Self-mixing interferometry based on a phase decomposition neural network

Abstract

In spite of the robustness and simplicity of the self-mixing interferometry (SMI) concept, interpreting the SMI signal is often complicated in practice, which is detrimental to its measurement value. SMI based on machine learning is presented to extract the phase directly from pure SMI signals. A simple phase decomposition neural network (PDNN) was constructed to realize direct phase extraction. A special feature of the PDNN is that it can use the simulation data directly to train the model, and the trained model can be used directly for measurement. In the training process of the PDNN model, the simulated cosine-like signal and the flag signal were used as inputs, and the simulated phase was used as the label. Thus, the training set was easy to prepare, the model structure was simple, and the training speed was very fast. Experimentally, we measured targets with cosine-like movements directly using the trained model, and the results obtained were consistent with the simulation. This contributes to simplifying the signal processing of interferometry in practical engineering applications.