《A survey of artificial intelligence for prognostics.docx》由会员分享，可在线阅读，更多相关《A survey of artificial intelligence for prognostics.docx（43页珍藏版）》请在三一文库上搜索。
1、A survey of artificial intelligence for prognosticsA Survey of Artificial Intelligence for PrognosticsMark Schwabacher and Kai GoebelNASA Ames Research CenterMS 269-3Moffett Field, CA 94035mark.a.schwabacherhttp:/ kai.f.goebelhttp:/ Systems Health Management includes as keyelements fault detection,
2、fault diagnostics, and failure prognostics. Whereas fault detection and diagnostics havebeen the subject of considerable emphasis in the ArtificialIntelligence (AI) community in the past, prognostics has notenjoyed the same attention. The reason for this lack ofattention is in part because prognosti
3、cs as a discipline hasonly recently been recognized as a game-changing technology that can push the boundary of systems healthmanagement. This paper provides a survey of AI techniquesapplied to prognostics. The paper is an update to our previously published survey of data-driven prognostics.Introduc
4、tionNASA is currently planning long-duration human space exploration missions to the Moon and Mars. Reliability of the spacecraft will be extremely important for these missions, since they will be away from the Earth for months or years at a time. An important contributor to that reliability will be
5、 an on-board Integrated Systems Health Management (ISHM) system. ISHM can provide two advantages. First, it can increase safety, by detecting problems, quickly diagnosing them, and assessing remaining life before they become serious, so that controllers can respond rapidly and prevent major failures
6、. Second, it can reduce costs by enabling corrective action to be scheduled more efficiently. Corrective action such as maintenance scheduling is most important for reusable systems, such as aircraft or the Space Shuttle, but even expendable piloted spacecraft, such as Apollo or Soyuz, have had some
7、 maintenance actions that can be performed by the astronauts during a mission. Future air and space vehicle may also benefit from robotic or autonomic maintenance.An ISHM system takes as input sensor values and the command stream, and ideally performs fault detection (detecting that something is wro
8、ng), fault isolation (determining the location of the fault), fault identification (determining what is wrong; that is, determining the fault mode), and fault prognostics (determining when a failure will occur based conditionally on anticipated future usage).This material is declared a work of the U
9、.S. Government and is notsubject to copyright protection in the United States.We define diagnostics to include fault isolation and fault identification, so that full diagnostics requires determining the specific fault mode, rather than just reporting which sensor has an unusual value. We define prog
10、nostics to be detecting the precursors of a failure, and predicting how much time remains before a likely failure. Prognostics is the most difficult of these tasks. One must be able to detect faults before one can diagnose them. Similarly, one must be able to diagnose faults before one can perform p
11、rognostics. In addition to fault detection, diagnostics, and prognostics, ISHM also includes support for deciding what actions to take in response to a failure or a failure precursor. These actions can include reconfiguration of redundant or non-redundant hardware, maintenance actions performed by t
12、he crew, maintenance actions performed on the ground (for reusable vehicles), recalibration of sensor values or commanded values to compensate for degraded hardware, and mission replanning to accommodate degraded systems. The field of ISHM includes sensor development and optimization of sensor place
13、ment (Zhang, 2021), but this survey focuses only on the algorithms used for fault detection, diagnostics, and (especially) prognostics.A simple form of prognostics, known as a life usage model, is widely in use. This method is applicable to components that have been mass produced. It gathers statist
14、ical information about the amount of time that a component lasts before failure, and uses these statistics collected from a large sample of components to make remaining life predictions for individual components. These predictions are based solely on the passage of time and/or measures of usage of t
15、he system or component. For example, for a timing belt on an automobile, the manufacturer may recommend that the belt be replaced after five years or 60,000 miles. The recommendations from these life usage models are not based on any measured characteristics of the individual component. This survey
16、is primarily concerned with condition-based prognostic methods, i.e., methods that take advantage of measured characteristics of a particular system or component of interest in order to make predictions, and not on life usage models.Frameworks that illustrate the use of computational intelligence al
17、gorithms within ISHM have been discussed in the literature. For example, Bonissone (2021) defines this framework in the cross product of the ISHM decisions time horizon and domain knowledge type and structure.Within this framework, the full range of ISHM functions are defined. In contrast, the prese
18、nt paper classifies different types of ISHM algorithms in a taxonomy shown in Figure 1. With the strong caveat that the boundaries between the different classes are not crisp, we distinguish here between algorithms that are model-based and algorithms that are data-driven. We use a narrow definition
19、of the term “model-based” wherein algorithms encode human knowledge via a (more or less) hand-coded representation of the system. Such a model can be either physics-based (encapsulating first principles knowledge using systems of differential equations, for example), or based on techniques from Arti
20、ficial Intelligence (AI). Since AI is notoriously ill-defined, we adopt for the purpose of this paper a definition (in contrast to the more strict Turing test) that subsumes elements of learning and the ability to deal with ambiguity, including elements from soft computing, computational intelligenc
21、e, machine learning, etc. Model-based AI techniques include rule-based expert systems such as SHINE (James & Atkinson, 1990) and G2 (Gensym, 2021). Other examples of model-based AI techniques are finite-state machines, as in Livingstone (Williams & Nayak, 1996; Kurien & Nayak, 2021) and Qualitative
22、Reasoning (Weld & de Kleer, 1989), where a hand-coded model uses qualitative, rather than numerical, variables to describe the physics of the system.Data-driven approaches automatically fit a model of system behavior to historical data, rather than hand-coding a model. Data-driven approaches can eit
23、her use “conventional” numerical algorithms, such as linear regression or Kalman filters, or they can use algorithms from the machine learning and data mining AI communities, such as neural networks, decision trees, and support vector machines. The term “machine learning” isill-defined as well. We a
24、dopt here a definition of machine learning that imposes a degree of complexity on the learning aspect. That definition excludes linear regression and (marginally) Kalman filters, but it includes decision trees, case-based reasoning, clustering, and neural networks, for example.In Table 1, we have co
25、nstructed a matrix in which the rows represent the four types of algorithms from Figure 1, and the columns represent the three ISHM problems that we identified earlier in this section (fault detection, diagnostics, and prognostics). In each cell, we provide a representative (not exclusive) example o
26、f a method that uses the specified type of algorithm to solve the specified problem. Note that two cells are empty. There is little evidence of current activity in applying purely physics-based algorithms to diagnostics. This is not to say that it has not been done or could not be done. Indeed, one
27、could imagine a diagnostic system that has a physics-based model of the nominal operation of a system and physics-based models of several fault modes. When the sensor data fails to match the nominal model, the system would simulate several candidate failure modes in parallel, and compare the simulat
28、ed data from each failure mode with the sensor data. A match would result in a diagnosis. However, employing this approach to diagnostics may not be the most efficient way to accomplish diagnostics. One could also imagine a physics-based model augmented with if-then rules coded in a conventional pro
29、gramming language to perform diagnostics. Such a system would be considered a hybrid of a physics-based model and a very simple expert system.The second empty entry in Table 1 is for AI-model-based prognostics for which no specific references are cited here. Again, one could of course imagine a rule
30、-based Fault detection DiagnosticsPrognosticsPhysics-based System Theory Damage propagation models AI-model-basedExpert systems Finite state machinesConventional numerical Linear regression Logistic regression Kalman filtersMachine learningClustering Decision treesNeural networksFigure 1: Taxonomy o
31、f ISHM algorithms. Examples of each of the four types are shown at the bottom of the figure. Table 1: An example method for each pair of ISHM problem (columns) and algorithm type (rows)expert system being used for prognostics. For example, such a system might employ a set of rules that specify that
32、when certain sensor values first exceed a particular set of thresholds, a component has a given amount of remaining useful life. One could argue that rule-based systems are found in fuzzy logic systems. However, most of the fuzzy logic systems that are used for prognostics are encapsulated in a lear
33、ning paradigm so that the overall system looks more like a machine learning system than an expert system.Certainly, the work for which ample references are available in AI for prognosticsthe subject of this symposium and of this surveyis in the domain of machine-learning.This survey also includes hy
34、brid methods that combine the machine learning approaches with one or more of the other approaches. For all methods, we are interested in the full spectrum of technology readiness levels, from basic research to deployed systems.The next three sections are each devoted to one of the AI-related approa
35、ches described above. Since many systems use a combination of these approaches, they could fit into more than one of these sections. We have chosen, however, to include each system in the one section in which we feel it best fits.Most ISHM systems devote a large amount of effort to pre-processing th
36、e data using various algorithms including signal processing algorithms in order to extract the features that can be used for fault detection, diagnostics, and prognostics. While pre-processing is extremely important to the success of an ISHM system, it is not the focus of this study.We previously pu
37、blished a survey of data-driven prognostics in 2021 (Schwabacher, 2021). The present paper briefly summarizes that survey paper, and adds new work that has been published in the past two years. It also focuses more on work that uses the AI approach. Other recent survey papers have focused on the app
38、lication of prognostics and other parts of ISHM to particular applications, such as heating, ventilation, and air conditioning (Katipamula & Brambley, 2021a; Katipamula & Brambley, 2021b), electronics (Vichare & Pecht, 2021), manufacturing (Goh et al., 2021), and wheeled mobile robots (Luo et al., 2
39、021). Patterson-Hine, et al. (2021) presented a survey of diagnostic techniques for ISHM.Data-Driven PrognosticsOne of the most popular machine-learning approaches to prognostics is to use artificial neural networks to model the system (Bonissone & Goebel, 2021; Byington et al., 2021b; Byington et a
40、l., 2021c; Byington et al., 2021; Chinnam & Baruah, 2021; Chinnam & Mohan, 2021; Gebraeel et al., 2021; Goebel et al., 2021; Kallappa & Hailu, 2021; Khawaja et al., 2021; Kozlowski et al., 2021; Lavretsky & Chidambaram, 2021; Lee, 1996; Naipei et al., 2021; Roemer et al., 2021a; Shao & Nezu, 2021; S
41、harda, 1994; Stone & Jamshidid, 2021; Studer & Masulli, 1996; Watson & Byington, 2021; Weigend & Gershenfeld, 1993; Werbos, 1988). Artificial neural networks are a type of (typically non-linear) model that establishes a set of interconnected functional relationships between input stimuli and desired
42、 output where the parameters of the functional relationship need to be adjusted for optimal performance. This adjustment is typically accomplished by exposing the network to a set of examples, observing the response of the network, and readjusting the parameters to minimize the error. Several techni
43、ques can be employed to adjust (or “train”) these parameters, including a range of gradient descent techniques and optimization techniques (Bishop, 1995).Another machine-learning approach is anomaly detection algorithms (also known as novelty detection or outlier detection algorithms). These algorit
44、hms learn a model of the nominal behavior of the system, and then notice when new sensor data fail to match the model, indicating an anomaly that could be a failure precursor (Bock et al., 2021; Clifton, 2021; Volponi, 2021). Other machine-learning techniques used for prognostics include reinforceme
45、nt learning (Bock et al., 2021; Kalgren & Byington, 2021), classification (Watson & Byington, 2021), clustering (Byington et al., 2021), and Bayesian methods (Amin et al., 2021; Gebraeel, 2021).Data mining algorithms seek to discover hidden patternsin large data sets (Hand & Smyth, 2021). Some autho
46、rs have addressed the use of data mining algorithms to assemble and process the data needed to train data-driven prognostic algorithms (Reichard et al., 2021b; Sandborn et al., 2021).Another popular AI technique that is used for prognostics is fuzzy logic (Amin et al., 2021; Bonissone & Goebel, 2021
47、; Byington et al., 2021b; Byington et al., 2021c; Byington et al., 2021; Chinnam & Baruah, 2021; Frelicot, 1996; Kozlowski et al., 2021; Studer & Masulli, 1996; Volponi, 2021). Fuzzy logic provides a language (with syntax and local semantics) into which one can translate qualitative knowledge about
48、the problem to be solved. In particular, fuzzy logic allows the use of linguistic variables to model dynamic systems. These variables take fuzzy values that are characterized by a sentence and a membership function. The meaning of a linguistic variable may be interpreted as an elastic constraint on
49、its value. These constraints are propagated by fuzzy inference operations. The resulting reasoning mechanism has powerful interpolation properties that in turn give fuzzy logic a remarkable robustness with respectto variations in the systems parameters, disturbances, etc. When applied to prognostics, fuzzy logic is typically applied in conjunction with a machine lea