Integrating embedded neural networks and self-mixing interferometry for smart sensors design

TL;DR

This paper presents a novel integrated sensor combining self-mixing interferometry with embedded neural networks on a microcontroller, enabling real-time target displacement measurement with robustness to signal variations.

Contribution

It introduces the first prototype of an embedded self-mixing interferometry sensor with neural network processing on a microcontroller, demonstrating practical real-time displacement reconstruction.

Findings

Neural network effectively reconstructs target displacement from interferometric signals.

Sensor maintains robustness despite signal amplitude variations.

Quantization impacts neural network performance and is optimized for embedded deployment.

Abstract

Self-mixing interferometry is a measurement approach in which a laser beam is re-injected into the emitting laser itself after reflection on a target. Information about the position of the target can be obtained from monitoring the voltage across the laser. However, analyzing this signal is difficult. In previous works, neural networks have been used with great success to process this data. In this article, we present the first prototype of an integrated sensor based on self-mixing interferometry with embedded neural networks. It consists of a semiconductor laser (acting both as light emitter and detector) equipped with an embedded platform for data processing. The platform includes an ADC (Analog-to-Digital Converter) and an STM32L476RG microcontroller. The microcontroller runs the neural network in charge of reconstructing the displacement of a target from the interferometric signal entering the ADC. We assess the robustness of the neural network to unwanted signal amplitude variations and the impact of different network weights quantization choices required to run the network on the microcontroller. Finally, we provide a demonstration of target displacement reconstruction fully running on the embedded platform. Our results pave the way towards robust, low power and versatile sensors based on self-mixing interferometry and embedded neural networks.