Automated Diagnosis of Physical Systems

“If anything can go wrong, it will.” (Murphy’s Law) Although deceptive in its simplicity, the above quote has proved to be a profound insight into how things happen. Things are likely to fail no matter how well the system is designed. When a high degree of reliability and safety is desired the effects of these failures must be mitigated, and control must be maintained under all fault scenarios. Faults need to be detected close to their onset so that quick action can be taken by resetting control parameters to compensate for the fault or by reconfiguring the system to minimize the effects of the fault and thus prevent damage. In this paper we provide a brief introduction to variety of automated techniques for diagnosing faults and then discuss in more detail one specific technology called HyDE. WHAT IS AUTOMATED DIAGNOSIS? Fault diagnosis involves the detection of anomalous system behavior and the identification of the cause for the deviant behavior. Automated diagnosis refers to the use of software technologies to assist in diagnosing faults in a system. This is to be contrasted with autonomous diagnosis which deals with software technologies that operate autonomously to detect, isolate and compensate for faults in a system. We would like to introduce some definitions (taken from IFAC Technical Committee SAFEPROCESS [1]) to set the context for the rest of the paper. Fault A fault is an unpermitted deviation of at least one characteristic property or parameter of the system from acceptable, usual or standard conditions. Fault Detection Fault detection is monitoring measured variables to determine if a fault has occurred in the plant. If a fault has occurred, it may be important to determine the time at which the fault occurred. Fault Isolation Fault Isolation is determining the type and location of a fault once it is known that a fault has occurred. It typically follows fault detection but the two processes are often combined for additive faults. Fault Identification Fault Identification is determining the size and time-variant behavior of a fault. It follows fault isolation. Fault may be of many types including: • Plant, actuator, sensor or controller faults • Additive or multiplicative faults • Abrupt or incipient faults • Persistent or intermittent faults Diagnosis is made harder by several factors including and not limited to: • Typical only a few sensors are placed leading to limited observability into system behavior. • The data from sensors are noisy due to inherent properties of the sensors • There are unknown inputs acting on the systems, due to lack of complete knowledge about conditions in which system operates. • The knowledge about how the system functions may be limited. • The effects of faults may be non-local and may take some time to manifest. In the rest of this paper we will introduce a sampling of automated diagnosis techniques that deal with some of the fault characteristics described above. We will briefly describe these techniques and then focus on one specific technique called HyDE. AUTOMATED DIAGNOSIS TECHNIQUES The diagnosis problem has been dealt with in several domains from several angles resulting in a wide variety of diagnosis approaches. Some of these include Expert Systems, Case-based reasoning, Data-driven techniques and Model-based reasoning. Expert System Diagnosis [2,3] Traditionally diagnosis was performed by human troubleshooting experts who built up diagnostic knowledge based on their expertise and experience. The natural extension to this was to encode the diagnostic knowledge in a machine storable structure. These structures took the form of associations between observed symptoms and probable fault occurrences. Tools were built to assist in the creation of these structures and then to use these structures for the diagnosis task. Once these structures are built, users and non-expert operators could troubleshoot the system in case of faults. Some of the more commonly used structures to encode expert diagnostic knowledge are rules and fault trees. A rule describes the action(s) that should be taken if a symptom is observed. A set of rules describing the symptoms of all the possible faults is incorporated into a rule-based reasoning system. The reasoning may use a backward-chaining algorithm which starts at the hypothesis (consequents of rules) and collects and verifies evidence (antecedents of rules) that supports the hypothesis. Alternately forward-chaining may be used where rules whose antecedents match observed symptoms are examined. When several rules match, a chain of rule firings (based on pre-defined rule priority) is used to establish the diagnosis. A fault tree (decision tree) encodes diagnostic knowledge as a sequence of questions that trace a path from the root of the tree to its leaf nodes which represent diagnoses. Advantages of expert systems are: • Diagnosis structure is “certified” by experts and can be trusted to produce “correct” results. • In many cases, deep understanding of the physical properties of the system is either unavailable or too costly to obtain. • The diagnostic reasoning is fast and bounded both temporally and computationally. Disadvantages of expert systems are: • All fault-symptom manifestations have to be encoded for “correct” diagnosis. • Troubleshooting experts may not be available. Even when they are available it takes years of experience for them to gather all the diagnostic knowledge. • The whole process has to be started from scratch for each new application. Case-based Diagnosis [4,5] Case-based Reasoning Systems exploit knowledge about solutions developed for past problems to solve current problems. In this approach, experiences are stored in the form of diagnostic cases. When a new case needs to be diagnosed, stored cases are scanned to find any matches with the new case. The new case is then added to the library of stored cases under an appropriate category. Like rule-based systems, past experience with normal and abnormal behavior of a system are essential to building effective case-based diagnosis systems. In addition, case-based reasoning systems include a learning component which makes possible adaptation of a past solution to fit other, similar situations. This technique is well suited for poorly understood problem areas for which structured data are available to characterize operating scenarios. A case-based reasoning system consists of a case library containing features that describe the problem, outcomes, solutions, methods used and an assessment of their efficacy. A coding mechanism is used to index the case information so that the cases can be organized into meaningful structures, such as clusters, enabling efficient retrieval. Advantages of case-based reasoners are: • Only experience in the form of past solutions is needed rather than a deep understanding of the physical properties of the system. • Diagnosis of cases that have been seen before is very fast. • It can be directly applied to any new application. Disadvantages of case-based reasoners are: • New “cases” cannot be diagnosed. • A lot of prior experience is needed to build a large database of cases. Data-driven Diagnosis [6-10] • Data-driven approaches are based on the assumption that statistical characteristics of monitored data from the system indicate abnormal events in the system. These techniques transform the high-dimensional noisy data into lower-dimensional information for detection and diagnostic decisions allowing the ability to handle highly collinear data of high dimensionality, substantially reduce the dimensionality of the monitoring problem, and compress the data for archiving purposes. • Diagnosis using these approaches typically involves two steps: 1. Learning Step – Prior to diagnosis, data from various operational scenarios of the system is fed to a learning algorithm which computes characteristics of the data sets (in a much lowerdimensional form). A classifier then tries to classify the lower dimensional characteristics into groups with similar properties. The classification may be based on prior knowledge about the actual operational scenarios. 2. Diagnosis Step – The learning algorithm is applied on the data from real-time operation of the system to compute lower-dimensional characteristics. These characteristics are compared against learned groups (from the classifier) to select the closest group as the diagnosis. Some of the approaches that use data-driven techniques are Regression analysis, Principal components analysis, Artificial neural networks, Filters, Harmonic analyzers, Auto and cross-correlation functions, Fast Fourier transform (FFT), Pattern Recognition, Feature Selection, Model Selection, Ensemble Learning, and Support Vector Machines. Advantages of data-driven diagnosis are: • No understanding of system properties is necessary. Only data from operation of system is necessary. • Very high dimensional and noisy data can be handled without any problems. • The diagnostic inference is bounded and can be very fast. • Generic algorithms can be applied on any data set from any application. Disadvantages of data-driven diagnosis are: • A lot of data about various diagnostic scenarios is necessary for proper classification. • The diagnosis results are very sensitive to the data used. Model-based Diagnosis [11-22] Model-Based Diagnosis (MBD) departs from these approaches by using a model of the system configuration and behavior. MBD exploits the analytical redundancy (functional relationships) between the model and the system measurements. In principle, this means that a model runs in parallel to the real process, and discrepancies between the model outputs and real outputs are utilized to detect and isolate faults. Some of the major categories of model-based diagnosis techniques are consistency-based approaches, control-theory based approaches, and stochastic approaches. Consistency-based Diagnosis [11-15] In this approach an abstract model of the system is used for diagnosis. The model may be discrete, discrete-event, or continuous in form and typically includes only information necessary to diagnose faults. The faults are represented as changes in operational modes of components of the system. The model is used to predict the expected behavior of the system under hypothesized conditions for operational modes of the system. The predictions are compared against observations to determine discrepancies. Discrepancies are used to generate conflicts which drive a search process for alternate hypothesis for operation modes. The hypothesized operation modes that predict behavior closest to observations are reported as diagnosis. Some of the technologies that use this approach are GDE/Sherlock, TRANSCEND, and Livingstone/L2. Control-theory based approaches [16-18] Quantitative diagnostic algorithms derived from the control theory community use state estimation and parameter estimation techniques for diagnosis. A mathematical model of the system is used to perform diagnosis as a two step process: 1. Residual Generation Residuals are generated from the model and information about inputs and outputs of the system. Residuals are variables with zero value when the plant operation is nominal and non-zero otherwise, and 2. Decision Making Decision procedure is applied to discriminate non-zero residuals that are result of modeling errors, unknown inputs, measurement errors and those which reflect abnormal behavior. The residual generator enhances the raw residuals in one of the following two ways: 1. In response to a fault, only a specific set of residual vector elements become non-zero (structured residuals), or 2. In response to a fault, the residual vector lies in a specific direction (fixed directional residuals). The decision maker takes these enhanced residuals and evaluates them to determine the fault. In the case of structured residuals, this involves mapping the specific set of non-zero residuals to faults. Some of the approaches that use this method are structured parity equations, structured residuals from state equations, diagnostic observer design with direct eigen-structure assignment, and unknown input observer in Kronecker canonical form. In the case of the fixed directional residuals, the decision maker finds pre-defined fault residual direction that is closest the direction of the actual residuals. Some of the approaches that use this method are detection filter design by eigen-structure assignment, fixed direction residuals from parity equations, and matched filters. Stochastic Approaches [19-21] In recent years there has been a push towards the use of probabilistic approaches for diagnosis. This is motivated by the fact that a lot of uncertainty exists in the diagnosis process. The sources of uncertainty may be the models, the sensors, the environment, the diagnosis algorithms etc. These techniques maintain a belief state (a set of possible diagnoses ranked by probabilities) about the system and update the beliefs using probabilistic update mechanisms. Some examples include Bayes Nets, Kalman Filters and Particle Filters. Advantages of model-based diagnosis are: • Model libraries can be re-used. • Reasoning algorithms remain the same for all applications. Disadvantages of model-based diagnosis are: • Someone has to build models. • Reasoning time is not bounded and my take very long.