsistent data formats or insufficient feature extraction, which in turn affects the reliability of the parameter adjustment suggestions. In addition, the dynamic adaptation mechanism between the historical sample library and the current optimization needs has not yet been perfected, which may cause the adjustment suggestions to lack specificity.

[0134] This application further proposes to encode the parameters of the wall thickness distribution, reinforcement configuration and support plate size of the rocker structure into the current structural parameter tensor, input the current structural parameter tensor into the trained deep learning model to obtain the corresponding structural performance prediction index, and obtain parameter adjustment suggestions based on the difference between the structural performance prediction index and the preset target performance value and combined with historical structural samples.

[0135] Among them, the wall thickness distribution is converted into a multi-channel matrix through three-dimensional grid discretization processing, and each grid cell records the thickness value of the corresponding area as the spatial dimension feature of the tensor. The rib configuration adopts the coordinate grid marking method, and the layout position of the rib is mapped into a binary matrix, in which the grid cells with ribs are marked as 1 and the rest are marked as 0, which serves as the channel dimension feature of the tensor. The support plate size is represented by geometric parameter vectorization, including length, width and installation angle, which are spliced ​​into an additional feature layer of the tensor after normalization. The deep learning model extracts local structural features in the tensor through the convolution layer and maps the features into prediction results of maximum equivalent stress, maximum displacement, first-order modal frequency and overall structural stiffness value through the fully connected layer. During the generation of parameter adjustment suggestions, the predicted results are compared with the target performance values. If the maximum displacement exceeds the threshold, the combination of wall thickness adjustment and rib layout optimization is retrieved from the historical sample library, and the top 10% samples with the largest displacement reduction are selected as the recommended solutions.

[0136] Specifically, during the parameter encoding stage, the wall thickness distribution is discretized on the bolster 3D model at a grid resolution of 5mm×5mm, generating a 3D matrix of size 128×128×64. Each cell stores the thickness value at the corresponding position, and the tensor dimensions are aligned by zero padding. The rib layout is decomposed into two 128×128 matrices in the horizontal and vertical directions using a 2D projection method, recording the position information of the ribs in different directions respectively. The support plate size parameters are converted into a standardized four-dimensional vector, including the length percentage, width percentage, installation angle cosine value, and relative position offset. The encoded tensor is input into a deep learning model consisting of a residual convolution layer and a spatial attention module. The model output layer generates predicted values ​​for four performance indicators. When the predicted maximum displacement exceeds the preset safety threshold of 2.5mm, the system extracts all sample data from the historical sample library where the wall thickness increases by more than 15% and the rib density increases by more than 20%. It calculates the correlation coefficient between the displacement reduction and the stiffness increase, and selects samples with a correlation coefficient greater than 0.85 to generate wall thickness change recommendations. Specifically, this includes increasing the wall thickness in the stress concentration area from 12mm to 16mm and adding cross-ribs with a spacing of 80mm in the area where the displacement exceeds the standard. After the optimized parameter recommendations pass finite element verification, if the displacement index still does not meet the standard, the model incremental training process is triggered, the newly generated simulation data is added to the training set, and the channel attention weight is adjusted to improve the accuracy of subsequent recommendations.

[0137] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0138] The parameters of the bolster structure's wall thickness distribution, rib configuration, and support plate dimensions are encoded into the current structural parameter tensor. Specifically, the wall thickness distribution can be represented by a 3x3x3 matrix, where each element corresponds to the wall thickness value of a different region of the bolster structure. The rib configuration can be represented by a binary matrix whose size is the number of grid cells on the bolster surface, where 1 indicates the presence of a rib at that location and 0 indicates the absence. The support plate dimensions can be represented by a one-dimensional vector consisting of the three parameters: length, width, and thickness.

[0139] The current structural parameter tensor is fed into a trained deep learning model to obtain corresponding structural performance prediction indicators. The deep learning model utilizes a residual convolutional neural network architecture, comprising multiple convolutional, pooling, and fully connected layers. The model outputs four indicators: maximum equivalent stress, maximum displacement, first-order modal frequency, and overall structural stiffness.

[0140] Based on the discrepancies between predicted structural performance indicators and preset target performance values, and combined with historical structural samples, parameter adjustment recommendations are generated. Furthermore, the sensitivity of each structural parameter to the performance indicator is calculated using a gradient descent method, prioritizing adjustments to those with high sensitivity. For example, if the predicted maximum equivalent stress exceeds the target value, a recommendation might be made to increase the wall thickness in the high-stress area or to add stiffeners near that area. This generates specific parameter adjustment recommendations, including changes to wall thickness, adjustments to stiffener layout, and corrections to support plate positions.

[0141] Through the above-mentioned technical solution, this application achieves rapid optimization of railway bogie bolster structures based on a deep learning model. This reduces the time and workload of manual iterative optimization and improves structural design efficiency. Furthermore, by introducing comprehensive predictions of multiple performance indicators, the optimization results can meet performance requirements under various operating conditions, enhancing the reliability of structural optimization. Furthermore, a parameter adjustment suggestion mechanism based on historical samples improves the targetedness and effectiveness of the optimization process and accelerates the convergence of structural optimization.

[0142] In some of the above-mentioned solutions of this application, when the mechanical indicators of the optimized structure still exceed the safety threshold, the existing methods lack an effective model optimization mechanism, which may cause the structural optimization process to fall into an inefficient cycle and cannot effectively correct the impact of the deep learning model prediction error on the optimization results.

[0143] The present application further proposes that when any mechanical index of the optimized structure still exceeds the safety threshold, the model should be optimized and the structure should be re-optimized.

[0144] The error between the finite element simulation results of the optimized structure and the predicted values ​​of the deep learning model is compared. If the error exceeds a set threshold, the structure sample is added to the training dataset, and the deep learning model is incrementally trained. During incremental training, the channel weights of the multi-attention mechanism are adjusted according to the error location. Based on the updated deep learning model, parameter adjustment suggestions are re-obtained and a new bolster structure model is constructed. The new model is imported into the finite element simulation platform for re-verification to determine whether it meets the performance objectives. If it still does not meet the requirements, the model optimization and structural optimization process are repeated until the preset number of times is reached or the target performance indicators are met.

[0145] Specifically, when the error between the finite element simulation results and the model prediction values ​​exceeds 5%, the current structural parameters and the corresponding simulation results are added to the training data set. During the incremental training process, the network parameters are updated through the back-propagation algorithm, and the weight coefficients of the channel attention layer in the multi-attention module are adjusted for areas with large prediction errors, such as increasing the channel weights in high-stress areas to above 0.8. The updated model regenerates parameter adjustment suggestions, such as increasing the wall thickness from 12mm to 15mm, or adjusting the rib spacing to 200mm. The new structural model is again verified by finite element analysis. If the maximum equivalent stress still exceeds 250MPa, the model training and structural optimization steps are repeated until the number of iterations reaches 10 or the maximum equivalent stress drops below 200MPa. Through the cyclic optimization mechanism, the model prediction accuracy is gradually improved, and the reliability of the structural parameter adjustment is enhanced, ultimately ensuring that the optimized rocker meets the target performance indicators.

[0146] As a preferred embodiment, the solution of this application is specifically implemented as follows: When the maximum equivalent stress of the optimized bolster structure still exceeds the safety threshold, the finite element simulation platform outputs a maximum equivalent stress value of 325 MPa, while the deep learning model predicts a value of 298 MPa, with the difference between the two exceeding the set threshold. The wall thickness distribution parameters, rib layout parameters, and corresponding simulation results of the optimized structure are automatically added to the training dataset, and the neural network is updated using an incremental training algorithm. During training, the channel weights related to stress characteristics in the multi-attention mechanism are specifically adjusted, with the channel attention coefficient of the third convolution layer increased from 0.46 to 0.71. The updated model is re-invoked to generate parameter adjustment recommendations, including shifting the support plate forward by 12 mm and increasing the wall thickness of the ear area by 1.5 mm. The reconstructed bolster model is re-injected into the finite element platform for verification. The results show that the maximum equivalent stress is reduced to 293 MPa and the first-order modal frequency is increased to 42 Hz, meeting the target performance requirements.

[0147] Through the above technical solution, this application achieves closed-loop control of the structural optimization process. By automatically identifying model prediction errors and dynamically correcting neural network parameters, it effectively solves the problem of low iteration efficiency caused by model prediction deviations in traditional optimization methods. This solution enables the structural optimization process to autonomously adapt to performance deviations under different working conditions, improving the effectiveness and reliability of the optimization solution and avoiding the time and cost of manual repeated adjustments.

[0148] In some of the above-mentioned solutions of the present application, when the mechanical indicators of the optimized rocker structure still exceed the safety threshold, the existing model optimization method lacks a dynamic adjustment mechanism, resulting in low optimization efficiency and inability to effectively narrow the gap between the prediction error and the actual simulation results, making it difficult to quickly converge to a structural solution that meets the performance goals.

[0149] The present application further proposes that when any mechanical indicator of the optimized structure still exceeds the safety threshold, model optimization and re-structural optimization are performed, including the following steps: comparing the error between the finite element simulation results of the optimized structure and the predicted value of the deep learning model. If the error exceeds the set threshold, the structural sample is added to the training data set, and the deep learning model is incrementally trained. During the incremental training process, the channel weights of the multi-attention mechanism are adjusted according to the error position. Based on the updated deep learning model, the parameter adjustment suggestions are re-obtained and a new bolster structure model is constructed. The new model is imported into the finite element simulation platform for re-verification to determine whether it meets the performance target. If it still does not meet the target, the model optimization and structural optimization process are continued to cycle until the preset number of times is reached or the target performance indicator requirements are met.

[0150] The error comparison step sets an error threshold of 5%. When the relative error between the predicted and simulated values ​​exceeds this threshold, the model update mechanism is triggered. During incremental training, the training dataset is expanded using a dynamic batching approach, with the number of new samples added at each time not exceeding 10% of the original data volume. The multi-attention mechanism automatically optimizes channel weights using a backpropagation algorithm, prioritizing the enhancement of feature channels with high correlation with the error. The number of structural optimization cycles is preset to 5; if the target is still not met, manual intervention is triggered.

[0151] Specifically, the error analysis between the finite element simulation results and the model prediction values ​​is achieved by calculating the root mean square error of each mechanical index. When the error exceeds the threshold, the current structural parameters and the corresponding simulation results are marked as new training samples. Incremental training adopts a transfer learning strategy, freezing the parameters of the first three layers of the backbone network, and only updating the parameters of the multi-attention module and the fully connected layer. The training cycle is controlled within 50 times. The adjustment of the channel weight is achieved through the gradient descent algorithm. The weight of the feature channel corresponding to the high error area is increased by 20%-30%. The updated model generates new parameter adjustment suggestions, such as increasing the wall thickness by 2mm or adjusting the rib spacing to 150mm. The reconstructed model is verified by finite element again. If the maximum equivalent stress drops below 250MPa and the displacement amplitude is less than 3mm, the optimization is judged to be successful. If the optimization fails to meet the standard for three consecutive times, the system automatically switches to the optimization mode based on the genetic algorithm.

[0152] As a preferred embodiment, the solution of the present application is specifically implemented as follows: during the optimization process of the bolster structure, when the maximum displacement error between the optimized finite element simulation results and the predicted value of the deep learning model exceeds 5%, the three-dimensional geometric parameters, load condition data and corresponding finite element simulation results of the structure are stored in the training database. During the incremental training process, the channel weight coefficient of the high-stress area is increased in the spatial attention layer, and the weight distribution ratio of the low-stress area is reduced. The updated deep learning model generates a new support plate position correction scheme, which is imported into the parametric modeling tool to generate an improved structure. After the new structure completes the vertical load condition verification in the finite element platform, if its first-order modal frequency is still lower than the target value, the current structural parameter tensor is extracted again and input into the model, and the parameter adjustment and simulation verification are performed cyclically until all performance indicators meet the railway industry standard requirements in the fifth iteration.

[0153] Through the above technical solution, this application effectively solves the problem of iteration failure caused by the accumulation of model prediction errors in traditional optimization processes. By dynamically updating training data and adjusting attention weights, the deep learning model's adaptability to complex working conditions is improved. This solution can automatically identify the source of error and modify model parameters in a targeted manner, avoiding local optimal solutions caused by manual experience adjustments, ensuring that the structural optimization process converges to a safe threshold within a limited number of iterations, and continuously enhancing the model's generalization performance through a closed-loop feedback mechanism.

[0154] The intelligent optimization framework developed in this embodiment, combining deep learning with finite element simulation, addresses the challenges of traditional railway bogie bolster structure optimization, which relies on manual experience, exhibits a disconnect between design and simulation, and lacks adaptability to multiple operating conditions. By acquiring a three-dimensional geometric model and applying combined loads to perform multi-operational finite element simulations, the stress, deformation, and modal response of the bolster structure during actual operation are comprehensively extracted, enabling accurate assessment of structural performance. A deep neural network model incorporating a multi-attention mechanism establishes a high-dimensional mapping between structural parameters and performance responses, enabling more refined performance predictions and, in particular, enhanced identification of high-stress areas and structural weaknesses. The trained model generates wall thickness adjustments, rib optimization, and support layout recommendations tailored to the current structural state, enabling automated and personalized structural optimization. A closed-loop iterative optimization mechanism based on the prediction results improves design efficiency, shortens development cycles, and ensures that the optimized structure meets performance requirements under multiple operating conditions, thereby enabling efficient, coordinated optimization of multiple performance objectives for the railway bogie bolster structure, including strength, stiffness, modal characteristics, and fatigue life.

[0155] In another preferred embodiment based on the above embodiment, refer to Figure 2 As shown, this embodiment provides a railway bogie structure optimization system based on deep learning, which is used to apply the railway bogie structure optimization method based on deep learning, including:

[0156] The geometric modeling unit is used to acquire the three-dimensional geometric model of the railway bogie bolster to be optimized and to construct a parametric model for finite element analysis.

[0157] The simulation analysis unit is used to apply a combined load condition including longitudinal load, transverse load and vertical load to the bolster structure in the finite element analysis platform to generate stress maps, displacement response maps and modal characteristic maps.

[0158] The structural evaluation unit is used to extract the maximum equivalent stress, maximum displacement, first-order modal frequency and structural stiffness value in the graph, and determine whether to perform structural optimization based on the preset safety threshold.

[0159] The model training unit is used to construct a training data set based on historical structural samples and train a deep learning neural network model that introduces a multi-attention mechanism. The input of the deep learning neural network model is a structural parameter tensor, and the deep learning neural network model outputs a structural performance prediction indicator.

[0160] The parameter optimization unit is used to input the parameter tensor of the structure to be optimized into the trained neural network model to obtain the structural performance prediction results, and output parameter adjustment suggestions for structural improvement in combination with the historical sample library. The parameter adjustment suggestions include wall thickness changes, rib layout adjustments, and support plate position corrections.

[0161] The verification unit is used to generate a new rocker structure model based on the parameter adjustment suggestions and perform finite element simulation to verify whether the optimized structure meets the target performance index requirements.

[0162] It can be understood that a railway bogie structure optimization system based on deep learning includes a geometric modeling unit, a simulation analysis unit, a structural evaluation unit, a model training unit, a parameter optimization unit and a verification unit.

[0163] The geometric modeling unit uses parametric modeling technology to convert the bolster's three-dimensional geometric model into the parametric model required for finite element analysis, ensuring model editability and data compatibility. The simulation analysis unit integrates a finite element solver, automatically applying combined loads based on preset load conditions and generating multidimensional mechanical maps. Load combinations are calculated based on vehicle limit loads specified in railway industry standards. The structural assessment unit incorporates built-in safety threshold judgment logic, triggering the optimization process by comparing mechanical indicators with preset thresholds. The model training unit utilizes a multi-attention neural network for supervised training using a historical structural sample dataset containing structural parameter tensors and corresponding simulation performance metrics. The parameter optimization unit generates parameter adjustment recommendations based on the trained model, including adjustments to wall thickness distribution, optimized rib layout, and corrected support plate positions. The verification unit uses automated scripts to feed the optimized parameters back to the geometric modeling unit, generating a new structural model and performing finite element verification.

[0164] Specifically, the geometric modeling unit receives the original 3D bolster model and parametrically processes it to generate an editable model that includes wall thickness, rib location, and support plate dimensions. The simulation analysis unit uses a finite element solver to mesh the model, performing local mesh refinement in the ear connection area, rib transition zone, and hollow cavity to ensure computational accuracy in stress concentration areas. The structural assessment unit extracts maximum equivalent stress, displacement amplitude, and modal frequency, comparing them against pre-set safety thresholds. If any metric exceeds the threshold, an optimization process is triggered. The model training unit utilizes a residual convolutional neural network combined with spatial and channel attention mechanisms to enhance the learning of features in high-stress areas and improve the model's prediction accuracy of structural performance. The parameter optimization unit encodes the current structural parameters as tensors and inputs them into the trained model. It then outputs predicted performance metrics, compares them with target values, and generates specific parameter adjustment plans based on a historical sample library. The verification unit automatically modifies model parameters based on the adjustment plan, generates a new structural model, and re-runs finite element simulation. If the verification results do not meet the requirements, an incremental training process is initiated, adding new samples to the training set to optimize the model weights. This process continues iteratively until performance requirements are met.

[0165] As a preferred embodiment, the solution of the present application is specifically implemented as follows: The railway bogie structure optimization system consists of six functional modules. The geometric modeling unit constructs a three-dimensional parametric model of the bolster through SolidWorks software. The model contains adjustable wall thickness parameters, stiffener topology parameters and support plate position parameters, and is imported into ANSYS Workbench to establish a finite element mesh model. The simulation analysis unit sets the combined working conditions of longitudinal load, transverse load and vertical load based on the TB / T 1335-1996 standard, performs static and modal analysis in ANSYS Mechanical, and generates an equivalent stress cloud map with a maximum stress of 235MPa, a displacement vector map with a maximum displacement of 1.8mm, and a vibration mode diagram with a first-order modal frequency of 32Hz. The structural evaluation unit extracts key indicators from the simulation results through Python scripts, and triggers the optimization instruction when it is detected that the maximum displacement exceeds the 2mm threshold. The model training unit uses the TensorFlow framework to build a residual convolutional neural network. Its input layer receives a four-dimensional tensor consisting of a wall thickness distribution matrix, rib position coordinates, and support plate dimensions. The network's middle layer embeds a spatial attention module to focus on the high-stress features of the ear connection area. The output layer predicts the structure's maximum stress, displacement, modal frequency, and stiffness. The parameter optimization unit calls the trained model for forward reasoning and generates an adjustment plan for the current structure, increasing the side wall thickness from 15mm to 18mm and adding three radial ribs in the stress concentration area. The verification unit sends the optimized parameters back to SolidWorks to automatically rebuild the geometric model and re-perform finite element verification. When the maximum displacement of the new model drops to 1.5mm and the modal frequency increases to 38Hz, it is determined that the performance requirements are met.

[0166] Through the above technical solutions, this application realizes a closed-loop optimization mechanism based on data-driven, which effectively solves the problem of low efficiency of traditional manual iteration. By establishing an end-to-end mapping relationship between structural parameters and performance indicators, the design cycle is shortened. The introduction of the multi-attention mechanism strengthens the ability to learn the mechanical characteristics of key areas, making the optimization suggestions more in line with actual engineering needs. The automated verification process ensures that each structural modification can accurately evaluate the comprehensive performance under multiple working conditions, avoiding the limitations of single working condition optimization. The closed-loop link of forward design-simulation feedback-model update formed by the system provides reliable technical support for the intelligent optimization of complex load-bearing structures.

[0167] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0168] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems) and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0169] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0170] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0171] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A railway bogie structure optimization method based on deep learning, characterized in that: include: Collect the 3D geometric model of the railway bogie bolster to be optimized and build a finite element simulation model; Apply combined loads to the bolster structure based on finite element simulation to obtain the stress spectrum, displacement response spectrum, and modal characteristic spectrum of the bolster structure; Extracting mechanical indicators from the atlas to determine whether to perform structural optimization, the mechanical indicators include maximum equivalent stress, maximum displacement, first-order modal frequency, and structural stiffness value; When it is determined that structural optimization is to be performed, a deep learning model is trained based on historical structural samples, wherein the deep learning model includes a neural network model that introduces a multi-attention mechanism, the input of the deep learning model is a structural parameter tensor, and the output of the deep learning model is a structural performance prediction index; Inputting the bolster structural parameters to be optimized into the deep learning model to obtain parameter adjustment suggestions for optimization, the parameter adjustment suggestions including wall thickness changes, rib layout adjustments, and support plate position corrections; The bolster structure model is rebuilt based on the parameter adjustment suggestions, and finite element verification is performed to determine whether the optimized structure meets the target performance index requirements.

2. The railway bogie structure optimization method based on deep learning according to claim 1, characterized in that: When applying a combined load to the bolster structure based on finite element simulation to obtain a stress spectrum, a displacement response spectrum, and a modal characteristic spectrum of the bolster structure, the following steps are included: The 3D geometric model of the railway bogie bolster was meshed in the finite element analysis platform, and local mesh refinement was performed on the ear connection area, stiffener transition area, and hollow cavity. Setting material properties of the bolster structure, wherein the material properties include elastic modulus, Poisson's ratio, and yield strength; applying a combined load condition including a longitudinal load, a transverse load, and a vertical load to the bolster structure; Solve the finite element model and obtain the stress spectrum, displacement response spectrum and modal characteristic spectrum.

3. The railway bogie structure optimization method based on deep learning according to claim 2, characterized in that: Extracting mechanical indicators from the atlas to determine whether to perform structural optimization includes: Extract the maximum displacement amplitude and compare it with the allowable displacement tolerance of the structure installation; Extract the first-order modal frequency value to determine whether it falls within the operating excitation frequency range; Obtain the overall stiffness value of the structure based on the load-deformation relationship under loading conditions; If any mechanical index exceeds the safety threshold, it is determined that the railway bogie bolster structure needs to be optimized.

4. The railway bogie structure optimization method based on deep learning according to claim 3 is characterized in that: When training a deep learning model based on historical structure samples, it includes: Collect several bolster structure samples with different structural parameters and simulation results to construct a training dataset containing structural parameters, load conditions, and simulation performance indicators; The structural parameters include wall thickness distribution, rib positions, and the number and size of support plates, and the structural parameters are encoded in tensor form as model input.

5. The railway bogie structure optimization method based on deep learning according to claim 4, characterized in that: When training a deep learning model based on historical structure samples, it also includes: The deep learning model includes a backbone network and a multi-attention module. The backbone network is a residual convolutional neural network, and the multi-attention module includes a spatial attention mechanism and a channel attention mechanism, which are used to enhance the feature expression ability of high stress areas.

6. The railway bogie structure optimization method based on deep learning according to claim 5, characterized in that: When training a deep learning model based on historical structure samples, it also includes: The deep learning model is trained through supervised learning, and the optimization goal is to minimize the error between the structural performance prediction value and the finite element simulation result. The structural performance prediction value includes the maximum equivalent stress, maximum displacement, first-order modal frequency and overall structural stiffness value.

7. The railway bogie structure optimization method based on deep learning according to claim 6, characterized in that: Inputting the bolster structural parameters to be optimized into the deep learning model to obtain parameter adjustment suggestions for optimization includes: Encode the parameters of the bolster structure's wall thickness distribution, reinforcement configuration, and support plate size into the current structural parameter tensor; Inputting the current structural parameter tensor into the trained deep learning model to obtain the corresponding structural performance prediction index; The parameter adjustment suggestion is obtained based on the difference between the structural performance prediction index and the preset target performance value and in combination with historical structural samples.

8. The railway bogie structure optimization method based on deep learning according to claim 7, characterized in that: Reconstruct the bolster structure model based on the parameter adjustment suggestions and perform finite element verification to determine whether the optimized structure meets the target performance index requirements, including: When all indicators of the optimized structure meet the target performance indicator requirements, the structural optimization is considered successful; When any mechanical index of the optimized structure still exceeds the safety threshold, the model is optimized and the structure is optimized again.

9. The railway bogie structure optimization method based on deep learning according to claim 8, characterized in that: When performing model optimization and re-optimization of the structure, including: Compare the error between the finite element simulation results of the optimized structure and the predicted value of the deep learning model. If the error exceeds the set threshold, the structure sample is added to the training dataset and the deep learning model is incrementally trained. During incremental training, adjusting the channel weights of the multi-attention mechanism according to the error location; Based on the updated deep learning model, parameter adjustment suggestions are re-obtained and a new bolster structure model is constructed; The new model is imported into the finite element simulation platform for re-verification to determine whether it meets the performance target. If it still does not meet the target, the model optimization and structural optimization processes are continued until the preset number of times is reached or the target performance index requirements are met.

10. A railway bogie structure optimization system based on deep learning, used to apply the railway bogie structure optimization method based on deep learning according to any one of claims 1 to 9, characterized in that: include: A geometric modeling unit, used to acquire a three-dimensional geometric model of the railway bogie bolster to be optimized and to construct a parametric model for finite element analysis; a simulation analysis unit, configured to apply a combined load condition including a longitudinal load, a transverse load, and a vertical load to the bolster structure in a finite element analysis platform, and generate a stress spectrum, a displacement response spectrum, and a modal characteristic spectrum; a structural evaluation unit, configured to extract the maximum equivalent stress, maximum displacement, first-order modal frequency, and structural stiffness value from the atlas, and determine whether to perform structural optimization based on a preset safety threshold; A model training unit is used to construct a training data set based on historical structural samples and train a deep learning neural network model that introduces a multi-attention mechanism. The input of the deep learning neural network model is a structural parameter tensor, and the deep learning neural network model outputs a structural performance prediction indicator; The parameter optimization unit is used to input the parameter tensor of the structure to be optimized into the trained neural network model to obtain the structural performance prediction results, and output parameter adjustment suggestions for structural improvement based on the historical sample library. The parameter adjustment suggestions include changing the wall thickness, adjusting the rib layout, and correcting the support plate position; A verification unit is used to generate a new rocker structure model according to the parameter adjustment suggestion and perform finite element simulation to verify whether the optimized structure meets the target performance index requirements.

Images