An event-driven Spike-DBN model for fault diagnosis using reward-STDP

Area of Science:

• Artificial Intelligence
• Machine Learning
• Robotics

Background:

• Deep neural networks (DNNs) excel at fault diagnosis but struggle with time-series data and high resource usage.
• Spike deep belief networks (spike-DBNs) offer lower resource consumption and better temporal analysis but often sacrifice diagnostic accuracy.

Purpose of the Study:

• To enhance spike-DBNs for improved fault diagnosis accuracy and efficiency in multivariate time-series data.
• To address the limitations of existing spike-DBN models in capturing temporal dynamics and maintaining high accuracy.

Main Methods:

• Integration of an event-driven approach into spike-DBNs using Latency-Rate coding for enhanced event representation.
• Application of the reward-STDP learning rule to focus on the global behavior of event-triggered spiking neurons.
• Experimental validation of the proposed method on manipulator fault classification tasks.

Main Results:

• The proposed method maintains low resource consumption while improving the fault diagnosis capabilities of spike-DBNs.
• Experimental results demonstrate enhanced accuracy in fault classification of manipulators.
• Significant reduction in learning time (nearly 76%) compared to spike-CNN under identical conditions.

Conclusions:

• The novel integration of event-driven mechanisms significantly boosts the performance of spike-DBNs for fault diagnosis.
• This approach offers a promising solution for accurate and resource-efficient analysis of time-varying industrial data.
• The developed method presents a viable alternative for complex fault classification tasks in robotics and industrial monitoring.