UC Berkeley

The rapid progress in computation power has led to drastic advances in the field of machine learning technologies to carry out data-driven predictions without explicitly programmed instructions. Different from the deterministic algorithm that heavily depends on the prior knowledge and human expertise, the self-evolving machine learning approaches have the potential to greatly reduce the human labor for superior performances. This work focuses on two goals: (1) data-driven MEMS structural designs with machine learning techniques for next generation sensing systems; (2) the combination of advanced sensing technologies and machine learning techniques inspired by the Traditional Chinese Pulse Diagnosis for mobile health applications. Accelerating MEMS design processes by machine learning from pixelated binary images have been developed using binary pixelated 2D images to represent MEMS device structure, including disk-shaped resonators and piezoelectric energy harvesters. Numerical simulation results are used to train the deep neural network (DNN) to accurately predict the target physical properties, such as resonant frequency, quality factor, and output voltage. After the adequate training process, the obtained machine learning (ML) analyzers can precisely predict the physical properties from images with more than 95% accuracy. Compared to the traditional numerical simulation approaches, the ML analyzers can achieve 4.6*10^3 , 2.6*10^4 and 3*10^4 times faster computation speed for frequency, quality factor and output voltage calculation respectively. The wrist pulse diagnosis is based on computational feature extraction and selection, as well as machine learning techniques. In this study, conformal ultra-flexible piezoelectret sensor (with ~112 μm thickness) are attached to volunteers for the data acquisition. The signal was then preprocessed, extracting the identical features via the wavelet-based computational methods, filtering out redundant features with statistical hypothesis and the suitable statistical learning algorithms are selected for the classification tasks. It has been proved that the proposed system can correctly identify every user (~99% F1 score on each tested category) simply based on wrist pulses.