CN101275900A - Road surface type recognition method based on wheel vibration - Google Patents

Abstract

The invention discloses a road surface type identification method based on wheel vibration. After the wheel vibration model is established, when the vehicle is running, the current wheel vibration signal is collected to obtain the current wheel vibration high-frequency spectrum feature vector, which is different from the typical road wheel The vibration high-frequency spectrum feature vectors are compared to identify the road surface type. The method is simple and easy, and the equipment used is simple and easy to install.

Description

Road surface type recognition method based on wheel vibration

technical field

The invention belongs to the technical field of vehicle driving safety, and in particular relates to a method for identifying road surface types during vehicle driving.

Background technique

The force necessary for the safe and smooth running of the vehicle (such as ABS, ASR, EBD, ESP, etc.) comes from the contact surface between the tire and the road surface, and the force depends on the skid resistance of the road surface. Therefore, the evaluation of road skid resistance is a necessary part of vehicle safety control.

Pavement skid resistance is determined by pavement surface characteristics, which include macrostructure and microstructure (Sha Qinglin, "Early Damage and Prevention of Expressway Asphalt Pavement", People's Communications Press, January 2005). The microstructure of the pavement refers to the roughness of the pavement aggregate surface, that is, the microstructure of the aggregate surface is 0-0.5mm in the horizontal direction and 0-0.2mm in the vertical direction. Anti-skid performance plays a decisive role. The macroscopic structure of the road surface refers to the gap or drainage capacity between the pavement aggregates. The structure formed between the pavement aggregates, that is, the horizontal direction of the road surface is 0.5-50 mm, and the vertical direction is 0.2-10 mm. Anti-slip ability.

At present, common high-speed pavements include cement concrete pavement and SMA asphalt pavement. These two pavement surfaces have different particle sizes of aggregates and different macrostructures of pavement surface characteristics. The surface aggregates of the cement concrete pavement are medium sand, medium coarse sand and finer coarse sand with a fineness modulus of 2.3 to 3.2. The characteristic macroscopic structure of the cement concrete pavement surface is shown in Figure 1. SMA asphalt pavement skeleton is coarse aggregate larger than 4.75mm (accounting for about 70% of the mixture), mastic filler composed of asphalt, mineral powder, fine aggregate and fiber (accounting for about 30% of the mixture), SMA The macroscopic structure of asphalt pavement surface features is shown in Fig. 2.

To sum up, different types of pavements have different surface features and macrostructures. The macrostructure affects the anti-sliding force of the car at high speed, and is the main factor affecting the safety of the car. Therefore, the skid resistance of the pavement can be evaluated by measuring the macroscopic structure of the pavement surface characteristics, which is essentially pavement type identification.

At present, the common methods for measuring the macroscopic structure of pavement surface characteristics are as follows:

1. Surface profile Tire sensors recognize the macroscopic structure of road surface features (Han Jianbao, Zhang Lubin, Li Bangguo, "Real-time Sensing and Recognition System for Tire-Pavement Adhesion Coefficient", Vehicle and Power Technology, Issue 2, 2005)

The surface profile tire sensor, see Figure 3, its core component is a micro-shaped flat piezoelectric crystal, through the potential generated by the piezoelectric crystal, it senses the deformation of the tire rubber (essentially obtains the macroscopic structure of the road surface characteristics), and accordingly identifies the wheel drive. Force, braking force, lateral force, wheel load, tire pressure, size and position of tire footprint, adhesion coefficient between tire and ground, etc. The tire sensor itself does not have a power supply, it is excited by external wireless electromagnetic waves, and the information is transmitted to the vehicle electronic receiving equipment through wireless electromagnetic waves.

This method has the following disadvantages:

(1) It is difficult to accurately locate the sensor in the tire rubber;

(2) The wireless power supply and wireless information transmission of sensors are relatively complicated;

(3) Because the sensor is very small, the information obtained is relatively small, which cannot reflect the overall situation of the road surface.

2. Acoustic scattering measurement of pavement surface characteristics (Pieter L.Swart, BeatrysM.Lacquet, "An Acoustic Sensor System for Determination of Macroscopic Surface Roughness", IEEE Transactions on Instrumentation and Measurement, Vol.45, No.5, October 1996)

Schematic diagram of the macroscopic structure of the pavement surface characteristics measured by acoustic scattering, as shown in Figure 4. The principle of measuring the pavement surface characteristics is that sound waves are emitted to the road surface. The rougher the pavement surface characteristics, the more sound waves with smaller wavelengths are scattered. The transmitting device continuously emits sound waves to the road surface with a specific sound wave (frequency and power) and emission angle, and the sound wave receiving device receives the reflected sound waves from the road surface in a specific range. By analyzing the received sound waves, it can be known that sound waves of a certain frequency are reflected and sound waves of a certain frequency are scattered. According to the frequency and degree of reflected and scattered sound waves, the macrostructure of pavement surface characteristics is estimated.

This method has the following disadvantages:

(1) The received sound waves are easily interfered by external sound waves;

(2) The installation of the equipment is inconvenient, and it is easy to collide when it is close to the road surface, and the accuracy is reduced when it is far from the road surface;

(3) The transmitting device and the receiving device are easily polluted by dust, rainwater, etc.

Contents of the invention

The object of the present invention is to provide a road surface type identification method based on wheel vibration, so as to overcome the defects in the above method.

The technical scheme of the road surface type recognition method based on wheel vibration in the present invention is as follows: after the wheel vibration model is established, when the vehicle is running, the vibration signal of the current wheel is collected to obtain the current wheel vibration high-frequency spectrum feature vector, which is different from the typical road wheel The vibration high-frequency spectrum feature vectors are compared to identify the road surface type.

Firstly, the wheel vibration model is established to prove that the high-frequency frequency of wheel vibration is equal to the excitation frequency of the macroscopic characteristic of the road surface; secondly, multiple groups of typical road surface (such as cement concrete pavement, SMA asphalt pavement, etc.) wheel vibration signals are collected, and wavelet analysis and fast Fu Lie transform to obtain multiple groups of typical road wheel vibration high-frequency spectrum feature vectors; use the typical road wheel vibration high-frequency spectrum feature vectors to construct a road RBF neural network classifier; finally, collect the road wheel vibration signals to be identified to obtain their spectrum feature vectors , input the road surface RBF neural network classifier to obtain the road surface type recognition result.

The method is simple and easy, and the equipment used is simple and easy to install.

The main functions of the present invention in vehicle safety control (such as ABS, ASR, EBD, ESP, etc.) are as follows:

(1) Determine the best slip ratio

Different types of roads have different optimal slip ratios, so the optimal slip ratio of the vehicle can be determined according to the type of road surface, as shown in Figure 5.

(2) Determine the maximum braking force and driving force

After determining the type of road surface and the optimum slip ratio, the maximum adhesion coefficient of the road surface can be determined, as shown in Figure 5. The maximum adhesion coefficient multiplied by the vehicle weight is the maximum friction that the road surface can provide. Determine the maximum braking force and driving force of the vehicle according to the maximum friction force.

Description of drawings

Figure 1. The characteristic macroscopic structure of cement concrete pavement surface.

Fig. 2 SMA asphalt pavement surface characteristic macrostructure.

Figure 3 Surface profile sensor compared to a match.

Fig. 4 Measure the road surface acoustic scattering to identify the road surface.

Figure 5 μ-λ curves of different road surfaces.

Fig. 6. Road surface type identification process based on wheel vibration.

Figure 7 Single degree of freedom wheel vibration model.

Figure 8 Vibration signals of wheels on cement concrete pavement and asphalt pavement.

Figure 9 Improved single subband reconstruction algorithm.

Figure 10. Eigenvectors of vibration spectrum of wheels on cement concrete pavement and asphalt pavement.

Figure 11 RBF neural network structure.

Figure 12 Principle of wheel vibration neural network road surface type recognition equipment.

Detailed ways

The road surface type identification process based on wheel vibration in the present invention is shown in FIG. 6 . It consists of the following parts: establish a wheel vibration model, collect wheel vibration signals, obtain multiple groups of typical road wheel vibration high-frequency spectrum feature vectors, construct a neural network classifier for road type recognition, and develop a wheel vibration neural network road type recognition device.

1.1 Establish wheel vibration model

The research object of this method is wheel vibration, and the vehicle vibration system is appropriately simplified according to the research needs (Jin Xiaoxiong, Zhang Lijun, Jiang Hao, "Automobile Vibration Analysis", Tongji University Press, first edition in May 2002). When the vehicle mass distribution coefficient ε=1, the wheel vibrations are not related to each other, and the wheel vibration model is simplified to a single-degree-of-freedom wheel vibration model composed of wheels m1, as shown in Figure 7. Wheel vibration includes low-frequency wheel vibration caused by road surface roughness and high-frequency wheel vibration caused by road surface characteristics macroscopic structure. High-frequency vibration is the research object of the present invention, and low-frequency vibration can be identified and eliminated.

The following proves that the high-frequency spectrum of the wheel vibration is the same as the excitation spectrum of the macrostructure of the pavement surface features:

In order to simplify the analysis, the excitation of the macrostructure of pavement surface features is simplified as

x _s = a sin ωt (1)

The differential equation of the wheel vibration model is

${\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; k</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; k</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

The solution of this vibrational differential equation is

${\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; a</span>}\sqrt{\frac{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; +</span>{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; \&xi;\&lambda;</span>}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; \&xi;\&lambda;</span>}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}}\mathrm{<span\; class="notranslate">\; sin</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; \&omega;t</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; \&psi;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 3</span>}<span\; class="notranslate">\; )</span>$

Among them, x _s is the characteristic excitation of the road surface, a is the excitation amplitude of the road surface characteristic, ω is the excitation frequency of the macroscopic structure of the road surface characteristic, m ₁ is the tire mass, x ₁ is the wheel vibration, c ₁ is the tire damping, k ₁ is the tire Stiffness, ξ is the damping ratio, λ is the frequency ratio, and ψ is the lag angle.

From the solution of differential equation (3), it is proved that the high frequency frequency spectrum of wheel vibration is the same as the excitation frequency spectrum of pavement surface characteristic macrostructure.

1.2 Acquisition of wheel vibration signals

An acceleration sensor is installed vertically on the axle close to the wheel to measure the vibration in the vertical direction of the wheel. The vehicle is driving on a typical high-speed road, and multiple sets of road wheel vibration signals are collected. One set of wheel vibration signals on cement concrete road and SMA asphalt road is shown in Figure 8.

1.3 Obtain multiple groups of typical road wheel vibration high-frequency spectrum feature vectors

(1.3.1) Wavelet analysis of wheel vibration signals

Select the appropriate wavelet function and scale, and use the improved single subband reconstruction algorithm (Yang Jianguo, "Wavelet Analysis and Its Engineering Application", Machinery Industry Press, first edition in July 2005) to obtain the wheel vibration high frequency subband signal. The improved single subband reconstruction algorithm is shown in Figure 9.

(1.3.2) Obtain high frequency spectrum of wheel vibration

Perform fast Fourier transform on the wheel vibration high-frequency sub-band signal to obtain the high-frequency frequency spectrum of wheel vibration.

(1.3.3) Obtaining the feature vector of high-frequency spectrum of wheel vibration

In order to reduce interference and reduce the amount of data, the high-frequency spectrum of wheel vibration is equally divided by a fixed frequency band, and the average value of the spectrum amplitude is used as the spectrum amplitude of the frequency band, and the spectrum is normalized to obtain the feature vector of the high-frequency spectrum of wheel vibration.

$\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}& {\mathrm{<span\; class="notranslate">\; <figure-callout\; id="0"\; label="f"\; filenames="A20081004761200132.png,A20081004761200141.png"\; state="\{\{state\}\}">f</figure-callout></span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; <</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}\\ {\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}& {\mathrm{<span\; class="notranslate">\; <figure-callout\; id="1"\; label="f"\; filenames="A20081004761200132.png,A20081004761200133.png"\; state="\{\{state\}\}">f</figure-callout></span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; <</span>{\mathrm{<span\; class="notranslate">\; <figure-callout\; id="2"\; label="f"\; filenames="A20081004761200142.png"\; state="\{\{state\}\}">f</figure-callout></span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}\\ & \mathrm{<span\; class="notranslate">\; \&Lambda;</span>}\\ {\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}& {\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; <</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}\end{array}\right.<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; )</span>$

Collect multiple sets of wheel vibration signals on various types of road surfaces on typical road surfaces, and obtain multiple sets of wheel vibration spectrum feature vectors on various types of road surfaces. Among them, a group of spectral feature vectors of cement concrete pavement and SMA asphalt pavement are shown in Figure 10 (the upper figure is the spectral feature vector of cement concrete pavement, and the lower figure is the spectral feature vector of SMA asphalt pavement).

1.4 Construction of neural network classifier for pavement type recognition

(1.4.1) Establishment of RBF neural network for pavement type recognition

The RBF neural network consists of two layers, including a hidden layer and an output layer. The hidden layer has several neurons. The most commonly used node function is a Gaussian function. The output layer has several neurons. The node function is a simple linear function. Suppose the network input X is an n-dimensional vector, and the output Y is an L-dimensional vector, and its structure is shown in Figure 11.

(1.4.2) Training road surface type recognition RBF neural network

The learning process of the RBF network is divided into two stages, the first stage is learning without a teacher, and the second stage is learning with a teacher.

No teacher learning phase

Cluster the input of all samples (Li Guoyong, "Intelligent Control and Its Realization with MATLAB", Electronic Industry Press, May 2005), and obtain the center vector c _i of each hidden layer node. The following is the K-means clustering algorithm, and the algorithm steps are as follows:

Given the initial center vector c _i (0) of each hidden layer node and the ε that determines to stop the calculation;

(i) Calculate the distance (Euclidean distance) and find the minimum distance node;

$\left\{\begin{array}{c}{\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; q</span>}\\ {\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; min</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; min</span>}{\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>\end{array}\right.<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; )</span>$

Where i=1, 2, 3...q, q is the number of hidden layer nodes, c _i is the center vector of the Gaussian kernel function of the i-th hidden layer node.

(ii) Adjustment center

$\left\{\begin{array}{c}{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; r</span>}\\ {\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; \&beta;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span>\end{array}\right.<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 6</span>}<span\; class="notranslate">\; )</span>$

In the formula, β(k) is the learning rate, β(k)=β(k-1)/1+int(k/q) ^1/2 , and the learning rate decreases gradually.

Determining the quality of clustering

Repeat steps i and ii above for all samples k until the following conditions are met, then the clustering ends.

${\mathrm{<span\; class="notranslate">\; J</span>}}_{\mathrm{<span\; class="notranslate">\; e</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; q</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; c</span>}}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; \&epsiv;</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 7</span>}<span\; class="notranslate">\; )</span>$

There are teachers to learn

When c _i is determined, train the weights from the hidden layer to the output layer. It is a linear equation system, then finding the weight value becomes a linear optimization problem, and the global minimum point can definitely be obtained. The learning algorithm for finding the weight W _i of the hidden layer and the output layer is

${\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; the\; ki</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; the\; ki</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; \&eta;</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; t</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; /</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; u</span>}}}^{\mathrm{<span\; class="notranslate">\; T</span>}}\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; u</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 8</span>}<span\; class="notranslate">\; )</span>$

In the formula, u _i (k+1) is a Gaussian function, and η is the learning rate.

The RBF neural network is trained with multiple groups of wheel vibration spectrum feature vectors on various road surfaces, and finally a neural network classifier for road surface type recognition is obtained.

1.5 Develop wheel vibration neural network road surface type recognition equipment

The principle of the wheel vibration neural network road surface type recognition equipment is shown in Figure 12. The sensor is used to collect the wheel vibration signal of the road surface to be recognized. The data acquisition system completes the conversion of the wheel vibration signal from analog to digital. The DSP completes 1.3 to obtain the high frequency spectrum characteristics of the wheel vibration Vector, load the road surface RBF neural network classification constructed in 1.4 and communicate with the vehicle safety control ECU data.

In the experiment of the present invention, the identification objects are cement concrete pavement and SMA asphalt pavement, and the vehicle speed is 40Km/h. In the implementation process, firstly, the high-speed dynamic acquisition instrument is used to collect the wheel vibration signal, and the high-frequency frequency spectrum feature vector of the wheel vibration is obtained by Matlab simulation, and the road surface type recognition neural network classifier is constructed; secondly, the road surface type recognition equipment is developed, including the wheel with the single-chip microcomputer as the core. Vibration data acquisition system, signal analysis with DSP as the core, RBF neural network road surface type identification and communication system.

2.1 Acquisition of wheel vibration signals

Shanghai GM Buick Sail SL1.6 is selected as the car; wavebook516E and its expansion module wbk18 are selected as the high-speed dynamic acquisition instrument, the sampling frequency is set to 16KHz, and the anti-aliasing filter frequency is set to 5KHz; the acceleration sensor is selected from CA-YD-181, and its The effective vibration frequency is 1-10KHZ, and it is installed vertically on the rear axle at a distance of 30mm from the wheel.

Collect 50 sets of wheel vibration signals on cement concrete pavement and SMA asphalt pavement. To ensure real-time and accuracy, the number of collection points is 4096. One set of wheel vibration signals on cement concrete pavement and SMA asphalt pavement is shown in Figure 3.

2.2 Obtain multiple sets of typical road wheel vibration high-frequency spectrum feature vectors

(2.2.1) Wavelet analysis of wheel vibration signals

The improved wavelet single subband reconstruction algorithm is used to decompose and reconstruct the wheel vibration signal. The wavelet function selects db10, carries out 2-scale single sub-band reconstruction, and obtains the 2K-4KHz high-frequency vibration sub-band signal of the wheel. The filter directly uses the biorthogonal wavelet filter bank function provided by Matlab:

[H_D, G_D, h_R, g_R] = wfilters('db10')

(2.2.2) Obtain high-frequency spectrum of wheel vibration

Fast Fourier transform is performed on the obtained high-frequency sub-band signal to obtain the high-frequency sub-band spectrum. The fast Fourier transform directly calls the function provided by Matlab:

X2=fft(d2, 4096)

Among them, d2 is the 2K～4KHz high-frequency vibration sub-band signal of the wheel, and X2 is the frequency spectrum in the 2K～4KHz frequency band of the vibration signal.

(2.2.3) Obtaining the feature vector of high-frequency spectrum of wheel vibration

Each vibration spectrum is equally divided into 20 frequency bands, and the average amplitude of each frequency band is used as the spectrum amplitude of the frequency band, and then the spectrum is normalized to obtain the feature vector of the wheel vibration spectrum. Among them, a set of spectral eigenvectors of cement concrete pavement and SMA asphalt pavement are shown in Figure 6.

50 groups of cement concrete pavement and SMA asphalt pavement wheel vibration signals are analyzed through steps (2.2.1), (2.2.2) and (2.2.3) to obtain 50 groups of spectral feature vectors, each with 20 dimensions. Among them, 40 groups of vectors are used to train the neural network classifier for road surface type recognition, and the other 10 groups are used to test the neural network classifier.

2.3 Construction of road surface neural network classifier

(2.3.1) Establish road surface type recognition RBF neural network, the network input vector is 20 dimensions, the hidden layer neuron number is 20, and the network output vector is 1 dimension;

(2.3.2) To train the RBF neural network for road type recognition, first input 40 high-frequency frequency spectrum feature vectors of cement concrete road wheel vibration into the RBF neural network, so that the network output is 0; The vector is input into the RBF neural network, and the network output is 1.0. By training the RBF neural network, the road surface neural network classifier is obtained.

2.4 Neural Network Pavement Type Identification Verification

The other 10 sets of road surface spectrum feature vectors are input into the road surface neural network classifier, and the network output results are shown in Table 1. The output of the network classifier close to 0 is the cement concrete pavement, and the output of the network classifier close to 1.0 is the SMA asphalt pavement. From the experimental results of pavement type recognition, it can be seen that for cement concrete pavement and SMA asphalt pavement, the recognition accuracy can reach 100%.

Table 1. Road surface neural network classifier output

Concrete Pavement SMA Asphalt Pavement

1 -0.15743 6 -0.16787 1 0.84977 6 0.82057

2 -0.16787 7 -0.16053 2 0.97070 7 0.79684

3 0.028965 8 -0.15645 3 0.75505 8 0.59715

4 -0.17188 9 -0.19633 4 0.79872 9 0.87721

5 -0.09944 10 -0.16654 5 0.79484 10 0.98537

2.5 Develop wheel vibration neural network road surface type recognition equipment

The principle of wheel vibration neural network road surface type recognition equipment is shown in Figure 12. The equipment includes:

(2.5.1) The wheel vibration sensor is used to collect the vibration signal of the road wheel to be identified, and its model is CA-YD-181.

(2.5.2) The wheel vibration data acquisition system with the single-chip microcomputer as the core completes the A/D conversion of the wheel vibration signal. The single-chip microcomputer uses AT89C2051, the A/D converter uses AD7574, and the data memory uses dual-port RAMIDT7134.

(2.5.3) Signal analysis, neural network road surface type recognition system and communication system with DSP as the core, internally includes data analysis algorithm and road surface neural network classifier constructed in 2.3, completes signal spectrum analysis, road surface classification, and finally passes the data The bus is transmitted to the vehicle safety electronic control unit. DSP chooses TMS320LF2407.

Claims (7)

1, a kind of method for recognizing road surface types based on unsteadiness of wheels, it is after setting up the unsteadiness of wheels model, when vehicle ', gather vibration signal when front vehicle wheel, obtain current unsteadiness of wheels high frequency spectrum proper vector, compare with typical road surface unsteadiness of wheels high frequency spectrum proper vector, thus the identification road surface types.

2, according to claim 1 based on the method for recognizing road surface types of unsteadiness of wheels, it is characterized in that the described method of obtaining unsteadiness of wheels high frequency spectrum proper vector is: adopt the reconstruct of list band to improve algorithm and decompose and reconstruct unsteadiness of wheels signal, obtain unsteadiness of wheels high-frequency sub-band signal; Unsteadiness of wheels high-frequency sub-band signal is carried out fast fourier transform, obtain the unsteadiness of wheels high frequency spectrum; With branch unsteadiness of wheels high frequency spectrums such as fixed frequency bands, and with spectral magnitude mean value as this band spectrum amplitude, frequency spectrum is carried out normalization, obtain unsteadiness of wheels high frequency spectrum proper vector.

3, according to claim 1 based on the method for recognizing road surface types of unsteadiness of wheels, it is characterized in that current unsteadiness of wheels high frequency spectrum proper vector and each quasi-representative road surface unsteadiness of wheels high frequency spectrum proper vector are compared, concrete grammar is: structure road surface types identification neural network classifier, utilize the neural network classifier road pavement to discern.

4, as described in the claim 3 based on the method for recognizing road surface types of unsteadiness of wheels, it is characterized in that the method for constructing road surface types identification neural network classifier is: set up road surface types identification RBF neural network, the RBF neural network is made up of two-layer, comprises hidden layer and output layer; Utilize the unsteadiness of wheels spectrum signature vector training RBF neural network of unsteadiness of wheels model, realize road surface types identification thereby finally obtain road surface types identification neural network classifier.

5, as described in the claim 4 based on the method for recognizing road surface types of unsteadiness of wheels, it is characterized in that described training RBF neural network comprises that the phase one is the teacherless learning, subordinate phase is that teacher learning is arranged; Described teacherless learning carries out cluster to the input of all samples, tries to achieve the center vector c of each hidden node _iDescribed teacher learning is to work as c _iAfter determining, train by hidden layer to the weight w between the output layer _i

6, as described in the claim 2 based on the method for recognizing road surface types of unsteadiness of wheels, it is characterized in that and will obtain the method dsp program of unsteadiness of wheels high frequency spectrum proper vector, and download in the dsp chip.

7, as described in the claim 3 based on the method for recognizing road surface types of unsteadiness of wheels, it is characterized in that the RBF neural network classifier dsp program that will obtain, and in the download dsp chip.

Images