Intelligent Real-Time Monitoring

Thomas J. Laffey, Scott M. Weitzenkamp, Jackson Y. Read, Simon A. Kao, James L. SchmidtAAAI 1988

Abstract

In the late 1980s, the Hubble Space Telescope (HST) presented a significant challenge for real-time telemetry monitoring due to its complexity and high data rates. Traditional expert systems were often too slow and lacked the temporal reasoning necessary for such environments. This paper details the development of LSTAR (Lockheed Satellite Telemetry Analysis in Real Time), a distributed, multi-tasking architecture designed to bridge the gap between symbolic AI and the rigorous demands of spacecraft monitoring. By utilizing a modular framework and optimized inference techniques, LSTAR successfully demonstrated the ability to process hundreds of data changes per second in an operational setting.

Introduction

Presented at AAAI-88, the LSTAR system represents a pioneering effort in applying artificial intelligence to high-stakes engineering environments. As satellite complexity increased, human operators faced the daunting task of monitoring thousands of telemetry points sampled at high frequencies. LSTAR was developed to automate the detection of subtle trends and non-nominal behaviors that would otherwise be impossible for humans to track in real-time.

The Real-Time Monitoring Challenge

The Hubble Space Telescope features nearly 5,000 telemetry monitors. During mission-critical phases, the volume of data—often hundreds of changes per second—exceeds the cognitive capacity of human operators. While expert systems were a popular AI solution at the time, standard shells were fundamentally ill-suited for real-time use. They typically lacked temporal reasoning, could not handle asynchronous data streams, and were designed for batch processing rather than continuous operation. L*STAR was designed to overcome these hurdles by prioritizing speed, asynchrony, and temporal awareness.

The L*STAR Architecture

The core innovation of L*STAR is its distributed architecture, which decomposes the monitoring task into three independent, communicating processes. This modularity prevents the "freezing" of the reasoning engine during bursts of sensor data.

The L*STAR distributed architecture, illustrating the flow of filtered and unfiltered telemetry data between specialized processes.
  1. Data Management Process (DMP): Acting as the gatekeeper, the DMP acquires, scales, and compresses raw telemetry. It employs knowledge-based filtering to send only relevant data to the inference engine, significantly reducing computational overhead.
  2. Inference Process (IP): The "brain" of the system, the IP uses a high-performance engine to analyze spacecraft states via frames and rules. It supports both forward and backward chaining.
  3. I/O Process: This module provides the human-machine interface, offering operators graphical schematics and real-time plots that highlight alerts generated by the IP.

Optimizing for Performance

To meet the demands of "soft" real-time performance, L*STAR implemented several technical optimizations:

  • Rule Compilation: Rules are compiled into an efficient intermediate postfix format or C code, eliminating the overhead of pattern-matching during execution.
  • Temporal Reasoning: The system utilizes ring buffers to store historical data, enabling rules to evaluate trends over time (e.g., monitoring voltage stability over a 10-second window).
  • Context Partitioning: By partitioning the ruleset into mission-specific "contexts" (such as Maneuver or Science Mode), the engine only evaluates rules pertinent to the current mission phase.

Operational Deployment and Results

L*STAR was evaluated using live engineering telemetry from the HST Pointing Control System and Flight Software. The system was deployed at the HST Test Control Center in Sunnyvale, California, where it successfully integrated into the existing hardware pipeline.

System hardware configuration for the Hubble Space Telescope monitoring environment, showing the data path from the telescope to the VAX-based inference engines.

Operational results were highly encouraging. The system comfortably handled 200 data changes per second and achieved inference speeds of nearly 1,000 rules per second on a VAX 8650. Most importantly, it proved capable of detecting anomalies in the Rate Gyro Assembly and electrical power systems that were difficult for human operators to identify manually.

Conclusion and Future Directions

While LSTAR successfully demonstrated the utility of distributed AI for spacecraft monitoring, the authors noted several areas for future improvement. The system lacked formal guarantees for "hard" real-time deadlines during catastrophic "event storms," and the verification of large, real-time knowledge bases remained a challenge. Future work pointed toward a "lattice of parallel processes" and more robust temporal reasoning to handle noisy or missing data. Despite these limitations, LSTAR provided a vital blueprint for the next generation of intelligent real-time monitoring systems.

Summary updated

BibTeX
@inproceedings{laffey1988intelligent,
  author = {Thomas J. Laffey and Scott M. Weitzenkamp and Jackson Y. Read and Simon A. Kao and James L. Schmidt},
  title = {{Intelligent Real-Time Monitoring}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1988},
}
Figure