Fatigue Detection Using Deep Long Short-Term Memory Autoencoders

Abstract

Efficient time series data mining techniques are an essential part of real world measurement systems and can yield meaningful results from unlabeled data by taking advantage of feature extraction principles. In this paper, we perform kinematic analysis on time series data from IMU sensors for fatigue detection on runners, using several unsupervised machine learning techniques. We propose a robust feature extraction scheme composed of an LSTM Autoencoder, to exploit the advantages of recurrent neural networks and the data compression capabilities of an Autoencoder. The proposed model combines the advantages of several clustering algorithms for accurate fatigue detection in real time, making it suitable for implementation in an embedded device. Experimental evaluation of the feature extraction algorithms showcased their capabilities to produce meaningful features, overcoming the obstacle of extremely limited training data. The inference procedure yielded successful detection in 43% of our representative sample, indicating the efficiency of our model in extracting robust features from unseen kinematic data.