A New Time-Series Classification Approach for Human Activity Recognition with Data Augmentation

Abstract

Accurate classification of multivariate time series data represents a major challenge for scientists and practitioners exploring time series data in different domains. LSTM-Autoencoders are Deep Learning models that aim to represent input data efficiently while minimizing information loss during the reconstruction phase. Although they are commonly used for Dimensionality Reduction and Data Augmentation, their potential in extracting dynamic features and temporal patterns for temporal data classification is not fully exploited in contrast to the tasks of time-series prediction and anomaly detection. In this article, we present a multi-level hybrid TSC-LSTM-Auto-Encoder architecture that takes full advantage of the incorporation of temporal labels to capture comprehensively temporal features and patterns. This approach aims to improve the performance of temporal data classification using this additional information. We evaluated the proposed architecture for Human activity Recognition (HAR) using the UCI-HAR and WISDM public benchmark datasets. The achieved performance outperforms the current state-of-the-art methods.