Method for evaluating noise of steering system based on neural network-Monte CarloTechnical Field
The invention relates to the technical field of steering systems, in particular to a method for evaluating noise of a steering system based on a neural network-Monte Carlo.
Background
Noise is becoming more and more important to various automobile manufacturers and component suppliers as the most direct perception object for consumers, and more than half of the cost after sales of steering systems is currently derived from noise of steering systems, and noise is mainly derived from two aspects, namely steady-state noise such as running, zip, transient noise such as rattle, clunk, click, knocking and the like. Because of the transient noise characteristics, both transient and intermittent, continuous and accurate measurement is very difficult. Subjective assessment is still a currently commonly used method, and has the advantages of intuitiveness, rapidity and the like. There are also some defects, such as difficulty in judging the specific position where Rattle noise occurs, and the evaluation result is limited by many factors. Moreover, there are many sources of noise in the steering system, such as impacts between racks and pinions, impacts between worm gears, friction inside bearings, impacts between a press block and a housing, etc., sometimes a single factor contributing mainly to noise and sometimes as a result of several factors being coupled together. Once the potentially noisy products are marketed, after several months of running-in, noise is likely to occur, which can lead to customer discomfort and complaints, even after-market claims.
In order to fully understand the mechanism of steering noise generation, and avoid the occurrence of the above noise of products on the market from the design source, it is necessary to invent an objective prediction method which is highly relevant to subjective evaluation and can accurately express the contribution ratio of each component in the steering system to the noise.
Disclosure of Invention
The invention provides a method for estimating the noise of a steering system based on a neural network-Monte Carlo, which aims to overcome the defects of the prior art and predicts and optimizes the noise of the steering system by combining a machine learning and reliability theory.
In order to achieve the aim, the method for estimating the noise of the steering system based on the neural network-Monte Carlo is designed and is characterized by comprising the following steps of:
S1, building a dynamic model of the steering system according to boundary conditions provided by customers and internal structural parameters of the steering system, wherein the dynamic model can output impact energy in all areas possibly generating noise so as to fully reflect the mechanism of impact noise between parts;
S2, customizing a steering system sample, and ensuring that the input information of an actual sample is consistent with the input information of a dynamic model;
S3, mounting the steering system sample in the step S2 on a corresponding test whole vehicle, and collecting data in a test yard according to the designed vehicle speed and the road surface, wherein the data mainly comprise acceleration data of a noise generating region of the steering system, pull rod force data and corresponding subjective evaluation scores;
S4, discretizing and normalizing the collected acceleration data, the pull rod force data, the impact energy obtained by simulation calculation and the subjective evaluation score, and classifying the data into a training set and a testing set;
S5, BP or RBF neural network learning is carried out, internal parameters of the neural network are optimized through cross verification, and finally a neural network model between simulation impact energy and acceleration and a neural network model between acceleration and subjective evaluation score are established;
s6, acquiring field data of a production line for the key size and key performance parameters in the steering system, and performing parameter distribution analysis for the acquired field data so as to facilitate setting of the key size and distribution estimation of the key performance parameters during Monte Carlo sampling;
S7, in the noise prediction process of the new project, firstly writing an algorithm program of a Monte Carlo sampling method, embedding the algorithm program into a dynamics model built in the S1, and carrying out noise simulation based on Monte Carlo sampling on the key size and the key performance parameters according to the parameter distribution obtained in the S6 so as to obtain the impact energy of the new project;
S8, calling the neural network model trained in S5 by the impact energy obtained in the S7 in the new project to predict the acceleration of the noise generating region, and then predicting the subjective evaluation score of the steering system under the corresponding whole vehicle environment and the corresponding road condition vehicle speed.
In the step S1, the dynamics model of the steering system is Newton' S law, and a dynamics differential equation of motion can be deduced, wherein the equation isWherein Js is steering wheel moment of inertia, Bs is steering wheel damping coefficient, θS is steering wheel input angle, Td is torque acting on the steering wheel, Ts is torsion bar torsional rigidity coefficient, θe is lower steering column angle, Je is lower steering column moment of inertia, Be is lower steering column damping coefficient, Tω is reaction torque on the output shaft, rp is pinion radius, Fδ is road surface random signal, i is worm gear ratio, Km is booster motor shaft torsional rigidity, Jm is booster motor moment of inertia, θm is booster motor angle, Bm is booster motor damping coefficient, Tm is booster motor electromagnetic torque, mr is rack mass, Br is rack damping coefficient, Kr is rack rigidity, xr is rack lateral displacement.
In the step S5, the method for cross verification is specifically that when a neural network model is trained, the calculated simulated impact energy data, the acquired whole vehicle acceleration data and the subjective components are divided into two parts, one part is used for training the model, the other part is used for evaluating the quality of the training model, each sub sample in a sample set used for training the model is used as a training target and also used as a test target, the limited data set is utilized to fully find the optimal training model in the whole cycle test verification process, and the trained model can reduce overfitting in the prediction of new projects, so that the accuracy is ensured.
Compared with the prior art, the invention provides a method for estimating the noise of the steering system based on the neural network-Monte Carlo, which predicts and optimizes the noise of the steering system by combining machine learning and reliability theory.
Firstly, a dynamic model between the interiors of a steering system is established, so that the dynamic model can well reflect the real mechanism of noise generation, then, a neural network model of impact energy and acceleration and a neural network between acceleration and subjective evaluation components are established, and the cross verification between normalization and training data is adopted, so that the data association between simulation data and subjective components can be well ensured, finally, the tolerance and performance fluctuation of the actual part size are considered in a certain range, and the real performance of the noise of the actual batch production parts is reflected by a Monte Carlo sampling method on the basis of the established dynamic model. The method has very high prediction precision and reliability, and is very simple, convenient and quick through two years of practical verification.
Drawings
FIG. 1 is a flow chart of the method of the present invention.
Fig. 2 is a schematic diagram of a steering system.
Fig. 3 is a diagram of an equivalent dynamics model of the steering system.
FIG. 4 is a graph of a compact gap.
Fig. 5 is a graph of pull rod angle versus travel.
Fig. 6 is a graph of drag versus travel for a tie rod in the Parking state.
Fig. 7 is a table of measured acceleration and subjective evaluation based on a certain item.
FIG. 8 is a graph of a neural network model of trained acceleration versus impact energy.
FIG. 9 is a neural network model graph of trained acceleration versus subjective score.
FIG. 10 is a graph of drag versus travel for a pull rod in a park condition in an embodiment.
FIG. 11 is a graph of link angle versus travel in an embodiment.
FIG. 12 is a table of impact energy calculated for a noise-generating region under known input information.
Fig. 13 is a table showing acceleration and subjective scores of random 5 samples in the same state on the whole vehicle.
Fig. 14 is a table of simulated impact energy for 5 sets of noise-producing regions obtained using monte carlo simulation.
Fig. 15 is a graph of predicted acceleration versus later measured acceleration (element 1).
Fig. 16 is a graph of predicted acceleration versus later measured acceleration (element 2).
Fig. 17 is a graph of predicted acceleration versus later measured acceleration (element 3).
Fig. 18 is a graph of predicted acceleration versus later measured acceleration (element 4).
Fig. 19 is a graph of predicted acceleration versus later measured acceleration (element 5).
Fig. 20 is a subjective score comparison of predicted subjective and later measured subjective scores (5).
Detailed Description
The invention is further described below with reference to the accompanying drawings.
As shown in fig. 1, a method for estimating noise of a steering system based on a neural network-monte carlo is as follows:
S1, building a dynamic model of the steering system according to boundary conditions provided by customers and internal structural parameters of the steering system, wherein the dynamic model can output impact energy in all areas possibly generating noise so as to fully reflect the mechanism of impact noise between parts.
The steering system model consists of a mechanical steering system and an electronic control unit, and the structure diagram of the steering system model simulates the mutual collision of internal parts of the steering machine under the actual input condition, and is shown in fig. 2. The mechanical steering system mainly comprises a steering sensor, a sensing end gear, a motor, a worm and worm wheel, a rack and a left pull rod and a right pull rod. The purpose of the hand-force torque conversion device 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. Meanwhile, the power-assisted torque output by the motor is amplified through a worm and a worm wheel, then the power-assisted torque is converted into rack power-assisted through the engagement of a sensing end gear and a rack, finally, the rack is pushed to move left and right together by the hand force of the rack and the resultant force of the rack power-assisted force, and then the tire steering of the whole vehicle is realized through a pull rod. The electronic control unit comprises an ECU, a motor control and the like, and aims to calculate the needed power assistance according to the signals of the angle sensor by the ECU, then control the motor to output necessary power assistance torque and ensure that the provided power assistance torque meets the design requirement. Based on the transmission principle and the mechanical structure of the steering system, the mechanical steering system can be equivalently transformed. The steering column is a rigid body with energy dissipation, and is equivalent to a damping element, the angle sensor has certain elastic deformation, and can be equivalent to an elastic element, and the meshing position of the worm gear and the gear rack has energy loss in the transmission process and elastic deformation when each tooth is meshed, so that the steering column is equivalent to a spring damping element, and according to an equivalent dynamics model, a dynamics motion differential equation can be deduced by utilizing Newton's law, and the specific form of a mathematical model is that
TABLE 1
By setting up the kinetic model, given input information such as the gap of the press block, the angle of the tie rod, the resistance of the tie rod in the working state, etc. are then set in the model, as shown in fig. 4 to 6.
S2, customizing a steering system sample, and ensuring that the input information of the actual sample is consistent with the input information of the dynamic model.
And S3, the steering system is installed on a corresponding whole vehicle, data acquisition is carried out in a test yard according to the designed vehicle speed and the road surface, and acceleration data, pull rod force data and corresponding subjective evaluation scores of a noise generating area of the steering system are mainly acquired, as shown in fig. 7.
S4, discretizing and normalizing the collected acceleration data, the pull rod force data, the impact energy obtained by simulation calculation and the subjective evaluation score, and classifying the data into a training set and a testing set.
Because the unit of acceleration is m/s2, the unit of pull rod force is N, the unit of impact energy is (m/s)2, the subjective division is a numerical value without unit, in order to avoid the numerical calculation problem and the network prediction error is larger because the magnitude difference of input and output data is larger, the convergence speed of the network is accelerated, the learning rate of the neural network is improved, and the like, all the data are subjected to maximum and minimum linear conversion normalization processing.
And S5, performing BP or RBF neural network learning, optimizing internal parameters of the neural network through cross verification, and finally establishing a neural network model between simulation impact energy and acceleration and a neural network model between acceleration and subjective evaluation score, as shown in fig. 8 and 9.
To fully exploit limited experimental data to train our neural network model, we use a cross-validation approach here. The method is characterized in that when a neural network model is trained, the calculated simulated impact energy data, the acquired whole vehicle acceleration data and the subjective data are divided into two parts, wherein one part is used for training the model, and the other part is used for evaluating the quality of the training model. Each sub-sample of the sample set used for training the model is used as a training target and a testing target, the optimal training model is fully found by utilizing a limited data set in the whole cyclic test verification process, and the trained model can reduce overfitting in the prediction of a new project, so that the precision is ensured.
S6, acquiring field data of a production line for the critical dimension and the critical performance parameter in the steering system, and performing parameter distribution analysis for the acquired field data so as to facilitate the setting of the critical dimension and the distribution estimation of the critical performance parameter during Monte Carlo sampling.
And S7, in the noise prediction process of the new project, firstly writing an algorithm program of a Monte Carlo sampling method, embedding the algorithm program into the dynamics model built in the S1, and carrying out noise simulation based on Monte Carlo sampling on the key size and the key performance parameters according to the parameter distribution obtained in the S6 so as to obtain the impact energy of the new project.
When a new project is assembled on a production line, the gap of the initial pressing block needs to be adjusted, and the production line only ensures that the gap of the pressing block does not exceed a certain specified value in order to produce the beat. As for the specific press block clearance value, the influence of the press block clearance value on the moving force and the wear after durability caused by the different initial values of the press block clearance value have great influence on the noise of the steering system.
Meanwhile, the steering system is formed by assembling a plurality of sub-parts into a whole, each part has a certain tolerance zone, the rigidity and the damping of each part are different among different batches of parts and among the same batch of parts, and finally, the gaps, the rigidities, the damping and the like which are shown by the steering engine assembled by the different parts can further amplify the noise performance differences which are shown by the whole steering engine.
In order to sufficiently evaluate the range of the steering noise and when the evaluated noise does not meet the demands of the market, it is necessary to sufficiently optimize the present design. At the moment, firstly, monte Carlo simulation sampling is carried out on key input information of a dynamic model, then calculation is carried out by combining the previous dynamic model to obtain impact energy as the response output of Monte Carlo, and finally, a boundary range of the impact energy is obtained.
Implementation and effect of actual case:
When the press block clearance is 0.2mm, the relationship between the tie rod force and the stroke in the parking state is shown in fig. 10. Fig. 10 is a graph of drag versus travel for a pull rod in a working condition, and fig. 11 is a graph of angle versus travel for a pull rod. As shown in fig. 12, the input information is input into the kinetic model, and the impact energy is calculated. 5 steering machines are prepared, the press block gap of which is controlled to be 0.2mm, and the steering machines are installed on the corresponding whole vehicle. The acceleration measurement and subjective evaluation were performed on these 5 steering machines, and the specific results are shown in fig. 13.
In the dynamic model, the difference between the dimensional tolerance and the sample assembly is considered, and the calculation is carried out by using a Monte Carlo simulation sampling method, so that a series of impact energy is obtained. And 5 groups of simulation results are randomly selected from the upper boundary, the lower boundary and the middle region, and specific numerical values are shown in fig. 14.
And predicting the acceleration values of the sensors in 5 states by using a trained neural network model of impact energy and acceleration, and comparing the sensor acceleration values with the acceleration values of the actual whole vehicle test in xyz three directions. As shown in fig. 15 to 19.
The subjective scores of the 5 states were predicted and compared with the subjective scores obtained from the actual whole vehicle test using the trained neural network model of subjective and acceleration, as shown in fig. 20.
The invention predicts and optimizes the noise of the steering system by combining machine learning and reliability theory. Firstly, a dynamic model between the interiors of a steering system is established, so that the dynamic model can well reflect the real mechanism of noise generation, then, a neural network model of impact energy and acceleration and a neural network between acceleration and subjective evaluation components are established, and the cross verification between normalization and training data is adopted, so that the data association between simulation data and subjective components can be well ensured, finally, the tolerance and performance fluctuation of the actual part size are considered in a certain range, and the real performance of the noise of the actual batch production parts is reflected by a Monte Carlo sampling method on the basis of the established dynamic model. The method has very high prediction precision and reliability, and is very simple, convenient and quick through two years of practical verification.