Related Experiment Videos
A Multimodal TinyML-Based Predictive Maintenance Architecture for Industrial IoT in the 6G Era
Carlos Exequiel Garay1,2, Fernando Alberto Miranda Bonomi2, Gonzalo Nicolás Mansilla1
1CIASUR (Centro de Investigación de Atmósfera Superior y Radiopropagación), Facultad Regional Tucumán (FRT), Universidad Tecnológica Nacional (UTN), Rivadavia 1050, San Miguel de Tucumán 4000, Argentina.
Sensors (Basel, Switzerland)
|July 28, 2026
Summary
This study introduces a TinyML edge system for predictive maintenance in rotating machinery, achieving high accuracy with low latency. The multimodal approach enhances robustness, paving the way for future 6G networks.
Area of Science:
- Engineering
- Computer Science
- Artificial Intelligence
Background:
- Predictive maintenance (PdM) is crucial for Industry 5.0, aiming to minimize unplanned downtime in rotating machinery.
- The evolution towards 6G networks necessitates edge computing solutions compatible with future application planes.
Purpose of the Study:
- To propose and evaluate a multimodal TinyML edge architecture for PdM.
- To ensure compatibility with evolving 6G network standards.
- To benchmark anomaly detection models on a Cortex-M4F microcontroller.
Main Methods:
- A multimodal TinyML edge architecture using vibration, acoustic, and thermography sensors.
- Local inference on smart sensor nodes and an embedded gateway.
- Benchmarking of five anomaly detection models on real vibration data using frequency-independent time-domain features.
- INT8-quantized fully connected autoencoder evaluated for reconstruction error.
Main Results:
- The INT8-quantized autoencoder achieved an F1 score of 0.9807 with 254 µs latency and 6056 B Flash footprint.
- The model maintained high performance (F1 = 0.975) after recalibration without retraining.
- Late fusion of multimodal confidence scores showed potential for improved baseline performance.
- Edge-first inference was confirmed as essential for 6G URLLC gains.
Conclusions:
- The proposed TinyML edge architecture is effective for PdM, offering high accuracy and low latency.
- Multimodal sensing enhances robustness and redundancy for predictive maintenance.
- Edge-based processing is critical for realizing the benefits of future high-speed networks like 6G.