Noise reduction in self-mixing interferometric systems with the help of artificial neural networks

The phenomenon called a self-mixing interferometer, can be explained by the fact that light emitted from a laser diode enters its cavity after being reflected from a vibrating target, creating an interference pattern. Such interferometers are compact and cost-effective because they contain few optical components. In addition to its use for detecting various parameters such as displacement and vibration, it can also be used for material characterization such as refractive index and surface roughness. The interference signal generated in the cavity contains various noises that affect the accuracy of the measurements. The quality of the signals obtained is extremely important for many potential application areas. The ability of artificial neural networks to easily learn and predict complex patterns offers a promising approach to improving signal sensitivity by applying them to such interferometers. In this thesis, a deep auto-encoder network model for interferometers is developed and a pre-processing step is proposed to facilitate the analysis of such interference signals. For this purpose, clean synthetic data signals are first generated and then white Gaussian and speckle noise are added. The model is then trained with these data. In the meantime, the training data were compared with the error results of the validation data generated with entirely different values, and the training was completed when a satisfactory result was reached. The model performance was then evaluated on a controlled test data set, and this performance was analyzed in detail for various displacement signals. The model is also compared to a conventional signal filter.