Imagine troubleshooting a complex machine. The traditional approach relies on a "cheat sheet"—a pre-defined library of known failure modes and their associated symptoms. But what happens when the system encounters a novel, completely unexpected fault? The cheat sheet fails because it cannot recognize what it has never been programmed to see.
To solve this, researchers developed Consistency-Based Model-Based Diagnosis (MBD). Instead of modeling how a system fails, MBD models how the system is supposed to work when everything is healthy. By comparing real-world observations against this model of correct behavior, the system can deduce exactly which components are malfunctioning based purely on logical contradictions.
The Core Intuition: The Logic of Contradiction
The fundamental power of consistency-based diagnosis lies in a simple logical inversion. Instead of looking for a matching fault signature, we look for deviations from normal operation.
Every component in a system can be described by a simple rule: if the component is healthy, then given a specific input, it must produce a specific output. Formally, we can express this as:
$$\text{Healthy}(C) \land \text{Input} \implies \text{Output}$$
In an ideal state, we assume every component is healthy ($\forall C, \text{Healthy}(C)$). When we feed real-time sensor data into our system model, we generate expected behaviors. If the actual observed output contradicts the predicted output, a logical conflict emerges.
As Davis et al. demonstrated in their foundational work, Diagnosis Based on Description of Structure and Function, this allows us to detect novel or unforeseen errors. Because we only need a description of the system's structure and intended function, any deviation provides immediate logical evidence that something is wrong, without requiring a pre-existing database of fault symptoms.
[Expected State (Model)] <-- Contradiction! --> [Observed State (Sensors)]
│ │
└───────────────────► Generates ◄──────────────────┘
Conflict Set
When a contradiction occurs, it isolates a conflict set—a group of components that cannot all be healthy simultaneously. In Reasoning about Multiple Faults, de Kleer et al. leveraged this mechanism to build the General Diagnostic Engine (GDE). By treating component health as a set of assumptions, any discrepancy between predicted and actual outputs identifies exactly which assumptions must be dropped to restore logical consistency.
Isolating the Culprits: Minimal Diagnoses
In complex systems, a single failure can cascade, or multiple components might break at the same time. This creates an exponential search space of potential explanations. To handle this efficiently, consistency-based diagnosis focuses on finding minimal diagnoses—the smallest sets of component failures that can fully account for the observed discrepancies.
Poole clarified this paradigm in Normality and Faults in Logic-Based Diagnosis, showing that consistency-based diagnosis works by identifying the smallest set of component failures that allows the remaining system model to coexist with the observations without contradiction. Rather than explicitly predicting symptoms, it simply carves away the assumptions of normality until logic is satisfied. The authors of Diagnosing framed this as a constraint satisfaction problem, demonstrating that leveraging the underlying network structure of the system can drastically accelerate the discovery of these minimal sets.
Once these minimal diagnoses are found, how do we interpret the status of an individual component? Biteus et al. provided a rigorous mathematical framework for this in DETERMINING A COMPONENT'S FAULT STATUS AND THE STATUS’ READINESS:
- Faulty: The component is present in all minimal diagnoses. Its failure is logically necessary to explain the problem.
- Suspected: The component is present in some, but not all, minimal diagnoses. It might be broken, but more information is needed.
- Normal: The component is present in none of the minimal diagnoses. It is completely cleared of blame.
Scaling to the Real World: Qualitative Abstractions
While pure logic works perfectly for discrete systems like digital circuits, real-world applications—like nuclear power plants, automotive engines, or robotics—deal with continuous, noisy, and time-dependent signals. Building exact numerical models for these systems is often impossible or computationally prohibitive.
To bridge this gap, researchers substitute precise numerical values with qualitative models. For example, instead of tracking an exact voltage or temperature, the model might reason about values as simply being High, Normal, or Low, or track their derivatives (Increasing, Stable, Decreasing).
In Troubleshooting: When Modeling Is the Trouble, Dague et al. introduced order-of-magnitude reasoning to isolate faults in analog circuits even when precise parameters are unknown. Similarly, Lackinger et al. developed the DIAMON framework in Integrating Model-Based Monitoring and Diagnosis of Complex Dynamic Systems, which pairs real-time sensor monitoring with qualitative simulations. When the physical system drifts away from the qualitative predictions of the healthy model, the engine immediately flags the minimal set of components that could cause the deviation.
To keep the diagnostic search space manageable in massive systems, researchers also utilize hierarchical abstractions and declarative models:
- Hierarchical Design: Genesereth proposed using multi-level design models in Diagnosis Using Hierarchical Design Models. The system diagnoses faults at a high level (e.g., an entire subsystem) before drilling down into specific sub-components, drastically reducing computational overhead.
- Declarative Software Abstractions: Steinbauer et al. applied this to autonomous robots in Detecting and locating faults in the control software of autonomous mobile robots. Instead of monitoring low-level code execution, they used high-level logic to reason about software component health, simplifying runtime fault isolation and recovery.
Cross-Disciplinary Reach and Modern Extensions
The mathematical elegance of consistency-based diagnosis extends far beyond physical machinery. Because it is fundamentally about resolving contradictions between data and assumptions, it maps cleanly to other domains in computer science.
Database Causality and Repairs
In database theory, an unexpected query result can be viewed as a "fault." Salimi et al. established a formal bridge between these fields in From Causes for Database Queries to Repairs and Model-Based Diagnosis and Back (and their subsequent 2017 work). They proved that finding the minimal set of database tuples responsible for an anomalous query result is mathematically identical to finding the minimal set of faulty components in model-based diagnosis.
Active Fault Identification
Rather than passively waiting for a system to exhibit a contradiction, modern frameworks take an active approach. In Bayesian Diagnosability and Active Fault Identification, Kong et al. combined Bayesian belief updates with active control design. If a fault is subtle or hidden, their system strategically "pokes" the physical system with specific control inputs designed to maximize the visibility of any underlying logical inconsistencies, forcing unmodeled failures to reveal themselves.
Key Foundational Work
A Diagnosis Method of Dynamic System Using the Knowledge on System Description (1983) — contributed 2 concept elements to this direction.
Diagnosis Based on Description of Structure and Function (1982) — contributed 1 concept elements to this direction.
Diagnosis Using Hierarchical Design Models (1982) — contributed 1 concept elements to this direction.