On-Device IoT-Based Predictive Maintenance Analytics Model: Comparing TinyLSTM and TinyModel from Edge Impulse

Area of Science:

• Industrial IoT and Predictive Maintenance
• Machine Learning for Equipment Prognostics
• Edge Computing for Real-time Analytics

Background:

• Accurate prediction of industrial equipment health is crucial for reliability and lifespan management.
• Data-driven prognostic models for Remaining Useful Life (RUL) estimation require continuous updates.
• Real-time monitoring and prediction can prevent failures and optimize maintenance costs.

Purpose of the Study:

• To evaluate the efficacy of two real-time tiny predictive analytics models for equipment RUL prediction.
• To compare TinyLSTM and an Edge Impulse TinyModel in terms of performance, development, and deployment.
• To assess the potential of on-device deployment for predictive maintenance.

Main Methods:

• Developed and evaluated two real-time predictive models: TinyLSTM and an Edge Impulse TinyModel.
• Utilized real-time operational equipment data for assessing degradation insights.
• Employed fuzzy logic based on expert knowledge to label datasets and compute actual RUL.
• Converted models into TinyModels for on-device deployment and simulated performance on unseen data.

Main Results:

• Both TinyLSTM and the Edge Impulse TinyModel showed strong performance in real-time predictive maintenance.
• Evaluation loss was 0.01 for TinyLSTM and 0.11 for the Edge Impulse TinyModel.
• The Edge Impulse TinyModel proved significantly easier for development, conversion, and deployment.

Conclusions:

• TinyLSTM and the Edge Impulse TinyModel are effective for real-time predictive maintenance.
• The Edge Impulse TinyModel offers a more streamlined development and deployment process.
• On-device deployment of these tiny models enhances predictive maintenance capabilities.