阅读说明：本技术 一种基于chde和pwfp的共轨喷油器敏感故障特征提取方法 (Common rail fuel injector sensitive fault feature extraction method based on CHDE and PWFP ) 是由 宋恩哲 柯赟 姚崇 于 2019-12-30 设计创作，主要内容包括：本发明的目的在于提供一种基于CHDE和PWFP的共轨喷油器敏感故障特征提取方法,首先利用高精度压力传感器收集高压油管压力信号；然后计算燃油压力信号的复合层次离散熵；接着计算各层次的离散熵之间的邻近度,并依照邻近度打分,分数按照升序排列,分数越低,层次的离散熵对故障特征更敏感；最后将测试样本输入训练后二叉树支持向量机多分类器进行故障诊断和模式识别,并输出故障诊断结果。本发明适应于复杂工况的共轨喷油器敏感故障特征的提取,具有较好的故障诊断效果。(The invention aims to provide a method for extracting sensitive fault characteristics of a common rail fuel injector based on CHDE and PWFP, which comprises the steps of firstly collecting pressure signals of a high-pressure fuel pipe by using a high-precision pressure sensor; then calculating the composite level discrete entropy of the fuel pressure signal; then, calculating the proximity between the discrete entropies of all levels, and scoring according to the proximity, wherein the scores are arranged in an ascending order, and the lower the score is, the more sensitive the discrete entropies of the levels to fault characteristics are; and finally, inputting the test sample into the trained binary tree support vector machine multi-classifier for fault diagnosis and pattern recognition, and outputting a fault diagnosis result. The method is suitable for extracting the sensitive fault characteristics of the common rail fuel injector under the complex working conditions, and has a good fault diagnosis effect.)

1. A common rail fuel injector sensitive fault feature extraction method based on CHDE and PWFP is characterized by comprising the following steps:

(1) collecting fuel pressure fluctuation signals of a high-pressure oil pipe through a pressure sensor, and dividing the collected signals into training signals and testing signals;

(2) respectively calculating the composite level discrete entropy of the training signal and the test signal;

(3) calculating the proximity between discrete entropies of each layer, scoring by taking the proximity as a reference, and arranging the scores according to an ascending order;

(4) selecting composite level discrete entropy in a training sample to form a feature vector subset, and inputting the feature vector subset into a binary tree support vector machine multi-classifier for training;

(5) and carrying out fault diagnosis and pattern recognition on the test sample by adopting the trained binary tree support vector machine multi-classifier, and outputting a fault diagnosis result.

2. The method for extracting the sensitive fault characteristics of the common rail injector based on the CHDE and the PWFP according to claim 1, characterized by comprising the following steps: the pressure fluctuation signals of the high-pressure oil pipe in the step (1) comprise three types of normal state of the oil sprayer, clamping stagnation of a needle valve of the oil sprayer and blockage of a spray hole of the oil sprayer.

3. The method for extracting the sensitive fault characteristics of the common rail injector based on the CHDE and the PWFP according to claim 1, characterized by comprising the following steps: the calculation steps of the composite level discrete entropy in the step (2) are as follows:

A. let a time series { x (i), i ═ 1, 2., N }, of length N, define an operatorAndthe following were used:

j＝0,1,...,2^n-1，

andrepresenting the low frequency components of the time series decomposed at the first layer,andrepresenting the high frequency components of the time series decomposition at the first layer,andrepresenting two different layering modes of time series under the same scale;

B. defining the time series x (i) the node components of each layer decomposition are as follows:

C. calculating the discrete entropy of the hierarchical sequence obtained by each node, and then averaging the entropy values of different k of the same node to obtain the composite level discrete entropy of each level, which is marked as CHDE_n,e。

4. The method for extracting the sensitive fault characteristics of the common rail injector based on the CHDE and the PWFP according to claim 1, characterized by comprising the following steps: in the step (3), scores are assigned according to the proximity degree of the samples in the same category, the maximum distance between the samples in other categories is kept, then each feature is assigned with a score to perform feature selection, and the flow of the PWFP algorithm is described as follows:

a. keeping β features out of d:

b. let q be_jk＝[b₁,b₂,...,b_d]^T,b_i∈ {0,1} is (x)_j,x_k) Features of the opposite edge; similar features are found by:

c. information is collected from all possible pairs, denoted by P and Q respectively, as:

d. the criterion for selecting the feature is minimization, as follows:

5. the method for extracting the sensitive fault characteristics of the common rail injector based on the CHDE and the PWFP according to claim 1, characterized by comprising the following steps: in the steps (4) and (5), the binary tree SVM adopts an RBF kernel function to classify, and the penalty factor C is 1000.

Technical Field

The invention relates to a diesel engine fault extraction method, in particular to a diesel engine common rail fuel injector fault extraction method.

Background

The electric control high-pressure common rail fuel injection technology is taken as the third diesel engine technology after the high-pressure injection technology and the supercharging technology, and becomes the hot spot of the competition of countries in the world in the aspect of the marine diesel engine technology. As the common rail fuel system has increasingly complex functions and structures and severe working environment, the reliability of the common rail diesel fuel system becomes an important research point for the electric control fuel system. The engine management research of the Japan Ship east Association shows that the failure rate of the fuel injector accounts for 17.1% of the main engine failure, and the failure of the fuel injector causes the combustion deterioration, the power performance, the economic performance and the reliability performance of the diesel engine to be reduced and the harmful emissions to be increased. Therefore, the method has important significance in timely and accurately diagnosing the fault of the common rail oil injector.

The concept of Composite Hierarchical Dispersion Entropy (CHDE) is used to measure the complexity of a fuel pressure wave time series at different scales or frequencies. The information of all sequences under the same scale is fully considered in the CHDE method, the entropy value of the node is the mean value of the entropy values of all the sequences, and the problem of entropy value mutation caused by sequence shortening can be well restrained. Then, when the fault information of the original time series is reflected by taking the composite-level discrete entropy as a feature, redundant information and insensitive information are often contained in the fault feature, so that the selection of the fault feature is indispensable. For high-dimensional low-sample data, the processing effect of the existing dimension reduction method is not obvious enough.

Disclosure of Invention

The invention aims to provide a method for extracting sensitive fault characteristics of a common rail injector based on CHDE and PWFP, which solves the problems that the fault characteristics of the common rail injector are difficult to extract or the extraction precision is not high in the complex working condition environment.

The purpose of the invention is realized as follows:

the invention relates to a CHDE and PWFP-based common rail injector sensitive fault feature extraction method, which is characterized by comprising the following steps:

(1) collecting fuel pressure fluctuation signals of a high-pressure oil pipe through a pressure sensor, and dividing the collected signals into training signals and testing signals;

(2) respectively calculating the composite level discrete entropy of the training signal and the test signal;

(3) calculating the proximity between discrete entropies of each layer, scoring by taking the proximity as a reference, and arranging the scores according to an ascending order;

(4) selecting composite level discrete entropy in a training sample to form a feature vector subset, and inputting the feature vector subset into a binary tree support vector machine multi-classifier for training;

(5) and carrying out fault diagnosis and pattern recognition on the test sample by adopting the trained binary tree support vector machine multi-classifier, and outputting a fault diagnosis result.

The present invention may further comprise:

1. the pressure fluctuation signals of the high-pressure oil pipe in the step (1) comprise three types of normal state of the oil sprayer, clamping stagnation of a needle valve of the oil sprayer and blockage of a spray hole of the oil sprayer.

2. The calculation steps of the composite level discrete entropy in the step (2) are as follows:

A. let a time series { x (i), i ═ 1, 2., N }, of length N, define an operatorAndthe following were used:

j＝0,1,...,2^n-1，

andrepresenting the low frequency components of the time series decomposed at the first layer,andrepresenting the high frequency components of the time series decomposition at the first layer,andrepresenting two different layering modes of time series under the same scale;

B. defining the time series x (i) the node components of each layer decomposition are as follows:

C. calculating the discrete entropy of the hierarchical sequence obtained by each node, and then averaging the different k entropy values of the same node to obtain the composite level discrete entropy of each level, which is marked as CHDE_n,e。

3. In the step (3), scores are distributed according to the proximity degree of the samples in the same category, the maximum distance between the samples in other categories is kept, then each feature is distributed with scores to perform feature selection, and the flow of the PWFP algorithm is described as follows:

a. keeping β features out of d:

b. let q be_jk＝[b₁,b₂,...,b_d]^T,b_i∈ {0,1} is (x)_j,x_k) Features of the opposite edge; similar features are found by:

c. information is collected from all possible pairs, denoted by P and Q respectively, as:

d. the criterion for selecting the feature is minimization, as follows:

4. in the steps (4) and (5), the binary tree SVM adopts an RBF kernel function to classify, and the penalty factor C is 1000.

The invention has the advantages that: the method is suitable for extracting the sensitive fault characteristics of the common rail fuel injector under the complex working conditions, and has a good fault diagnosis effect.

Drawings

FIG. 1 is a flow chart of the present invention;

FIG. 2 is a block diagram of a high pressure common rail fuel system test platform;

FIG. 3 is a signal diagram of fuel pressure in the high pressure rail for three fuel injector states under different operating conditions;

FIG. 4 is a flowchart of CHDE calculation;

FIG. 5 is a chart of CHDE calculations for three injectors under different operating conditions.

Detailed Description

The invention will now be described in more detail by way of example with reference to the accompanying drawings in which:

with reference to fig. 1-5, a method for extracting sensitive fault characteristics of a common rail injector based on a channel and a power supply unit (PWFP) includes the following steps:

s1, collecting fuel pressure fluctuation signals of the high-pressure oil pipe by using a high-precision pressure sensor, and dividing the collected signals into training signals and testing signals;

s2, respectively calculating the composite level discrete entropy of the training signal and the test signal;

s3, calculating the proximity between discrete entropies of each layer, and scoring by taking the proximity as a reference, wherein the scores are arranged in an ascending order, and the lower the score is, the more sensitive the fault characteristics are;

s4, selecting the composite level discrete entropy in the training sample with the top rank to form a feature vector subset, and inputting the feature vector subset into a binary tree support vector machine multi-classifier for training;

and S5, performing fault diagnosis and pattern recognition on the test sample by adopting the trained binary tree support vector machine multi-classifier, and outputting a fault diagnosis result.

The high-pressure oil pipe pressure fluctuation signals in the step S1 comprise three types of normal state of the oil sprayer, clamping stagnation of the oil sprayer needle valve and blockage of the oil sprayer spray hole.

The calculation steps of the composite-level discrete entropy in step S2 are as follows:

the first step is as follows: let a time series { x (i), i ═ 1, 2., N }, of length N, define an operatorAndthe following were used:

j＝0,1,...,2^n-1

in fact, it is possible to use,andrepresenting the low frequency components of the time series decomposed at the first layer,andrepresents the high frequency components of the time sequence decomposed at the first layer, andandrepresenting two time series at the same scaleDifferent hierarchical approaches.

The second step is that: defining the time series x (i) the node components of each layer decomposition are as follows:

the third step: calculating the discrete entropy of the hierarchical sequence obtained by each node, and then averaging the different entropy values of k of the same node to obtain the composite hierarchical discrete entropy of each hierarchy, which is marked as CHDE_n,e。

The core idea of the pairwise proximity (PWFP) in step S3 is to assign scores based on proximity to samples of the same class while maintaining maximum distance to samples of other classes, and then assign scores to each feature for feature selection. The PWFP algorithm flow can be described as follows:

first step β features need to be kept out of d for (x)_j,x_k) Close to each other:

the second step is that: similarly, let q be_jk＝[b₁,b₂,...,b_d]^T,b_i∈ {0,1} is (x)_j,x_k) Features of the opposite edge; if b_i＝1，y_j≠y_kIs the most distant. The method for finding similar features can be found by the following ways:

the third step: information is collected from all possible pairs, denoted by P and Q respectively, as:

the fourth step: is well characterized in thatAndthe feature with higher probability appears in the table. A reasonable criterion for selecting good features is to minimize the difference between the following equation:

in steps S4 and S5, the binary tree SVM performs classification by using an RBF kernel function, and the penalty factor C is 1000.

The invention discloses a CHDE and PWFP-based common rail injector sensitive fault feature extraction method, which comprises the following specific steps:

s1, collecting fuel pressure fluctuation signals of the high-pressure oil pipe by using the high-precision pressure sensor, dividing the collected signals into training signals and testing signals, wherein the test platform is shown in figure 2, and the collected signals are shown in figure 3.

And S2, respectively calculating the composite level discrete entropy of the training signal and the test signal. The CHDE calculation flow is shown in FIG. 4, the calculation result is shown in FIG. 5, and the specific steps are as follows:

the first step is as follows: let a time series { x (i), i ═ 1, 2., N }, of length N, define an operatorAndthe following were used:

whereinAndthe form depends on the length of the time series, j being 0 or 1. Will be provided withAndacting on time series x (i) respectively, thus having

j＝0,1,...,2^n-1

In fact, it is possible to use,andrepresenting the low frequency components of the time series decomposed at the first layer,andrepresents the high frequency components of the time sequence decomposed at the first layer, andandrepresenting two different hierarchical modes of time series under the same scale.

The second step is that: constructing an n-dimensional vectorThe integer e can be represented as

Wherein the vector corresponding to the positive integer e is

The third step: based on vectorsDefining the time sequence x (i) the node components of each layer decomposition are as follows

The fourth step: calculating the discrete entropy of the hierarchical sequence obtained by each node, and then averaging the different entropy values of k of the same node to obtain the composite hierarchical discrete entropy of each hierarchy, which is marked as CHDE_n,e。

For the low frequency part, the time series is defined by the following formula for the time series { x (i), i ═ 1,2,.. N }, with scale factor τNamely, it is

j＝1,2,...,[N/τ],p＝1,2,...,τ

Then, for each scale factor τ, each time series is calculatedThen calculating the average value of the scale to obtain the discrete entropy under the scale factor, namely

And S3, calculating the proximity between the discrete entropies of each layer, and scoring by taking the proximity as a reference, wherein the scores are arranged in an ascending order, and the lower the score is, the more sensitive the fault characteristics are. The method comprises the following specific steps:

first step β features need to be kept out of d for (x)_j,x_k) Close to each other:

the second step is that: similarly, let q be_jk＝[b₁,b₂,...,b_d]^T,b_i∈ {0,1} is (x)_j,x_k) Features of the opposite edge; if b_i＝1，y_j≠y_kIs the most distant. The method for finding similar features can be found by the following ways:

the third step: information is collected from all possible pairs, denoted by P and Q respectively, as:

the fourth step: good characteristics are that when P ═ P¹,p²,...,p^d]And Q ═ Q¹,q²,...,q^d]The feature with higher probability appears in the middle. A reasonable criterion for selecting good features is to minimize the difference between the following equation:

s4, selecting the composite level discrete entropy in the training sample with the top rank to form a feature vector subset, and inputting the feature vector subset into a binary tree support vector machine multi-classifier for training;

and S5, performing fault diagnosis and pattern recognition on the test sample by adopting the trained binary tree support vector machine multi-classifier, and outputting a fault diagnosis result.