CN112085202A

CN112085202A - Automobile fault diagnosis method based on hybrid Bayesian network

Info

Publication number: CN112085202A
Application number: CN202010941854.5A
Authority: CN
Inventors: 杨晛; 潘春茹; 关展旭
Original assignee: Yanshan University
Current assignee: Yanshan University
Priority date: 2020-09-09
Filing date: 2020-09-09
Publication date: 2020-12-15

Abstract

The application discloses an automobile fault diagnosis method based on a hybrid Bayesian network, which comprises the following steps: and obtaining a hybrid Bayesian network by adopting a hybrid Bayesian network structure learning method combining a sequence generation method based on conditional entropy and a K2 algorithm fusing expert information, and then carrying out automobile fault diagnosis by using the hybrid Bayesian network. The performance effect same as that of the mutual information sequencing can be obtained under the condition of no expert knowledge, but the calculation amount is simpler and more convenient. A small amount of expert knowledge is blended into the K2 algorithm to obtain a better performance effect, and the deviation possibly caused by the missing of sample data or large difference of input sequence can be corrected to learn a better network structure. The problem of high computational complexity in the automobile fault diagnosis method is solved, so that the automobile fault diagnosis is quicker and the diagnosis result is more accurate.

Description

Automobile fault diagnosis method based on hybrid Bayesian network

Technical Field

The invention relates to the field of information technology and machine learning, in particular to an automobile fault diagnosis method based on a hybrid Bayesian network.

Background

With the development of automobile technology, the automobile structure and the electric control part are complex, the fault phenomena are diversified, the fault causes are complex, and randomness and ambiguity are caused, so that the fault has certain uncertainty.

The Bayesian network is a causal graph model which can express the mutual relation between random variables based on probability theory and graph theory. The Bayesian network has a solid theoretical foundation, is an effective method for processing uncertain knowledge expression and reasoning, can be used for representing and reasoning uncertain problems in automobile fault diagnosis, effectively integrates the prior knowledge in the field and the distribution characteristics of real-time sensing data, and realizes the self-adaptation of the automobile fault diagnosis system.

When the automobile fault diagnosis is realized based on the Bayesian network, firstly, a structure of the Bayesian network needs to be established, and the method for establishing the structure of the Bayesian network mainly comprises the following steps: score search based methods and dependency analysis based methods. The scoring-based search method mainly comprises two parts, namely a scoring function and a search algorithm, the search algorithm is adopted to continuously search the network structure, each structure is scored through the scoring function, and the network structure with the highest score is output after the algorithm is finished, wherein the most classical network structure is the K2 algorithm. The K2 algorithm takes sample data and a given sequence as input, starts to gradually increase connecting edges with father nodes by an empty node set and calculates scores, and outputs a network structure after the algorithm is finished. The dependency analysis based method mainly judges the interrelationship between nodes through independence detection and determines the direction of the interrelationship. According to the method, as the number of nodes is increased, the independence detection times of the algorithm are exponentially increased, so that the structural difference is large.

The classic K2 algorithm has the limitation that the order of nodes needs to be determined by manual input, at present, an initial network graph is generated and a graph is searched to obtain a corresponding node sequence as the input of the K2 algorithm by adopting a mutual information-based method, however, the mutual information-based method needs to generate an initial structure graph and judge the direction of edges, and the algorithm complexity is high.

Disclosure of Invention

The invention aims to provide an automobile fault diagnosis method based on a hybrid Bayesian network, which aims to overcome the problem of high calculation complexity in the existing automobile fault diagnosis method, so that the automobile fault diagnosis is faster and the diagnosis result is more accurate.

In order to achieve the above object, the following solutions are proposed:

1. a vehicle fault diagnosis method based on a hybrid Bayesian network is characterized by comprising the following steps:

collecting information of relationships among all nodes in the automobile fault diagnosis problem provided by a group expert, fusing the obtained node relationships by adopting a D-S evidence theory, making a decision on a fusion result and converting effective information after the decision into an adjacency matrix;

combining the prior structure with a K2 algorithm according to the obtained adjacency matrix as the prior structure, and limiting a search space of the K2 algorithm through the prior structure;

calculating entropy functions of all nodes in the automobile fault data and sequencing to obtain a node sequence;

inputting the node sequence into a K2 algorithm fused with a prior structure for Bayesian network structure learning to obtain a mixed Bayesian network structure;

and carrying out automobile fault diagnosis based on the hybrid Bayesian network structure to obtain an automobile fault diagnosis result.

Preferably, the collecting information of relationships among all nodes in the automobile fault diagnosis problem provided by the group experts, and fusing the obtained node relationships by using a D-S evidence theory includes:

in the process of collecting expert knowledge, three causal relationships exist between any two nodes (i, j), including: i → j represents that the expert determines that a directed edge pointing to j exists, j → i represents that the expert determines that a directed edge pointing to i exists, and i ≠ j indicates that no pointing relation exists between the nodes;

collecting all expert knowledge to obtain a credibility allocation table of all experts;

fusing the credibility allocation table by applying a D-S fusion rule to obtain a final fusion result;

wherein, if m is known₁And m₂The D-S fusion rule is as follows:

wherein m (A) is a fusion result, A, B and C represent propositions or assumptions in the automobile fault diagnosis problem; phi denotes an empty set, K denotes a collision coefficient, and when K is 0, m is described₁And m₂There is no conflict between them.

Preferably, the making a decision on the fusion result includes:

after obtaining the final fusion result m (A)_h) And after h is more than or equal to 1 and less than or equal to 4, deciding whether a corresponding directed edge exists according to a preset threshold value theta in the following decision mode:

if m (A)₁) If > theta, determining i → j;

if m (A)₂) If > theta, determining j → i;

if m (A)₃) If the value is larger than theta, determining that i is not equal to j;

if m (A)₄) > theta or m (A)_h) If the values are all smaller than theta, the relation between the two nodes cannot be determined, and then the subsequent processing is carried out according to the information without experts.

Preferably, the calculating the entropy function of each node in the vehicle fault data and sequencing to obtain the node sequence includes:

respectively calculating and sequencing the information entropy of each node in the automobile fault data, and selecting a first node as an initial node;

calculating the conditional entropy of other nodes under the known condition of the selected node, sequencing, and selecting one node as an Nth sequence node;

repeatedly executing the calculation of the conditional entropy until the last sequential point is obtained, wherein the selected nodes in each calculation of the conditional entropy comprise the nodes selected before the current calculation, and N is a positive integer which is more than 1 and less than the total number of the nodes;

and obtaining a node sequence according to the starting node and each sequence node.

Preferably, when the Bayesian network structure learning is performed, a Bayesian information metric BIC score is used as a scoring function.

Preferably, the inputting the node sequence into the K2 algorithm of the fusion prior structure for the bayesian network structure learning includes:

and correcting the illegal structure obtained in the searching process, wherein the illegal structure is a structure which does not accord with the attribute of the acyclic graph.

Preferably, the correcting comprises:

if a bidirectional arc exists between two nodes in the matrix, one of the two nodes is randomly selected to be set to be 0;

if the matrix has a closed loop structure, deleting edges or reversing edges are randomly selected in the loop so as to meet the requirement that the Bayesian network is a directed acyclic graph.

Compared with the prior art, the technical scheme of the invention has the beneficial effects that:

1) the invention adopts the entropy function principle to obtain the node sequence order from the data, thereby avoiding the uncertainty of the manual input order and the complexity of obtaining the node sequence order by adopting the directed spanning tree algorithm in the prior art.

2) The invention adopts the idea of integrating the initial structure into the classic k2 algorithm, thereby greatly reducing the network search and the algorithm running time.

Drawings

In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained according to the drawings without creative efforts.

FIG. 1 is a flow chart of a Bayesian network structure learning method based on knowledge and data mixing according to an embodiment of the present invention;

FIG. 2 is a diagram illustrating a correspondence between a network structure and an adjacency matrix according to an embodiment of the present invention;

FIG. 3 is a diagram of a Bayesian network in an embodiment of the present invention;

fig. 4 is an exemplary diagram of an illegal structure in an embodiment of the present invention.

Detailed Description

The invention provides a novel knowledge and data fusion-based hybrid Bayesian network structure learning method by integrating expert knowledge and node pre-sequencing based on a K2 algorithm, and the obtained hybrid Bayesian network can be used for automobile fault diagnosis.

The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application, and it is obvious that the described embodiments are only a part of the embodiments of the present application, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present application.

The information entropy (entropy function) is an effective method for processing uncertain information, and when the uncertain information is expressed by probability, the information entropy can be converted into the fragrance entropy. The fragrance concentration entropy is widely applied to the field of information processing at present.

The D-S evidence theory (Dempster-Shafer) is also called belief function theory, which has been greatly developed in theory and application since being proposed, and is an effective tool for processing uncertain information at present.

Referring to fig. 1, a flowchart of a hybrid bayesian network-based vehicle fault diagnosis method according to an embodiment of the present invention is shown. The method comprises the steps that an engine is difficult to start in a cold mode at the temperature of 20-25 ℃ and serves as a father node, and related child nodes of the engine are respectively fuel quality, a spark plug, a water temperature sensor, a fuel pressure regulator, a fuel pump circuit, an oil pipe and cylinder compression pressure, and the above components form a Bayesian network diagram.

The method comprises the following steps:

and S1, collecting information of relationships among all nodes in the automobile fault diagnosis provided by the group experts, fusing the obtained node relationships by adopting a D-S evidence theory, making a decision on a fusion result, and converting the effective information after the decision into an adjacency matrix.

The method comprises the steps of collecting group expert knowledge of the relation among all nodes of the automobile fault diagnosis problem, fusing by a D-S evidence theory, and obtaining accurate prior information as the known cause-effect relation.

In the process of collecting expert knowledge, three kinds of causal relationships exist between any two nodes (i, j), namely i → j represents that a directed edge pointing to j is determined by an expert to exist by i, j → i represents that a directed edge pointing to i is determined by the expert to exist by j, and i ≠ j indicates that no directional relationship exists between the nodes. After all knowledge has been collected, the confidence scores for all experts can be obtained as shown in table 1.

TABLE 1

	i→j	j→i	i≠j	Uncertainty
					E₁	m₁(A₁)	m₁(A₂)	m₁(A₃)	m₁(A₄)
E₂	m₂(A₁)	m₂(A₁₂)	m₂(A₃)	m₂(A₄)
					…	…	…	…	…
E_n	m_n(A₁)	m_n(A₂)	m_n(A₃)	m_n(A₄)
					DS	m(A₁)	m(A₂)	m(A₃)	m(A₄)

Wherein E_k(1. ltoreq. k. ltoreq.n) represents the kth expert, A₁，A₂......A_hRepresenting any hypothesis or proposition, m, in the vehicle fault diagnosis problem₁，m₂......m_nExpressing probability distribution of hypothesis or proposition, applying D-S rule for fusion after obtaining opinion distribution of all n-bit experts, and applying D-S rule for opinion of n-bit expertsAfter fusion, the final fusion result is obtained, and the D-S fusion rule is as follows:

let the recognition framework (sample space) Ω be a set of mutually independent but overall exhaustive events, as follows:

Ω＝{θ₁，θ₂，…，θ_i，…，θ_n}；

power set 2 of omega^ΩRepresents all assumptions or propositions, expressed as:

for a specific recognition frame Ω, its basic information distribution function is 2^θIn [0, 1 ]]A mapping of (2);

m：2^θ→[0，1](ii) a Satisfies the following conditions:

where phi denotes an empty set.

Known m₁And m₂The combination rule of the D-S evidence theory is as follows.

Wherein m (A) is a fusion result, A, B and C respectively represent propositions or assumptions in the automobile fault diagnosis problem; phi denotes an empty set, K denotes a collision coefficient, and when K is 0, m is described₁And m₂There is no conflict between them.

if m (A)₁) If > theta, determining i → j;

if m (A)₂) If > theta, determining j → i;

The explicit relationship between the nodes after fusion is given in the form of an adjacency matrix, and the corresponding relationship between the network structure and the adjacency matrix is shown in fig. 2.

If the directed edge of the node 1 pointing to the node 2 is accurately obtained after the multi-expert information is fused, no connecting edge exists between the node 3 and the node 5 and between the node 4 and the node 8, namely 1 → 2, the node 3 is not equal to 5, and the node 4 is not equal to 8. Converting the multi-expert knowledge fusion result into a form of an adjacency matrix A:

and S2, combining the prior structure with the K2 algorithm according to the obtained adjacent matrix as the prior structure, and limiting the search space of the K2 algorithm through the prior structure.

In the original K2 algorithm implementation process, for each node, starting from an empty node set, continuously increasing connecting edges with a father node according to a node sequence relation input in advance and judging the change of network scores, and if the scores are increased, inputting the node into the empty father node set. In the embodiment of the present invention, 1 → 2 in the fused adjacency matrix indicates that the parent node set of node 2 necessarily includes node 1. Therefore, in the process of searching parent nodes for partial nodes in the improvement process, the process is not started from an empty node set any more. And for the nodes 3 and 5, and between the node 4 and the node 8, there is definitely no directed edge, and if the whole network score is higher by adding the directed edge in the subsequent searching process, the directed edge is forcibly cancelled according to the preset of the prior network.

And S3, calculating entropy functions of all nodes in the automobile fault data and sequencing to obtain a node sequence.

The specific implementation can be as follows:

Wherein, the information entropy can be calculated as follows:

let X be a discrete random variable, with X possibly taking the value { X }₁，x₂...x_nAnd expressing the fragrance concentration entropy as follows:

where p (X) is the edge probability distribution function of X.

The conditional entropy can be calculated as follows:

let p (X, Y) be a joint distribution function of discrete variables X and Y, which may take the values { (X, Y)₁，y₁)，(x₂，y₂)，...(x_n，y_n) And then the joint distribution function is expressed as:

given X, the information entropy of Y is called conditional entropy, and then the conditional entropy can be expressed as:

where p (X, Y) is the joint probability distribution function of X and Y, and p (X) is the edge probability distribution function of X.

Conditional entropy can also be expressed as:

H(Y|X)＝H(X，Y)-H(X)；

referring to fig. 3, a bayesian network structure for vehicle fault diagnosis in an embodiment of the present invention is shown, where the bayesian network structure includes 8 nodes. Taking the bayesian network structure as an example to explain a specific process of obtaining a node sequence, the following is shown:

(1) respectively calculating the information entropies of 8 nodes and sequencing, wherein the obtained result is as follows: h (1) < H (3) < H (5) < H (4) < H (7) < H (6) < H (8) < H (2), and if it is calculated that H (1) is minimum, the node 1 is selected as the start node.

(2) The conditional entropy of the remaining nodes under known conditions at node 1 is calculated and sorted. The results obtained were: h (2|1) < H (4|1) < H (3|1) < H (5|1) < H (6|1) < H (7|1) < H (8|1), and node 2 is selected as the second sequential node.

(3) Calculating the conditional entropy under the known conditions of the nodes 1 and 2 and sequencing, wherein the obtained result is as follows: h (4|1,2) < H (3|1,2) < H (5|1,2) < H (6|1,2) < H (7|1,2) < H (8|1,2), and the node 4 is selected as the third sequential point.

(4) Calculating the conditional entropy under the condition that the nodes 1,2 and 4 are known and sequencing, wherein the obtained result is as follows: h (3|1,2,4) < H (5|1,2,4) < H (6|1,2,4) < H (7|1,2,4) < H (8|1,2,4), the node 3 is a fourth sequential point.

(5) Calculating the conditional entropy under the condition that the nodes 1,2,4 and 3 are known and sequencing, wherein the obtained result is as follows: h (6|1,2,4,3) < H (5|1,2,4,3) < H (7|1,2,4,3) < H (8|1,2,4,3), the node 6 is the fifth sequential point.

(6) Calculating the conditional entropy under the condition that the nodes 1,2,4,3 and 6 are known and sequencing, wherein the obtained result is as follows: h (5|1,2,4,3,6) < H (8|1,2,4,3,6) < H (7|1,2,4,3,6), the node 5 is a sixth sequential point.

(7) Calculating the conditional entropy under the condition that the

nodes

1,2,4,3,6 and 5 are known and sequencing, wherein the obtained result is as follows: h (7|1,2,4,3,6,5) < H (8|1,2,4,3,6,5), then node 7 is the seventh order point and node 8 is the last order point.

The resulting node sequence is: node 1-node 2-node 4-node 3-node 6-node 5-node 7-node 8.

S4, inputting the node sequence into a K2 algorithm of a fusion prior structure to learn the Bayesian network structure, and obtaining a mixed Bayesian network structure;

the resulting node sequence is input into the K2 algorithm that fuses expert knowledge. (the learning result of the K2 algorithm is not influenced by changing the order of the nodes in the same layer), the embodiment of the invention takes Bayesian information measurement BIC score as a scoring function:

score_BIC(G：D)＝score_MLE(G：D)-0.5logM*DIM[G]；

wherein score_MLE(G: D) is a maximum likelihood score function,

representing the dimensions of the network structure.

In the structure learning by using the improved K2 algorithm, the structure state obtained in the search process may become infeasible because the structure state no longer conforms to the attributes of the acyclic graph, and then the illegal structure (such as fig. 4) needs to be modified, and the steps are as follows:

if two nodes in the matrix are mutually dependent, the existence of bidirectional arcs is unreasonable, and one of the two nodes can be randomly selected to be set to be 0;

if a closed loop structure exists in the matrix, deleting edges or reversing edges should be randomly selected in the loop so as to meet the requirement that the Bayesian network is a directed acyclic graph.

And S5, carrying out automobile fault diagnosis based on the hybrid Bayesian network structure to obtain an automobile fault diagnosis result.

The embodiment of the invention adopts a mixed Bayesian network structure learning method combining a sequence generation method based on conditional entropy and a K2 algorithm fusing expert information. The performance effect same as that of the mutual information sequencing can be obtained under the condition of no expert knowledge, but the calculation amount is simpler and more convenient. On the basis, the method provided by the invention is obtained by adding expert knowledge in the matrix A, compared with the MWST and MMHC algorithms which only use data to carry out sequencing search, a small amount of expert knowledge is blended into the K2 algorithm to obtain a better performance effect, and the deviation possibly caused by sample data missing or large difference of input sequence can be corrected to learn a better network structure.

The following is a verification of the improved K2 algorithm in the practice of the present invention:

in order to avoid process randomness, the algorithm is run for ten times, the number (RE) of the added edges (IE), the number (ME) of the reversed edges (ME), the Hamming distance (SHD) and the like are selected as measurement and evaluation indexes, and the advantages of the hybrid Bayesian network structure learning method for fusing knowledge and data provided by the invention are verified by comparing the algorithm with the currently popular maximum minimum hill climbing method (MMHC), maximum spanning tree algorithm (MWST), mutual information sorting (MI) and conditional entropy sorting (Con-En), as shown in Table 2.

TABLE 2

On the basis, by continuously increasing the amount of expert knowledge, more relationships among nodes can be determined after fusion, corresponding results are recorded, and the invention principle is further verified, as shown in table 3.

TABLE 3

Determining a number of relationships	IE	ME	RE
				3	0.2	0.1	1.2
5	0.3	0	0.8
				7	0	0	0.2

As can be seen from the learning results in table 3, as the number of experts increases, the relationship between nodes that can be determined is also greater, the performance of bayesian network structure learning is improved continuously, and the network structure obtained by learning is closer to the original structure, that is, the number of expert knowledge is increased, so that the accuracy of the learning result can be improved. Therefore, it can be seen from the above analysis that expert knowledge plays an important role in the learning of the bayesian network structure.

Finally, it should also be noted that, herein, relational terms such as first and second, and the like may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Also, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising an … …" does not exclude the presence of other identical elements in a process, method, article, or apparatus that comprises the element.

The embodiments in the present description are described in a progressive manner, each embodiment focuses on differences from other embodiments, and the same and similar parts among the embodiments are referred to each other.

The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the application. Thus, the present application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

2. The method according to claim 1, wherein the collecting information of relationships among all nodes in the automobile fault diagnosis problem provided by the group experts and fusing the obtained node relationships by using a D-S evidence theory comprises:

wherein, if m is known₁And m₂The D-S fusion rule is as follows:

wherein m (A) is a fusion result, A, B and C respectively represent propositions or assumptions in the automobile fault diagnosis problem; Φ represents an empty set, K represents a collision coefficient, and when K is 0, m is described₁And m₂There is no conflict between them.

3. The method of claim 2, wherein the deciding the fusion result comprises:

if m (A)₁) If > theta, determining i → j;

if m (A)₂) If > theta, determining j → i;

4. The method according to claims 1 to 3, wherein the calculating an entropy function of each node in the vehicle fault data and sorting to obtain a node sequence comprises:

5. The method of claim 1, wherein a bayesian information metric BIC score is used as a scoring function in the bayesian network structure learning.

6. The method of claim 1, wherein the inputting the node sequence into a K2 algorithm for fusing prior structures for bayesian network structure learning comprises:

7. The method of claim 6, wherein the modifying comprises: