Computational learning for sensor signal analysis

Abstract

Objective- The general objective is to build accurate machine learning models to solve practical challenges like training data scarcity, compact model construction and data privacy preservation for diverse set of sensor signal analysis tasks. With the proliferation of Internet of Things (IoT), advancements of sensing technologies, incredible enhancements towards computing power along with the outstanding progress of Artificial Intelligence algorithms and tools, researchers are finding new avenues to build different useful applications and novel research directions. The research work focuses on the construction of models for computational learning of analysis tasks involving different types of sensor signals from sensors like Electrocardiogram, Phonocardiogram, accelerometer, energy meter etc. In general, we can consider sensors as the micro-representation of our ambient world. Given that sensors capture near-human information, they usually contain sensitive data. Hence, our foremost task is the enablement of privacy preserving techniques as part of the computational sensing models that analyze the sensor signals and infer critical decision. Methodology- It is understood that remote healthcare is one of the critical applications of IoT and we solve the problem of data privacy protection by proposing de-risking of sensitive data management using differential privacy, where user-enabled controlled privacy protection on sensitive healthcare data can be employed. We propose a novel data privacy preservation method that obfuscates the sensitive component of the sensor data while utility is not severely compromised, while user controls the quantum of privacy. The proposed machine learning algorithm requires subtly hand-crafted feature engineering, which not only restricts the scalability of the computational learning, but also depends on the expensive process of expert or domain-knowledge aided feature generation and selection. We develop intelligence-embedded sensing that does supervised classification tasks using novel deep learning (DL) method of hyperparameter-adjusted convolutional neural network without feature engineering efforts. We extend research to address the integral problem of training data scarcity in DL model generation. It is known that DL models demand substantial training examples for reliable construction of the computational model. Practical sensor signal analysis tasks are often provided with limited number of training examples mainly due to the costs associated with expert annotation. We propose a novel method of effective learning under training data limitation using Shapley-attributed discovery of subset of positively influencing inputs to construct an effective Residual network-based DL model. Results- Our novel privacy preserving method proposes sensor data uncertainty principle, such that controlled statistical uncertainty is employed to the sensitive information with the definition of privacy protection that the prior and posterior probabilities of finding private information does not change beyond a pre-defined threshold and the adversary's gain of sensitivity data access becomes insignificant. The proposed hyperparameter estimation from the input signal characteristics facilitates compact CNN model construction. We demonstrate that our model consistently performs superior over the relevant state-of-the-art algorithms for the given computational learning task of Atrial Fibrillation condition detection from single-lead ECG recordings. We propose an unique push-pull DL architecture, where, firstly Shapley value attributed input subset selection pushes the model parameters towards lower dimension and subsequently, we augment the learnability of the model through adversarial training. We demonstrate the efficacy of proposed model that empirically outperforms the current state-of-the-art algorithms in diverse set of time series sensor signal classification tasks. Conclusion- We have proposed a holistic framework to solve the practical and research challenges of computational analysis of sensor signals including the data privacy preservation, deep learning algorithm for compact model generation, effective computational model under training data scarcity issue. In summary, the research work provides a unified approach to develop practical computational analysis for diverse set of sensor data.