Pervasive Diagnosis: The Integration of Diagnostic Goals into Production Plans

Lukas D. Kuhn, Bob Price, Johan de Kleer, Minh Binh Do, Rong ZhouAAAI 2008

Abstract

Traditional diagnostic methods for complex automated systems often force a choice between passive observation, which may fail to detect intermittent faults, and explicit testing, which incurs significant downtime. This paper presents Pervasive Diagnosis, a novel paradigm that integrates diagnostic testing directly into standard production plans. By intelligently selecting production paths that fulfill work goals while maximizing information about component health, the system drastically reduces downtime and resource waste. Experimental results on industrial printing press simulations demonstrate that this approach can reduce loss rates by orders of magnitude compared to traditional strategies.

Introduction

In complex systems such as industrial printing presses or spacecraft, component failures are inevitable. Engineers typically rely on two strategies: passive diagnosis, which observes normal operation but often fails to "stress" the system enough to reveal intermittent faults, and active (explicit) diagnosis, which halts production for dedicated testing.

Pervasive Diagnosis addresses the efficiency gap between these extremes. It is specifically designed to handle intermittent faults—errors that appear and disappear unpredictably—by treating "information gathering" as a primary objective of the planning process. Instead of choosing the safest path, the system selects paths that are likely to reveal the state of suspect components without halting the production line.

Methodology: The Pervasive Diagnosis Loop

The core of the approach is a model-based integration of planning and diagnosis. The system maintains a Belief State, representing a probability distribution over potential component faults.

The Pervasive Diagnosis system architecture, integrating the planner and diagnosis engine through a shared system model.

The diagnostic engine calculates a Target Probability ($T$) for the next plan. To maximize information gain (entropy), the system does not necessarily seek the safest plan. Instead, it might target a specific failure probability (e.g., $T=0.5$) to most effectively isolate a suspect component. This target represents the point where a success or failure provides the most definitive evidence regarding the system's health.

Heuristic Search for Diagnostic Plans

To find a plan that meets production goals while approaching $T$, the authors utilize a modified A* search. Because the objective function (closeness to $T$) is non-monotonic, standard search algorithms require adaptation. The researchers developed recursive decomposition and pruning techniques to enable real-time planning by evaluating how close a partial plan's probability is to the target.

Heuristic search for a plan that achieves a target failure probability T to maximize information gain.

Experimental Evaluation

The approach was evaluated using a high-fidelity simulation of a modular industrial digital printing press featuring 48 controllable actions, including feeders, print engines, and finishers. A "failed" plan was defined by physical defects like "dog ears" (bent corners) or wrinkles in the output. The study compared Pervasive Diagnosis against Passive and Explicit strategies across varying levels of fault intermittency ($q$).

Key Results

The primary metric was the Loss Rate, representing resources wasted on failed production or dedicated test sheets.

Intermittency ($q$) Passive Loss Explicit Loss Pervasive Loss
0.01 (Rare) 0.431% 4.123% 0.032%
0.1 (Occasional) 3.640% 0.757% 0.034%
1.0 (Persistent) 53.061% 0.041% 0.031%

Pervasive Diagnosis consistently outperformed other methods. For rare intermittent faults ($q=0.01$), it achieved a loss rate an order of magnitude lower than passive diagnosis and two orders of magnitude lower than explicit testing. It effectively "hides" the cost of diagnosis within the production cycle.

Discussion and Insights

Pervasive Diagnosis treats system redundancy as a diagnostic resource. If multiple paths can achieve a production goal, the system selects the one that provides the most utility for fault isolation. This "buys" information at the cost of a calculated risk.

This approach is most effective when the cost of a single failure (e.g., one sheet of paper) is low compared to the cost of machine downtime. While finding a globally optimal sequence of plans is computationally intensive, a "greedy" strategy—maximizing information gain for the immediate next step—proved highly effective in manufacturing contexts.

Conclusion and Future Work

The current model assumes a single-fault hypothesis and is best suited for high-volume, low-stakes production environments. Future research will explore generalizing the approach to multiple simultaneous faults and integrating SAT-solver techniques to manage more complex logical constraints. By embedding diagnostic goals into production plans, Pervasive Diagnosis offers a path toward truly self-maintaining autonomous systems.

Summary updated

BibTeX
@inproceedings{kuhn2008pervasive,
  author = {Lukas D. Kuhn and Bob Price and Johan de Kleer and Minh Binh Do and Rong Zhou},
  title = {{Pervasive Diagnosis: The Integration of Diagnostic Goals into Production Plans}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2008},
}
Figure