Abstract
In complex industrial environments, the reliability of sensor data is critical for accurate diagnosis. Traditional hardware redundancy is often insufficient or prohibitively expensive. This paper presents a "diagnostic expectation-based validation" method that utilizes hierarchical classification to identify faulty sensor data. By incorporating a "suspicious data" flag within diagnostic knowledge groups, the system can maintain diagnostic integrity even when faced with conflicting or erroneous sensor inputs.
Introduction
Reliable monitoring in complex systems, such as nuclear power plants or chemical refineries, is often hampered by the inherent unreliability of sensors. Sensors are prone to failure, drift, and conflicting readings. While engineering traditionally addresses this through hardware redundancy—installing multiple sensors for the same variable—this approach is costly and susceptible to common-mode failures.
Human experts, however, exhibit a high tolerance for "bad data" by employing mental models of system behavior. If multiple indicators suggest a specific failure while a single gauge reports normal conditions, an expert will likely suspect a sensor malfunction rather than dismissing the diagnosis. This research aims to formalize this "common sense" approach within an expert system framework.
Methodology: Hierarchical Classification and Expectation-Based Validation
The proposed approach is built upon Hierarchical Classification, where diagnostic knowledge is structured as a tree of increasingly specific hypotheses.
Modified Knowledge Groups
The authors utilized the Conceptual Structures Representation Language (CSRL) to define "Knowledge Groups" at each node of the hierarchy. Unlike traditional systems that strictly match data patterns, these groups were modified to include a "Suspicious Data" column. This allows the system to establish a diagnosis based on a preponderance of evidence while simultaneously flagging contradictory data points as suspicious.
The Establish-Refine Strategy
The system employs an "Establish-Refine" strategy, moving top-down through the hierarchy. If a high-level node (e.g., Pressure System Failure) is established despite a conflicting sensor, this "expectation" of failure is passed to child nodes. This context allows the system to evaluate suspicious sensors more effectively as it narrows down the specific malfunction.
The complete validation cycle involves:
- Traversing the hierarchy to identify questionable values.
- Attempting to resolve these values through cross-variable relationships (e.g., physical laws like $PV=nRT$).
- Updating the system's database and re-evaluating the diagnosis if necessary.
Insights and Results
The core of this method is Diagnostic Redundancy. Redundancy is treated as an informational property derived from the relationships between different physical variables (temperature, pressure, flow), providing a "software-based" alternative to hardware redundancy.
The system was tested using a simplified reactor vessel model and simulations of a Boiling Water Nuclear Power Plant (BWR). Results demonstrated that the hierarchical system could successfully reach correct diagnoses, such as a Relief Valve Failure, even when primary sensors reported incorrect "Normal" values. This high-level validation acts as a robust "second line of defense" alongside traditional hardware checks.
Limitations and Future Directions
While effective, the system relies on "compiled knowledge"—pre-defined hierarchies of known malfunctions. It may struggle with novel or un-modeled failures. Additionally, re-running the entire hierarchy after flagging data can be computationally expensive; future work suggests using backtracking to re-evaluate only affected nodes. The authors also highlight the potential for "deep models" based on first-principles physics to provide even more accurate expectations for data validation.
Summary updated