Deep Learning Enhanced Time-Domain Microwave Photonic Sensor

In this article, a new deep learning enhanced microwave photonic (MWP) sensor based on optical sideband processing with linear frequency modulated (LFM) pulse to enable fast and accurate sensing in the presence of noise and interference, is proposed and demonstrated. The LFM pulse produces a fast sweep of the optical sideband across the resonance, causing an envelope dip in the transmitted LFM pulse which enables fast interrogation. By optimizing the optical modulator bias point, the depth of this envelope dip can be increased, thus enhancing the interrogation resolution and facilitating higher accuracy. To enhance the noise resistance for practical uses and avoid adding unnecessary system complexity, we adopt a deep neural network (DNN) to process the raw interrogation output of this time-domain MWP sensor. The DNN can compress input data into a lower-dimensional representation and be trained to accurately estimate the target measurand despite noise and interference. Beside presenting a detailed theoretical model, as a proof-of-concept, we demonstrate the proposed scheme with a convolutional neural network (CNN) in measuring the glycerol solution concentration amidst thermal interference and system noise. The interrogation speed reached 2.5 MHz. Compared to the conventional MWP interrogation without bias control, using optimized bias voltages improves the accuracy of the estimation model based on the linear fitting of the manually extracted envelope dip by 2-fold. The CNN based prediction model achieves a root-mean-square error of 0.05%, which demonstrates an overall 4-fold higher accuracy than that of the reference model using traditional interrogation and filtering methods.