Active Diagnosis by Self-Organization: An Approach by The Immune Network Metaphor

Yoshiteru Ishida

Abstract: Traditional diagnostic systems often fail in complex, high-speed environments due to syndrome overload and environmental shifts. This paper introduces "Active Diagnosis," a bio-inspired paradigm where a distributed network of agents continuously monitors a system's "normal" state. By employing an immune network metaphor, the author demonstrates how local consistency checks between sensors can self-organize to identify complex process faults in real-time.

Introduction

In industrial settings like chemical plants or blast furnaces, passive, event-driven diagnostic systems are prone to two major flaws. First, a single fault can trigger a "syndrome" of simultaneous abnormal reports, obscuring the root cause. Second, absolute thresholds often fail when the "normal" state shifts due to external factors like seasonal changes. Yoshiteru Ishida proposes a system that behaves like a biological immune system—always "on," defining "self" (the normal state), and using mutual monitoring to detect "non-self" (faults).

The Immunity-Based Agent Approach

The proposed framework treats sensors and process constraints as "agents" (analogous to B-cells or T-cells). These agents do not merely report data; they perform mutual evaluations based on qualitative physical relationships rather than absolute values.

A network of agents performing mutual evaluations based on qualitative constraints.

Each agent is assigned a continuous reliability measure ($R_i$). The system's state is governed by a set of differential equations where an agent's reliability is updated based on feedback from its neighbors. This dynamical system approach allows the network to reach a stable equilibrium that reflects the health of the overall process.

Case Study: Controlled Tank Simulation

The theory was validated through a simulation of a tank system with level and temperature controls, utilizing eight sensors and eight physical constraints (e.g., mass and heat balances).

Schematic of the controlled tank system used for simulation, featuring sensors for flow, temperature, and level.

The simulation demonstrated successful fault discrimination, distinguishing between simple sensor failures and process faults like pipe leaks. In the event of a leak, the reliability measure for the specific constraint agent involved showed a sharp, significant decrease, proving the system's dynamic responsiveness.

Key Insights and Self-Organization

The "Active Diagnosis" model offers several advantages:

  • Consistency over Absolute Values: By focusing on relative constraints, the system is robust to global environmental shifts.
  • Self-Organization: Diagnosis is an emergent property; the network identifies faults through local interactions without a central controller.
  • Pro-activeness: Continuous monitoring of the "normal" state allows for faster filtering of noise during a crisis.

Limitations and Future Work

The primary challenge remains "knowledge engineering"—the requirement to pre-define qualitative relations based on deep domain expertise. Future work will explore adaptive agents that can learn interaction rules and the integration of more rigorous probabilistic tools to complement the dynamical system approach.

Summary updated

BibTeX
@inproceedings{ishida1997active,
  author = {Yoshiteru Ishida},
  title = {{Active Diagnosis by Self-Organization: An Approach by The Immune Network Metaphor}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1997},
}
Figure