A self-learning prediction method for road noise of an automotive suspension structure

Abstract

The present invention belongs to the technical field of transportation vehicles, and particularly relates to a self-learning prediction method for road noise of an automotive suspension structure. The present invention provides a self-learning prediction method for road noise of an automotive suspension structure, aiming to enhance the robustness and interpretability of the automotive road noise prediction method. Based on a knowledge-driven model for suspension road noise and a data-driven model built using LSTM, the present invention iteratively fuses knowledge and data, aiming to enhance the robustness and interpretability of the automotive road noise prediction method, and can realize the self-learning of the knowledge-driven model and the data-driven model as the data samples are expanded, thereby providing effective guidance for automotive NVH engineers and improving the efficiency of automotive road noise prediction and analysis.

Description

Technical Field

[0001] The present invention relates to a method for self-learning prediction of road noise of an automotive suspension structure, and particularly to a method for self-learning prediction of road noise of an automotive suspension structure. Background Art

[0002] In recent years, the rapid development of the domestic automotive industry has not only improved people's living rhythm and quality, but also promoted the development of social productivity. Nowadays, the competition among various automotive companies is becoming increasingly fierce, and people's requirements for automobiles are also increasing. To improve competitiveness, in addition to an eye-catching appearance, the most crucial thing is the overall vehicle performance. The overall vehicle performance involves various professional departments, which requires one department to take the lead and be fully responsible for the overall organization, with the cooperation of other professional departments.

[0003] The overall vehicle performance of an automobile, also known as NVH performance, namely Noise, Vibration, and Harshness.

[0004] Automotive NVH performance has always been an important evaluation index during the development of passenger cars and is also one of the most concerned aspects of passengers' driving experience. With the popularization of new energy vehicles, since there is no longer the shielding of engine noise, the contribution of road noise to the overall vehicle noise has become more prominent. The so-called road noise is the noise response generated by the vibration transmitted to the vehicle body after the road surface excitation is attenuated by rubber components such as tires and suspension bushings and then acts jointly on the vehicle body wall panels and the interior acoustic cavity. For a normally running vehicle, the low-frequency noise below 300 Hz mainly occurs when the vehicle is running at a medium speed, and this operating condition is the most common and typical one on urban roads. When the vehicle is running under this condition, road noise is the main noise source inside the vehicle. Therefore, the analysis and prediction of the overall vehicle road noise are key issues of concern in automotive NVH design.

[0005] Traditional methods for analyzing vehicle road noise include Transfer Path Analysis, finite element modeling and simulation analysis, and dynamic modeling. Among the currently available research materials related to vehicle road noise, the invention patent "A Method for Testing the Structural Road Noise Transfer Function" by Lu Chihua, Zhu Yawei, etc. reduces the influence of strong coherent excitation input and background noise on the structural noise transfer function by applying periodic pulse excitation to each wheel in turn and windowing the measured signals (CN110243609B). In this invention, since multiple vehicle tests are required, the testing and debugging process has a long cycle and thus a high cost. Wang Dengfeng, Li Junming, etc. established a mathematical model of in-vehicle noise generated by random excitation of road surface unevenness and analyzed the influence of mount and tire parameters on in-vehicle noise based on the model (Automotive Engineering, 2000(02):93-96+142.). In this research, mainly mathematical analysis of the vibration mechanism was carried out. Since the vehicle was simplified to a linear system and many equivalences and simplifications were involved, the calculation results deviated greatly from the measured values. Huang Haibo, Huang Xiaorong, etc. conducted subjective evaluation and objective parameter calculation on in-vehicle noise signals under constant speed conditions, and carried out correlation analysis on the subjective and objective evaluation results. On this basis, an in-vehicle noise sound quality prediction model was established based on the Adaboost algorithm combined with BP neural network, ELM, and SVM. The prediction accuracy and precision of this model are higher than those of the GA-BP, GA-ELM, and GA-SVM sound quality prediction models with optimized parameters (Automotive Engineering, 2016, 38(09):1120-1125.). This research introduced intelligent algorithms on the basis of traditional methods, but it lacks the interpretability of the mechanism and the data model strongly depends on the sample quality. The difference between the present invention and the above papers and patents lies in that it has established a vehicle dynamics analysis model for suspension road noise analysis and a data-driven model based on data driving, and iteratively fused knowledge and data, enhancing the interpretability and robustness of road noise prediction while improving the efficiency of road noise prediction analysis. Summary of the Invention

[0006] Aiming at the problems existing in the background technology, the present invention provides a self-learning prediction method for vehicle suspension structure road noise, aiming to enhance the robustness and interpretability of vehicle road noise prediction methods. Based on a vehicle dynamics analysis model for suspension road noise and a data-driven model built using LSTM, the present invention iteratively fuses knowledge and data, aiming to enhance the robustness and interpretability of vehicle road noise prediction methods, and can realize the self-learning of the knowledge-driven model and the data-driven model as the data samples are expanded, so as to provide effective guidance for vehicle NVH engineers and improve the efficiency of vehicle road noise prediction analysis.

[0007] The present invention provides the following technical solutions to solve the above technical problems:

[0008] A self - learning prediction method for road noise of an automotive suspension structure, comprising the following steps:

[0009] Step 1: According to the different structural characteristics of different automotive suspensions, analyze the structural road noise transfer path from road excitation to the noise in the right ear of the driver in the vehicle for the corresponding suspension, and establish a multi - level decomposition tree of suspension structural road noise according to the hierarchical standard of "vehicle level - system level - subsystem level - component level";

[0010] Step 2: Embed an intelligent algorithm in the multi - level decomposition tree, and use the intelligent algorithm to learn the data relationship between each target node in the multi - level decomposition tree to obtain a data - driven model;

[0011] Step 3: Establish a vehicle dynamics analysis model from road excitation to the suspension body attachment point according to existing knowledge, and then obtain the transfer relationship from the vibration of the suspension body attachment point to the road noise in the vehicle based on experimental analysis, so as to construct a knowledge - driven model from road excitation to the noise in the right ear of the driver in the vehicle;

[0012] Step 4: Based on the established data - driven model and knowledge - driven model, use a multi - objective optimization algorithm, with the evaluation indexes of the two models and the mean square error of the prediction results between the two as the objectives, optimize the built - in hyperparameters of the data - driven model and the set parameters of the knowledge - driven model. Finally, multiply the prediction results of the two models by their respective set weights and sum them, and use the summation result as the final prediction result.

[0013] After adopting this technical solution, a vehicle dynamics analysis model for suspension road noise analysis is established, and a data - driven model based on data - driving is also established, and knowledge and data are fused and iterated, enhancing the interpretability and robustness of road noise prediction while improving the efficiency of road noise prediction analysis.

[0014] Preferably, it further includes Step 5 for continuing Step 4: Based on the multi - level decomposition tree of suspension structural road noise and the vehicle dynamics analysis model as the data relationship basis, establish a MySQL database according to experimental and simulation data, and connect the database to the data - driven model and the knowledge - driven model in real time.

[0015] After adopting this preferred solution, after the database is updated, the knowledge - driven model and the data - driven model can autonomously learn new sample data and improve their generalization ability.

[0016] Preferably, Step 2 includes the following steps:

[0017] Sort out and normalize the experimental and simulation data according to the multi - level decomposition tree of suspension structural road noise, and then divide them into a training set, a test set and a validation set according to a certain ratio;

[0018] Introduce the multi-layer LSTM algorithm and the Dropout method between each target node at the upper and lower levels in the multi-level decomposition tree of the road noise of the suspension structure to construct a data-driven model, and train the data-driven model based on the training set. The LSTM structure formula is as follows:

[0019]

[0020] In the formula, f is the forgetting gate, g is the memory cell, i is the input gate, o is the output gate, σ is the sigmoid function, x t is the input at this moment, h t-1 is the hidden state of the output at the previous moment, is the weight corresponding to x t ; is the weight corresponding to h t-1 , and b (f) is the bias corresponding to f;

[0021] Input the data of the test set into the trained data-driven model for model evaluation, and determine the prediction effect of the data-driven model according to the mean square error of the data-driven model evaluation index;

[0022] Adjust the number of LSTM layers of the data-driven model and the neuron loss percentage of Dropout, and repeat the training, testing, and evaluation of the data-driven model until the mean square error of the data-driven model converges.

[0023] After adopting this optimal solution, the Dropout method randomly deletes each layer of neurons during the learning process. The deleted neurons will no longer transmit signals, and overfitting of the model is prevented by reducing the complexity of the model itself.

[0024] Based on the multi-layer LSTM algorithm and the Dropout method, the learning ability of the model can be further improved by stacking the number of LSTM layers.

[0025] Preferably, step 3 includes the following steps:

[0026] Simplify and equivalent the vehicle system: The sprung mass and unsprung mass of the vehicle are regarded as rigid bodies. Considering the vertical degrees of freedom of the four wheels and the vertical, roll, and pitch degrees of freedom of the vehicle body as well as the local degrees of freedom of the four suspension body attachment points, after obtaining the equivalent stiffness and damping parameters of each component, establish an eleven-degree-of-freedom vehicle dynamics analysis model;

[0027] The frequency response function test is carried out in a semi-anechoic chamber. Using the reciprocity principle, a spatially omnidirectional sound source is placed at the right ear of the driver as an excitation, and then the vibration acceleration response of the suspension body attachment point is measured. Finally, the transfer relationship from the suspension body attachment point to the noise at the right ear of the driver inside the vehicle is obtained by the matrix inversion method. The matrix inversion formula is as follows:

[0028] {F N}=[H MN -1 ×{X M};

[0029] In the formula: {F N} is the system input excitation vector, {X M} is the response vector of the response point, and H MN is the transfer function from the input to the response;

[0030] Build a vehicle dynamics analysis model in Simulink, and introduce the transfer function from the suspension body attachment point to the noise at the right ear of the driver obtained from the test to construct a knowledge-driven model.

[0031] Preferably, step 4 includes the following steps:

[0032] Based on the data of the validation set in step 2, use the NSGA2 genetic algorithm as the multi-objective optimization algorithm, and take the evaluation indexes of the data-driven model and the knowledge-driven model and the deviation of the prediction results between the two as the optimization objectives, and iteratively update the weight parameters of the memory unit g in the data-driven model and the setting of the body dynamic stiffness in the knowledge-driven model. The formula is as follows:

[0033]

[0034] In the formula, y d is the output result of the data-driven model, t k is the true value of the data, y k is the output result of the knowledge-driven model, k is the number of samples, W is the weight parameter of the memory unit g, μ1, μ2, μ3, μ4 are the equivalent dynamic stiffness of the suspension body attachment point, and i1, i2 are the optimization intervals set according to the engineering practice;

[0035] According to the prediction effects of the data-driven model and the knowledge-driven model, allocate the weights of the two according to a certain proportion, and multiply the prediction results of the two by their respective weights and then sum them as the final prediction result.

[0036] After adopting this technical solution, the prediction results of the data-driven model and the knowledge-driven model are weighted to obtain the final prediction result, which further improves the prediction accuracy compared with using the data-driven model or the knowledge-driven model alone.

[0037] ​In summary, due to the adoption of the above technical solutions, the beneficial effects of the present invention are as follows:

[0038] 1. Based on test and simulation data, and combined with intelligent algorithms for predictive analysis of road noise in the automotive suspension structure, the present invention improves the robustness and accuracy of prediction.

[0039] 2. The present invention introduces a knowledge model of vehicle dynamics to characterize the mechanism of road noise in the suspension structure, fuses knowledge and data, and realizes double iterative learning of data-driven models and knowledge-driven models.

[0040] 3. The present invention establishes a database for road noise in the suspension structure. With the help of data-driven models and knowledge-driven models, it realizes the autonomous learning of data-driven models and knowledge-driven models for updated data, improves the generalization ability, and thus enhances the efficiency of predictive analysis of road noise in the suspension structure. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 is a technical flow chart.

[0042] Figure 2 is a schematic diagram of Dropout.

[0043] Figure 3 is a flow chart of the NSGA2 algorithm.

[0044] Figure 4 is a schematic diagram of the multi-level decomposition tree of the MacPherson suspension.

[0045] Figure 5 is a schematic diagram of the multi-level decomposition tree of the inverted E-type multi-link suspension.

[0046] Figure 6 is a schematic diagram of the vibration acceleration in the X direction of the front steering knuckle.

[0047] Figure 7 is a schematic diagram of the eleven-degree-of-freedom vehicle dynamics model.

[0048] Figure 8 is a comparison chart of prediction results.

[0049] Figure 9 is a schematic diagram of the architecture of the database for road noise in the automotive suspension structure. DETAILED DESCRIPTION OF THE INVENTION

[0050] In order to make the technical means, features and effects achieved by the present invention easier to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with specific embodiments and the accompanying drawings in the embodiments of the present invention.

[0051] As Figures 1-9 shown, the optimal embodiment of the present invention is described as follows:

[0052] See Figure 1 As shown, a self - learning prediction method for road noise of an automotive suspension structure includes the following steps:

[0053] Step 1: According to the different structural characteristics of different automotive suspensions, analyze the structural road - noise transfer path from road excitation to the noise in the right ear of the driver in the vehicle for the corresponding suspension, and establish a multi - level decomposition tree of suspension structural road noise according to the hierarchical standard of "vehicle - level—system - level—sub - system - level—component - level";

[0054] Step 2: Embed an intelligent algorithm in the multi - level decomposition tree, and use the intelligent algorithm to learn the data relationship between each target node in the multi - level decomposition tree to obtain a data - driven model;

[0055] Step 3: Establish a vehicle dynamics analysis model from road excitation to the suspension body attachment point according to existing knowledge, and then obtain the transfer relationship from the vibration of the suspension body attachment point to road noise in the vehicle based on experimental analysis, so as to construct a knowledge - driven model from road excitation to the noise in the right ear of the driver in the vehicle;

[0056] Step 4: Based on the established data - driven model and knowledge - driven model, use a multi - objective optimization algorithm. Taking the evaluation indexes of the two models and the mean square error of the prediction results between the two as the objectives, optimize the built - in hyperparameters of the data - driven model and the set parameters of the knowledge - driven model. Finally, multiply the prediction results of the two models by their respective set weights and sum them up, and take the sum result as the final prediction result.

[0057] In this embodiment, a vehicle dynamics analysis model for suspension road - noise analysis is established, and a data - driven model based on data - driven is also established, and knowledge and data are fused and iterated, enhancing the interpretability and robustness of road - noise prediction while improving the efficiency of road - noise prediction analysis.

[0058] See Figure 1 As shown, this embodiment also includes Step 5 for continuing the above - mentioned Step 4: Based on the multi - level decomposition tree of suspension structural road noise and the vehicle dynamics analysis model as the data relationship basis, establish a MySQL database according to experimental and simulation data, and connect the database to the data - driven model and the knowledge - driven model in real - time.

[0059] In this embodiment, after the database is updated, the knowledge - driven model and the data - driven model can autonomously learn new sample data and improve their generalization ability.

[0060] See Figure 3 As shown, the said Step 2 includes the following steps:

[0061] Sort out and normalize the experimental and simulation data according to the multi - level decomposition tree of suspension structural road noise, and then divide them into a training set, a test set and a validation set according to a certain ratio;

[0062] Introduce the multi - layer LSTM algorithm and Dropout method between each target node at the upper and lower levels in the multi - level decomposition tree of the suspension structure road noise to build a data - driven model, and train the data - driven model based on the training set. The LSTM structure formula is as follows:

[0063]

[0064] In the formula, f is the forget gate, g is the memory cell, i is the input gate, o is the output gate, σ is the sigmoid function, x t is the input at this moment, h t-1 is the hidden state of the output at the previous moment, is the weight corresponding to x t ; is the weight corresponding to h t-1 ; b (f) is the bias corresponding to f;

[0065] Input the data of the test set into the trained data - driven model for model evaluation, and determine the prediction effect of the data - driven model according to the mean square error of the data - driven model evaluation index;

[0066] Adjust the number of LSTM layers of the data - driven model and the neuron loss percentage of Dropout, and repeat the training, testing, and evaluation of the data - driven model until the mean square error of the data - driven model converges.

[0067] In this embodiment, the Dropout method randomly deletes each layer of neurons during the learning process. The deleted neurons will no longer transmit signals, and overfitting of the model is prevented by reducing the complexity of the model itself.

[0068] Based on the multi - layer LSTM algorithm and Dropout method, the learning ability of the model can be further improved by stacking the number of LSTM layers.

[0069] See Figure 1 As shown, step 3 includes the following steps:

[0070] Simplify and equivalent the vehicle system: The sprung mass and unsprung mass of the vehicle are regarded as rigid bodies. Consider the vertical degrees of freedom of the four wheels and the vertical, roll, and pitch degrees of freedom of the vehicle body as well as the local degrees of freedom of the four suspension body attachment points. After obtaining the equivalent stiffness and damping parameters of each component, establish an eleven - degree - of - freedom vehicle dynamics analysis model;

[0071] The frequency response function test is carried out in a semi-anechoic chamber. Using the reciprocity principle, a spatially omnidirectional sound source is placed at the right ear of the driver as the excitation, and then the vibration acceleration response of the suspension body attachment point is measured. Finally, the transfer relationship from the suspension body attachment point to the noise at the right ear of the driver inside the vehicle is obtained by the matrix inversion method. The matrix inversion formula is as follows:

[0072] {F N}=[H MN -1 ×{X M};

[0073] In the formula: {F N} is the system input excitation vector, {X M} is the response vector of the response point, and H MN is the transfer function from the input to the response;

[0074] Build a vehicle dynamics analysis model in Simulink, and introduce the transfer function from the suspension body attachment point to the noise at the right ear of the driver obtained from the test to construct a knowledge-driven model.

[0075] See Figure 2 shown, the step 4 includes the following steps:

[0076] Based on the data of the validation set described in step 2, use the NSGA2 genetic algorithm as the multi-objective optimization algorithm, and take the evaluation indexes of the data-driven model and the knowledge-driven model and the deviation of the prediction results between the two as the optimization objectives, and iteratively update the weight parameters of the memory unit g in the data-driven model and the setting of the body dynamic stiffness in the knowledge-driven model. The formula is as follows:

[0077]

[0078] In the formula, y d is the output result of the data-driven model, t k is the true data value, y k is the output result of the knowledge-driven model, k is the number of samples, W is the weight parameter of the memory unit g, μ1, μ2, μ3, μ4 are the equivalent dynamic stiffness of the suspension body attachment point, and i1, i2 are the optimization intervals set according to the engineering practice;

[0079] According to the prediction effects of the data-driven model and the knowledge-driven model, allocate the weights of the two according to a certain proportion, and multiply the prediction results of the two by their respective weights and then sum them as the final prediction result.

[0080] In this embodiment, the prediction results of the data-driven model and the knowledge-driven model are weighted to obtain the final prediction result, which further improves the prediction accuracy compared with using the data-driven model or the knowledge-driven model alone. ​

[0081] In this embodiment, based on test and simulation data, combined with intelligent algorithms, the road noise of the vehicle suspension structure is predicted and analyzed to improve the robustness and accuracy of the prediction.

[0082] In this embodiment, a knowledge model of vehicle dynamics is introduced to characterize the mechanism of road noise of the suspension structure, and knowledge and data are fused to achieve double iterative learning of data-driven models and knowledge-driven models.

[0083] In this embodiment, a road noise database of the suspension structure is established. With the help of data-driven models and knowledge-driven models, the data-driven models and knowledge-driven models are enabled to autonomously learn updated data, improve the generalization ability, and thus enhance the efficiency of predicting and analyzing the road noise of the suspension structure.

[0084] The working principle of the invention will be further described below in conjunction with the drawings so that those skilled in the art can fully understand the invention, which is specifically described as follows:

[0085] A self-learning prediction method for road noise of a vehicle suspension structure according to the invention takes a vehicle model with a front MacPherson suspension and a rear inverted E-type multi-link suspension as an example to predict the road noise of the vehicle suspension structure.

[0086] Refer to Figure 1 As shown, it includes the following steps:

[0087] Step 1: Refer to Figure 4 and Figure 5 As shown, analyze the different vibration transmission paths of the suspension according to the different structural compositions of the MacPherson suspension and the inverted E-type multi-link suspension, and uniformly divide the nodes in each transmission path according to the hierarchy of "vehicle level - system level - subsystem level - component level", and construct a multi-level decomposition tree of road noise of the suspension structure from the dynamic parameters of the bottom components to the noise of the driver's right ear in the top-level vehicle interior, and construct a multi-level decomposition tree of the MacPherson suspension and a multi-level decomposition tree of the inverted E-type multi-link suspension.

[0088] Step 2: Embed intelligent algorithms in the multi-level decomposition tree of road noise of the suspension structure, and use the intelligent algorithms to learn the data relationships between the various target nodes in the decomposition tree, so as to construct a data-driven model:

[0089] 2.1 Screen and divide the test and simulation data based on the upper and lower level relationships between the various target points in the multi-level decomposition tree of road noise of the suspension structure. Since the research on suspension structure noise mainly focuses on the low-frequency band, all vibration and noise data are sorted in the range of 0 Hz to 300 Hz with an interval of 1 Hz. Taking the vibration acceleration in the X direction of the front steering knuckle as an example of the sorted data result, the data samples are as Figure 6As shown, linear function normalization is then used to normalize the basic data according to the corresponding frequency points. Finally, all data samples are shuffled and divided into a training set, a test set, and a validation set in a ratio of 8:1:1. The normalization formula is as follows:

[0090]

[0091] In the formula, X is the original data, X min and X max are the minimum and maximum values of the original data set respectively, and X nom is the normalized data.

[0092] 2.2 For each target node in the multi-level decomposition tree of the suspension structure road noise, the LSTM algorithm is introduced to learn the internal relationship between nodes. First, an input gate i, an output gate o, a forget gate f, and a memory unit g are introduced based on the idea of the traditional RNN to establish an LSTM layer. On this basis, another LSTM layer is stacked to improve the model's learning ability for time series data. Then, the Dropout method is used to establish a Dropout layer with the neuron deletion ratio set to 20% to prevent model overfitting. Finally, all the parameters in the model are randomly initialized, and the model is trained based on the training set to construct a data-driven model for the suspension structure road noise.

[0093] 2.3 After the model training is completed, the data of the test set is input for model evaluation. According to the mean square error MSE of the data-driven model evaluation index, the prediction effect of the data-driven model is determined. The MSE formula is as follows:

[0094]

[0095] In the formula, E is the mean square error value, y k is the output result of the data-driven model, t k is the true data value, and k is the sample size of the data.

[0096] 2.4 Taking the mean square error as the model evaluation index, it is found that the prediction effect of this model still needs to be improved. Therefore, the number of stacked LSTM layers and the neuron deletion percentage of the Dropout layer in the data-driven model are appropriately adjusted. The comparison results of the model evaluation index and the calculation duration under different parameter adjustments are shown in Table 1. After comprehensively considering the model prediction accuracy and the calculation duration, finally, stacking 3 LSTM layers and setting the neuron deletion percentage of the Dropout layer to 30% are selected as the final model architecture.

[0097] Table 1 Comparison of data-driven models under different parameters

[0098]

[0099] Step 3: First, establish a vehicle dynamics analysis model from road excitation to the suspension body attachment points, and then obtain the transfer relationship from the vibration of the suspension body attachment points to the interior road noise based on experimental analysis, so as to construct a knowledge-driven model from road excitation to the noise in the right ear of the driver in the vehicle:

[0100] 3.1 Refer to Figure 7 As shown, simplify and equivalent the vehicle system. Consider the sprung mass and unsprung mass of the vehicle as rigid bodies, and consider the vertical degrees of freedom of the four wheels and the vertical, roll, and pitch degrees of freedom of the vehicle body, as well as the local degrees of freedom of the four suspension body attachment points. After obtaining the equivalent stiffness, damping, and other parameters of each component required for the model, establish an eleven-degree-of-freedom vehicle dynamics analysis model, and transform the kinetic energy, potential energy, and energy dissipation formulas of the model system based on Lagrangian transformation to obtain the following differential equations;

[0101]

[0102] In the formula, M is the sprung mass; Z1, Z2, Z3, Z4 are road excitations; Z5 is the body displacement; K a1 , K a2 , K a3 , K a4 is the equivalent dynamic stiffness of the suspension body attachment point; C a1 , C a2 , C a3 , C a4 is the damping of the suspension body attachment point; Z a , Z b , Z c , Z d is the attachment point displacement; Z w1 , Z w2 , Z w3 , Z w4 is the displacement between the attachment point and the suspension system; I X , I y are the roll and pitch moments of inertia; θ, φ are the roll angle and pitch angle; m1, m2, m3, m4 are the unsprung masses; Z s1 , Z s2 , Z s3 , Z s4 is the tire displacement; C b1 , C b2 , C b3 , C b4 is the suspension system damping; K b1 , K b2 , K b3 , K b4 is the suspension system stiffness; K t1 , K t2 , K t3 , Kt4 is the tire stiffness; a and b are the distances from the front and rear axles to the vehicle's center of mass respectively; d is the track width.

[0103] 3.2 Conduct the frequency response function test FRF in an anechoic chamber. Select two excitation points on the left and right of the front suspension and the rear suspension respectively. Pay attention to the X direction at each excitation point. There are a total of 4 noise transfer paths. Use the reciprocity principle to place a spatially omnidirectional sound source at the right ear of the driver as the excitation, then measure the vibration acceleration response at the attachment points of the suspension to the body, and finally obtain the transfer relationship from the attachment points of the suspension to the body to the noise at the right ear of the driver inside the vehicle through matrix inversion.

[0104] 3.3 Build a vehicle dynamics analysis model in Simulink, and introduce the transfer function from the attachment points of the suspension to the body to the noise at the right ear of the driver obtained from the test to construct a knowledge-driven model.

[0105] Step 4: Based on the established data-driven model and knowledge-driven model, use the multi-objective optimization algorithm. Take the evaluation indexes of the two models and the deviation of the prediction results between the two as the objectives, optimize the built-in hyperparameters of the data-driven model and the set parameters of the knowledge-driven model, and finally multiply the prediction results of the two models by their respective set weights and sum them as the final prediction result.

[0106] 4.1 Based on the data of the validation set, optimize the data-driven model and the knowledge model. The input of the data-driven model is the underlying parameters in the multi-level decomposition tree of the suspension structure road noise, such as the dynamic stiffness of the bushing and the vibration acceleration signal. The input of the knowledge-driven model is the road surface spectrum corresponding to the same test data. Use the NSGA2 genetic algorithm as the multi-objective optimization algorithm. Take the prediction evaluation indexes MSE d 、MSE k and the mean square error of the prediction results between the two as the optimization objectives. Expand the weight parameters in the memory unit g of the data-driven model up and down by 50% as the optimization interval, and set the floating range of the dynamic stiffness value of the suspension body in the knowledge-driven model according to engineering practical experience as the optimization interval. Finally, set the number of iterations to 500 times and optimize the two models simultaneously. According to the optimized results, the mean square error of the data-driven model is reduced by 17%, and the mean square error of the knowledge-driven model is reduced by 13%.

[0107] 4.2 Based on the current test data of this vehicle model, the mean squared error of the data-driven model is 0.078, and the mean squared error of the knowledge-driven model is 0.187. By comparison, it is found that the data-driven model has a better prediction effect than the knowledge-driven model. Therefore, according to the relationship between the evaluation indexes of the prediction results of the two models, their weights are set in a ratio of 0.54:0.46. Finally, the prediction results of the two models are multiplied by their respective weights and then summed as the final prediction result. Taking a set of the latest measured data of this vehicle model as an example, the prediction effects of the single model and the data-knowledge fusion dual-drive model are compared as Figure 8 shown. It can be clearly seen from the figure that the knowledge-data fusion dual-drive model exhibits better performance.

[0108] Step 5: Based on the multi-level decomposition tree of the suspension structure road noise and the overall vehicle dynamics analysis model architecture as the basis of the data relationship, use the MySQL relational database system to establish an automotive suspension structure road noise database. The overall architecture is as Figure 9 shown. Through this database, functions such as storage, deletion, modification, and query of vehicle test data and vehicle dynamics model simulation data are realized, and the database is docked with the knowledge-driven model and the data-driven model, so that after the database is updated, the knowledge-driven model and the data-driven model can automatically use the new data samples as the training set for training. Thus, by realizing the autonomous learning of new knowledge by the knowledge-driven model and the data-driven model, the model has the ability of infinite iterative upgrade.

[0109] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present invention. Any reference signs in the claims should not be regarded as limiting the claimed rights.

Claims

1. A self - learning prediction method for road noise of an automotive suspension structure, characterized in that: It includes the following steps: Step 1: According to the different structural characteristics of different suspensions of the vehicle, analyze the structural road noise transfer path from road surface excitation to the noise in the right ear of the driver in the vehicle, and establish a multi-level decomposition tree of suspension structural road noise according to the hierarchical standard of "vehicle level - system level - subsystem level - component level"; Step 2: Embed an intelligent algorithm in the multi-level decomposition tree, and use the intelligent algorithm to learn the data relationship between each target node in the multi-level decomposition tree to obtain a data-driven model; Step 3: Establish a vehicle dynamics analysis model from road surface excitation to the suspension body attachment point, and then based on experimental analysis, obtain the transfer relationship from the vibration of the suspension body attachment point to the road noise in the vehicle, so as to construct a knowledge-driven model from road surface excitation to the noise in the right ear of the driver in the vehicle; Step 4: Based on the established data-driven model and knowledge-driven model, use a multi-objective optimization algorithm, with the evaluation indexes of the two models and the mean square error of the prediction results between the two as the objectives, optimize the built-in hyperparameters of the data-driven model and the set parameters of the knowledge-driven model, and finally multiply the prediction results of the two models by their respective set weights and sum them, and take the sum result as the final prediction result; The said Step 2 includes the following steps: Sort out and normalize the experimental and simulation data according to the multi-level decomposition tree of suspension structural road noise, and then divide them into a training set, a test set and a validation set according to a certain ratio; Introduce a multi-layer LSTM algorithm and Dropout method between each target node at the upper and lower levels in the multi-level decomposition tree of suspension structural road noise to construct a data-driven model, and train the data-driven model based on the training set. The LSTM structure formula is as follows: where f is the forget gate, g is the memory cell, i is the input gate, o is the output gate, σ is the sigmoid function, x t is the input at this time step, h t-1 is the hidden state of the output at the previous time step, is the weight corresponding to x t , is the weight corresponding to h t-1 , b (f) is the bias corresponding to f; Input the data of the test set into the trained data-driven model for model evaluation, and determine the prediction effect of the data-driven model according to the mean square error of the data-driven model evaluation index; Adjust the number of LSTM layers of the data-driven model and the neuron loss percentage of Dropout, and repeat the training, testing and evaluation of the data-driven model until the mean square error of the data-driven model converges.

2. The self - learning prediction method for road noise of an automotive suspension structure according to claim 1, wherein: It also includes Step 5 for continuing the said Step 4: Based on the multi-level decomposition tree of suspension structural road noise and the vehicle dynamics analysis model as the data relationship basis, obtain the basic road noise data through vehicle road tests or CAE vehicle modeling and simulation, and then establish a MySQL database according to the basic road noise data, and connect the database with the data-driven model and the knowledge-driven model in real time.

3. The self - learning prediction method for road noise of an automotive suspension structure according to claim 1, wherein: The said Step 3 includes the following steps: Simplify and equivalent the vehicle system: Consider the sprung mass and unsprung mass of the vehicle as rigid bodies, consider the vertical degrees of freedom of the four wheels and the vertical, roll and pitch degrees of freedom of the vehicle body as well as the local degrees of freedom of the four suspension body attachment points. After obtaining the equivalent stiffness and damping parameters of each component, establish an eleven-degree-of-freedom vehicle dynamics analysis model; The frequency response function test is carried out in a semi-anechoic chamber. Using the reciprocity principle, a spatially omnidirectional sound source is placed at the right ear of the driver as the excitation, and then the vibration acceleration response of the suspension body attachment point is measured. Finally, the transfer relationship from the suspension body attachment point to the noise at the right ear of the driver inside the vehicle is obtained by the matrix inversion method. The matrix inversion formula is as follows: {F N} = [H MN -1 × {X M};​ Where: {F N} is the system input excitation vector, {X M} is the response vector of the response point, and H MN is the transfer function from the input to the response; Build a vehicle dynamics analysis model in Simulink, and introduce the transfer function from the suspension body attachment point to the noise at the right ear of the driver obtained from the test to construct a knowledge-driven model.

4. The self - learning prediction method for road noise of an automotive suspension structure according to claim 1, wherein: The step 4 includes the following steps: Based on the data of the validation set described in step 2, using the NSGA2 genetic algorithm as the multi-objective optimization algorithm, taking the evaluation indexes of the data-driven model and the knowledge-driven model and the deviation of the prediction results between the two as the optimization objectives, iteratively update the weight parameters of the memory unit g in the data-driven model and the setting of the body dynamic stiffness in the knowledge-driven model. The formula is as follows: where y d is the output result of the data-driven model, t k is the true data value, y k is the output result of the knowledge-driven model, k is the number of samples, W is the weight parameter of the memory unit g, μ1, μ2, μ3, μ4 are the equivalent dynamic stiffness of the suspension body attachment points, and i1, i2 are the optimization intervals set according to the engineering practice; According to the prediction effects of the data-driven model and the knowledge-driven model, allocate the weights of the two according to a certain proportion, multiply the prediction results of the two by their respective weights and then sum them as the final prediction result.

Images