AI-enabled MEMS Sensors and Applications

Why do we need ‘AI’ algorithms for MEMS sensors?

• Measurement Limitations of MEMS sensors?
  • Response to Physical Input
  • Signal to Noise Ratio
  • Drift over Time
• Examples of AI-enabled Sensor Applications
  • Introduction to AI-enabled Sensor Signal Processing and Classification
  • Examples of Applications

Limitation

Key Noise Sources in Electronic Circuits:

• Thermal noise
• Shot noise
• Flicker noise
• Amplifier noise
• Power-supply noise
• Quantization+sampling noise

Why Wearable IMU Position Tracking Drifts? Orientation can be tracked well; absolute position drifts fast because acceleration is integrated twice and bias accumulates as $t^2$.

Even with excellent noise density, bias and gravity-leakage dominate long-term drift. External references(GPS, ZUPT, map matching, UWB, vision) are needed for stable position

Application: behavior classification

Data Processing for Behavior Classification:

1. Raw data
2. Preprocessing: filter, normalization, Fourier transform...
3. Segmentation: sliding window, ...
4. Feature Extraction: time domain, frequency domain, heuristic domain, ...
5. Dimension Reduction: PCA, UMAP, ...
6. Classification: Naive bayes, Neural network, kNN, SVM, ...

Conclusion

• Cyber Physical Human Systems (CPHS): connections between physical world, cyber world, and humans/animals to create innovative systems and applications.
• Combining MEMS, Nanotechnology, and AI are fundamentally important to the development of human-centric, low power, ubiquitous, mobile, and intelligent sensing platforms:
  • Healthcare: Flexible, skin-like sensors for TCM-based diagnosis
  • Drug discovery: Injectable Motion Sensor for animal motion tracking and recognition
  • Sports: An IoT framework for next-generation sports training