Fringe slope discrimination in laser self-mixing interferometry using artificial neural network

• Published in: Optics and laser technology Vol. 132; p. 106499
• Main Authors: Kou, Ke, Wang, Cuo, Lian, Tianhong, Weng, Jun
• Format: Journal Article
• Language: English
• Published: Kidlington Elsevier Ltd 01.12.2020; Elsevier BV
• Subjects: Artificial neural network; Artificial neural networks; Fringe slope discrimination; Interferometry; Laser interferometry; Laser self-mixing interferometry; Lasers; Learning theory; Machine learning; Neural networks; Noise; Self alignment; Signal to noise ratio; Vibration measurement; Laser self-mixing interferometry; Fringe slope discrimination; Artificial neural network
• ISSN: 0030-3992; 1879-2545
• DOI: 10.1016/j.optlastec.2020.106499

•A method based on neural networks is proposed to identify self-mixing fringes’ slope.•Discrimination accuracy can be up to 96% in experiments without noise filtering.•Discrimination accuracy decreases as noise increases and feedback strength reduces.•The model trained under a certain condition is also applicable to other conditions. Due to high sensitivity, self-aligned configuration and laser type independence, Laser Self-Mixing Interferometry (SMI) has always been an interesting and attractive technique in the field of laser interferometry, with displacement and vibration measurements being the most frequently encountered applications. Fringe slope is the most distinctive merit which traditional two-beam interferometry does not have, and it is always utilized to indicate the moving direction of external targets. Due to the existence of noise and even speckle effect, the slope of SMI fringes may be difficult to determine and therefore attracts much more research attention. In order to bring about a robust and easy-to-use method in fringe detection, an Artificial Neural Network (ANN) model established by a well-known Machine Learning (ML) framework named pytorch is proposed to recognize SMI fringes, and through elaborate training, it can distinguish left or right tilted unseen fringes at the accuracy of about 100% in the simulations under good Signal-to-Noise Ratio (SNR) and 96% during the experiments without any noise filtering. Relatively satisfactory results can still be obtained with the feedback strength parameter C lowed to about 0.4 and the SNR reduced to 5 dB. The manuscript explores the application of ML in the field of SMI, with a focus on the accurate and effective SMI fringe recognition, while the subsequent displacement or vibration measurement is not in our research scope.