Towards the effective representation learning of signal : algorithm and application

Abstract

Signal processing is widely used in various industries, including finance, communication, remote sensing, and wearable sensing. Recent flourish in the deep learning community enables improved signal processing efficiency for various domains. However, most advanced approaches mainly focus on learning better coarse-grained feature representations by providing large and complex training systems while (1) ignoring the effect of implicit factors such as environmental noise on the recognition; (2) the various signal type making universal learning method not applicable to certain domains. All of these considerations lead to the necessity of designing suitable signal representation learning methods for fine-grained learning under different application domains. In this thesis, we study the fine-grained signal representation learning methods with respect to time-dependent and frequency-dependent signal type under the corresponding applications. We first propose a strategy to improve the representation learning ability-under wearable sensing domain-for time-dependent signals: the study starts by investigating the limitation of wearable based activity recognition, and then designs a disentangle algorithm to remove the noise brought by behaviour factors (e.g., age, height) to achieve the fine-grained activity recognition. Next, we investigate the representation learning ability for frequency-dependent signals (modulated/radio) signals, which is yet another crucial field in the signal processing community. The proposed solution provides a solution to transform the frequency-dependent signals (e.g., radio signals) to be better processed by the state-of-the-art representation learning framework when insufficient training samples are provided. The proposed framework can be generalised to various scenarios on wireless communication. As the last key contribution of this thesis, we develop a framework which is the first work to adapt the signal representation learning approaches to explore the privacy leakage problem on power signals. The proposed learning framework’s effectiveness, as well as robustness, has been evaluated, and the corresponding countermeasure has been discussed. The proposed framework can be generalised to other scenarios, such as information leakage on mobile power consumption signals.