CN104951661A - PCM (possibilistic C-means) clustering algorithm based online automotive suspension performance monitoring method - Google Patents

Abstract

The invention relates to a PCM (possibilistic C-means) clustering algorithm based online automotive suspension performance monitoring method and aims to solve the problems of excessive calculation amount and incapacity of online performance monitoring of a conventional method. The method comprises steps as follows: step one, when an automotive suspension is in a healthy state, N accelerometer samples are collected to obtain normal data; step two, a clustering center and a weight of PCM are obtained through calculation; step three, accelerometer values are collected every once in a while during running of the automotive suspension, and the membership degrees mu ki of the accelerometer samples relative to the normal data are calculated according to the clustering center and the weight of the PCM; step four, the number A of the accelerometer samples with the membership degrees lower than a threshold value Thr in a next period of time T is calculated; step five, the collected fault data and normal data are classified with an FDA (force directed algorithm) so as to obtain a classification feature vector wk. The PCM clustering algorithm based online automotive suspension performance monitoring method is applied to the field of automobile performance monitoring.

Description

Based on the online automotive suspension method for monitoring performance of possibility C means clustering algorithm

Technical field

The present invention relates to the online automotive suspension method for monitoring performance based on possibility C means clustering algorithm (PCM).

Background technology

Along with the develop rapidly of automobile industry, the comfort level of vehicle, security and reliability receive to be paid close attention to widely, and wherein automobile suspension system is the pith affecting automotive performance.The automobile suspension system of health status not only can ensure that doughnut firmly grasps ground closely, and advantageously in brake, and in driving, car body vibration amplitude is very little, improves comfort level.If automobile suspension system breaks down, the security comfortableness of automobile will be caused greatly to reduce.Therefore, the on-line performance monitoring of automobile suspension system is very important.At present, automotive suspension performance monitoring system is also immature, and the monitoring means based on model are limited to accurate model to be difficult to obtain, and is also in laboratory stage; Some are based on the policing algorithm of data, are limited to calculated amount comparatively large, are not easy to realize on-line performance monitoring.

Summary of the invention

The present invention will solve the problem that existing method calculated amount can not realize greatly on-line performance monitoring, and provides the online automotive suspension method for monitoring performance based on possibility C means clustering algorithm.

Based on the online automotive suspension method for monitoring performance of possibility C means clustering algorithm, it realizes according to the following steps:

Step one: gather N number of accelerometer sample and obtain normal data under automotive suspension is in health status; Wherein said each accelerometer sample is four dimensional vectors, and an accelerometer sample packages is containing four accelerometer sampled values, and each accelerometer sampled value is called an element of accelerometer sample, and each element represents an acceleration evaluation;

Step 2: the cluster centre calculating N number of normal data with FCM algorithm, using the cluster centre that calculates as the initial value of PCM, calculates the cluster centre υ of PCM _kwith weights η _i;

Step 3: in automotive suspension runs, ts gathers primary acceleration evaluation at set intervals, according to the cluster centre of PCM and this accelerometer sample of weight computing degree of membership μ relative to normal data _kiif degree of membership, lower than threshold value Thr (0.4<Thr<0.5), proceeds to step 4; If degree of membership is higher than threshold value, then show normal condition, do not need to take measures; Wherein, described in d _kirepresent sample x _iwith cluster centre υ _kbetween distance;

Step 4: add up in T of lower a period of time (T>>ts), degree of membership is lower than the accelerometer number of samples A of threshold value Thr, if accelerometer number of samples A reaches 80% of total sample number NN in T time, then automobile suspension system breaks down and collects fault data;

Step 5: by the fault data of collection and normal data FDA algorithm classification, obtains characteristic of division vector w _k; By proper vector w _kobtain the contribution of each acceleration evaluation to classification, contribute spring corresponding to maximum accelerometer to break down.

Invention effect: existing method for monitoring performance adopts the method being timed to the maintenance of vehicle maintenance point mostly, can not realize on-line performance monitoring.First this method achieves on-line performance monitoring; Secondly, this method only needs the value of four accelerometers, and in practical implementations, do not need the hardware that extra increase is a large amount of, realizability is strong; Finally, PCM off-line learning obtains cluster centre, during on-line operation, only need the simple distance calculating make new advances sampled point and cluster centre, and then by formulae discovery degree of membership, on-line calculation is very low, and realizability is high

The present invention proposes the online automotive suspension performance monitoring system based on possibility C means clustering algorithm.This system only utilizes the output valve of four accelerometers in automotive suspension, and whether the spring that just can realize in four angles of corresponding four tires in on-line monitoring automotive suspension breaks down.

Accompanying drawing explanation

Fig. 1 is the structural drawing of automotive suspension.

Embodiment

Embodiment one: the online automotive suspension method for monitoring performance based on possibility C means clustering algorithm of present embodiment, it realizes according to the following steps:

Step one: gather N number of accelerometer sample and obtain normal data under automotive suspension is in health status; Wherein said each accelerometer sample is four dimensional vectors, and an accelerometer sample packages is containing four accelerometer sampled values, and each accelerometer sampled value is called an element of accelerometer sample, and each element represents an acceleration evaluation;

Step 2: the cluster centre calculating N number of normal data with FCM algorithm, using the cluster centre that calculates as the initial value of PCM, calculates the cluster centre υ of PCM _kwith weights η _i; (the first two step is off-line learning process, after be on-line monitoring process)

Step 3: in automotive suspension runs, ts gathers primary acceleration evaluation at set intervals, i.e. an accelerometer sample, according to the cluster centre of PCM and this accelerometer sample of weight computing degree of membership μ relative to normal data _kiif degree of membership, lower than threshold value Thr (0.4<Thr<0.5), proceeds to step 4, illustrate that this accelerometer sample may not belong to normal data; If degree of membership is higher than threshold value, then show normal condition, do not need to take measures; Wherein, described in d _kirepresent sample x _iwith cluster centre υ _kbetween distance;

Step 4: add up in T of lower a period of time (T>>ts), degree of membership is lower than the accelerometer number of samples A of threshold value Thr, if accelerometer number of samples A reaches 80% of total sample number NN in T time, then automobile suspension system breaks down and collects fault data;

Step 5: by the fault data of collection and normal data FDA algorithm classification, obtains characteristic of division vector w _k; By proper vector w _kobtain the contribution of each acceleration evaluation to classification, contribute spring corresponding to maximum accelerometer to break down.

Embodiment two: present embodiment and embodiment one unlike: calculate with FCM in described step 2

The cluster centre that method calculates N number of normal data is specially:

FCM is insensitive to starting condition, X={x _i| i=1,2 ..., N} is the data set of N number of accelerometer sample composition, and one of them sample is expressed as x _i=[x _i1x _in], x _ibe a sampled point, have b data in a sampled point, x _inbe b data;

The cluster that FCM has come N number of sample by asking for the solution making objective function minimum, objective function is as follows,

$J=\underset{k=1}{\overset{c}{\&Sigma;}}\underset{i=1}{\overset{N}{\&Sigma;}}{\left({\mathrm{\&mu;}}_{ki}\right)}^m{d}_{ki}^2$

Wherein, μ _ki∈ [0,1], m represents Fuzzy Exponential, and c represents the number of cluster centre; K represents a kth cluster centre, and i represents i-th sample, μ _kirepresent the degree of membership of i-th sample relative to a kth cluster centre;

The solution of objective function is as follows:

${\mathrm{\&mu;}}_{ki}=\frac{1}{\underset{j=1}{\overset{c}{\&Sigma;}}{(d_{ki}/d_{ji})}^{2/(m-1)}}$

υ _krepresent the cluster centre of kth class, d _kirepresent sample x _iwith cluster centre υ _kbetween distance, j represents a jth cluster centre, μ _kirepresent the degree of membership of i-th sample for a kth cluster centre; d _jirepresent the distance of i-th sample for a jth cluster centre.

Other step and parameter identical with embodiment one.

Embodiment three: present embodiment and embodiment one or two unlike: described step 2 using the cluster centre that calculates as the initial value of PCM, calculate the cluster centre υ of PCM _kwith weights η _ibe specially:

PCM is comparatively responsive for starting condition, utilize the initial value of result as PCM of FCM preliminary clusters, PCM is when classifying to accelerometer sample point, sample point to the degree of membership of arbitrary cluster classification between 0-1, be not 1 to the degree of membership sum of all categories, therefore, one not belong to any known class data sample all very little to all kinds of degrees of membership;

PCM realizes cluster by the result asking for objective function minimum, and objective function is as follows,

$J=\underset{i=1}{\overset{N}{\&Sigma;}}\underset{k=1}{\overset{c}{\&Sigma;}}{\left({\mathrm{\&mu;}}_{ki}\right)}^m{d}_{ki}^2+\underset{k=1}{\overset{c}{\&Sigma;}}{\mathrm{\&eta;}}_i\underset{i=1}{\overset{N}{\&Sigma;}}{(1-{\mathrm{\&mu;}}_{ki})}^m$

Wherein, μ _ki∈ [0,1], η _ibe self-defining constant, the solution of objective function is as follows,

${\mathrm{\&eta;}}_i=\frac{\underset{i=1}{\overset{N}{\&Sigma;}}{\mathrm{\&mu;}}_{ki}^m{d}_{ki}^2}{\underset{i=1}{\overset{N}{\&Sigma;}}{\mathrm{\&mu;}}_{ki}^m}$

${\mathrm{\&mu;}}_{ki}=\frac{1}{1+{(d_{ki}^2/{\mathrm{\&eta;}}_i)}^{1/(m-1)}}$

${\mathrm{\&upsi;}}_i=\frac{\underset{i=1}{\overset{N}{\&Sigma;}}{\mathrm{\&mu;}}_{ki}^m{x}_i}{\underset{i=1}{\overset{N}{\&Sigma;}}{\mathrm{\&mu;}}_{ki}^m}.$

Other step and parameter identical with embodiment one or two.

Embodiment four: one of present embodiment and embodiment one to three unlike: described step 5 FDA algorithm is specially:

Define following matrix: overall scatter matrix S _t, scatter matrix within class S _wwith inter _ class relationship matrix S _b, overall scatter matrix S _tbe defined as follows:

$S_t=\underset{i=1}{\overset{n}{\&Sigma;}}(x_i-\stackrel{\&OverBar;}{x}){(x_i-\stackrel{\&OverBar;}{x})}^T$

Wherein, for the average value vector of total sample, the scatter matrix within class S of jth class sample _jbe defined as follows:

$S_j=\underset{x_i\&Element;X_j}{\&Sigma;}(x_i-\stackrel{\&OverBar;}{x_j}){(x_i-\stackrel{\&OverBar;}{x_j})}^T$

Wherein X _jit is the vector x belonging to jth class sample _iset, for the mean value of jth class sample;

Then scatter matrix within class S _wfor:

$S_w=\underset{j=1}{\overset{p}{\&Sigma;}}{S}_j$

Inter _ class relationship matrix S _bbe defined as follows:

$S_b=\underset{j=1}{\overset{p}{\&Sigma;}}{n}_j(\stackrel{\&OverBar;}{x_j}-\stackrel{\&OverBar;}{x}){(\stackrel{\&OverBar;}{x_j}-\stackrel{\&OverBar;}{x})}^T$

FDA objective function is as follows:

$\underset{v\&NotEqual;0}{max}\frac{v^T{S}_{b}v}{v^T{S}_{w}v}$

The calculating of FDA vector can be equivalent to the proper vector w asking for following generalized eigenvalue problem _k;

S _bw _k＝λ _kS _ww _k；

N is feature samples sum, and q is the dimension of a sample; P is sample class number, n _jthe number of samples of jth class observation sample, x _ibe i-th sample, by all training samples (training sample is used to the models such as estimation FDA, PCM, FCM) stored in matrix X ∈ R ^{n × q}in, easily know that the transposition of X i-th row is exactly vector x _i, λ _krepresentation eigenvalue, v ^trepresent the transposition of v, v represents the solution of FDA objective function.

Other step and parameter identical with one of embodiment one to three.

Claims (4)

1., based on the online automotive suspension method for monitoring performance of possibility C means clustering algorithm, it is characterized in that it realizes according to the following steps:

Step one: gather N number of accelerometer sample and obtain normal data under automotive suspension is in health status; Wherein said each accelerometer sample is four dimensional vectors, and an accelerometer sample packages is containing four accelerometer sampled values, and each accelerometer sampled value is called an element of accelerometer sample, and each element represents an acceleration evaluation;

Step 2: the cluster centre calculating N number of normal data with FCM algorithm, using the cluster centre that calculates as the initial value of PCM, calculates the cluster centre υ of PCM _kwith weights η _i;

Step 3: in automotive suspension runs, ts gathers primary acceleration evaluation at set intervals, according to the cluster centre of PCM and this accelerometer sample of weight computing degree of membership μ relative to normal data _kiif degree of membership, lower than threshold value Thr and 0.4<Thr<0.5, proceeds to step 4; If degree of membership is higher than threshold value, then show normal condition, do not need to take measures; Wherein, described in d _kirepresent sample x _iwith cluster centre υ _kbetween distance;

Step 4: add up in T and T>>ts of lower a period of time, degree of membership is lower than the accelerometer number of samples A of threshold value Thr, if accelerometer number of samples A reaches 80% of total sample number NN in T time, then automobile suspension system breaks down and collects fault data;

Step 5: by the fault data of collection and normal data FDA algorithm classification, obtains characteristic of division vector w _k; By proper vector w _kobtain the contribution of each acceleration evaluation to classification, contribute spring corresponding to maximum accelerometer to break down.

2. the online automotive suspension method for monitoring performance based on possibility C means clustering algorithm according to claim 1, is characterized in that being specially with the cluster centre that FCM algorithm calculates N number of normal data in described step 2:

FCM is insensitive to starting condition, X={x _i| i=1,2 ..., N} is the data set of N number of accelerometer sample composition, and one of them sample is expressed as x _i=[x _i1x _ib], x _ibe a sampled point, have b data in a sampled point, x _ibbe b data;

The cluster that FCM has come N number of sample by asking for the solution making objective function minimum, objective function is as follows,

$J=\underset{k=1}{\overset{c}{\&Sigma;}}\underset{i=1}{\overset{N}{\&Sigma;}}{\left({\mathrm{\&mu;}}_{ki}\right)}^m{d}_{ki}^2$

Wherein, μ _ki∈ [0,1], m represents Fuzzy Exponential, and c represents the number of cluster centre; K represents a kth cluster centre, and i represents i-th sample, μ _kirepresent the degree of membership of i-th sample relative to a kth cluster centre;

The solution of objective function is as follows:

${\mathrm{\&mu;}}_{ki}=\frac{1}{\underset{j=1}{\overset{c}{\&Sigma;}}{(d_{ki}/d_{ji})}^{2/(m-1)}}$

υ _krepresent a kth cluster centre, d _kirepresent sample x _iwith cluster centre υ _kbetween distance, j represents a jth cluster centre, μ _kirepresent the degree of membership of i-th sample for a kth cluster centre; d _jirepresent the distance of i-th sample for a jth cluster centre.

3. the online automotive suspension method for monitoring performance based on possibility C means clustering algorithm according to claim 1 and 2, is characterized in that described step 2 using the cluster centre that calculates as the initial value of PCM, calculates the cluster centre υ of PCM _kwith weights η _ibe specially:

PCM realizes cluster by the result asking for objective function minimum, and objective function is as follows,

$J=\underset{i=1}{\overset{N}{\&Sigma;}}\underset{k=1}{\overset{c}{\&Sigma;}}{\left({\mathrm{\&mu;}}_{ki}\right)}^m{d}_{ki}^2+\underset{k=1}{\overset{c}{\&Sigma;}}{\mathrm{\&eta;}}_i\underset{i=1}{\overset{N}{\&Sigma;}}{(1-{\mathrm{\&mu;}}_{ki})}^m$

Wherein, μ _ki∈ [0,1], η _ibe self-defining constant, the solution of objective function is as follows,

${\mathrm{\&eta;}}_i=\frac{\underset{i=1}{\overset{N}{\&Sigma;}}{\mathrm{\&mu;}}_{ki}^m{d}_{ki}^2}{\underset{i=1}{\overset{N}{\&Sigma;}}{\mathrm{\&mu;}}_{ki}^m}$

${\mathrm{\&mu;}}_{ki}=\frac{1}{1+{(d_{ki}^2/{\mathrm{\&eta;}}_i)}^{1/(m-1)}}$

${\mathrm{\&upsi;}}_i=\frac{\underset{i=1}{\overset{N}{\&Sigma;}}{\mathrm{\&mu;}}_{ki}^m{x}_i}{\underset{i=1}{\overset{N}{\&Sigma;}}{\mathrm{\&mu;}}_{ki}^m}.$

4. the online automotive suspension method for monitoring performance based on possibility C means clustering algorithm according to claim 3, is characterized in that described step 5 FDA algorithm is specially:

Define following matrix: overall scatter matrix S _t, scatter matrix within class S _wwith inter _ class relationship matrix S _b, overall scatter matrix S _tbe defined as follows:

$S_t=\underset{i=1}{\overset{n}{\&Sigma;}}(x_i-\stackrel{\&OverBar;}{x}){(x_i-\stackrel{\&OverBar;}{x})}^T$

Wherein, for the average value vector of total sample, the scatter matrix within class S of jth class sample _jbe defined as follows:

$S_j=\underset{x_i\&Element;X_j}{\&Sigma;}(x_i-\stackrel{\&OverBar;}{x_j}){(x_i-\stackrel{\&OverBar;}{x_j})}^T$

Wherein X _jit is the vector x belonging to jth class sample _iset, for the mean value of jth class sample;

Then scatter matrix within class S _wfor:

$S_w=\underset{j=1}{\overset{p}{\&Sigma;}}{S}_j$

Inter _ class relationship matrix S _bbe defined as follows:

$S_b=\underset{j=1}{\overset{p}{\&Sigma;}}{n}_j(\stackrel{\&OverBar;}{x_j}-\stackrel{\&OverBar;}{x}){(\stackrel{\&OverBar;}{x_j}-\stackrel{\&OverBar;}{x})}^T$

FDA objective function is as follows:

$\underset{v\&NotEqual;0}{\max }\frac{v^T{S}_{b}v}{v^T{S}_{w}v}$

The calculating of FDA vector can be equivalent to the proper vector w asking for following generalized eigenvalue problem _k;

S _bw _k＝λ _kS _ww _k；

N is feature samples sum, and q is the dimension of a sample; P is sample class number, n _jthe number of samples of jth class observation sample, x _ibe i-th sample, by all training samples stored in matrix X ∈ R ^{n × q}in, easily know that the transposition of X i-th row is exactly vector x _i, λ _krepresentation eigenvalue, v ^trepresent the transposition of v, v represents the solution of FDA objective function.