Hybrid Deep Learning Method for Solving the Inverse Problem of Heterogeneous Systems Detection Using Multidimensional Optical Sensor Data

Abstract

Modern machine and deep learning methods effectively handle direct sensor data processing tasks and identify complex dependencies, ensuring robustness to noise and data variations. The most pressing problem is the synthesis of the results of processing multidimensional data using hybrid learning methods to solve inverse problems, such as determining the components of dynamically changing complex heterogeneous systems without violating their properties in multisensory measurement platforms. This paper presents a hybrid deep learning method integrated into an optical sensor platform that combines convolutional neural networks, bidirectional long short-term memory networks, and transformer encoders to enable dynamic monitoring of heterogeneous system components. The hybrid method, when applied to a biosensor platform, demonstrates classification accuracy of over 98 % by solution type and recovery of protein component concentration values with a median accuracy of up to 2 ng/ml.