Information fusion is applied widely during fault diagnosis and there are many ways. Evidence theory or D-S theory (DST) [1, 2] can compound the uncertain information from several information sources and is a very effective uncertainty reasoning way in information fusion technology. However, the evidence theory itself has some problems [3], such as dependency on the evidence provided by expert knowledge, focus element explosion caused by evidence combination, rigorous demands to combination conditions (mutual independence among evidences), inefficiency to evidential conflicts and the subjectivity in the distribution of credibility and so on. Many scholars have improved the theory to overcome these limitations, but still there are some deficiencies [4,5,6]. Rough set theory is a new mathematical tool to deal with vagueness and uncertain knowledge and is very practical. It can extract unrelated essential characteristics to eliminate redundancies without any initial or additional data [7, 8]. Both the rough set and evidence theories are math tools for uncertain information and they focus on the group ability. There is a strong complementary relationship between these two theories [9,10,11]. This paper provides a new research thought for intelligent fault diagnosis by combining the evidence theory with rough set theory to solve the subjectivity, relativity, integrated focus element explosion and conflict problems of evidences, improve the accuracy of fault diagnosis and at last carry out online diagnostics.
The paper just gives two definitions about the importance of attribute and the basic ideas about rough set theory, which are referred in the literature [12].
The importance of any attribute
To
Here the paper just gives the D-S evidence theory synthesis formula and the related knowledge, which are referred in the literature [1, 2].
If
Among them:
From the D-S evidence theory synthesis formula: when
Among them:
Eq. (4) has some subjective factors, while the physical meaning of Eqs (2) and (3) is not clear. Since
Among them:
Suppose the conflict sum of evidence
Definition 4 shows the greater the amount of self-conflict for evidence, the smaller its weight factor is. Otherwise, the smaller the amount of self-conflict for evidence, the greater its weight factor is. The weight factor shows the important degree of evidences provided by the information source in the synthesis process and the incidence of the synthetic results.
Here is an example to prove the validity of the way. Suppose Θ = {
Three evidences are synthesised by applying the synthetic methods of DST, Yager [14], Sun Quan, Li Bicheng [15] and the synthesis method proposed in this paper, respectively (evidence weight
Comparison of results of various evidence synthesis methods.
DST | 0.99901 | 0 | 0 | 1 | 0 |
Yager | 0.99901 | 0 | 0 | 0.00099 | 0.99901 |
Sun Quan | 0.99901 | 0.3210 | 0.0030 | 0.1880 | 0.4880 |
Li Bicheng | 0.99901 | 0.6260 | 0.0067 | 0.3673 | 0 |
This paper | 0.99901 | 0.7321 | 0.0059 | 0.2620 | 0 |
Application of the attribute reduction of the rough set can effectively optimise the key features as the evidence for diagnosis decisions. Attributes significance of rough set can evaluate the weight of each evidence objectively. And then there comes the formula
This paper creates a fault diagnosis model based on inference strategy to fuse rough set and evidence theories, as shown in Figure 1. First, the basic thought is building an information decision table by discretising the fault sample data sets with continuous attributes. Then, optimising feature parameters suitable for fault diagnosis as theory body by applying the rough set theory for attribute reduction of the decision table. In practice, for the discretised sample set to diagnose, basic probability assignment of related evidences can be calculated based on the reduction and the diagnosis result can be obtained with the inductive decision by the use of the D-S combination rule.
A rough set theory can deal with discretised data, while original sample data are always continuous, so continuous data should be discretised first. There are many ways to discretise and each has its advantage. In practice, each field seeks a proper algorithm according to the characteristic [16, 17].
Symptom attribute reduction can reduce the relativity among evidence and decide the fault with attributes as less as possible, remaining the sorting quality invariant and avoiding focus element explosion. On the other hand, the weight of each attribute can be obtained from the information in the decision table and the subjectivity from experts can be avoided, overcoming the difficulties in practical application efficiently caused by the subjectivity and relativity of evidence during the fault diagnosis with evidence theory. About the reduction algorithm of the rough set theory, still there is no recognised and high-efficient algorithm. With the completion of the reduction algorithm considered, attribute reduction is applied in this paper by combining discernibility matrix, dependability of attribute and the heuristic reduction algorithm of information entropy by improving the importance of attribute. The detailed process of this algorithm is referred in the literature [18, 19].
Basic probability assignment of evidence is realised with the decision attribute
The proof of Theorems 1 and 2 can be referred in the literature [10]. From Theorems 1 and 2, it is possible to calculate belief function based on a decision table with rough set theory. It provides the theoretical basis for the fusion reasoning of evidence theory and rough set theory.
The basic probability assignment
Here
Next, a practical algorism of basic probability assignment is given based on a rough set decision table: {Input: decision table after reduction; samples to be diagnosed}; {Output: basic probability assignments of all evidences}.
Quantification of samples to be diagnosed: to discretise the samples to be diagnosed according to the discretisation criterion of original sample data;
Recognition framework Θ is decision attribute set and the reasoning evidence is condition attributes
To determine the division
To determine the division
To determine equivalence class in
To get the basic probability assignment of evidence
If there is only 1 −
The diagnosis can be concluded based on the decision of basic probability assignment after a combination of the D-S rule [20].
The engine is an important equipment of a ship, concerning the survivability and battle effectiveness of a ship. The performance of the engine can influence and restrict the performance of the technique and tactics of a ship. The structure of the engine is complex, while the fault analysis to it is difficult. To ensure the accuracy of fault diagnosis, more characteristic parameters should be used. The number of characteristics to be extracted gets larger with the increase of various kinds of parameters and this makes the amount of information to be processed too large and cannot meet the needs of online diagnosis. Only a few key characteristics are sensitive to a fault. They are independent and provide complementary information for each other to improve the accuracy of diagnosis. While the redundant characteristics are not sensible to fault or they are related to other characteristics but useless. The rough set theory can eliminate redundant information effectively, select the key characteristics, get probability assignment and make reasoning decisions with D-S combination rule to solve the problems such as subjectivity in evidence obtaining, relativity of evidence and focus element explosion of evidence combination.
After the experiments of five fault phenomena appearing in a kind of ship engine and the management to the diagnostic knowledge, this paper selects 10 sets of data samples to build. The common fault information is shown in Figure 2. In Figure 2,
The attribute value of the sample data in Table 2 is continuous and needs to be discretised. Here a discretisation for continuous attribute value based on SOFM network classification is used and the details and the computing process of this algorithm can be referred in the literature [21]. The decision table after discretisation is shown in Table 3. After attribute reduction with the method proposed in this paper, only three condition attributes,
Table of fault information.
Original samples | 1 | 0.95 | 1.40 | 1.20 | 5100 | 0.39 | |
2 | 0.70 | 1.74 | 0.94 | 3450 | 0.98 | ||
4 | 0.06 | 1.73 | 0.98 | 3900 | 0.91 | ||
5 | 0.65 | 3.31 | 0.63 | 1950 | 0.55 | ||
7 | 0.30 | 1.31 | 1.32 | 4350 | 0.40 | ||
8 | 0.28 | 1.23 | 1.11 | 4300 | 0.44 | ||
10 | 0.05 | 2.47 | 1.01 | 3800 | 0.20 | ||
Test samples | 3 | 1.02 | 1.73 | 0.57 | 3500 | 0.98 | |
6 | 0.67 | 3.71 | 0.77 | 1900 | 0.45 | ||
9 | 0.18 | 0.95 | 1.14 | 3600 | 0.30 |
Decision table after discretization.
Original samples | 1 | 3 | 1 | 3 | 3 | 2 | |
2 | 2 | 2 | 2 | 2 | 3 | ||
4 | 1 | 2 | 2 | 2 | 3 | ||
5 | 2 | 3 | 1 | 1 | 2 | ||
7 | 2 | 1 | 3 | 3 | 2 | ||
8 | 1 | 1 | 2 | 3 | 2 | ||
10 | 1 | 2 | 2 | 2 | 1 | ||
Test samples | 3 | 3 | 2 | 1 | 2 | 3 | verify |
6 | 2 | 3 | 1 | 1 | 2 | verify |
|
9 | 1 | 1 | 2 | 2 | 1 | verify |
From Definition 1, the importance of attribute
It shows
For evidence
Basic probability assignment about
0 | 0.9500 | 0 | 0 | 0 | 0 | 0.05 | ||
0 | 0.3167 | 0.3167 | 0 | 0 | 0.3167 | 0.05 | ||
0 | 0.4750 | 0.4750 | 0 | 0 | 0 | 0.05 | ||
DS | 0.9144 | 0.0397 | 0 | 0 | 0.0397 | 0.0062 | ||
Improve DS | 0.6017 | 0.8269 | 0.0703 | 0 | 0 | 0.0703 | 0.0325 | |
DS | 0.9522 | 0.0433 | 0 | 0 | 0 | 0.0045 | ||
Improve DS | 0.4512 | 0.8899 | 0.0850 | 0 | 0 | 0 | 0.0251 | |
DS | 0.4771 | 0.4771 | 0 | 0 | 0.0397 | 0.0062 | ||
Improve DS | 0.6017 | 0.4282 | 0.4282 | 0 | 0 | 0.1111 | 0.0325 | |
DS | 0.9488 | 0.0469 | 0 | 0 | 0.0039 | 0.0006 | ||
Improve DS | 0.7972 | 0.7534 | 0.1498 | 0 | 0 | 0.0569 | 0.0399 |
Basic probability assignment about
0 | 0 | 0.3167 | 0 | 0.3167 | 0.3167 | 0.05 | ||
0 | 0.3167 | 0 | 0 | 0.6334 | 0 | 0.05 | ||
0 | 0 | 0 | 0 | 0 | 0.9500 | 0.05 | ||
DS | 0.0531 | 0.0531 | 0 | 0.8326 | 0.0531 | 0.0081 | ||
Improve DS | 0.7020 | 0.0794 | 0.1746 | 0 | 0.5339 | 0.1746 | 0.0375 | |
DS | 0 | 0.0398 | 0 | 0.0398 | 0.9145 | 0.0060 | ||
Improve DS | 0.6017 | 0 | 0.1520 | 0 | 0.1520 | 0.6636 | 0.0324 | |
DS | 0.1624 | 0 | 0 | 0.3248 | 0.4872 | 0.0256 | ||
Improve DS | 0.9025 | 0.1587 | 0 | 0 | 0.3175 | 0.4762 | 0.0476 | |
DS | 0.0245 | 0.0245 | 0 | 0.3840 | 0.5634 | 0.0036 | ||
Improve DS | 0.9677 | 0.0688 | 0.1707 | 0 | 0.3191 | 0.3932 | 0.0482 |
From Tables 4 and 5, the case cannot be diagnosed and a diagnostic error will occur if the diagnosis is with single evidence. For example,
It shows that the diagnostic capability can be improved by reducing the uncertainty efficiently by fusion reasoning for all the evidences that are not redundant.
The fusion results of the method proposed in this text are not as good as those from D-S evidence theory, according to Tables 4 and 5. Because the former is based on solving the fusion of high conflict evidences (
An integrated diagnostic method with evidence theory and rough set fusion reasoning is put forward and it can overcome the problems of subjectivity, dependency and focus element explosion of traditional evidences. It has a significant theoretical significance and practical value.
This paper proposed a synthetic method by considering the weight factor. The method can solve the synthetic problems of the evidences with high conflict and is effective for synthesising normal evidences.
The application effect of an integrated diagnosis of rough set and evidence theories is proved to be good with diagnosis examples. With the increase of original sample data, the accuracy of diagnosis will improve. The method proposed in this text can be promoted and applied in the fault diagnosis to other devices if the sample data are large enough.
The research of this paper will be further deepened. It will focus on developing artificial intelligence diagnostic system for better application in industrial practice to achieve maximum production benefits in the direction of our subsequent research.