Software combines signal forecasting and prognostic reasoning methods to predict system failures.
This new generalized approach to prognostics can provide an automated early failure prediction of an engineering system or its components, often in time to prevent occurrence of hard failures. This approach has been demonstrated in a proof-of-concept software prototype, shown to accurately predict anomalies in the Mars Explorer Rover’s (MER) power systems using archived and model data. The approach differs from other attempted prognostic solutions in that it can interpret any sensed system trend, and not just specific failure modes with previously developed physicsof- failure models. The software employs an iterative reasoning process that implements (1) methods of forecasting signals represented by streams of sensor, telemetric, and other monitoring data and (2) new artificial intelligence methods for performing prognostic reasoning. This approach affords the following capabilities:
- The ability to predict future performance in a variety of systems;
- The ability to distinguish between normal variations in monitoring data and trends in the data representative of significant deterioration of the system or its components, through correlation and logical reasoning;
- The ability to prognose, relating trends to specific fault modes and failures of specific components that give rise to those fault modes;
- The ability to predict times and likelihoods of failures; and
- The ability to reason through diagnostic models with missing, delayed, contradictory, or intermittent symptoms, all of which are typical in degraded systems prior to failure.
Underlying this approach is the observation that a typical prognostic event is imprecisely known in its early stages. This means some trends are missing or inaccurately predicted until details emerge, and potentially important symptoms must be tracked as the event gradually progresses towards a failure. It is crucial to follow the progression and to produce an unambiguous conclusion as soon as the event can be confirmed. Therefore, it is necessary to reason in an iterative fashion, incorporating monitoring data and additional system knowledge as they are acquired. The iterative process can be summarized as follows:
- A system deviation is detected by monitoring functions. The system is still operating, and no faults have yet been indicated.
- Monitoring data are buffered and sent to a forecasting engine.
- The forecasting engine predicts signal values in the future, and estimates the probability that each signal will cross a predetermined operating threshold, and the time at which this is expected.
- Signals that show evidence of significant trends are grouped according to its estimated time of failure and measures of confidence and consistency.
- The groups generated in step 4 are used as the basis for automatically generating hypothetical scenarios, each of which is a partial match to the current system state estimate. The software creates variations based on the observed trends, gradually eliminating those trends that are unsupported by accumulating data or those found to be of low probability after repeated observations.
- Hypothetical scenarios are evaluated against a predictive diagnostic model. Scenarios containing sets of trending signals that show no causal correlation are rejected, as they represent separate or spurious events rather than a unified prognosis. The scenarios that remain are ranked according to their probability of plausibility, numbers of missing or conflicting symptoms, and consistency over time.
- Surviving hypothetical scenarios are compared against new state information using a possible-mode calculator, by first predicting the expected system state implied by each scenario and then comparing expectations against actual system knowledge as it becomes available. At each step, scenario probabilities are updated and conflicting scenarios discarded.
- The process continues until one or more scenarios are self-consistent and probable enough to justify corrective action. The prognostic reasoner outputs the expected failure mode, the likelihood and expected time of occurrence, and, in case of remaining ambiguity, specific measurements that will determine the exact prognosis.