Introduction
In the mid-1980s, the field of diagnostic expert systems faced a significant architectural divide. Developers typically chose between rule-based systems, which were easy to implement but difficult to maintain at scale, and deep causal models, which simulated first principles but suffered from high computational costs. This paper introduces TEST (Troubleshooting Expert System Tool), a specialized application shell that occupies a "weak" causal modeling niche. TEST provides a structured framework for building diagnostic systems for complex machinery by combining the clarity of schematic hierarchies with the flexibility of heuristic rules.
The Knowledge Architecture: Failure-Mode Hierarchies
At the core of TEST is a semantic network of "failure modes" organized into a "due-to" hierarchy. Unlike flat rule-based systems, TEST represents knowledge in a way that mirrors an engineer's mental model of a machine. A failure mode represents a specific state of malfunction, which is linked to its potential causes through directed relationships.
As shown in Figure 1, a high-level symptom like a "Raster Problem" can be decomposed into specific modes such as "Missing Raster" or "Bent Raster." These, in turn, point to underlying causes like "LV Power Supply" failure or "Hum in LV Power Supply." This hierarchical organization allows the system to narrow down faults systematically.
The Hybrid Problem Solver and Dynamic Rules
TEST employs a domain-independent engine that performs a recursive, depth-first search through the failure-mode tree. However, a simple tree search is often insufficient for real-world troubleshooting. To address this, TEST integrates heuristic rules that can dynamically modify the search space at runtime.
As illustrated in Figure 2, each failure mode acts as an aggregate for related information, including documentation, tests (T), and repairs (RP). Crucially, "Rule" objects are attached to these nodes to provide strategic control:
- Causal-modeling rules prune irrelevant branches based on context (e.g., ignoring electrical faults if the battery is confirmed functional).
- Procedural rules reorder the search based on the cost of testing or the historical likelihood of a specific failure.
- Relevance rules filter the model to match specific machine configurations or manufacturing runs.
Key Advantages and Results
The "weak" causal model approach offers several advantages over contemporary methods:
- Maintainability: By clustering tests and repairs around failure modes (Figure 2), the system avoids the "spaghetti" logic of large rule sets. Updates to a specific component only require changes to its local aggregate.
- Efficiency: Heuristic rules allow the system to bypass exhaustive simulations, providing faster diagnostic conclusions than pure causal-model systems.
- Knowledge Acquisition: The use of domain-familiar concepts—failure modes, tests, and repairs—enables design engineers and technicians to contribute directly to the knowledge base, reducing the reliance on specialized knowledge engineers.
Conclusion
TEST successfully bridges the gap between rigid heuristics and complex physical simulations. While it still requires manual knowledge engineering to map out hierarchies, its structured approach significantly improves the development and maintenance of diagnostic systems for industrial applications. Future work focuses on automating this knowledge acquisition process through specialized workbenches to further lower the barrier for expert participation.
Summary updated