Title             :  Service Failure Diagnostics and Prognostics: Probabilistic Event-Driven System Field Failure Prognostics

Sponsor        : General Electric Health Care

Collaborators: GE

Field failures of healthcare equipment are critically important during service phase of product life-cycle. Field failures are particularly important for expensive and complex products such as CT, MRI and vascular systems, for which the manufacturer is responsible for maintaining product functionality and minimizing breakdowns in field.
Current advances in data acquisition and storage technology has enabled monitoring of products field performance in the form of event logs wherein all events occurring during system operations are recorded. Events are generated automatically in response to various user and system requests and reflect changes in system state. All the generated events contain information pertaining to system operations which can be further used to predict system field failures.
However, current discrete-event approaches for system failure diagnostics are based on the designed templates which represent all system conditions and require that the design knowledge of the system is complete, i.e., represent all possible system states. These diagnostic approaches are based on matching the current system state with designed condition templates. The template-based diagnostic approaches are successfully applied to diagnose appliances and manufacturing systems, however, for complex equipment with many sub-systems such as CT and vascular devices, it is not computationally tractable to determine and represent all system states based on design knowledge. Furthermore, the template-based methods use deterministic templates which cannot handle large variation in occurrence of events in the system event log.
In this paper we propose a data-based approach for field failure diagnosis of complex systems. We use historical event logs to obtain fault patterns associated with various failure modes. The pattern extraction focuses on extracting infrequent patterns that are strongly associated with failures. The extracted fault patterns are used to construct a fault library.
The fault library is used in conjunction with event log of system suffering failure to identify a match between fault patterns and event log. The event log could have variations in terms of type of events occurring in the event log (not all events in the pattern occur in the event log). Further there is also variation in occurrence of events which causes the events identified in the pattern to be non-contiguous in the event log and separated by a non-deterministic gap of benign events in the event log. In this paper we propose a dynamic programming based matching approach to determine optimum alignment of the patterns with the event log and determine maximum attainable similarity between the patterns and the event logs. Further the statistical significance of the optimal alignment is determined to identifythe most probable pattern match for diagnosis.
The developed methodology is illustrated using a case study from CT system in which pattern extraction and diagnosis methodologies are applied to field failure causing disruptions in patient scan.

• Current discrete event based prediction research focuses on extracting patterns related to target events and the occurrence of the pattern is used to predict the target event. This provides prediction after all the event elements of the pattern has occurred. This has been applied in fields of market basket analysis, computer network intrusion detection etc. However in case system failure prediction additional factors need to be considered such as determining optimum prediction decision for monitoring multiple faults simultaneously, incorporating critical nature of different faults and finding optimum prediction point considering both increased accuracy as more events are observed and the required prediction horizon for the failure. These factors ensure that there is sufficient time for service intervention to prevent failure and also have optimum prediction error.

• We propose a Bayesian sequential prediction scheme that updates system failure state probabilities as new events are generated, determines optimum prediction decision after event is generated by determining the utility of each prediction decision option. Finally determines the optimum failure prediction point for the failure state monitored by considering both the expected value of future events in increasing prediction decision utility and deterioration of the system condition as more events are observed.

• The developed methodology is illustrated by using a case study of vascular scanning systems. The case study shows the pattern extracted for given failure considered and on-line monitoring of system for optimum failure prediction.