Research Article
Latest Progress of Research on Fault Diagnosis Based on Information Fusion
Wuhan Institute of Technology, School of Electronic and Mechanical Engineering, Wuhan 430074, China
With the rapid development of science and technology, especially, the development and application of control technology, the security and reliability of control system become more and more important. Fault diagnosis is one of the key technologies that can assure the system security and reliability. Seen from its history of fault diagnosis, for example, since 1970s, people have begun to pay attention to the fault diagnosis of control system, that is, fault diagnosis is a marginal and crossed subject based on neoteric mathematics, computer theory and technology, control theory, signal processing, simulation technology and reliability theory and so on. For example, in 1971, Dr. Beard from MIT at first put forward an idea that analytic redundancy was used to replace hardware redundancy and system was kept stable through self-organization and found out the system fault by comparing output results, which signed the beginning of this new technology. Willsky published the first review paper about fault diagnosis in magazine Automatica in 1976. Himmelblau published the first composition in 1978. During this period, some simple algorithms were proposed, such as detecting filter, generalized likelihood and maximum likelihood. The theory of fault diagnosis was not ripe and the cases of application were also less at this time. Fault diagnosis come through a vigorous developing phase in 1980s. During this period, many new theories were proposed; moreover, there were many application cases. However, its application range is very confined, which is mainly centralized in spaceflight, watercraft, power factory fields etc. Fault diagnosis methods mainly included observer/filter or system identification and parameter estimation. After 1990s, people have a deep cognition on fault diagnosis of control system. During this period, all kinds of methods mutually penetrated and integrated and its application field was also extended greatly. Whereas, observer/filter and parameter still is the main method, at the same time, neural network, fuzzy theory and their combination method also obtain rapid development in fault diagnosis field and there were also more researches on nonlinear fault diagnosis system.
Though fault diagnosis has made great progress in the pass, there is a great defect that the diagnosed result might be incorrect by diagnosing fault from single sensor or single information source, which might be polluted for some uncertainty. Therefore Intelligent fault diagnosis will make more progress with the help of multi-sensor fusion technology, which is regarded as the most popular information processing tool.
THE PRINCIPLE OF FAULT DIAGNOSIS BASED ON INFORMATION FUSION
Multi-sensor information fusion is a new technology developed from the end of 1970s, which based on multiple discipline process data by applying computer technology, sensor technology, mathematics and signal processing. It was applied to martial field, i.e., enemy-ourselves identification, flight path tracking, object orientation and so on, which has been a ripe theory and method and has shown its special advantage. With the rapid development of science and technology, especially, within recent more than decade years, information fusion technology has been transformed from martial application to civil one, moreover its application fields also was expanding gradually.
Information fusion is applied to fault diagnosis, in order to synthetically utilize all kinds of information, improver the correctness rate of diagnosis. There are three reasons to explain this problem: the first one is that multi-sensors produce many signals from different channels. Secondly, same signal can own different characteristics. Thirdly, diagnosis results owning errors can be reached through different diagnosis approach. Therefore, the principle of fault diagnosis based on information fusion is to acquire symptom by analyzing the characteristic signs from multi-sensors and associate, combine and select data from multiple levels and then obtain more reliable acquaintance on fault information from diagnosis object and situation estimation to potential tendency of fault development.
FAULT DIAGNOSIS METHODS BASED ON INFORMATION FUSION
DST- based fault diagnosis method: DST (Dempster-Shafer Theory) finds out the main reason of events occurrence, according to the results of events occurrence. For multi-characteristics diagnosis problems owning subjective uncertainty, DST is a valid way in fusing subjective uncertainty. In the course of fault diagnosis of equipment, when fault occurs, some symptoms relative to it also accompany to turn up. For each symptom, different faults maybe might have certain probability of occurrence. DST expresses the probability of identified object mode with basic belief assignment, which can represent the believable degree to hypothesis of each object mode. And then, through testing the diagnosis object with multiple sensors, the basic belief assignments can be got and then combine them with Dempster combinational rule and get the fused fault diagnosis information, which supplies with integrative and correct evidence for further decision of object mode recognition (Ye et al., 2006). This kind of procedure of fusion diagnosis (Fan and Ming, 2006) is shown in Fig. 1.
Of course, DST has many difficulties in dealing with high conflict because of its inherent nature. Many experts apply improved DST to fault diagnosis (Xue, 2006; Fan and Zuo, 2007). For example, Wang and Wang (2000) introduced fuzzy set theory based on DST and were sure of the evidences and then carried out data fusion with Dempster combinational rule. Finally they also give an application example in fault diagnosis of hydraulic pressure pump, where some experimental results are analyzed and testified to be the validity of this method.
Fuzzy NN-based fault diagnosis method: Fuzzy theory is a kind of self-organization network model, which adopts unsupervised and completive learning rule, doesn`t depend on samples basically and can correctly and rapidly identify the exception. However, this model finishes the mode classification task through clustering analysis and can`t carry out the evaluation of fault ponderance and forecast of developing tendency. Neural network owns powerful nonlinear approach ability and can identify fault mode and also carry out the evaluation of fault ponderance and forecast of developing tendency. However, neural network has low ability in identifying the exception, doesn`t have increment learning ability. So combine fuzzy theory with neural network, i.e., FNN method, which integrates supervision algorithm with un-supervision one and take the fuzzy value regards as input, so that overcome each shortcomings of fuzzy theory and neural network, keep their advantages and optimize the diagnosis procedure.
Fig. 1: | The procedure of DST-based fusion diagnosis |
Li et al. (2003) carried out the fault diagnosis of aluminum plane flaw by applying FNN. Based on the vibration model slope change rate of girder structure, crack diagnosis index of plane structure vibration model slope change rate in two directions was taken. By comparing the diagnosis result between FNN and FNN fusion method, it testified that FNN fusion diagnosis method had a high precision in detection, localization and property depiction about the crack and also testified the diagnosis precision of crack diagnosis index of vibration model slope change rate is the highest among five crack diagnosis index. Zhang et al. (2003) also integrated fuzzy theory with neural network and carried out the proper fusion procedure in diagnosis decision level, which improved the reliability of diagnosis and supplied with a new idea for intelligent diagnosis of mechanical fault. Some authors proposed a multi-level fault diagnosis system (Fig. 2), which solved the difficulty in training network, when fuzzy rules presented exponential increment tendency, with the input variables increasing. They also studied the specific application of fault diagnosis of scale catalyzing machines in certain refinery. In addition, Li et al. (2003, 2006) applied FNN fusion method to revolver (i.e., diesel engine).
Fig. 2: | Scheme of multi-level fuzzy NN fault diagnosis |
Agent-based fault diagnosis method: Multi-agent system is very popular in distributed artificial intelligence, which refers to multi-agent coordination, mutual service and completes a task together. At the same time, each agent is also independent, that is, its own object and behavior will be not restricted from other agents and they solve the contradiction and conflict through competition and negotiation. Diagnosis agent is a nesting multi-agent organization model, which introduces multi-sensor fusion arithmetic to agent design. It is the key to decompose and control the diagnosis task and cooperate and harmonize between multi-agent, when applying multi-agent theory in fault diagnosis system. This method can not only improve the diagnosis correctness rate, but also strengthen the adaptability to environment and find and dig knowledge in system running process and improve learning ability and then make diagnosis system performance self-perfection (Fig. 3). Presently there turn up many multi-agent fault diagnosis methods based on different theories, i.e., expert system agent, NN agent, fuzzy agent, genetic agent etc. Ren et al. (2001) applied multi-agent technology, established fuzzy multi-agent model and obtained primary experimental results.
Symbol reasoning-based fault diagnosis method: When considering the intuitionistic phenomena in occurring faults, by adopting certain search and conflict-solving policies, find out all preconditions satisfied with potential object, adjust the reasoning policies and improve the reasoning efficacies. For example, some ones proposed a kind of eclectic facility fault detection and isolation arithmetic based on symbol reasoning. This method fused all kinds of digital and symbolic information by applying fuzzy logic reasoning and decision technology, so that improved the performance of fault isolation (Lu and Huang, 2006). Yang et al. (1997) from Shanghai Jiaotong University introduced fault diagnosis methods based on symbol reasoning and artificial neural network respectively and at the same time also analyzes their limitations. He also proposed multiple fault diagnosis method integrating symbol reasoning with artificial neural network, which could realize the fault diagnosis, because he combined their advantage between both of them.
Fig. 3: | Structural model based on fuzzy multi-agent cooperation |
Wavelet analysis-based fault diagnosis method: Wavelet analysis as signal analysis tool has very good character. Its main idea is to project signals to subspace, which is composed of a set of normal wavelet functions. This is helpful to expand signals according to different sizes, so that acquire signal traits from different frequencies and at the same time save time-domain traits in different sizes. Wavelet analysis decomposes the low frequency signals and keep signals unchanged in high frequency, moreover, its frequency resolving ability is proportional to 2j and can observe the signals in low frequency in detail. Dong et al. (2008) proposed a fault diagnosis on diesel engine based on wavelet K-L information value, acquired the character of jar vibration signals, where wavelet decomposition was regarded as the preprocessing step and computed K-L information value from acquired signals. The procedure of this method was to regard character sequence relative to decomposition on referred and checked samples as the corresponding sequence to establish different sequence AR model and computed the K-L information value. And then justified system state according to K-L value, that is, smaller its value was, nearer the referred and checked states was. Otherwise, the other way round. Chen et al. (2007), because there were more or less noises in vibration signals of mechanical facility, which made weak fault information acquisition difficult, then they proposed a reducing noise method based on duality tree wavelet threshed and applied it to mechanical fault diagnosis.
In addition, someone integrated wavelet with NN, so that avoided from the blindness in NN structure design and nonlinear optimization (i.e., local optimization). Someone also proposed a fault diagnosis method integrating wavelet with least square support vector, which carried out wavelet decomposition on fault signals power spectrum and simplified the acquisition of fault character vectors and then carried out fault diagnosis with least square support vector. Mao et al. (2007) replace the conventional gradient descent method with PSO method to optimize each parameter in wavelet network.
Rough set based fault diagnosis method: Pawlak proposed rough set theory in 1982, which is a new method used in dealing with incomplete, imprecise knowledge. That is, when data is inadequate, incomplete, noisy or needs to be solved through fault-tolerance and has difficulty in establishing model, rough set method can overcome the difficulty in solving the complex system problem when using conventional methods and solve the problem of acquiring diagnosis sign and automatically pick up knowledge. However, when there are some differences of diagnosis cost, classical rough set method can`t take satisfying results because of no considering prior experience. Therefore, they introduced the subjective weight of sample object, in order to propose the weighted rough set method (Liu et al., 2007).
Of course, there also turn up many integrated methods by combining other theories (i.e., Fuzzy, Neural Network, Wavelet, etc.) with Rough set theory, in order to learn from others`s strong points to offset one`s weakness.
FRAMEWORK OF FAULT DIAGNOSIS
Layered framework of fault diagnosis: Layered framework of fault diagnosis sets out from special diagnosis object, optimize the status signal and parameter at first and then pick up features after dealing with them and then diagnose synthetically through layered information fusion methods. Because this framework only considers whether the information is from the same sensor, the organic relation among information is cut off, so that some hidden features are lost when combining information (Fig. 4).
Fig. 4: | Layered framework of fault diagnosis |
Fig. 5: | Method-synthesized framework of fault diagnosis |
Fig. 6: | Independently layered fusion framework |
Method-synthesized framework of fault diagnosis: This kind of framework integrates some diagnosis methods together, in order to obtain a more precise result than single method. Of course, if certain diagnosis method is very perfect, then the framework will lose its significance (Fig. 5).
Layered fusion framework of fault diagnosis: The layer of information fusion is marked off data, feature and decision layer. Fan and Ming (2006) proposed a framework in Fig. 6, where the fusion in every layer is independent, so that all information isn`t considered synthetically and only a layer result is obtained. Therefore, this kind of framework lacks of global diagnosis results and has no capacity in consultation.
In order to overcome the shortcoming of above framework and make the framework more general, someone proposed a kind of fusion framework in Fig. 7, where the fusion in the data layer includes direct data from multi-sensor and some processes of necessary pretreatment and analysis, i.e., signal filter. Feature layer includes the valid justification on the result of data layer fusion, which corresponds to all kinds of fault diagnosis methods. Decision layer corresponds to all kinds of tactics of dealing with fault diagnosis. This three-layer framework can satisfy with all kinds of needs of monitoring, diagnosing and isolating faults, so that its application is very wide. Of course, with the rapid development of fault diagnosis and information fusion, the fusion framework will be more perfect.
Fig. 7: | Layered fusion framework |
In this research, through the review of latest progress of research on fault diagnosis principle, method and framework based on information fusion, we know that information fusion technology is more and more important and even become a firm foundation in fault diagnosis field.
However, Information fusion theory has no unified theoretic framework and generalized fusion model and arithmetic presently. Because the output forms, environment descriptions of all kinds of datum are different, the approach of transforming all kinds of different data forms into same data ones needs to be further studied. Also because artificial intelligence makes system itself more flexible and comprehensible and also can solve more complex problems, the research on fault diagnosis based on more and more perfect intelligent information fusion.