Road unevenness estimation method based on vehicle acceleration signal

Abstract

The invention discloses a road unevenness estimation method based on a vehicle acceleration signal, and relates to the technical field of vehicle engineering. According to the method, a nonlinear whole vehicle model based on polynomial fitting tire and damper nonlinear dynamic characteristic experimental data is provided, a data set of the model is used for training an NARX neural network, and an efficient road unevenness estimation system is established; a G-NARX neural network model more suitable for random excitation is also used, and the G-NARX model shows better performance in the aspect of road unevenness estimation. The road unevenness classification system collects vertical acceleration signals through vehicle-mounted acceleration and then inputs the signals into a neural network to calculate the root-mean-square of the road space power spectral density so as to realize road unevenness estimation. On-site tests are carried out on three typical road surfaces of the Chengdami city of Hubei province in China, and the robustness of a road unevenness estimation system of the model is proved.

Description

Technical Field

[0001] The invention belongs to the technical field of vehicle engineering and relates to a road roughness estimation method based on vehicle acceleration signals. Background Art

[0002] The condition of the road is a key factor affecting vehicle performance. Uneven roads can reduce the vehicle's driving stability, which in turn affects the driving experience and even increases the risk of accidents. Therefore, estimating the roughness of the road is crucial for optimizing vehicle control systems, improving driving experience, and ensuring handling stability.

[0003] There are three main types of road roughness estimation, namely contact measurement, non-contact measurement, and estimation based on system response. In the research on road profile estimation based on artificial neural network (ANN) method of system response, most of the existing studies use data from linear 1 / 4 suspension model or linear multi-degree-of-freedom vehicle model to train the neural network. The linear 1 / 4 suspension model simplifies the vehicle model and only considers the interaction between a single wheel and the road surface, ignoring the influence of other parts of the vehicle. The linear multi-degree-of-freedom vehicle model considers the overall response of the vehicle to a certain extent, but due to the linearization of the model, it cannot accurately reflect the nonlinear characteristics of components such as the vehicle suspension system and tires. Summary of the invention

[0004] In order to make up for the shortcomings of the prior art, the present invention provides a method for estimating road roughness based on vehicle acceleration signals. The present invention proposes a cost-effective and efficient method for evaluating road roughness. The method combines the vertical acceleration data of the vehicle tires, frame and cab and the vehicle speed data to effectively estimate the roughness of the road surface. The present invention is implemented by the following technical solutions: The present invention provides a method for estimating road roughness based on vehicle acceleration signals, including Establish a nonlinear full vehicle model with ten degrees of freedom; According to Newton's second law, the vertical vibration equation of the tire assembly is established. Formulate differential equations for the vertical vibration, pitch, and roll motions of the vehicle frame, Establish the governing differential equations for the vertical vibration, pitch motion, and roll motion of the cab Gaussian white noise with unit intensity of 1 is used to establish the time domain model of the road surface; The bench test method is used to analyze the relationship between the damping force of the shock absorber and its compression or extension speed; Describe the nonlinear characteristics of tires and use quadratic curve fitting method to analyze experimental data obtained under different tire pressure conditions; Define the equations of the NARX neural network model, The open-loop NARX model is trained using pseudo output feedback using the G-NARX model, which minimizes the deviation between the initial training phase and the subsequent training phases by integrating Gaussian noise into the original target data to generate pseudo output feedback; output feedback of the G-NARX model; Determine the relationship between temporal and spatial frequencies of road inputs during vehicle motion; Calculate the spatial frequency power spectral density of the road profile; The trapezoidal rule is applied to calculate the integral of the discrete spatial power spectral density, and the root mean square value of the spatial power spectral density is further calculated; Calculate the root mean square value of the power spectrum density of the road disturbance in the current time window; For the same type of roads, the root mean square value of the spatial power spectrum density falls within a specific range; the roughness of the road is evaluated by estimating the root mean square value of the spatial power spectrum density of the road disturbance and defining its interval; The wheels, frame, cab and vehicle speed are selected as the input data of the neural network model G-NARX, and the root mean square value of the road power spectrum density is used as the output data for training the neural network, that is, the training set of four input variables and one output variable is used to train the neural network; The root mean square value of the road roughness power spectrum density output by the neural network is used to classify the road surface, thereby obtaining the road roughness interval of the current road.

[0005] The beneficial effects of the present invention are: The present invention proposes a cost-effective and efficient method for evaluating road roughness, which combines the vertical acceleration data of the vehicle tires, frame and cab and the vehicle speed data to effectively estimate the roughness of the road surface.

[0006] 1. The proposed vehicle model incorporates the nonlinear characteristics of shock absorbers and tires, has ten degrees of freedom, and has been experimentally validated. The model demonstrates a high degree of accuracy, as evidenced by the agreement between the acceleration responses at three different locations and the actual data.

[0007] 2. Increasing the time window length significantly reduces the error rate of road roughness estimation, and the estimation accuracy is optimal at a 10-second interval.

[0008] 3. Compared with the traditional NARX network, the G-NARX network proposed in the present invention achieves significant improvement in the accuracy of road roughness classification, with the accuracy increased by at least 26%.

[0009] 4. Through actual road tests, the estimation method based on acceleration data is proved to be economical and effective, demonstrating its strong performance under different vehicle speeds and road conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] The present invention will be further described below in conjunction with the accompanying drawings; Figure 1 It is a schematic diagram of a nonlinear full-vehicle model with ten degrees of freedom of the present invention; Figure 2 Fitting experimental data and fitting curves to cubic polynomials; Figure 3 for , Value varies with tire pressure Change graph; Figure 4 For the NARX model; Figure 5 A road roughness estimation method based on vehicle acceleration signal; Figure 6 is the time domain data and the root mean square value of the spatial power spectrum density of the mixed road surface; Figure 7 The following are the classification effects of mixed road surfaces under different time windows. DETAILED DESCRIPTION

[0011] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention, so the present invention is not limited by the specific embodiments disclosed below.

[0012] The core of the present invention is a road roughness estimation method based on a vehicle acceleration signal, which specifically includes the following steps: 1. Model building: Utilize rigorously established road surface models and vehicle nonlinear models.

[0013] 2. Data acquisition: The simulated acceleration signal and vehicle speed signal of the established model are used as data for neural network training.

[0014] 3. Train the neural network: Use the collected data to train the neural network model.

[0015] 4. Signal acquisition: Measure acceleration and speed signals from the vehicle’s cab, frame and wheels.

[0016] 5. Estimation of road roughness: Input the filtered acceleration signal into the trained neural network model and output the root mean square value of the power spectrum density of road roughness.

[0017] 6. Classification: The obtained root mean square value of the power spectrum density of road roughness is used to classify the road surface, thereby obtaining the range of road roughness.

[0018] The accompanying drawings are a specific embodiment of a road roughness estimation method based on a vehicle acceleration signal of the present invention. The embodiment includes Establish a nonlinear full vehicle model with ten degrees of freedom; As attached Figure 1 Nonlinear full vehicle model with ten degrees of freedom is shown.

[0019] The nonlinear full-vehicle model with ten degrees of freedom includes formulas 1-3; among them, formula 1 is the four vertical degrees of freedom of the wheels, formula 2 is the three vertical degrees of freedom of the frame's pitch and roll, and formula 3 is the three vertical degrees of freedom of the cab's pitch and roll, for a total of ten degrees of freedom.

[0020] According to Newton's second law, the vertical vibration equation of the tire assembly is established as shown in formula (1); The differential equations for the vertical vibration, pitch and roll motion of the vehicle frame are shown in formula (2): The control differential equations of the vertical vibration, pitch motion and roll motion of the cab are shown in formula (3): in, to Indicates the mass of the four tire components; and are the mass of the frame and the cab respectively; the geometric parameters ,and is the structure of the vehicle; ,and They correspond to tire force, suspension stiffness force, suspension damping force, cab suspension stiffness force and cab suspension damping force respectively; Indicates road motivation, is the vertical displacement under the suspension spring, is the vertical displacement of the frame, is the vertical displacement of the cab; Table 1 lists the specific values ​​of the vehicle's mass parameters and geometric dimensions.

[0021] Table 1: Mass and geometric parameters symbol unit value #timg# #timg# 2.4 #timg# #timg# 2.6 #timg# 0.6 #timg# 2.0 #timg# 3.2 #timg# 0.7 #timg# 0.8 #timg# 0.5 #timg# 90.0 #timg# 180.0 #timg# #timg# 2350.0 #timg# #timg# 980.0 The vertical velocity power spectrum density of road roughness is only related to the road standard and the reference spatial frequency. It remains constant over the entire frequency range and does not vary with frequency. This characteristic is consistent with the power spectrum definition of white noise. Therefore, the vertical displacement input due to road roughness can be modeled by integrating the white noise.

[0022] Gaussian white noise with unit intensity of 1 is used to establish the time domain model of the road surface, as shown in formula (4): in, is the lower cutoff spatial frequency, ; Represents the vehicle speed; is the reference spatial frequency, ; is the road roughness coefficient; is a Gaussian white noise with unit intensity 1; q(t) is the vertical height of the road surface; the signal of q(t) changing with time is the time domain model of the road surface.

[0023] The mechanical properties of the shock absorber show significant nonlinearity, which is crucial to the performance of the vehicle. In order to accurately determine these parameters, the present invention adopts a bench test method to analyze the relationship between the damping force of the shock absorber and its compression or extension speed; a cubic polynomial is used to fit the experimental data to obtain the damping force Fc As speed v The following expression of the function is shown in formula (5); in, represents the damping force of the shock absorber, determined experimentally; represents the damping coefficient obtained by polynomial fitting; The fitting results of the front suspension shock absorber are shown in Table 2; Table 2: Fitted polynomial damping coefficients Shock absorber location #timg# #timg# #timg# Cab Front Suspension - Compression 0.000031 0.027476 11.389424 Cab front suspension – extended -0.000017 0.011704 6.975855 Cab Rear Suspension - Compression -0.000023 -0.007452 4.482131 Cab front suspension – extended 0.000032 -0.022581 6.182971 Front Suspension - Compression 0.000022 0.021163 7.101091 Front suspension - extended 0.000073 -0.080829 29.347936 The cubic polynomial fitting experimental data and fitting curve are shown in the attached figure. Figure 2 shown.

[0024] In order to describe the nonlinear characteristics of the tire, the quadratic curve fitting method is used to analyze the experimental data obtained under different tire pressure conditions; a quadratic polynomial is selected for fitting, as shown in formula (6) for the relationship between the tire vertical force and displacement; in, represents the spring force; represents the spring stiffness coefficient;x represents the displacement change of the spring; through curve fitting, the tire pressure under different air pressure conditions is derived and The values ​​are shown in Table 3 below.

[0025] Table 3: Tire pressure stiffness coefficient table #timg# #timg# #timg# 480 55877 19717 550 58301 27217 620 61051 35460 690 63932 47447 760 66457 58395 830 69311 66560 Stiffness coefficient and shows a linear relationship with the tire inflation pressure, where a first-order fitting method is used to determine , With tire pressure p The numerical relationship between them; the selected polynomial is shown in formula (7); The coefficients obtained from the quadratic fit As shown in Table 4.

[0026] Table 4: Value Parameters #timg# #timg# #timg# #timg# 0.3857 372.18 0.0138 -4.8361 The correlation between experimental data and fitted values ​​is shown in Figure 3 shown.

[0027] The mechanical properties of leaf springs and coil springs are usually different from those of tires; tires exhibit obvious nonlinear mechanical behavior, while leaf springs and coil springs usually exhibit linear stiffness; this difference is particularly obvious in their mechanical model representation. The stiffness of leaf springs and coil springs can be described using a simple linear formula, as shown in formula (8), in, represents the spring force, k represents the stiffness coefficient of the spring, x Represents the change in displacement of the spring.

[0028] As shown in Table 5, the specific parameters of the leaf spring and coil spring are described in detail.

[0029] Table 5: Suspension and tire stiffness parameters parameter symbol value unit Front suspension stiffness #timg# 315000 #timg# Rear suspension stiffness #timg# 1042000 #timg# Front cab suspension stiffness #timg# 45000 #timg# Rear cab suspension stiffness #timg# 35000 #timg# The equation defining the NARX model is shown in equation (9).

[0030] in, f () represents a specific nonlinear function,y represents the output of the system, u represents the input of the system, is the time delay of the output, is the time delay of the input.

[0031] In NARX neural networks, external input is a key component, and the output is usually directed back to the input layer to form a closed-loop configuration, such as Figure 4 (a). However, during the training phase, an open-loop configuration may be used to facilitate training by leveraging the known target outputs, e.g. Figure 4 (b) is shown. In this architecture, the target output is directly integrated as an input to the network. This approach not only improves prediction accuracy, but also simplifies the network architecture, allowing static neural networks to be used for modeling. Initially, the NARX network is trained in an open-loop mode to enhance its predictive capabilities and stabilize the training process. After training is complete, these pre-trained weights are applied in a closed-loop configuration to improve the performance of the network.

[0032] The G-NARX model is used to train the open-loop NARX model using pseudo output feedback to avoid using the original target as the internal input of the approximate closed-loop neural network; by integrating Gaussian noise into the original target data to generate pseudo output feedback, this method minimizes the deviation between the initial training stage and the subsequent training stage; the output feedback of the G-NARX model is shown in formula (10); represents the output of the system, represents Gaussian noise.

[0033] For neural networks, adding noise to the input data for regularization during training can reduce variance. This regularization acts as a compensation mechanism for the network capacity, enhancing the generalization ability of the model. The improved pseudo-parallel training method proposed here is particularly suitable for handling road excitations caused by random stimuli. In the subsequent training stage, it has been shown that adding 10% noise can significantly improve the performance.

[0034] Design a method to estimate road roughness by analyzing acceleration signals collected from various parts of the vehicle: the cab, frame, and wheels, to achieve a rapid assessment of road conditions. Figure 5The G-NARX network is presented in , where the G-NARX network is trained using data from a rigorously established road surface model and a comprehensive vehicle nonlinear model. The G-NARX network is able to determine the RMS value of the road surface power spectral density from experimental acceleration signal data. Subsequently, the road type classification is achieved according to the ISO 8608 standard from the RMS value of the road roughness power spectral density, thereby determining the road roughness interval.

[0035] The spatial power spectral density of the road profile is crucial for estimating road roughness. In the process of calculating the spatial power spectral density of the road profile, the relationship between the temporal frequency and spatial frequency of the road input during vehicle motion must first be determined, as described in formula (11).

[0036] is the temporal frequency, n is the spatial frequency, and u is the vehicle speed; The spatial frequency power spectral density is described by formula (12).

[0037] in, Represents the road power spectral density band The power contained in represents the road surface spatial power spectrum density; represents the temporal power spectral density.

[0038] The temporal power spectral density is described by formula (13).

[0039] in: It's a signal The power spectral density of . is the number of sampling points of the signal. is the sampling frequency. Indicates signal The discrete Fourier transform of .

[0040] The present invention applies the trapezoidal rule to calculate the integral of the discrete spatial power spectral density, and further calculates the root mean square value of the spatial power spectral density, as shown in formula (14); in, Defined as the root mean square value, it represents a statistical measure of the magnitude of the change; Represents the spatial frequency corresponding to x The number of data points indicates the frequency of the pattern in the defined spatial domain; Indicates spatial frequency The power spectral density of the road disturbance at the location is used to quantify the power at different frequencies of the road disturbance. The root mean square value of the power spectral density of the road disturbance in the current time window can be calculated using the given formula.

[0041] For the same type of roads, the root mean square value of the spatial power spectrum density falls within a specific range. The road type is determined by the root mean square value of the spatial power spectrum density, and then the road roughness interval is determined to achieve the estimation of road roughness. Table 6 lists in detail the root mean square values ​​of the spatial power spectrum density of different types of roads and their corresponding road roughness intervals.

[0042] Table 6: Different road surface grades value The accuracy of the estimation in the spatial power spectral density calculation is closely related to the length of the time window. When using the root mean square value of the road disturbance power spectral density (PSD) to classify road roughness, it is crucial to balance the time window length to coordinate the trade-off between estimation accuracy and system response time.

[0043] Considering different speeds and road surfaces, different surface categories from Class A to Class E are selected. Generally, the worse the road surface, the lower the typical driving speed. Therefore, the present invention introduces the concept of mixed road surfaces, combining the typical driving speed of each surface category, each type has at least two speeds; a mixed road surface table is established, as shown in Table 7. Based on formula (15), the time domain signal of the mixed road surface is Figure 6 (a) is shown; According to formula (15), the root mean square value of the power spectral density of the mixed road surface is Figure 6 (b) is presented.

[0044] Table 7: Mixed road parameters Road surface grade Speed Gq (n0)#timg# B 65 0.000064 A 75 0.000016 D 40 0.001024 C 55 0.000256 B 70 0.000064 E 53 0.004096 D 60 0.001024 A 70 0.000016 C 60 0.000256 E 55 0.004096 The estimation accuracy of the road profile online estimation method directly affects the accuracy of the calculation method. However, the accuracy of these methods is affected by the length of the time window during the calculation process. Considering the mixed road surfaces of different types of road surfaces and vehicle speeds, the time window is set to t The time windows are set to 1 second, 2 seconds, 5 seconds and 10 seconds. The accuracy of road classification under different time windows is shown in the attached figure. Figure 7 shown.

[0045] As the time window length increases, the estimation error rate of the system decreases significantly, indicating that the estimation accuracy has been significantly improved. However, once the time window reaches a certain length, the change in the estimation error rate tends to stabilize, which shows that further extending the time window has limited effect on improving the estimation accuracy. On the contrary, this will increase the response time, thereby worsening the response speed of the system. Therefore, the present invention selects a time window length of 10 seconds and uses this time length as the final time window length for the study.

[0046] The present invention selects wheels, a vehicle frame, a cab and a vehicle speed as input data of a neural network model (NARX, G-NARX), uses the root mean square value of the road power spectrum density as output data for training the neural network, including four input variables and one output variable; thereby achieving road roughness classification.

[0047] Considering the impact of vehicle speed on dynamic response, the vehicle speed range is set to 5 to 20 meters per second, and data is collected every 1 meter per second under different road surface conditions. The sampling frequency of each data set is determined to be 1 / 10 of the vehicle speed, expressed as 1 / 10u, where u represents the vehicle speed. Based on the power spectral density, the acquisition duration of each set of data is set to 10 seconds. To ensure the consistency of input and output data lengths, the root mean square value signal of the speed signal and the power spectral density is expanded to match the length of the acceleration signal. After normalizing the original data, all data sets are merged into a large data set for model training. The training algorithm uses trainbr, the maximum number of iterations is set to 1000, and the number of nodes in the three hidden layers is configured as [15,15,10]. The data set is divided into training set, validation set, and test set, with a ratio of 7:2:1. The training set is used for model training. The validation set is used to monitor the training progress. The hidden layer activation function uses the Tansig function, and the output layer applies the Purelin function.

[0048] The embodiments of the present invention are carried out on a computer equipped with an NVIDIA 2060s GPU and an AMD 5700x processor, the operating system is Windows 10, and the calculation is performed using Matlab software. The mean square logarithmic loss function defined in formulas (15) to (17) is used in the training process. The model evaluation indicators include MS E , RMSE and Rc , which is defined in the corresponding equation.

[0049] Mean squared error (MSE) refers to the average of the squares of the differences between the predicted value and the true value. RMSE (Root Mean Squared Error) is the square root of MSE; Rc (R-squared), also known as the coefficient of determination, is an indicator used to assess the degree of total variation in the model's explanation of variables; its value ranges from 0 to 1, and the closer the value is to 1, the stronger the explanatory power of the model.

[0050] Where n is the sample size, is the ith actual value, is the ith predicted value, is the mean of the actual values, represents the sum of all samples.

[0051] The input of the neural network model is the acceleration signal and vehicle speed obtained during the vehicle's driving process, and the output is the root mean square value of the spatial power spectrum density, which is closely related to the road roughness. The root mean square value of the spatial power spectrum density is an important indicator for determining the range of road roughness. Through this model, we convert the acceleration signal and vehicle speed information into an effective estimate of road roughness.

[0052] In order to evaluate the performance of the closed-loop neural network models NARX and G-NARX in a real-world application using neural networks, ten seconds of vertical acceleration and velocity data must be provided to the trained neural network models. This evaluation includes a comparative analysis comparing the acceleration response signals in the closed-loop neural network with different speeds, specifically 5, 10, 15, and 20 m / s on Class A roads. Table 8 below illustrates the performance details, highlighting the superior accuracy of G-NARX on the three evaluation metrics, with a performance improvement of more than 26% compared to NARX.

[0053] Table 8: Comparative analysis of NARX and G-NARX models at different speeds In addition to considering different vehicle speeds, the performance differences of the neural network on different road surfaces should also be evaluated. The present invention specifically evaluates the performance of the network on four different road surface types A, B, C and D at a speed of 6 m / s. The results are shown in Table 9, indicating that the G-NARX model exhibits higher accuracy, with an improvement of more than 40% compared to the traditional NARX model.

[0054] Table 9: Comparative analysis of NARX and G-NARX models for different pavement grades.

Claims

1. A road roughness estimation method based on vehicle acceleration signal, characterized in that: The following steps are included: Establish a nonlinear full vehicle model with ten degrees of freedom; including formulas 1-3; According to Newton's second law, the vertical vibration equation of the tire assembly is established as shown in the following formula (1); The differential equations for the vertical vibration, pitch and roll motion of the vehicle frame are established as shown in the following formula (2): The control differential equations of the vertical vibration, pitch motion and roll motion of the cab are established as follows: As shown; in, m 1 to m 4 represents the mass of the four tire components; mf and mc are the mass of the frame and the cab respectively; the geometric parameters ai , bi ,and d The structure of the vehicle; FWf , FqV , F , FqV ,and F They correspond to tire force, suspension stiffness force, suspension damping force, cab suspension stiffness force and cab suspension damping force respectively; q Indicates road motivation, z is the vertical displacement under the suspension spring, zf is the vertical displacement of the frame, z is the vertical displacement of the cab; Gaussian white noise with unit intensity of 1 is used to establish the time domain model of the road surface; The bench test method is used to analyze the relationship between the damping force of the shock absorber and its compression or extension speed; Describe the nonlinear characteristics of tires and use quadratic curve fitting method to analyze experimental data obtained under different tire pressure conditions; The equation defining the NARX neural network model is shown in formula (9): in, f () represents a specific nonlinear function, y represents the output of the system, u represents the input of the system, is the time delay of the output, is the time delay of the input; The G-NARX model is used to train the open-loop NARX model using pseudo output feedback. The Gaussian noise is integrated into the original target data to generate pseudo output feedback, minimizing the deviation between the initial training stage and the subsequent training stage. The output feedback of the G-NARX model is shown in the following formula (10): represents the output of the system, represents Gaussian noise; Determine the relationship between the temporal frequency and spatial frequency of the road input during vehicle motion, as shown in formula (11); is the temporal frequency, n is the spatial frequency, and u is the vehicle speed; Calculate the spatial frequency power spectrum density of the road profile, as shown in formula (12); Among them, σ −Δ n Represents the road surface power spectrum density band Δ n The power contained in G ( n ) represents the road surface spatial power spectrum density; G ( f ) represents the temporal power spectral density; The temporal power spectral density is described by the following formula (13); in: It's a signal The power spectral density of is the number of sampling points of the signal, is the sampling frequency, Indicates signal The discrete Fourier transform of The trapezoidal rule is applied to calculate the integral of the discrete spatial power spectral density, and the root mean square value of the spatial power spectral density is further calculated, as shown in formula (14); in, Defined as the root mean square value, it represents a statistical measure of the magnitude of the change; Represents the spatial frequency corresponding to x The number of data points indicates the frequency of the pattern in the defined spatial domain; Indicates spatial frequency The power spectrum density of the road disturbance at the location is used to quantify the power at different frequencies of the road disturbance; the root mean square value of the power spectrum density of the road disturbance in the current time window is calculated using the given formula; For the same type of roads, the root mean square value of the spatial power spectrum density falls within a specific range; the roughness of the road is evaluated by estimating the root mean square value of the spatial power spectrum density of the road disturbance and defining its interval; The wheels, frame, cab and vehicle speed are selected as the input data of the neural network model G-NARX, and the root mean square value of the road power spectrum density is used as the output data for training the neural network, that is, the training set of four input variables and one output variable is used to train the neural network; Considering the influence of vehicle speed on dynamic response, the vehicle speed range is set to 5 to 20 meters per second, and data is collected every 1 meter per second under different road surface conditions; the sampling frequency of each data set is determined to be 1 / 10 of the vehicle speed, expressed as 1 / 10u, where u represents the vehicle speed; Based on the power spectral density, the acquisition duration of each set of data was set to 10 seconds; The original data was normalized, and all data sets were merged into a large data set for model training. The training algorithm used trainbr, the maximum number of iterations was set to 1000, and the number of nodes in the three hidden layers was configured to [15, 15, 10]. The data set was divided into training set, validation set, and test set, with a ratio of 7:2:

1. The training set was used for model training, and the validation set was used to monitor the training progress. The hidden layer activation function uses the Tansig function, and the output layer applies the Purelin function; The mean square logarithmic loss function defined in formulas (15) to (17) is used during training; the model evaluation indicators include MSE , RMSE and Rc , which is defined in the corresponding equation; Where n is the number of samples, is the ith actual value, is the ith predicted value, is the mean of the actual values, represents the sum of all samples; The mean square error (MSE) is the average of the squares of the differences between the predicted value and the true value. RMSE is the square root of MSE; Rc is the coefficient of determination, which is an indicator used to evaluate the total variation of the model's explanatory variables; its value ranges from 0 to 1, and the closer the value is to 1, the stronger the explanatory power of the model; The root mean square value of the road roughness power spectrum density output by the neural network is used to classify the road surface, thereby obtaining the road roughness interval of the current road.

2. The method for estimating road roughness based on vehicle acceleration signal according to claim 1, characterized in that: Using Gaussian white noise with unit intensity of 1, the time domain model of the road surface is established, as shown in the following formula (4): in, is the lower cutoff spatial frequency, ; Represents the vehicle speed; is the reference spatial frequency, ; is the road roughness coefficient; is a Gaussian white noise with unit intensity 1.

3. The method for estimating road roughness based on vehicle acceleration signal according to claim 1, characterized in that: The bench test method is used to analyze the relationship between the damping force of the shock absorber and its compression or extension speed; the damping force is obtained by fitting the experimental data with a cubic polynomial. Fc As speed v The expression of the function is shown in formula (5); in, represents the damping force of the shock absorber, determined experimentally; represents the damping coefficient obtained by polynomial fitting.

4. The method for estimating road roughness based on vehicle acceleration signal according to claim 1, characterized in that: To describe the nonlinear characteristics of the tire, the quadratic curve fitting method is used to analyze the experimental data obtained under different tire pressure conditions; a quadratic polynomial is selected for fitting, as shown in formula (6) for the relationship between the vertical force and displacement of the tire; in, represents the spring force; represents the spring stiffness coefficient; x represents the displacement change of the spring; Use a first-order fitting method to determine and and tire pressure p The numerical relationship between them is shown in formula (7); The coefficients obtained from the quadratic fit ; The stiffness of leaf springs and coil springs is expressed using a linear formula, as shown in formula (8), in, represents the spring force, k represents the stiffness coefficient of the spring, x Represents the change in displacement of the spring.

Images