A Probabilistic Approach to Hierarchical Model-based Diagnosis

Sampath SrinivasUAI 1994

TL;DR

This paper presents a probabilistic framework for diagnosing faults in complex engineering systems by translating their functional schematics into Bayesian networks. It extends this to hierarchical models, enabling more efficient diagnosis by "compiling away" lower-level details and focusing inference on relevant levels of abstraction.

What Problem Does This Paper Address?

Diagnosing faults in complex engineering systems is crucial for their reliability and maintenance. Model-based diagnosis aims to pinpoint faulty components by reasoning backward from observed anomalous behavior to a system's functional model. However, as systems grow larger, the number of potential fault explanations explodes, making diagnosis computationally intractable. Furthermore, engineering systems are often designed hierarchically, with components themselves composed of sub-components. Existing diagnostic approaches struggle to effectively leverage this hierarchical structure for computational gains. This paper addresses the challenge of performing efficient, probabilistic diagnosis in both flat and hierarchical system models.

Methods

The core of the paper is a method for translating a system's functional schematic into a Bayesian network. Each component in the schematic becomes a fragment in the Bayesian network, with nodes representing inputs, outputs, and a "mode" variable (e.g., "ok" or "broken"). These fragments are then interconnected based on the system's signal flow. Diagnostic inference is then performed using standard Bayesian network algorithms.

The key novelty lies in extending this to hierarchical models. When a component is modeled at a lower level of detail (as a collection of sub-components), the paper describes a "compilation" process. This process effectively marginalizes out the detailed internal variables and sub-component modes, creating a higher-level Bayesian network that represents the component's behavior without explicit lower-level detail. This is achieved using topological transformations (arc reversal and node absorption) on the Bayesian network. This compilation allows diagnostic inference to operate at different levels of abstraction, providing coarse-grained or fine-grained diagnoses as needed, and crucially, offering computational savings.

Datasets

Not discussed in the paper. The paper uses a Boolean circuit example to illustrate its concepts.

Key Insights and Ideas

  • Bayesian Networks as a Unified Representation: Translating functional schematics into Bayesian networks provides a coherent probabilistic framework for diagnosis, naturally handling uncertainty.
  • Hierarchical "Compilation" for Efficiency: The ability to "compile away" lower-level details in hierarchical models is a significant computational advantage. It allows focusing inference on the relevant level of abstraction, avoiding unnecessary computation.
  • Leveraging Hierarchy for Granular Diagnosis: Users can choose the level of detail for diagnosis. They can get a high-level overview or drill down into specific sub-components as needed, with the system dynamically providing more detail.
  • Abstraction Function is Key: The "abstraction function" that maps lower-level component modes to higher-level component modes is central to the hierarchical compilation process.

Key Results

The paper demonstrates that its approach can provide computational gains by exploiting hierarchical models. While specific quantitative results (e.g., percentage speedup) are not presented with empirical data, the theoretical framework for "compiling away" detail and tailoring inference to specific levels of a hierarchy is the main contribution. The example of the XOR gate illustrates how a complex component can be represented at a higher level without exposing its internal structure, and how diagnosis can be performed on this compiled model.

Limitations and Future Work

The authors acknowledge several simplifying assumptions made for exposition:

  • Non-correlated faults.
  • Full fault models (i.e., the behavior of a component in a "broken" state is fully specified).
  • Fully specified input distributions.
  • Components with single outputs.
  • A restricted form of hierarchy.
  • Systems without dynamics or feedback.

They mention that their general approach can be extended to relax these assumptions. Future work could explore the temporal aspect of "prior probability of failure" and its relation to reliability literature, and integrate diagnosis with other aspects like reliability engineering, test generation, and optimal repair.

Summary updated

BibTeX
@inproceedings{srinivas1994probabilistic,
  author = {Sampath Srinivas},
  title = {{A Probabilistic Approach to Hierarchical Model-based Diagnosis}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1994},
}
Figure