A bearing roller intelligent reshaping design method

Abstract

The application discloses a kind of bearing roller intelligent modification design method, belong to bearing intelligent design technical field, including the following steps: step 1, the geometric parameter of bearing roller, working condition parameter, material parameter and process constraint parameter are collected, and input sample set is generated after preprocessing;Step 2, feature extraction is carried out to input sample set, and output comprehensive feature set;Step 3, the comprehensive feature set is handled, and the initial optimal modification parameter combination is output;Step 4, the initial optimal modification parameter combination is imported into mechanical field-thermal field-flow field coupling model and is simulated and verified, until the performance evaluation result meets the preset target.The application is through the multi-scale feature extraction of convolution attention feature aggregation module and the fine parameter optimization of self-adaptive transformer model, accurately identifies key geometric and physical characteristics, realizes the accurate matching of modification curve and actual working condition, to significantly improve the modification geometric precision and stress control precision.

Description

Technical Field

[0001] This invention relates to the field of intelligent bearing design technology, and in particular to an intelligent shaping design method for bearing rollers. Background Technology

[0002] Current bearing roller shaping primarily relies on traditional geometric shaping strategies and conventional optimization methods. Traditional approaches, such as logarithmic shaping, circular arc shaping, and elliptical shaping, essentially use pre-set fixed geometric curves to correct the roller generatrix shape to alleviate end stress concentration. These methods, based on empirical formulas and simplified mechanical models, can only adapt to ideal contact scenarios under a single operating condition, failing to consider the multi-physics coupling effects of mechanical, thermal, and flow fields, as well as the dynamic evolution characteristics of the operating conditions. Alternatively, a single basic deep learning model can be introduced into the shaping-aided design, but this only achieves simple geometric feature extraction or parameter adaptation under a single operating condition, exhibiting limitations across multiple scales. Bottlenecks such as insufficient capture of geometric feature associations, lack of modeling accuracy of the coupling relationship between shaping parameters and multiphysics fields, and fixed model structure that cannot be adaptively adjusted according to the complexity of shaping make it difficult to meet the dynamic optimization requirements under complex working conditions. In summary, existing technologies generally suffer from problems such as limited shaping accuracy, insufficient optimization of load distribution, and weak adaptability and generalization ability under working conditions. Traditional methods are prone to residual stress concentration phenomena, and uneven load distribution is likely to occur under complex working conditions. Conventional intelligent methods are difficult to take into account both local features and global associations, and cannot achieve adaptive shaping control under multiple working conditions, making it difficult to match the high-precision and long-life performance requirements of high-end equipment for bearings.

[0003] Chinese patent application CN110765617A, published on February 7, 2020, discloses a design method for logarithmic profile modification of cylindrical roller bearings based on lubrication theory. The method calculates the load distribution inside the cylindrical roller bearing using a static model; then, based on the worst-case operating conditions of the roller friction pair, it evaluates the lubrication state using elastohydrodynamic lubrication theory. Elastohydrodynamic lubrication theory mainly involves solving a set of nonlinear equations, including the Reynolds equation, geometric clearance equation, viscous-pressure equation, and load balance equation; finally, it optimizes the logarithmic profile of the rollers using a modification coefficient to achieve the best lubrication and contact state.

[0004] The patent application discloses a design method for logarithmic profile modification of cylindrical roller bearings based on lubrication theory, which can reduce bearing vibration and noise, and improve the load-carrying capacity and service life of cylindrical roller bearings. However, it is difficult to achieve precise matching between the profile modification curve and actual working conditions, and the geometric accuracy and stress control accuracy of the profile modification are not ideal. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, this invention provides an intelligent shaping design method for bearing rollers. This invention accurately identifies key geometric and physical features through multi-scale feature extraction of the convolutional attention feature aggregation module and fine parameter optimization of the adaptive transformer model, thereby achieving precise matching between the shaping curve and the actual working conditions, and significantly improving the geometric accuracy of shaping and stress control accuracy.

[0006] This invention is achieved through the following technical solution: A method for intelligent profile modification design of bearing rollers includes the following steps: Step 1: Collect the geometric parameters, operating parameters, material parameters, and process constraint parameters of the bearing rollers, and generate the input sample set after preprocessing; Step 2: Extract features from the input sample set using the convolutional attention feature aggregation module, and output a comprehensive feature set; Step 3: Process the comprehensive feature set through the adaptive transformer model, and output the initial optimal combination of shaping parameters by combining the multi-objective optimization screening strategy; Step 4: Import the initial optimal shaping parameter combination into the mechanical field-thermal field-flow field coupled model for simulation verification. Obtain performance deviation data through multi-dimensional performance evaluation, and use the performance deviation data as feedback information for iterative optimization until the performance evaluation results meet the preset target.

[0007] In step 2, the convolutional attention feature aggregation module is a two-stage cascaded moving inverted bottleneck convolution module, including a pre-stage moving inverted bottleneck convolution and a post-stage moving inverted bottleneck convolution. The pre-stage moving inverted bottleneck convolution extracts the macroscopic geometric features and basic physical features of the roller through a combination of large convolutional kernels and dilated convolutions. The post-stage moving inverted bottleneck convolution extracts microscopic features through a combination of small convolutional kernels. A convolutional block attention module is embedded between the pre-stage and post-stage moving inverted bottleneck convolutions, and cross-scale feature fusion is completed through residual connections to output a comprehensive feature set.

[0008] The large convolution kernel in the preceding moving inverted bottleneck convolution is a 3×3 convolution kernel. Its macroscopic geometric features include the basic shape of the busbar, the convexity distribution trend, the overall size adaptability, and the initial shape deviation. Its basic physical features include the load transfer path, the range of the initial contact area, the stress distribution trend, and the adaptability of the material mechanical properties.

[0009] The micro-features in the subsequent moving inverted bottleneck convolution include surface profile fluctuations, local geometric abrupt changes, characteristic signals of potential stress concentration regions, material micro-property heterogeneity, and processing residual defect characteristics.

[0010] In step 3, the adaptive transformer model uses trainable relative position encoding, mines the coupling relationship between shaping parameters and the nonlinear mapping relationship between parameters and performance through a multi-head self-attention mechanism, and dynamically adjusts the encoding level based on the credibility evaluation index.

[0011] The convolutional block attention module dynamically allocates feature weights through the cascading effect of channel attention branches and spatial attention branches. The channel attention branch quantifies the weights of each feature channel through the fusion operation of global average pooling and max pooling, and the spatial attention branch generates a two-dimensional spatial weight map through convolution operations.

[0012] The residual connection refers to dimensional alignment and information fusion of the macroscopic features extracted by the preceding moving-inverted bottleneck convolution and the microscopic features extracted by the subsequent moving-inverted bottleneck convolution, forming a multi-dimensional comprehensive feature set that includes global geometric patterns and local key details.

[0013] The credibility evaluation index is composed of a weighted average of feature matching degree, constraint satisfaction degree, and performance target fit degree.

[0014] The multi-objective optimization screening strategy uses a non-dominated sorting algorithm to screen candidate shaping parameter schemes.

[0015] The shaping parameter combination includes the mathematical expression of the shaping curve, discrete coordinate points, convexity distribution curve, and transition region size.

[0016] In step 4, the multi-dimensional performance evaluation includes load distribution uniformity, stress control, lubrication stability, fatigue life prediction, and geometric fit.

[0017] The geometric parameters include the nominal diameter of the roller, the effective contact length, the radius of curvature of the raceway, the nominal dimensions of the inner and outer rings of the bearing, the assembly clearance, and the initial shape deviation data of the roller generatrix.

[0018] The operating parameters include rated radial load, rated axial load, instantaneous impact load range, operating speed range, ambient temperature, roller operating temperature threshold, lubricating medium type and viscosity grade.

[0019] The beneficial effects of this invention are mainly reflected in the following aspects: 1. This invention, through multi-scale feature extraction of the convolutional attention feature aggregation module and fine parameter optimization of the adaptive transformer model, accurately identifies key geometric and physical features, and achieves precise matching between the shaping curve and the actual working conditions, thereby significantly improving the geometric accuracy of shaping and the accuracy of stress control.

[0020] 2. This invention can effectively alleviate edge stress concentration and optimize load distribution uniformity. Through mechanical field-thermal field-flow field coupling verification and closed-loop iterative feedback mechanism, the shaping scheme can fully adapt to complex and variable working conditions, significantly extend bearing fatigue life and reduce total life cycle cost.

[0021] 3. In this invention, the convolutional attention feature aggregation module can accurately locate the potential stress concentration area and contact boundary on the roller surface. Combined with feedback iterative adjustment through multi-physics field coupling verification, the contact stress distribution of the roller after shaping is more uniform, effectively eliminating the edge stress concentration phenomenon that is difficult to avoid by traditional methods and reducing local peak stress.

[0022] 4. This invention uses multi-physics coupling simulation to restore the bearing operating state under real working conditions, so that the modification scheme can adapt to the bearing requirements under different loads, speeds and lubrication conditions, effectively improving the bearing stability and operational reliability of bearings under complex working conditions such as heavy loads and high speeds.

[0023] 5. In this invention, the adaptive transformer model dynamically adjusts the coding level based on the credibility evaluation index and automatically adapts the model depth according to the shaping complexity. It does not require redesigning the model structure for different working conditions and achieves a dynamic balance between shaping accuracy and computational efficiency. 6. This invention, through precise shaping, can effectively reduce the risk of localized excessive wear and contact fatigue damage on the roller surface, which is beneficial to extending the fatigue life of the bearing throughout its entire life cycle and ensuring operational reliability.

[0024] 7. This invention directly feeds back the multiphysics simulation verification results to the model, reducing the repetitive experiments and processing verification steps that rely on experience trial and error in traditional methods, thereby effectively shortening the design and development cycle of the modification scheme. Attached Figure Description

[0025] The present invention will now be further described in detail with reference to the accompanying drawings and specific embodiments, wherein: Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0026] Example 1 See Figure 1 A method for intelligent profile modification design of bearing rollers includes the following steps: Step 1: Collect the geometric parameters, operating parameters, material parameters, and process constraint parameters of the bearing rollers, and generate the input sample set after preprocessing; Step 2: Extract features from the input sample set using the convolutional attention feature aggregation module, and output a comprehensive feature set; Step 3: Process the comprehensive feature set through the adaptive transformer model, and output the initial optimal combination of shaping parameters by combining the multi-objective optimization screening strategy; Step 4: Import the initial optimal shaping parameter combination into the mechanical field-thermal field-flow field coupled model for simulation verification. Obtain performance deviation data through multi-dimensional performance evaluation, and use the performance deviation data as feedback information for iterative optimization until the performance evaluation results meet the preset target.

[0027] This embodiment is the most basic implementation method. Through multi-scale feature extraction of the convolutional attention feature aggregation module and fine parameter optimization of the adaptive transformer model, key geometric and physical features are accurately identified, and the precise matching between the shaping curve and the actual working conditions is achieved, thereby significantly improving the shaping geometric accuracy and stress control accuracy.

[0028] Example 2 See Figure 1 A method for intelligent profile modification design of bearing rollers includes the following steps: Step 1: Collect the geometric parameters, operating parameters, material parameters, and process constraint parameters of the bearing rollers, and generate the input sample set after preprocessing; Step 2: Extract features from the input sample set using the convolutional attention feature aggregation module, and output a comprehensive feature set; Step 3: Process the comprehensive feature set through the adaptive transformer model, and output the initial optimal combination of shaping parameters by combining the multi-objective optimization screening strategy; Step 4: Import the initial optimal shaping parameter combination into the mechanical field-thermal field-flow field coupled model for simulation verification. Obtain performance deviation data through multi-dimensional performance evaluation, and use the performance deviation data as feedback information for iterative optimization until the performance evaluation results meet the preset target.

[0029] Preferably, in step 2, the convolutional attention feature aggregation module is a two-stage cascaded moving inverted bottleneck convolution module, including a pre-stage moving inverted bottleneck convolution and a post-stage moving inverted bottleneck convolution. The pre-stage moving inverted bottleneck convolution extracts the macroscopic geometric features and basic physical features of the roller through a combination of large convolutional kernels and dilated convolutions. The post-stage moving inverted bottleneck convolution extracts microscopic features through a combination of small convolutional kernels. A convolutional block attention module is embedded between the pre-stage and post-stage moving inverted bottleneck convolutions, and cross-scale feature fusion is completed through residual connections to output a comprehensive feature set.

[0030] This embodiment is a preferred implementation method, which can effectively alleviate edge stress concentration and optimize load distribution uniformity. Through mechanical field-thermal field-flow field coupling verification and closed-loop iterative feedback mechanism, the shaping scheme can fully adapt to complex and variable working conditions, significantly extend bearing fatigue life and reduce total life cycle cost.

[0031] Example 3 See Figure 1 A method for intelligent profile modification design of bearing rollers includes the following steps: Step 1: Collect the geometric parameters, operating parameters, material parameters, and process constraint parameters of the bearing rollers, and generate the input sample set after preprocessing; Step 2: Extract features from the input sample set using the convolutional attention feature aggregation module, and output a comprehensive feature set; Step 3: Process the comprehensive feature set through the adaptive transformer model, and output the initial optimal combination of shaping parameters by combining the multi-objective optimization screening strategy; Step 4: Import the initial optimal shaping parameter combination into the mechanical field-thermal field-flow field coupled model for simulation verification. Obtain performance deviation data through multi-dimensional performance evaluation, and use the performance deviation data as feedback information for iterative optimization until the performance evaluation results meet the preset target.

[0032] In step 2, the convolutional attention feature aggregation module is a two-stage cascaded moving inverted bottleneck convolution module, including a pre-stage moving inverted bottleneck convolution and a post-stage moving inverted bottleneck convolution. The pre-stage moving inverted bottleneck convolution extracts the macroscopic geometric features and basic physical features of the roller through a combination of large convolutional kernels and dilated convolutions. The post-stage moving inverted bottleneck convolution extracts microscopic features through a combination of small convolutional kernels. A convolutional block attention module is embedded between the pre-stage and post-stage moving inverted bottleneck convolutions, and cross-scale feature fusion is completed through residual connections to output a comprehensive feature set.

[0033] The large convolution kernel in the preceding moving inverted bottleneck convolution is a 3×3 convolution kernel. Its macroscopic geometric features include the basic shape of the busbar, the convexity distribution trend, the overall size adaptability, and the initial shape deviation. Its basic physical features include the load transfer path, the range of the initial contact area, the stress distribution trend, and the adaptability of the material mechanical properties.

[0034] The micro-features in the subsequent moving inverted bottleneck convolution include surface profile fluctuations, local geometric abrupt changes, characteristic signals of potential stress concentration regions, material micro-property heterogeneity, and processing residual defect characteristics.

[0035] This embodiment is a preferred implementation. The convolutional attention feature aggregation module can accurately locate the potential stress concentration area and contact boundary on the roller surface. Combined with feedback iterative adjustment of multi-physics field coupling verification, the contact stress distribution of the roller after shaping is more uniform, effectively eliminating the edge stress concentration phenomenon that is difficult to avoid by traditional methods and reducing local peak stress.

[0036] Example 4 See Figure 1 A method for intelligent profile modification design of bearing rollers includes the following steps: Step 1: Collect the geometric parameters, operating parameters, material parameters, and process constraint parameters of the bearing rollers, and generate the input sample set after preprocessing; Step 2: Extract features from the input sample set using the convolutional attention feature aggregation module, and output a comprehensive feature set; Step 3: Process the comprehensive feature set through the adaptive transformer model, and output the initial optimal combination of shaping parameters by combining the multi-objective optimization screening strategy; Step 4: Import the initial optimal shaping parameter combination into the mechanical field-thermal field-flow field coupled model for simulation verification. Obtain performance deviation data through multi-dimensional performance evaluation, and use the performance deviation data as feedback information for iterative optimization until the performance evaluation results meet the preset target.

[0037] In step 2, the convolutional attention feature aggregation module is a two-stage cascaded moving inverted bottleneck convolution module, including a pre-stage moving inverted bottleneck convolution and a post-stage moving inverted bottleneck convolution. The pre-stage moving inverted bottleneck convolution extracts the macroscopic geometric features and basic physical features of the roller through a combination of large convolutional kernels and dilated convolutions. The post-stage moving inverted bottleneck convolution extracts microscopic features through a combination of small convolutional kernels. A convolutional block attention module is embedded between the pre-stage and post-stage moving inverted bottleneck convolutions, and cross-scale feature fusion is achieved through residual connections to output a comprehensive feature set. The large convolutional kernel in the pre-stage moving inverted bottleneck convolution is a 3×3 convolutional kernel. The macroscopic geometric features include the basic shape of the busbar, the convexity distribution trend, the overall size adaptability, and the initial shape deviation. The basic physical features include the load transfer path, the initial contact area range, the stress distribution trend, and the material mechanical property adaptability. The microscopic features in the post-stage moving inverted bottleneck convolution include surface contour fluctuations, local geometric abrupt changes, feature signals of potential stress concentration areas, material microscopic property heterogeneity, and processing residual defect features.

[0038] Preferably, in step 3, the adaptive transformer model uses trainable relative position encoding, mines the coupling relationship between shaping parameters and the nonlinear mapping relationship between parameters and performance through a multi-head self-attention mechanism, and dynamically adjusts the encoding level based on the credibility evaluation index.

[0039] The convolutional block attention module dynamically allocates feature weights through the cascading effect of channel attention branches and spatial attention branches. The channel attention branch quantifies the weights of each feature channel through the fusion operation of global average pooling and max pooling, and the spatial attention branch generates a two-dimensional spatial weight map through convolution operations.

[0040] This embodiment is a preferred implementation method. By using multiphysics coupling simulation to restore the bearing operating state under real working conditions, the modification scheme can be adapted to the bearing requirements under different loads, speeds and lubrication conditions, effectively improving the bearing stability and operational reliability of the bearing under complex working conditions such as heavy load and high speed.

[0041] Example 5 See Figure 1 A method for intelligent profile modification design of bearing rollers includes the following steps: Step 1: Collect the geometric parameters, operating parameters, material parameters, and process constraint parameters of the bearing rollers, and generate the input sample set after preprocessing; Step 2: Extract features from the input sample set using the convolutional attention feature aggregation module, and output a comprehensive feature set; Step 3: Process the comprehensive feature set through the adaptive transformer model, and output the initial optimal combination of shaping parameters by combining the multi-objective optimization screening strategy; Step 4: Import the initial optimal shaping parameter combination into the mechanical field-thermal field-flow field coupled model for simulation verification. Obtain performance deviation data through multi-dimensional performance evaluation, and use the performance deviation data as feedback information for iterative optimization until the performance evaluation results meet the preset target.

[0042] In step 2, the convolutional attention feature aggregation module is a two-stage cascaded moving inverted bottleneck convolution module, including a pre-stage moving inverted bottleneck convolution and a post-stage moving inverted bottleneck convolution. The pre-stage moving inverted bottleneck convolution extracts the macroscopic geometric features and basic physical features of the roller through a combination of large convolutional kernels and dilated convolutions. The post-stage moving inverted bottleneck convolution extracts microscopic features through a combination of small convolutional kernels. A convolutional block attention module is embedded between the pre-stage and post-stage moving inverted bottleneck convolutions, and cross-scale feature fusion is completed through residual connections to output a comprehensive feature set.

[0043] The large convolution kernel in the preceding moving inverted bottleneck convolution is a 3×3 convolution kernel. Its macroscopic geometric features include the basic shape of the busbar, the convexity distribution trend, the overall size adaptability, and the initial shape deviation. Its basic physical features include the load transfer path, the range of the initial contact area, the stress distribution trend, and the adaptability of the material mechanical properties.

[0044] The micro-features in the subsequent moving inverted bottleneck convolution include surface profile fluctuations, local geometric abrupt changes, characteristic signals of potential stress concentration regions, material micro-property heterogeneity, and processing residual defect characteristics.

[0045] In step 3, the adaptive transformer model uses trainable relative position encoding, mines the coupling relationship between shaping parameters and the nonlinear mapping relationship between parameters and performance through a multi-head self-attention mechanism, and dynamically adjusts the encoding level based on the credibility evaluation index.

[0046] The convolutional block attention module dynamically allocates feature weights through the cascading effect of channel attention branches and spatial attention branches. The channel attention branch quantifies the weights of each feature channel through the fusion operation of global average pooling and max pooling, and the spatial attention branch generates a two-dimensional spatial weight map through convolution operations.

[0047] The residual connection refers to dimensional alignment and information fusion of the macroscopic features extracted by the preceding moving-inverted bottleneck convolution and the microscopic features extracted by the subsequent moving-inverted bottleneck convolution, forming a multi-dimensional comprehensive feature set containing global geometric patterns and local key details. The credibility evaluation index is a weighted average of feature matching degree, constraint satisfaction degree, and performance target fit degree. The multi-objective optimization screening strategy uses a non-dominated ranking algorithm to screen candidate shaping parameter schemes. The shaping parameter combination includes the mathematical expression of the shaping curve, discrete coordinate points, convexity distribution curve, and transition region size.

[0048] This embodiment is a preferred implementation. The adaptive transformer model dynamically adjusts the coding level based on the credibility evaluation index and automatically adapts the model depth according to the shaping complexity. It does not require redesigning the model structure for different working conditions and achieves a dynamic balance between shaping accuracy and computational efficiency. Example 6 See Figure 1 A method for intelligent profile modification design of bearing rollers includes the following steps: Step 1: Collect the geometric parameters, operating parameters, material parameters, and process constraint parameters of the bearing rollers, and generate the input sample set after preprocessing; Step 2: Extract features from the input sample set using the convolutional attention feature aggregation module, and output a comprehensive feature set; Step 3: Process the comprehensive feature set through the adaptive transformer model, and output the initial optimal combination of shaping parameters by combining the multi-objective optimization screening strategy; Step 4: Import the initial optimal shaping parameter combination into the mechanical field-thermal field-flow field coupled model for simulation verification. Obtain performance deviation data through multi-dimensional performance evaluation, and use the performance deviation data as feedback information for iterative optimization until the performance evaluation results meet the preset target.

[0049] In step 2, the convolutional attention feature aggregation module is a two-stage cascaded moving inverted bottleneck convolution module, including a pre-stage moving inverted bottleneck convolution and a post-stage moving inverted bottleneck convolution. The pre-stage moving inverted bottleneck convolution extracts the macroscopic geometric features and basic physical features of the roller through a combination of large convolutional kernels and dilated convolutions. The post-stage moving inverted bottleneck convolution extracts microscopic features through a combination of small convolutional kernels. A convolutional block attention module is embedded between the pre-stage and post-stage moving inverted bottleneck convolutions, and cross-scale feature fusion is completed through residual connections to output a comprehensive feature set.

[0050] The large convolution kernel in the preceding moving inverted bottleneck convolution is a 3×3 convolution kernel. Its macroscopic geometric features include the basic shape of the busbar, the convexity distribution trend, the overall size adaptability, and the initial shape deviation. Its basic physical features include the load transfer path, the range of the initial contact area, the stress distribution trend, and the adaptability of the material mechanical properties.

[0051] The micro-features in the subsequent moving inverted bottleneck convolution include surface profile fluctuations, local geometric abrupt changes, characteristic signals of potential stress concentration regions, material micro-property heterogeneity, and processing residual defect characteristics.

[0052] In step 3, the adaptive transformer model uses trainable relative position encoding, mines the coupling relationship between shaping parameters and the nonlinear mapping relationship between parameters and performance through a multi-head self-attention mechanism, and dynamically adjusts the encoding level based on the credibility evaluation index.

[0053] The convolutional block attention module dynamically allocates feature weights through the cascading effect of channel attention branches and spatial attention branches. The channel attention branch quantifies the weights of each feature channel through the fusion operation of global average pooling and max pooling, and the spatial attention branch generates a two-dimensional spatial weight map through convolution operations.

[0054] The residual connection refers to dimensional alignment and information fusion of the macroscopic features extracted by the preceding moving-inverted bottleneck convolution and the microscopic features extracted by the subsequent moving-inverted bottleneck convolution, forming a multi-dimensional comprehensive feature set that includes global geometric patterns and local key details.

[0055] The credibility evaluation index is composed of a weighted average of feature matching degree, constraint satisfaction degree, and performance target fit degree.

[0056] The multi-objective optimization screening strategy uses a non-dominated sorting algorithm to screen candidate shaping parameter schemes.

[0057] The shaping parameter combination includes the mathematical expression of the shaping curve, discrete coordinate points, convexity distribution curve, and transition region size.

[0058] More preferably, in step 4, the multi-dimensional performance evaluation includes load distribution uniformity, stress control, lubrication stability, fatigue life prediction, and geometric adaptability.

[0059] The geometric parameters include the nominal diameter of the roller, the effective contact length, the radius of curvature of the raceway, the nominal dimensions of the inner and outer rings of the bearing, the assembly clearance, and the initial shape deviation data of the roller generatrix.

[0060] The operating parameters include rated radial load, rated axial load, instantaneous impact load range, operating speed range, ambient temperature, roller operating temperature threshold, lubricating medium type and viscosity grade.

[0061] This embodiment is a preferred implementation method. By precisely shaping the roller, the risk of localized excessive wear and contact fatigue damage on the roller surface can be effectively reduced, which is beneficial to extending the fatigue life of the bearing throughout its entire life cycle and ensuring operational reliability.

[0062] By directly feeding the multiphysics simulation verification results back to the model, the repetitive experiments and processing verification steps that rely on experience-based trial and error in traditional methods are reduced, thereby effectively shortening the design and development cycle of the modification scheme.

[0063] The working principle of the adaptive transformer model is as follows: Abandoning the limitations of traditional fixed absolute position encoding, this method introduces trainable relative position encoding. By learning the relative spatial relationships between feature points on the roller surface, it accurately captures the correlation and smoothness requirements of continuous geometric features. Compared to absolute position encoding, it does not rely on a fixed coordinate system and can adaptively adapt to the geometric modeling needs of rollers of different sizes and types. It also avoids disrupting the continuity of the shaping curve, ensuring that the optimized shaping curve meets the smoothness requirements of ultra-precision machining, thus balancing geometric rationality and machining feasibility.

[0064] A dynamic coding hierarchy control mechanism is constructed based on a credibility evaluation index to achieve a dynamic balance between shaping accuracy and computational efficiency. The specific logic is as follows: based on the initial setting of shaping complexity, the initial coding hierarchy is preset. Typically consisting of 2-4 layers, the model inference outputs reshaping parameter schemes and calculates their credibility. Credibility is determined by a weighted average of feature matching degree, constraint satisfaction degree, and performance target fit degree; the coding level is gradually increased to Repeat the reasoning process and calculate the credibility. Solve for the difference in confidence levels between the two tests. ;like If the value exceeds the preset threshold, it indicates that the current shaping complexity is high, and the encoding depth needs to be increased to strengthen the feature association modeling capability. Continue iteratively increasing the layer until the difference meets the requirements. If the value is below the threshold, the current level is maintained to avoid over-computation leading to efficiency loss, thus achieving intelligent control of complexity adaptation and dynamic level adjustment.

[0065] A multi-parallel attention head structure is employed, with each attention head focusing on capturing a specific type of parameter correlation, enabling simultaneous mining of multi-dimensional coupling relationships. Specifically, some attention heads focus on the coupling effects between shaping parameters, clarifying the influence of parameter adjustments on each other; the remaining attention heads focus on the nonlinear mapping relationship between shaping parameters and core performance indicators, establishing a precise parameter-performance correlation model. Through dynamic allocation of attention weights, multiple sets of candidate shaping parameter schemes that meet basic process constraints are generated, providing sufficient basis for subsequent screening.

[0066] By constructing a multi-objective optimization function with optimal load distribution uniformity, minimum maximum contact stress, and maximum fatigue life as core objectives, and introducing processing difficulty and manufacturing cost as auxiliary constraints, a non-dominated sorting algorithm is used to screen candidate solutions from multiple dimensions. During the screening process, invalid solutions that violate process constraints and assembly boundary conditions are first eliminated. Then, the valid solutions are sorted according to Pareto optimality, and the solution with the best overall performance is selected as the initial optimal combination of shaping parameters. The mathematical expression of the shaping curve, discrete coordinate points, convexity distribution curve, transition region size, and surface accuracy control requirements are clearly defined.

Claims

1. A method for intelligent profile modification design of bearing rollers, characterized in that, Includes the following steps: Step 1: Collect the geometric parameters, operating parameters, material parameters, and process constraint parameters of the bearing rollers, and generate the input sample set after preprocessing; Step 2: Extract features from the input sample set using the convolutional attention feature aggregation module, and output a comprehensive feature set; Step 3: Process the comprehensive feature set through the adaptive transformer model, and output the initial optimal combination of shaping parameters by combining the multi-objective optimization screening strategy; Step 4: Import the initial optimal shaping parameter combination into the mechanical field-thermal field-flow field coupled model for simulation verification. Obtain performance deviation data through multi-dimensional performance evaluation, and use the performance deviation data as feedback information for iterative optimization until the performance evaluation results meet the preset target.

2. The intelligent profile design method for bearing rollers according to claim 1, characterized in that: In step 2, the convolutional attention feature aggregation module is a two-stage cascaded moving inverted bottleneck convolution module, including a pre-stage moving inverted bottleneck convolution and a post-stage moving inverted bottleneck convolution. The pre-stage moving inverted bottleneck convolution extracts the macroscopic geometric features and basic physical features of the roller through a combination of large convolutional kernels and dilated convolutions. The post-stage moving inverted bottleneck convolution extracts microscopic features through a combination of small convolutional kernels. A convolutional block attention module is embedded between the pre-stage and post-stage moving inverted bottleneck convolutions, and cross-scale feature fusion is completed through residual connections to output a comprehensive feature set.

3. The intelligent profile design method for bearing rollers according to claim 2, characterized in that: The large convolution kernel in the preceding moving inverted bottleneck convolution is a 3×3 convolution kernel. Its macroscopic geometric features include the basic shape of the busbar, the convexity distribution trend, the overall size adaptability, and the initial shape deviation. Its basic physical features include the load transfer path, the range of the initial contact area, the stress distribution trend, and the adaptability of the material mechanical properties.

4. The intelligent profile design method for bearing rollers according to claim 2, characterized in that: The micro-features in the subsequent moving inverted bottleneck convolution include surface profile fluctuations, local geometric abrupt changes, characteristic signals of potential stress concentration regions, material micro-property heterogeneity, and processing residual defect characteristics.

5. The intelligent profile design method for bearing rollers according to claim 1, characterized in that: In step 3, the adaptive transformer model uses trainable relative position encoding, mines the coupling relationship between shaping parameters and the nonlinear mapping relationship between parameters and performance through a multi-head self-attention mechanism, and dynamically adjusts the encoding level based on the credibility evaluation index.

6. The intelligent profile design method for bearing rollers according to claim 2, characterized in that: The convolutional block attention module dynamically allocates feature weights through the cascading effect of channel attention branches and spatial attention branches. The channel attention branch quantifies the weights of each feature channel through the fusion operation of global average pooling and max pooling, and the spatial attention branch generates a two-dimensional spatial weight map through convolution operations.

7. The intelligent profile design method for bearing rollers according to claim 2, characterized in that: The residual connection refers to dimensional alignment and information fusion of the macroscopic features extracted by the preceding moving-inverted bottleneck convolution and the microscopic features extracted by the subsequent moving-inverted bottleneck convolution, forming a multi-dimensional comprehensive feature set that includes global geometric patterns and local key details.

8. The intelligent profile design method for bearing rollers according to claim 5, characterized in that: The credibility evaluation index is composed of a weighted average of feature matching degree, constraint satisfaction degree, and performance target fit degree.

9. The intelligent profile design method for bearing rollers according to claim 1, characterized in that: The multi-objective optimization screening strategy uses a non-dominated sorting algorithm to screen candidate shaping parameter schemes.

10. The intelligent profile design method for bearing rollers according to claim 1, characterized in that: The shaping parameter combination includes the mathematical expression of the shaping curve, discrete coordinate points, convexity distribution curve, and transition region size.

11. The intelligent profile design method for bearing rollers according to claim 1, characterized in that: In step 4, the multi-dimensional performance evaluation includes load distribution uniformity, stress control, lubrication stability, fatigue life prediction, and geometric fit.

12. The intelligent profile design method for bearing rollers according to claim 1, characterized in that: The geometric parameters include the nominal diameter of the roller, the effective contact length, the radius of curvature of the raceway, the nominal dimensions of the inner and outer rings of the bearing, the assembly clearance, and the initial shape deviation data of the roller generatrix.

13. The intelligent profile design method for bearing rollers according to claim 1, characterized in that: The operating parameters include rated radial load, rated axial load, instantaneous impact load range, operating speed range, ambient temperature, roller operating temperature threshold, lubricating medium type and viscosity grade.

Images