A method for evaluating noise of a steering system based on a neural network-Monte Carlo

Abstract

The present application relates to the technical field of steering system, in particular to a method for evaluating noise of steering system based on neural network-Monte Carlo, which is characterized by the following specific method: according to the boundary conditions and internal structure parameters of the steering system provided by the customer, a dynamic model of the steering system is built, which outputs impact energy in all possible noise generating areas, so as to fully reflect the mechanism of impact noise between parts; custom-made steering system sample is ensured to be consistent with the input information of the dynamic model. Compared with the prior art, the method for evaluating noise of steering system based on neural network-Monte Carlo is used to predict and optimize the noise of the steering system by combining machine learning and reliability theory.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of steering system, in particular to a method for evaluating noise of steering system based on neural network-Monte Carlo. BACKGROUND

[0002] Noise is the most direct perception of consumers and is paid more and more attention by automobile manufacturers and parts suppliers. More than half of the cost of steering system after sale comes from noise of steering system, and noise mainly comes from two aspects: steady noise such as running, zipper, and transient noise such as rattle, clunk, click, knocking, etc. Due to the transient and intermittent characteristics of transient noise, it is difficult to continuously and accurately measure it. Subjective evaluation is still the commonly used method, which has the advantages of intuition and rapidness. However, it also has some defects, such as it is difficult to determine the specific position of rattle noise, and the evaluation results are affected by many factors. Moreover, the noise of steering system comes from many sources, such as the impact between gear and rack, the impact between worm and worm gear, the friction inside bearing, the impact between pressure block and shell, etc. Sometimes, a single factor contributes to noise, and sometimes several factors work together. Once the product with potential noise problem goes to market, noise may occur after several months of running-in, which easily leads to customer discomfort and complaints, and even after-sales compensation.

[0003] In order to fully understand the mechanism of steering noise and avoid the above noise in the market from the design source, an objective prediction method highly related to subjective evaluation and accurately expressing the contribution ratio of each part in steering system to noise is needed. SUMMARY

[0004] In order to overcome the shortcomings of the prior art, the present application provides a method for evaluating noise of steering system based on neural network-Monte Carlo, which combines machine learning and reliability theory to predict and optimize noise of steering system.

[0005] To achieve the above purpose, a method for evaluating noise of steering system based on neural network-Monte Carlo is designed, and the specific method is as follows:

[0006] S1, according to the boundary conditions and internal structure parameters of the steering system provided by the customer, a dynamic model of the steering system is built, which outputs impact energy in all possible noise generating areas, so as to fully reflect the mechanism of impact noise between parts;

[0007] S2, customizing a steering system sample to ensure that the input information of the actual sample is consistent with the input information of the dynamic model;

[0008] S3, install the steering system sample in step S2 on the corresponding test vehicle, and collect data at the designed vehicle speed and road surface, mainly collect acceleration data, rod force data and corresponding subjective evaluation score in the noise area of the steering system;

[0009] S4, discretize and normalize the collected acceleration data, rod force data, impact energy obtained by simulation calculation and subjective evaluation score, and classify the data into training set and test set;

[0010] S5, perform BP or RBF neural network learning, optimize the internal parameters of the neural network through cross-validation, and finally establish a neural network model between the simulated impact energy and acceleration and a neural network model between acceleration and subjective evaluation score;

[0011] S6, obtain the on-site data of the key dimensions and key performance parameters in the steering system, and perform parameter distribution analysis on the obtained on-site data to set the distribution estimation of the key dimensions and key performance parameters during Monte Carlo sampling;

[0012] S7, in the noise prediction process of a new project, first write an algorithm program of Monte Carlo sampling method and embed it into the dynamic model built in S1, and perform noise simulation based on Monte Carlo sampling of the key dimensions and key performance parameters according to the parameter distribution obtained in S6 to obtain the impact energy of the new project;

[0013] S8, call the trained neural network model in S5 to predict the acceleration in the noise area and then predict the subjective evaluation score of the steering system under the corresponding vehicle environment and corresponding road speed.

[0014] In step S1, the dynamic model of the steering system is derived from Newton's law, and the equation is Where J s is the steering wheel moment of inertia, B s is the steering wheel damping coefficient, θ S is the steering wheel input angle, T d is the torque acting on the steering wheel, T s is the torsion bar torsional stiffness coefficient, θ e is the lower steering column angle, J e is the lower steering column moment of inertia, B e is the lower steering column damping coefficient, T ω is the reaction torque on the output shaft, r p is the pinion radius, F δ is the road random signal, i is the worm gear ratio, Km J is the torsional stiffness of the assist motor shaft, m I is the rotational inertia of the assist motor, m B is the rotation angle of the assist motor, m T is the damping coefficient of the assist motor, m T is the electromagnetic torque of the assist motor, r b is the rack mass, r K is the damping coefficient of the rack, r K is the stiffness of the rack, r x is the lateral displacement of the rack.

[0015] In the step S5, the cross-validation method is specifically that, when training the neural network model, the calculated simulation impact energy data, the collected vehicle acceleration data and the subjective score are divided into two parts, one part is used for training the model, and the other part is used for evaluating the training model; each sub-sample in the sample set used for training the model is used as both a training target and a test target, in the whole cycle test and verification process, the limited data set is used to fully find the optimal training model, and the trained model can reduce overfitting in the prediction of a new project, so that the accuracy is ensured.

[0016] Compared with the prior art, the method for evaluating the noise of the steering system based on the neural network-Monte Carlo is provided, the noise of the steering system is predicted and optimized by combining the method of machine learning and the reliability theory.

[0017] Firstly, since the dynamic model between the internal parts of the steering system is established, the real mechanism of noise generation can be well reflected, then the neural network model of the impact energy and acceleration and the neural network between the acceleration and the subjective evaluation score are established, the cross-validation between the normalized and training data is adopted, the data correlation between the simulation data and the subjective score is well ensured, finally, considering the tolerance of the actual part size and the fluctuation of the performance within a certain range, the noise real performance of the actual batch production parts is reflected based on the established dynamic model and combined with the Monte Carlo sampling method. The method has very high prediction accuracy and reliability through two years of practical verification, and is very simple and fast. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 The method flowchart of the application.

[0019] Figure 2 The structure diagram of the steering system.

[0020] Figure 3 The equivalent dynamic model diagram of the steering system.

[0021] Figure 4 The gap curve diagram of the pressing block.

[0022] Figure 5 Figure for the relationship between the pull rod angle and the stroke.

[0023] Figure 6 Figure for the relationship between the pull rod resistance and the stroke in the parking state.

[0024] Figure 7 Table for the measured acceleration and subjective score based on a certain project.

[0025] Figure 8 Figure for the trained neural network model of the acceleration and the impact energy.

[0026] Figure 9 Figure for the trained neural network model of the acceleration and the subjective score.

[0027] Figure 10 Figure for the relationship between the pull rod resistance and the stroke in the parking state in the embodiment.

[0028] Figure 11 Figure for the relationship between the pull rod angle and the stroke in the embodiment.

[0029] Figure 12 Table for the calculated impact energy of the noise generating area under the known input information.

[0030] Figure 13 Table for the acceleration and subjective score of 5 random samples in the same state on the whole vehicle.

[0031] Figure 14 Table for the simulated impact energy of 5 groups of noise generating areas obtained by Monte Carlo simulation.

[0032] Figure 15 Comparison figure of the predicted acceleration and the later measured acceleration (piece 1).

[0033] Figure 16 Comparison figure of the predicted acceleration and the later measured acceleration (piece 2).

[0034] Figure 17 Comparison figure of the predicted acceleration and the later measured acceleration (piece 3).

[0035] Figure 18 Comparison figure of the predicted acceleration and the later measured acceleration (piece 4).

[0036] Figure 19 Comparison figure of the predicted acceleration and the later measured acceleration (piece 5).

[0037] Figure 20 Comparison figure of the predicted subjective score and the later measured subjective score (5 kinds). DETAILED DESCRIPTION

[0038] The application is further described below with reference to the accompanying drawings.

[0039] As Figure 1 shown, a method for evaluating the noise of a steering system based on a neural network-Monte Carlo, the specific method is as follows:

[0040] S1, according to the boundary conditions provided by the customer and the internal structure parameters of the steering system, a dynamic model of the steering system is built, which outputs impact energy in all possible noise generating areas, so as to fully reflect the mechanism of impact noise between parts.

[0041] The steering system model is composed of a mechanical steering system and an electronic control unit, which simulates the mutual impact of the internal parts of the steering machine under actual input conditions, and the structural diagram is as Figure 2 shown. The mechanical steering system is mainly composed of a steering sensor, a sensing end gear, a motor, a worm and a gear, a rack and left and right tie rods. Its purpose is to transmit the torque of the steering wheel to the sensing end gear and then convert the hand force torque into rack hand force through the meshing of the gear and the rack. At the same time, the assist torque output by the motor is amplified through the worm and the gear, and then the assist torque is converted into rack assist torque through the meshing of the sensing end gear and the rack, and finally the combined force of the rack hand force and the rack assist force moves the rack left and right, and then the tie rods realize the steering of the tires of the whole vehicle. The electronic control unit includes ECU, motor control, etc., and its purpose is that ECU calculates the required assist force according to the angle sensor signal, and then controls the motor to output the necessary assist torque, and ensures that the assist torque provided meets the design requirements. Based on the transmission principle and mechanical structure of the steering system, the mechanical steering system can be equivalent transformed. Among them, the steering column is a rigid body with energy dissipation, which is equivalent to a damping element, the angle sensor has a certain elastic deformation, which can be equivalent to an elastic element, and the meshing place of the worm and the gear and the rack has energy loss in the transmission process and elastic deformation when the teeth mesh, so it is equivalent to a spring damping element. According to the equivalent dynamic model, the kinetic differential equation can be derived by using Newton's law, and the specific form of the mathematical model is

[0042] Table 1

[0043]

[0044]

[0045] Through the built dynamic model, given input information such as the gap of the pressing block, the tie rod angle and the tie rod resistance in the parking state is set in the model, as Figures 4 to 6 shown.

[0046] S2, custom steering system sample, ensure that the actual sample input information and the input information of the dynamic model consistent.

[0047] S3, the steering system is installed on the corresponding vehicle, and the data is collected in the test field according to the designed vehicle speed and road surface, mainly collecting acceleration data, rod force data and corresponding subjective evaluation points in the noise area of the steering system, such as Figure 7 .

[0048] S4, the acceleration data, rod force data and impact energy obtained by simulation calculation and subjective evaluation points are discretized and normalized, and the data is classified into training set and test set.

[0049] Since the unit of acceleration is m / s 2 , the unit of rod force is N, the unit of impact energy is (m / s) 2 , and the subjective score is a value without unit, in order to avoid numerical calculation problems and because the number of input and output data is different, which can cause large network prediction error, at the same time, it can speed up the convergence speed of network and improve the learning rate of neural network, all data are normalized by maximum and minimum linear conversion.

[0050] S5, BP or RBF neural network learning, through cross validation, the internal parameters of neural network are optimized, and finally the neural network model between simulation impact energy and acceleration and the neural network model between acceleration and subjective evaluation points are established, as shown in Figure 8 , Figure 9 .

[0051] In order to make full use of limited experimental data to train our neural network model, we use cross validation method here. The specific method is that when training the neural network model, the calculated simulation impact energy data and the collected vehicle acceleration data and subjective points are divided into two parts, one part is used to train the model, and the other part is used to evaluate the goodness of the trained model. The sample set used to train the model is used as both the training target and the test target. In the whole cycle test verification process, the limited data set is used to find the optimal training model, and the trained model can reduce over fitting in the prediction of new project, so as to ensure the accuracy.

[0052] S6, the key size and key performance parameters of the steering system are obtained on the production line, and the parameter distribution analysis is carried out on the obtained field data, so as to set the distribution estimation of key size and key performance parameters when monte carlo sampling.

[0053] S7, in the noise prediction process of the new project, first write the algorithm program of Monte Carlo sampling method, embedded in the dynamic model built in S1, and the key size and key performance parameters according to the parameter distribution obtained in S6 are simulated based on Monte Carlo sampling to obtain the impact energy of the new project.

[0054] When the new project is assembled in the production line, the initial gasket gap needs to be adjusted, and the production line only ensures that the gasket gap does not exceed a specified value for the need of production rhythm. As for the specific gasket gap value, the influence of this gasket gap value on the moving force, and the different initial values of the gasket gap leading to different wear sizes after durability, etc. all have a great influence on the steering system noise.

[0055] At the same time, the steering system is assembled by many sub-parts, each part has a certain tolerance, and the stiffness and damping of each part between different batches and the same batch are different. The gap, stiffness, damping, etc. of the steering machine assembled by different parts will further amplify the noise performance difference of the whole steering.

[0056] In order to fully evaluate the interval range of the steering machine noise, and when the evaluated noise does not meet the requirements of putting into the market, the current design needs to be fully optimized. At this time, we first perform Monte Carlo simulation sampling on the key input information of the dynamic model, and then combine the previous dynamic model to calculate the impact energy as the response output of Monte Carlo, and finally obtain a boundary range of the impact energy.

[0057] Actual case implementation and effect:

[0058] The relationship between the pull rod force and the stroke in the parking state of a certain steering machine project with a gasket gap of 0.2mm is shown in Figure 10 . Figure 10 The relationship between the pull rod resistance and the stroke in the parking state is shown in Figure 11 . Figure 12 The input information is input into the dynamic model, and the impact energy calculated is obtained. The impact energy in the noise area is calculated under the known input information; 5 steering machines are prepared, and the gasket gap is controlled at 0.2mm, and they are installed on the corresponding vehicles. Acceleration measurement and subjective evaluation are performed on the 5 steering machines, and the specific results are shown in Figure 13 .

[0059] In the dynamic model, consider the difference between the size tolerance and the assembly of the sample, and use the Monte Carlo simulation sampling method to calculate a series of impact energies. 5 simulation results are randomly selected in the upper and lower boundaries and the middle region, and the specific values are shown in Figure 14 .

[0060] Using the trained neural network model of impact energy and acceleration, the sensor acceleration values in 5 states are predicted and compared with the actual vehicle test acceleration values in the xyz three directions, as shown in Figures 15 to 19 .

[0061] Using the trained neural network model of subjective score and acceleration, the subjective scores in 5 states are predicted and compared with the actual vehicle test subjective scores, as shown in Figure 20 .

[0062] The present application uses a method combining machine learning and reliability theory to predict and optimize the noise of the steering system. First, since the dynamic model between the internal parts of the steering system is established, it can well reflect the real mechanism of noise generation. Then, the neural network model of impact energy and acceleration and the neural network between acceleration and subjective evaluation score are established, and the cross validation between normalized and training data is adopted, which can well guarantee the data correlation between the simulation data and the subjective score. Finally, considering the actual part size tolerance and the performance fluctuation within a certain range, the noise real performance of the actual batch production parts is reflected on the basis of the established dynamic model combined with the Monte Carlo sampling method. This method has been verified for two years, and has very high prediction accuracy and reliability, and is very simple and fast.

Claims

1. A method for evaluating steering system noise based on a neural network-Monte Carlo method, characterized by: The specific method is as follows: S1: Based on the boundary conditions and internal structural parameters of the steering system provided by the customer, a dynamic model of the steering system is built. This dynamic model outputs impact energy in all areas where noise may be generated, so as to fully reflect the mechanism of impact noise between parts. S2, customize the steering system prototype to ensure that the input information of the actual prototype is consistent with the input information of the dynamic model; S3, installing the steering system sample in step S2 on the corresponding test vehicle, and collecting data at the test track at the designed vehicle speed and road surface, collecting acceleration data of the steering system noise-generating area, tie rod force data, and corresponding subjective evaluation scores; S4, discretize and normalize the collected acceleration data, pull rod force data, impact energy obtained by simulation calculation, and subjective evaluation scores, and classify the data into training sets and test sets; S5, perform BP or RBF neural network learning, optimize the internal parameters of the neural network through cross-validation, and finally establish a neural network model between the simulated impact energy and acceleration, and a neural network model between acceleration and subjective evaluation score; S6, acquire on-site data of the production line for key dimensions and key performance parameters within the steering system, and perform parameter distribution analysis on the acquired on-site data to facilitate setting distribution estimates of key dimensions and key performance parameters during Monte Carlo sampling; S7, in the noise prediction process of the new project, first write the algorithm program of the Monte Carlo sampling method, embed it into the dynamic model built in S1, and perform noise simulation based on Monte Carlo sampling on the key dimensions and key performance parameters according to the parameter distribution obtained in S6 to obtain the impact energy of the new project; S8 uses the impact energy obtained in S7 in the new project to call the neural network model trained in S5 to first predict the acceleration of the noise-generating area, and then predict the subjective evaluation score of the steering system under the corresponding vehicle environment and corresponding road conditions and speed.

2. The method for evaluating steering system noise based on neural network-Monte Carlo according to claim 1, characterized in that: In step S1, the dynamic model of the steering system is based on Newton's law, and the dynamic motion differential equation can be derived. The equation is: ;in, J s is the steering wheel moment of inertia, B s is the steering wheel damping coefficient, θ S Enter the steering angle for the steering wheel, T d is the torque acting on the steering wheel, T s is the torsional stiffness coefficient of the torsion bar, θ e is the lower steering column angle, J e is the moment of inertia of the lower steering column, B e is the damping coefficient of the lower steering column, T ω is the reaction torque on the output shaft, r p is the radius of the pinion, F δ is a random signal on the road surface, i is the worm gear ratio, K m To assist the motor shaft torsional stiffness, J m To assist the motor's rotational inertia, θ m To assist the motor turning angle, B m is the damping coefficient of the power assist motor, T m is the electromagnetic torque of the assist motor, m r is the rack mass, b r is the rack damping coefficient, K r is the rack stiffness, x r is the lateral displacement of the rack.

3. The method for evaluating steering system noise based on neural network-Monte Carlo according to claim 1, characterized in that: In step S5, the cross-validation method is specifically to divide the calculated simulated impact energy data, the collected vehicle acceleration data and the subjective score into two parts when training the neural network model, one part is used to train the model, and the other part is used to evaluate the quality of the training model; Each subsample in the sample set used to train the model serves as both a training target and a test target. During the entire cycle of testing and verification, a limited data set is used to fully find the optimal training model. Moreover, the trained model can reduce overfitting in the prediction of new projects, thereby ensuring accuracy.

Images