Statistical feature learning and signal generation for time-series sensor signals

Abstract

The Human Activity Recognition (HAR) problem has attracted substantial attention from academia. HAR has many applications like smart home assisted living systems, healthcare monitoring systems, sports activity monitoring, and monitoring indoor and outdoor activities. HAR applications involve advanced machine learning techniques to identify and classify human activities by leveraging video cameras, wearable sensors, or any other signal like Wi-Fi or radar which eventually encodes the human activity. Human activities are encoded in signals and signal processing techniques are required to pre-process raw signals to filter out high-frequency components and to frame the signals into the fixed-length window. Wearable smart electronics are widely used in human daily life. Those smart devices contain sensors like accelerometer and gyroscope to measure triaxial acceleration and angular velocity respectively. Smartwatches, smartphones, or any such wearable sensor devices contain out-of-the-box sensors embedded in the device. Identification and classification of human activities from such signals by leveraging machine learning techniques require features to be extracted from the signal which represents the corresponding human activity. Many feature extraction techniques from such time-series signals exist in the literature. Time and frequency domain-based feature extraction is a widely used technique for sensor-based human activity classification. To train deep learning models, one needs features to be extracted from the signal. Though time and frequency domain feature extraction techniques are very efficient, the selection of the time and frequency domain features may have a significant impact on the overall classification accuracy. Alternatively, energy-based generative models eliminate the need for a feature extraction layer in the learning pipeline. Deep Belief Networks are alternatives to deep learning models eliminating the need for time and frequency-based feature extraction for sensor-based human activity classification: Restricted Boltzmann Machines (RBM) are the building blocks of Deep Belief Networks. RBMs are energy-based probabilistic graphical models which factorize the probability distribution of a random variable over a binary probability distribution. The visible layer of RBMs represents the real-valued random variable and the hidden layer represents the corresponding binary valued probability distribution. Conditional Restricted Boltzmann Machine (CRBM) is an extension to RBMs and is strong in capturing temporal dependency information encoded in time-series signals. They can be used in the classification of sensor-based human activities. The capacity of CRBM by factorizing a real-valued random variable probability distribution over a binary valued probability distribution eliminates the need for feature extraction from the signal by applying certain feature extraction techniques. This work shows how CRBM is trained to learn signal features. Once trained the signal is generated and reconstructed by the trained model. Along with CRBM, the results of other generative models RBM, GAN, WGAN-GP, and predictive model LSTM are also presented. To compare the performance of the models, similarity metrics are used as a performance criterion to show the performance of the generative models in generating the signals closest to the real signals. Euclidean, Canberra, and Dynamic Time Warping (DTW) distances are used as performance criteria. The results indicate that CRBM outperforms GAN, WGAN-GP, and RBM generative models in generating the signal closest to the original signal. LSTM performs close to CRBM. The capacity of the CRBM in generating signals closest to the original signal indicates that CRBM can learn features from the signal and can also be used in supervised classification.