A neural network training method, electronic device, and readable storage medium for preparing axles of special vehicles with high coaxiality

Abstract

The present invention discloses a neural network training method for preparing a high-coaxiality special vehicle axle. The method uses a neural network model to obtain the pre-welding pre-assembly angle of the axle body and the shaft sleeve, offsets the coaxiality deviation caused by welding deformation, and directly produces a high-coaxiality special vehicle axle. The method includes the steps of obtaining a test data training set and a simulation data training set as mixed training samples, establishing an initial neural network model, training the neural network model, obtaining the pre-assembly angle, guiding actual production, and obtaining a high-coaxiality axle. The present invention utilizes the welding deformation that inevitably exists during the processing process, adopts a neural network training method, uses mixed data as training samples, realizes the prediction of the pre-assembly angle, and then adjusts the pre-welding posture to directly obtain a high-coaxiality axle with an error of only 15 to 30 degrees. This reduces the difficulty and cost of obtaining training samples, and can achieve accurate prediction of the pre-assembly angle, weakening the measurement of welding deformation, which is more conducive to accurately guiding the production of high-coaxiality axles.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence manufacturing and relates to the preparation of high-coaxiality special vehicle axles. Specifically, it relates to a neural network training method, electronic equipment, and readable storage medium for the preparation of high-coaxiality special vehicle axles. Background Art

[0002] The axle is a crucial component of the vehicle chassis, supporting and protecting the powertrain and buffering road impacts. The quality of the connection between the bushing and the axle body directly impacts the overall axle assembly. Coaxiality between the two is a key quality indicator, affecting not only the operation of the gears within the axle housing but also the bearing capacity and internal stress distribution of the axle housing under vehicle load.

[0003] Compared to the design and optimization of automotive axles, the machining level of the connection between the rear axle body and the bushing for special-purpose vehicles is relatively low. The coaxiality between the bushing and the axle body, as assembled and corrected by welding, can only reach approximately 150μm. Due to the longer and wider dimensions of special-purpose vehicles, the greater load carrying capacity, and the harsher road conditions they operate in, the existing coaxiality accuracy, while sufficient for most light or narrow-body vehicles, cannot meet the axle coaxiality requirements of large-tonnage or wide-body special-purpose vehicles.

[0004] There are currently two main ways to improve the coaxiality of axles. One is through manual post-weld correction, and the other is through machine cutting and reprocessing. Both methods perform welding operations on the basis of ensuring high-precision coaxiality before welding. However, the welding process is extremely energy-intensive, which inevitably produces welding deformation, and the deformation is extremely difficult to accurately control and correct through existing adjustment methods. Therefore, after welding, the main body of the axle and the bushing will inevitably deform at the welding surface, causing the coaxiality of the two to deviate again, requiring correction again. However, on this basis, both correction methods inevitably affect the overall performance of the axle. Among them, the manual post-weld correction method is to correct the coaxiality error by manual tapping while the weld is still in a high-temperature state. The process requires repeated moving, tapping, and measuring of the bridge shell, which is not only inefficient and unable to accurately control the coaxiality error, but also may introduce stress when tapping the weld, affecting the performance of the axle. Although the machine tool cutting and reprocessing method can control the coaxiality of the bridge housing more accurately, surface processing is prone to defects such as work hardening, stress concentration, and the effective penetration layer thickness on the axle surface not meeting the standards, which affects the overall performance of the bridge housing. In addition, the equipment cost is directly related to the size of the bridge housing. The larger the axle size, the more expensive the machine tool equipment required. Summary of the Invention

[0005] (1) Technical issues

[0006] The technical problem to be solved by the present invention is: the existing coaxiality is mostly improved by mechanical correction after welding, but it is extremely difficult to accurately control and correct the welding deformation through the existing adjustment method. The correction process is prone to introduce additional stress, which affects the quality of the axle. It is difficult to prepare an axle with high coaxiality, high stability and high reliability after welding, and it is especially difficult to prepare an axle that meets the use requirements of large-tonnage or wide-body special vehicles.

[0007] (2) Technical solution

[0008] The present invention is achieved through the following technical solutions:

[0009] The present invention proposes a neural network training method for the preparation of high-coaxiality special vehicle axles. By using a neural network model to obtain the pre-assembly angle between the special vehicle axle body and the shaft sleeve before welding, the method is used to offset the coaxiality deviation caused by welding deformation during actual production, and directly manufacture high-coaxiality special vehicle axles. The method includes the following steps:

[0010] S1, obtain the data training set for neural network model training:

[0011] When obtaining actual test data, the welding starting position and pre-weld coaxiality are used as the main test factors. Combined with the materials and dimensions of the axle body and the bushing, the weld surface shape, and the welding parameters, a series of single-variable tests are designed. The actual post-weld coaxiality values of the axle are collected as test results. The test factors and test results are mapped one-to-one to generate a test data training set for neural network training.

[0012] When acquiring simulation data, the same test factors as those used to acquire actual test data are used as input conditions for the welding heat source model, and simulated values of post-weld coaxiality under the corresponding conditions are obtained. The difference between the actual and simulated values of post-weld coaxiality is calculated, and the welding heat source model is corrected until the error between the actual and simulated values is within 5%, and the correction is stopped. Based on the corrected welding heat source model, simulated values of post-weld coaxiality under different simulation test conditions are obtained, and a simulation data training set for neural network training is generated.

[0013] S2, establish the initial neural network model:

[0014] Using the experimental data training set obtained by S1 as the training sample, the initial neural network model is constructed, and the initial weights and thresholds of the neural network are obtained;

[0015] S3, training the neural network model to obtain the pre-assembly angle prediction value;

[0016] The experimental data training set and the simulation data training set obtained by S1 are used as training samples and imported into the initial neural network model. The welding process parameters, post-weld coaxiality and welding starting position coordinates are used as input data, and the pre-assembly angle is used as output data. The neural network model is trained and optimized until convergence, and the pre-assembly angle is finally predicted.

[0017] S4, guides actual production and obtains high coaxiality special vehicle axles:

[0018] According to the pre-assembly angle prediction value obtained in step S3, the pre-assembly angle between the axle body and the bushing is adjusted, and welding is performed to directly obtain a special vehicle axle with high coaxiality.

[0019] Based on the above technical solution, by adopting the neural network training method, it is only necessary to collect a limited number of test data for axles of different specifications, and combine the simulation model to obtain a mixed data set of experiments and simulations as training samples. By using easily accessible welding process parameters, welding starting position coordinates and post-weld coaxiality data, the prediction of the pre-assembly angle that completely offsets the welding deformation can be achieved. Then, before the actual welding process, the pre-assembly angle of the axle body and the shaft sleeve can be adjusted, and a special vehicle axle with high coaxiality can be directly obtained without secondary correction. This not only reduces the difficulty and cost of obtaining the training data set, but also improves the accuracy of the pre-assembly angle prediction, thereby accurately eliminating the post-weld coaxiality deviation caused by welding deformation.

[0020] Furthermore, it also includes step S5, continuous optimization of the neural network model: when executing S4, the real data of the pre-assembly angle, welding starting position coordinates and post-weld coaxiality are synchronously collected to generate a production data optimization set. The production data optimization set is used as a training sample to perform online training and optimization on the neural network model to obtain a continuously optimized neural network model. The purpose of this step is to use the production data optimization set to provide real-time feedback on welding parameter fine-tuning, welding head wear and other equipment status changes, and to perform online training and optimization on the neural network model so that the neural network model always fits the pre-assembly angle value accurately predicted by the current processing status.

[0021] Furthermore, before generating the production data optimization set, the neural network model automatically checks whether the post-weld coaxiality obtained based on the predicted pre-assembly angle meets the requirements. When the requirements are met, part of the actual production data is intelligently selected and included in the production data optimization set; when the requirements are not met, all the collected actual production data are included in the production data optimization set for online optimization training of the neural network model, and the optimization of the neural network model is performed once. This process can effectively reduce the amount of duplicate data in the production data optimization set while ensuring real-time feedback of state changes, thereby reducing data storage requirements and ensuring the long-term stability of the algorithm.

[0022] Furthermore, when obtaining the simulated data training set, on the basis of the modified welding heat source model, the simulated value of the post-weld coaxiality is again obtained with the same experimental factors as the actual experimental data as the input conditions; and the simulated value of the post-weld coaxiality is obtained by adjusting the different variable values. The two parts together constitute the simulated data training set. This process is conducive to obtaining more and more comprehensive training data samples in the neural network training stage, and eliminating the prediction result deviation that exists when pure simulated data is used as the training sample.

[0023] Preferably, in step S1, when obtaining simulation data, the welding heat source model is corrected by using a neural network model to obtain correction boundary conditions, and the specific steps are:

[0024] S101, using a test data training set generated when obtaining actual test data as a training sample, and establishing a neural network model that is correlated with a welding heat source model;

[0025] S102, training a neural network model: using the welding starting position, pre-weld coaxiality, and the materials and dimensions of the axle body and the bushing, the weld surface shape, and the welding process parameters in the training samples as input data, using the simulated value of the post-weld coaxiality and the model boundary conditions as output data, and using the difference between the simulated value and the actual value of the post-weld coaxiality as a control condition, the neural network model is trained and optimized to ultimately obtain the optimal boundary condition value of the welding heat source model;

[0026] S103, correcting the welding heat source model: correcting the welding heat source model using the optimal boundary condition value obtained in S102, to obtain a welding heat source model with the smallest deviation between the simulated value and the actual value of the post-weld coaxiality;

[0027] Based on the above steps, the process of correcting the welding heat source model can be simplified, the boundary conditions for model correction can be quickly obtained with less test data, and the simulation accuracy of the corrected model can be improved, with higher processing efficiency.

[0028] Preferably, in step S3, the test data training set and the simulation data training set obtained in step S1 are pre-divided into different category subsets according to the axle specifications. During training, a subset corresponding to a certain specification of axle is pre-selected as a training sample, and the neural network model is trained and optimized to obtain the predicted value of the pre-assembly angle of the axle of the current specification. When the number of training samples of the neural network model meets the model convergence requirement, the subset corresponding to another specification of axle is replaced, and the training and optimization are repeated on the existing model until convergence again. The training is repeated for multiple specifications of axles until the neural network model is applicable to the prediction of the pre-assembly angle of axles of any specification.

[0029] The above process helps to make the neural network model global and can be applied to the prediction of the pre-assembly angle of any specification of axle. At the same time, it can also optimize the local neural network model for a certain specification of axle to improve the accuracy of pre-assembly angle prediction for specific objects.

[0030] Furthermore, the method for obtaining the pre-assembly angle used to guide actual production in step S4 is as follows: the corresponding subset of the current axle specification to be processed is intelligently retrieved as the data source for neural network prediction, and finally the pre-assembly angle applicable to the current specification axle is obtained; the welding starting pre-assembly angle, starting position coordinates and post-weld coaxiality data generated in actual production are classified into the corresponding subset, which is used for the optimization of the local neural network model of this category, further improving the accuracy of the pre-assembly angle of the processing of axles of specific specifications, and thus ensuring the accuracy of the post-weld coaxiality.

[0031] The present invention also provides an electronic device comprising a processor and a memory, wherein the memory is used to store computer programs; and the processor is used to implement any of the aforementioned neural network training methods for preparing axles for high-coaxiality special vehicles when executing the programs stored in the memory.

[0032] The present invention also provides a computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions are used to execute any of the aforementioned neural network training methods for preparing axles for high-coaxiality special vehicles.

[0033] (3) Beneficial effects

[0034] At least one technical solution in the present invention has the following advantages or beneficial effects:

[0035] 1) The present invention utilizes the welding deformation that inevitably occurs during the welding process between the axle body and the shaft sleeve. By adopting a neural network training method, a limited number of test data and simulation data are used as training samples to predict the pre-assembly angle that can completely offset the welding deformation. Based on this, the pre-assembly angle guides the pre-welding posture adjustment of the axle body and the shaft sleeve before welding, thereby directly obtaining a high-coaxiality axle. The axle prepared according to this method can achieve a coaxiality error of only 15 to 30 degrees after welding, without the need for post-weld correction, and has high stability and reliability. It is not limited by axle specifications and machine tool equipment, and is particularly suitable for the preparation of axles for large-tonnage or wide-body special-purpose vehicles with high precision requirements.

[0036] 2) The neural network training method of the present invention only needs to collect a limited number of test data for axles of different specifications. By combining the simulation model to obtain a simulated data set, the mixed data of simulated data and test data is used as the training sample of the neural network model. The welding process parameters, welding starting position coordinates, post-weld coaxiality and other data that are easy to obtain in real time can be used to achieve accurate prediction of the pre-assembly angle. This reduces the difficulty and cost of obtaining training samples, eliminates the prediction result deviation that exists when pure simulation data is used as training samples, weakens the need to measure the welding deformation, and enhances the accuracy of the pre-assembly angle prediction, which is more conducive to accurately guiding production and reducing product defective rates.

[0037] 3) The neural network training method of the present invention uses a production data optimization set consisting of real-time detected pre-assembly angles, welding starting position coordinates, and post-weld coaxiality as a continuously optimized training sample, and provides real-time feedback on equipment state changes such as welding parameters and welding head wear, to train and optimize the neural network model online, so that the neural network model always fits the current processing state and accurately predicts the required pre-assembly angle, thereby always maintaining high coaxiality of the axle after welding.

[0038] 4) The present invention also obtains the corrected boundary conditions of the welding heat source model through a neural network training method, simplifies the process of correcting the welding heat source model, quickly and accurately obtains the boundary conditions with less test data, reduces manual correction errors, improves the simulation accuracy of the corrected model, and has higher processing efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments with reference to the following drawings:

[0040] Figure 1 This is a flow chart of the neural network training for predicting the pre-assembly angle of the present invention; DETAILED DESCRIPTION

[0041] The present invention will be further described in detail below with reference to the examples, but the embodiments of the present invention are not limited thereto.

[0042] Example 1

[0043] This embodiment provides a neural network training method for manufacturing a high-coaxiality special vehicle axle. By using a neural network model to obtain the pre-assembly angle between the special vehicle axle body and the shaft sleeve before welding, the method is used to offset the coaxiality deviation caused by welding deformation during actual production, thereby directly manufacturing a high-coaxiality special vehicle axle. The method includes the following steps:

[0044] S1, obtain the data training set for neural network model training:

[0045] When obtaining actual test data, the welding starting position and pre-weld coaxiality are used as the main test factors. Combined with the materials and dimensions of the axle body and the bushing, the weld surface shape, and the welding parameters, a series of single-variable tests are designed. The actual post-weld coaxiality values of the axle are collected as test results. The test factors and test results are mapped one-to-one to generate a test data training set for neural network training.

[0046] It should be noted that: during the actual test design, the materials, dimensions and welding surface shapes of the axle body and the shaft sleeve are determined according to the axle specifications. When the axle specifications are not changed, these are relatively fixed parameters. Welding parameters include process parameters that affect welding quality, including conventional parameters such as welding power and time. When the axle specifications are not changed, these are generally relatively fixed parameters. The coaxiality before welding and the coaxiality after welding mainly refer to the deflection angle and deflection direction between the axis of the axle body and the shaft sleeve before or after welding, which are generally characterized by coaxiality error. The smaller it is, the higher the coaxiality. The deviation between the coaxiality before welding and the coaxiality after welding is the required pre-assembly angle value. The deflection direction of the pre-assembly angle is opposite to the deflection direction of the coaxiality before and after welding. The welding starting position mainly affects the superposition of welding energy at the welding surface and affects the welding deformation direction, which is directly related to the deflection direction of the pre-assembly angle. Therefore, the welding starting position and the coaxiality before welding are the main test factors, and the axle body and sleeve material, size, welding surface shape, and welding parameters are auxiliary test factors to reduce the difficulty and cost of obtaining training samples.

[0047] When acquiring simulation data, the same test factors as those used to acquire actual test data are used as input conditions for the welding heat source model, and simulated values of post-weld coaxiality under the corresponding conditions are obtained. The difference between the actual and simulated values of post-weld coaxiality is calculated, and the welding heat source model is corrected until the error between the actual and simulated values is within 5%, and the correction is stopped. Based on the corrected welding heat source model, simulated values of post-weld coaxiality under different simulation test conditions are obtained, and a simulation data training set for neural network training is generated.

[0048] Preferably, the simulation data training set includes two parts: the first is to import the revised welding heat source model again with the same test factors as the actual test data as the input conditions, and re-obtain the simulation value of the coaxiality after welding; the second is to adjust the welding starting position and the coaxiality before welding to values different from the test data, and use this as the input condition of the revised welding heat source model to obtain the simulation value of the coaxiality after welding; this process is used to obtain more and more comprehensive training data samples in the neural network training stage, and eliminate the prediction result deviation that exists when pure simulation data is used as the training sample.

[0049] S2, establish the initial neural network model:

[0050] Using the experimental data training set obtained by S1 as the training sample, the initial neural network model is constructed, and the initial weights and thresholds of the neural network are obtained;

[0051] S3, training the neural network model to obtain the pre-assembly angle prediction value;

[0052] The experimental data training set and the simulation data training set obtained by S1 are used as training samples and imported into the initial neural network model. The welding process parameters, post-weld coaxiality and welding starting position coordinates are used as input data, and the pre-assembly angle is used as output data. The neural network model is trained and optimized until convergence, and the pre-assembly angle is finally predicted.

[0053] S4, guides actual production and obtains high coaxiality special vehicle axles:

[0054] According to the pre-assembly angle prediction value obtained in step S3, the pre-assembly angle between the axle body and the bushing is adjusted, and welding is performed to directly obtain a special vehicle axle with high coaxiality.

[0055] Furthermore, S5 is continuous optimization of the neural network model: when executing S4, the real data of the pre-assembly angle, welding starting position coordinates and post-weld coaxiality are synchronously collected to generate an optimized production data set. The optimized production data set is used as a training sample to train and optimize the neural network model online, and a continuously optimized neural network model is obtained. The neural network model is trained and optimized online in real time according to equipment status changes such as welding parameter fine-tuning and welding head wear, so that the neural network model always fits the pre-assembly angle accurately predicted by the current processing status.

[0056] Furthermore, before generating the production data optimization set, the neural network model automatically checks whether the post-weld coaxiality obtained based on the predicted pre-assembly angle meets the requirements. When the requirements are met, part of the actual production data is intelligently selected and included in the production data optimization set; when the requirements are not met, all the collected actual production data are included in the production data optimization set for online optimization training of the neural network model, and the optimization of the neural network model is performed once. This process can effectively reduce the amount of duplicate data in the production data optimization set while ensuring real-time feedback of state changes, thereby reducing data storage requirements and ensuring the long-term stability of the algorithm.

[0057] The advantages of this embodiment are:

[0058] Compared with the existing axle coaxiality correction method, the welding deformation that is bound to exist during the welding process of the axle body and the shaft sleeve, and the phenomenon that the welding deformation will inevitably lead to coaxiality deviation, are used in reverse to reversely utilize the coaxiality deviation amount and deviation angle before and after welding to weaken the coaxiality accuracy requirement before welding, so that the axle body and the shaft sleeve present a pre-assembly angle that is opposite to the welding deformation before welding, thereby offsetting the influence of the welding deformation on the coaxiality during the welding process, thereby avoiding secondary processing or correction of the axle after welding, avoiding the introduction of additional stress, and effectively improving the coaxiality, stability and reliability of the axle.

[0059] Based on the uncertainty of the value and direction of welding deformation, by adopting the neural network training method, only a limited number of test data need to be collected for axles of different specifications. Combined with the simulation model, a simulated data set is obtained as hybrid data training and neural network model, which reduces the difficulty and cost of obtaining training samples, eliminates the prediction result deviation that exists when pure simulation data is used as training samples, enhances the accuracy of pre-assembly angle prediction, and is more conducive to accurate production guidance; at the same time, by using welding process parameters that are easy to obtain in real time, welding starting position coordinates and post-weld coaxiality data, it is possible to achieve accurate prediction of the pre-assembly angle that completely offsets the welding deformation, and is conducive to the continuous online optimization of the neural network model, so that the neural network model always fits the current processing status, eliminates the prediction deviation caused by inevitable equipment status changes such as welding head wear, and further improves the accuracy of the pre-assembly angle, thereby always maintaining high coaxiality of the axle after welding. It is especially suitable for the preparation of special-purpose vehicle axles with large specifications and high coaxiality accuracy requirements.

[0060] Example 2

[0061] This embodiment provides a neural network training method for preparing a high-coaxiality special vehicle axle. The difference from the first embodiment is that:

[0062] On the basis of the solution of Example 1, when executing step S3, the test data training set and the simulation data training set obtained in step S1 are pre-divided into different category subsets according to the axle specifications. During training, a subset corresponding to a certain specification of axle is pre-selected as a training sample, and the neural network model is trained and optimized to obtain the predicted value of the pre-assembly angle of the current specification of the axle; when the number of training samples of the neural network model meets the model convergence requirements, the subset corresponding to another specification of the axle is replaced, and the training and optimization are repeated on the existing model until convergence again. The training of multiple specifications of axles is repeated until the neural network model is suitable for the prediction of the pre-assembly angle of axles of any specification.

[0063] Furthermore, when obtaining the pre-assembly angle to guide actual production in step S4, the corresponding subset of the current axle specification to be processed is intelligently retrieved as the data source for neural network prediction, and finally the pre-assembly angle applicable to the current specification axle is obtained; the welding starting pre-assembly angle, starting position coordinates and post-weld coaxiality data generated in actual production are classified into the corresponding subset for optimization of the local neural network model of this category.

[0064] The advantage of this embodiment is that by training and optimizing the neural network model globally and locally, the neural network model has both global and local specificity. It can be used to predict the pre-assembly angle of any specification of axle, and can also be locally optimized for a certain specification of axle, thereby improving the accuracy of the pre-assembly angle prediction of specific objects, and the applicability of the neural network model is better.

[0065] Example 3

[0066] This embodiment provides a neural network training method for preparing a high-coaxiality special vehicle axle. The difference from the first and second embodiments is that:

[0067] Based on the solution of embodiment 1 or 2, in step S1, when obtaining simulation data, the welding heat source model is corrected using a neural network model to obtain corrected boundary conditions. The specific steps are:

[0068] S101, using a test data training set generated when obtaining actual test data as a training sample, and establishing a neural network model that is correlated with a welding heat source model;

[0069] S102, training a neural network model: using the welding starting position, pre-weld coaxiality, and the materials and dimensions of the axle body and the bushing, the weld surface shape, and the welding process parameters in the training samples as input data, using the simulated value of the post-weld coaxiality and the model boundary conditions as output data, and using the difference between the simulated value and the actual value of the post-weld coaxiality as a control condition, the neural network model is trained and optimized to ultimately obtain the optimal boundary condition value of the welding heat source model;

[0070] S103, correcting the welding heat source model: correcting the welding heat source model using the optimal boundary condition value obtained in S102, to obtain a welding heat source model with the smallest deviation between the simulated value and the actual value of the post-weld coaxiality;

[0071] The advantages of this embodiment are: obtaining the corrected boundary conditions of the welding heat source model through the neural network training method, simplifying the process of correcting the welding heat source model, quickly and accurately obtaining the boundary conditions with less test data, reducing manual correction errors, improving the simulation accuracy of the corrected model, and improving processing efficiency.

[0072] Example 4

[0073] This embodiment provides an electronic device, including a processor and a memory, wherein the memory is used to store a computer program; the processor is used to implement the aforementioned neural network training method for preparing axles for high-coaxiality special vehicles as described in Examples 1 to 3 when executing the program stored in the memory.

[0074] Example 5

[0075] This embodiment provides a computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions are used to execute the neural network training method for preparing a high-coaxiality special vehicle axle as described in the first to third embodiments.

[0076] The computer-readable storage medium of this embodiment can be loaded onto any special vehicle axle manufacturing equipment, and can also be used in conjunction with welding models, automatic control systems, etc. to improve the intelligence of axle manufacturing.

[0077] The above description is merely a preferred embodiment of the present invention and does not limit the technical scope of the present invention. Therefore, any minor modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention still fall within the scope of protection of the present invention.

Claims

1. A neural network training method for preparing axles for special vehicles with high coaxiality, characterized by: A neural network model is used to obtain the pre-assembly angle between the axle body and the sleeve before welding for a special vehicle. This angle is used to offset the coaxiality deviation caused by welding deformation during actual production, and directly produce a high-coaxiality special vehicle axle. The process includes the following steps: S1, obtain the data training set for neural network model training: When obtaining actual test data, the welding starting position and pre-weld coaxiality are used as the main test factors. Combined with the materials and dimensions of the axle body and the bushing, the weld surface shape, and the welding parameters, a series of single-variable tests are designed. The actual post-weld coaxiality values of the axle are collected as test results. The test factors and test results are mapped one-to-one to generate a test data training set for neural network training. When obtaining simulation data, the same test factors as those used to obtain actual test data are used as input conditions for the welding heat source model to obtain the simulated value of post-weld coaxiality under the corresponding conditions; Calculate the difference between the actual and simulated values of the post-weld coaxiality, and correct the welding heat source model until the error between the actual and simulated values is within 5%; Based on the revised welding heat source model, the simulated values of post-weld coaxiality under different simulation test conditions are obtained, and a simulated data training set for neural network training is generated; S2, establish the initial neural network model: Using the experimental data training set obtained by S1 as the training sample, the initial neural network model is constructed, and the initial weights and thresholds of the neural network are obtained; S3, training the neural network model to obtain the pre-assembly angle prediction value; The experimental data training set and the simulation data training set obtained by S1 are used as training samples and imported into the initial neural network model. The welding process parameters, post-weld coaxiality and welding starting position coordinates are used as input data, and the pre-assembly angle is used as output data. The neural network model is trained and optimized until convergence, and the pre-assembly angle is finally predicted. S4, guides actual production and obtains high coaxiality special vehicle axles: According to the pre-assembly angle prediction value obtained in step S3, the pre-assembly angle between the axle body and the bushing is adjusted, and welding is performed to directly obtain a special vehicle axle with high coaxiality.

2. A neural network training method for preparing a high-coaxiality special vehicle axle according to claim 1, characterized in that: It also includes step S5, continuous optimization of the neural network model: When executing S4, the real data of the pre-assembly angle, welding starting position coordinates and post-weld coaxiality are collected synchronously to generate an optimized production data set. The optimized production data set is used as a training sample to perform online training and optimization on the neural network model to obtain a continuously optimized neural network model.

3. The neural network training method for preparing a high-coaxiality special vehicle axle according to claim 2, characterized in that: Before generating the production data optimization set, the neural network model automatically checks whether the post-weld coaxiality obtained based on the predicted pre-assembly angle meets the requirements. When the requirements are met, part of the actual production data is intelligently selected and included in the production data optimization set; when the requirements are not met, all the collected actual production data are included in the production data optimization set for online optimization training of the neural network model, and the optimization of the neural network model is performed once.

4. The neural network training method for preparing a high-coaxiality special vehicle axle according to claim 1, characterized in that: When obtaining the simulation data training set, based on the modified welding heat source model, the simulated value of the post-weld coaxiality is again obtained with the same experimental factors as the actual test data as the input conditions; and the simulated value of the post-weld coaxiality is obtained by adjusting different variable values. The two parts together constitute the simulation data training set.

5. The neural network training method for preparing a high-coaxiality special vehicle axle according to claim 1, characterized in that: In step S1, when obtaining simulation data, the welding heat source model is corrected using a neural network model to obtain correction boundary conditions. The specific steps are: S101, using a test data training set generated when obtaining actual test data as a training sample, and establishing a neural network model that is correlated with a welding heat source model; S102, training a neural network model: using the welding starting position, pre-weld coaxiality, and the materials and dimensions of the axle body and the bushing, the weld surface shape, and the welding process parameters in the training samples as input data, using the simulated value of the post-weld coaxiality and the model boundary conditions as output data, and using the difference between the simulated value and the actual value of the post-weld coaxiality as a control condition, the neural network model is trained and optimized to ultimately obtain the optimal boundary condition value of the welding heat source model; S103, correcting the welding heat source model: correcting the welding heat source model with the optimal boundary condition value obtained in S102, and obtaining a welding heat source model with the smallest deviation between the simulated value and the actual value of the coaxiality after welding.

6. The neural network training method for preparing a high-coaxiality special vehicle axle according to claim 1, characterized in that: In step S3, the test data training set and simulation data training set obtained in step S1 are pre-divided into different category subsets according to the axle specifications. During training, a subset corresponding to a certain specification of axle is pre-selected as a training sample, and the neural network model is trained and optimized to obtain the predicted value of the pre-assembly angle of the current specification of the axle; when the number of training samples of the neural network model meets the model convergence requirements, the subset corresponding to another specification of the axle is replaced, and the training and optimization are repeated on the existing model until convergence again. The training of multiple specifications of axles is repeated until the neural network model is suitable for the prediction of the pre-assembly angle of axles of any specification.

7. The neural network training method for preparing a high-coaxiality special vehicle axle according to claim 6, characterized in that: The method for obtaining the pre-assembly angle used to guide actual production in step S4 is: intelligently retrieve the corresponding subset as the data source for neural network prediction based on the specifications of the current axle to be processed, and finally obtain the pre-assembly angle applicable to the axle of the current specification; the pre-welding pre-assembly angle, initial position coordinates and post-weld coaxiality data generated by actual production are classified into the corresponding subset for optimization of the local neural network model of the corresponding category.

8. An electronic device, characterized in that: It comprises a processor and a memory, wherein the memory is used to store computer programs; the processor is used to implement the neural network training method for preparing axles of special vehicles with high coaxiality as described in any one of claims 1 to 7 when executing the programs stored in the memory.

9. A computer-readable storage medium, characterized in that: Computer-executable instructions are stored, and the computer-executable instructions are used to: execute the neural network training method for preparing a high-coaxiality special vehicle axle according to any one of claims 1 to 7.

Images