Deep Learning in Development of Model-Dependent Diagnostic: Recognition of Detector Characteristics in Measured Responses

Abstract

Deep learning has successfully been implemented in various domains, including photoacoustics. The collection and creation of massive datasets creates new possibilities. Deep learning methods, when applied on massive datasets, are able to extract very useful patterns. This can lead to solutions to many problems. In this paper we discuss and develop deep learning application for the recognition of a detector influence pattern on recorded responses of a measurement chain in model-dependent experimental measurements. This enables the fast calibration of the method, which is necessary for its further application in the characterization or scanning of the examined objects with satisfactory accuracy. Frequency gas-microphone photoacoustic measurements were taken as the case study. The paper presents three models for the solution of instrument influence on true signals in photoacoustic experiments. We analyze the influence of neural network depth and the number of outputs on the prediction accuracy, and then we discuss the choice of the optimal solution.