Learning faults in time: sequential behavioural modelling for complex fault detection in multi-robot systems

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Multi-robot systems require robust fault detection for reliability.
• Current data-driven methods often fail to capture time-dependent fault signatures.
• Memoryless classifiers are theoretically insufficient for faults with temporal precursors.

Purpose of the Study:

• To demonstrate the insufficiency of memoryless classifiers for complex fault detection.
• To introduce a Long Short Term Memory (LSTM) model for learning temporal fault dependencies.
• To enable real-time, distributed fault detection in multi-robot systems.

Main Methods:

• Formalized a theoretical impossibility result for memoryless classifiers.
• Developed and applied a Long Short Term Memory (LSTM) model using primitive behavioral features.
• Evaluated the model on progressive actuator degradation and intermittent actuator dropout fault types.
• Tested with varying swarm movement patterns to represent temporal complexity.

Main Results:

• LSTM models are essential for fault signatures defined by ordered temporal precursors.
• Memoryless models showed partial sufficiency for simpler temporal fault structures.
• Experimental results validated the necessity of sequential modeling for complex faults.
• Compute benchmarks confirmed the real-time feasibility of the LSTM approach for onboard deployment.

Conclusions:

• Sequential models, like LSTM, are crucial for advanced fault detection in multi-robot systems.
• The proposed LSTM model can learn latent temporal dependencies without domain-specific knowledge.
• The approach is computationally feasible for real-time, distributed applications.