Abstract
This paper presents a novel framework for accelerating real-time system monitoring in model-based diagnosis (MBD). By treating time as a logical context within an Assumption-based Truth Maintenance System (ATMS), qualitative inferences can be shared across different time points. This approach significantly reduces computational overhead, enabling sampling rate improvements from minutes to seconds in complex engineering applications such as ballast water systems.
1. Introduction
Model-based diagnosis utilizes logical models of a system to predict expected behavior. Discrepancies between these predictions and actual sensor data indicate faults. However, the computational cost of re-running complex predictive models at high frequencies is a major bottleneck for real-time monitoring. In many industrial systems, qualitative states (e.g., a valve remaining "open") persist for long durations, yet traditional MBD systems often re-calculate identical inferences at every clock tick. This research addresses this inefficiency by proposing a caching mechanism that allows predictions to be shared across valid time intervals.
2. Methodology
The proposed framework relies on two primary pillars: Temporally Generic Formulae and ATMS Integration.
Temporally Generic Formulae: System descriptions are defined independently of specific time points. For instance, the rule "if a valve is functioning and commanded to close, it will be closed" remains true regardless of the specific time of execution.
ATMS Integration: The system leverages an Assumption-based Truth Maintenance System (ATMS) where time indices are treated as a type of assumption. This allows the system to compute "temporal labels"—sets of time intervals or points where a statement holds true. When the system encounters a state it has previously analyzed, it retrieves the result from the ATMS cache rather than re-deriving it. The authors demonstrate that logical contexts (fault hypotheses) and temporal contexts (time points) are orthogonal, allowing the logic used for multi-context diagnosis to be seamlessly reused for temporal reasoning.
Delayed Consequences: To handle time-lagged effects, such as a tank filling over several minutes after a valve opens, the system employs a "demon" mechanism to trigger future justifications.
3. Experimental Evaluation
The approach was evaluated using a ballast water tank system, a critical engineering application for offshore platforms and crane ships.
The full system comprises 40 tanks. Experiments were conducted on 3-tank and 20-tank subsystems using ten different test vectors—sequences of sensor data and commands—to measure the effectiveness of inference reuse over time.
4. Results and Discussion
The efficiency gains were substantial as the system "learned" and cached qualitative states. For a 3-tank system, the initial prediction required 3.04 seconds, while subsequent cached predictions averaged only 0.21 seconds. In a larger 20-tank configuration, the initial 36.3-second calculation dropped to an average of 3.8 seconds for additional time points.
In a real-world crane ship application, the monitoring sampling rate improved from 2 minutes to 2 seconds. The results indicate that prediction time is directly tied to the number of new justifications the system must learn, rather than the total volume of data processed.
5. Conclusion
While the caching mechanism provides significant speedups, the system exhibits a slow linear increase in runtime over extended periods due to the indefinite accumulation of justifications in the ATMS. Future work suggests implementing a "time window" to archive or discard distant, irrelevant data to maintain constant performance. Furthermore, while qualitative models are highly efficient for initial monitoring, they may be best utilized as a trigger for more detailed quantitative diagnostic sessions when subtle anomalies are detected.
Summary updated