Machine Learning and Signal Processing Algorithms for a CPS with Chemical and Infrared Sensors

In this dissertation, we present our deep learning-based solutions to crowd-sourced cyber-physical systems (CPS) with low-cost chemical sensors to detect ammonia and methane gas leakages. The first algorithm is a gas-leak detection method using uncalibrated sensors. The idea is to construct one-dimensional time-varying waveforms that represent the temporal history of sensor measurements at a location and feed the one-dimensional signals to a convolutional neural network. Our main observation is that while ammonia and other volatile organic compounds can be detected using chemical sensors, sensor measurements decrease over time due to deterioration in the sensor conditions known as sensor drift. We show that it is possible to detect gas leaks even with “below-the-threshold values” using data-driven machine learning algorithms. Furthermore, we show that the detection process can be improved by using the data from multiple sensors. We develop a multiplication-free neural network that is more suitable for energy-constrained devices than the vanilla network, and we show it can achieve good detection and drift correction accuracy. We also show that integrating Hadamard-transform-based layers into the deep learning structure achieves better results, thanks to the regularization effect of the transform. We also investigate a novel l1-inducing kernel metric for drift correction in a multiple-sensor system in unsupervised settings. Our results show that the kernel-based approach achieves better robustness than the baseline methods. The second algorithm aims to detect malfunctioning units in a sensor array system. The deep-learning algorithm aims at learning contrast between normal and abnormal sensors to achieve better sensitivity than direct similarity comparison. We also devise a similarity score based on the aforementioned l1-inducing operator. Our results show that learning new representations via a contrastive learning scheme improves the ROC score from direct methods by 3.5%. Our third algorithm estimates the location of a gas leak source using sparse unreliable spatio-temporal chemical sensor data. This algorithm is based on deep learning and classical inverse problem methods. The neural network has a Fourier-domain layer that models the smoothness of the gas plume. In the transform domain, we project the feature map values onto a low-pass region whose boundary is determined during training using the backpropagation algorithm. This operation is equivalent to making a projection onto a convex set representing the smoothness of the data, and it is embedded into the non-linear structure of the convolutional neural network. We considered both VOC source leak detection and the ammonia vapor leak detection problems. In practice, we use the Discrete Cosine Transform (DCT) instead of the Fourier transform to take advantage of real arithmetic. Our method produces better results than the classical Papoulis-Gerchberg-based interpolation and Gaussian-mixture model-based interpolation methods.