The challenge for ultrasonic (US) power transfer systems, in implanted/wearable medical devices, is to determine when misalignment occurs (e.g., due to body motion) and apply directional correction accordingly. In this study, a number of machine learning algorithms were evaluated to classify US transducer misalignment, based on data signal transmissions between the transmitter and receiver. Over seven hundred US signals were acquired across a range of transducer misalignments. Signal envelopes and spectrograms were used to train and evaluate machine learning (ML) algorithms, classifying misalignment extent. The algorithms included an autoencoder, convolutional neural network (CNN) and neural network (NN). The best performing algorithm, was deployed onto a TinyML device for evaluation. Such systems exploit low power microcontrollers developed specifically around edge device applications, where algorithms were configured to run on low power, restricted memory systems. TensorFlow Lite and Edge Impulse, were used to deploy trained models onto the edge device, to classify signals according to transducer misalignment extent. TinyML deployment, demonstrated near real-time (<350 ms) signal classification achieving accuracies > 99%. This opens the possibility to apply such ML alignment algorithms to US arrays (capacitive micro-machined ultrasonic transducer (CMUT), piezoelectric micro-machined ultrasonic transducer (PMUT) devices) capable of beam-steering, significantly enhancing power delivery in implanted and body worn systems.
Keywords: TinyML; machine learning; ultrasonic.