In recent times, there has been a notable surge in the exploration of studying human body movements through the utilization of inertial measurement units that can be worn. This trend stems from its substantial impact on academic and industrial circles, encompassing applications ranging from portable healthcare solutions to sports and interfacing with computers through human interaction. Developing human activity recognition (HAR) systems that achieve remarkable accuracy in identification while relying on just one sensor remains an ongoing technological hurdle. In this paper, we explore both supervised and partially supervised approaches using convolutional neural networks and denoising autoencoders approaches. The goal of our study revolves around increasing the classification precision pertaining to previous related works and decreasing reliance on human-engineered features. Since the raw IMU data is composed of time series, it is split using a running window approach and segments of roughly one second of data are fed to the neural networks. The above approaches are tested with two different variations of the original dataset, obtained with data augmentation approaches, to balance uneven class distribution. The dataset contains measurements of one single IMU sensor positioned in the belt of different users performing seven actions: running, jumping, walking, falling, sitting, standing, and lying. The score of best 8-layer CNN-based system was 0.972. The worst represented class, falling, with just 2 min of recorded data, has an F1-score of 0.919.