Time Series Data Augmentation Methods for Deep Learning Models in Conveyor Belt Load Classification

Abstract

This work explores time series data augmentation methods for deep learning and their application for conveyor belt tension signals. In this research, conveyor belt load data was collected and analyzed for five different weights: 0.5 kg, 1 kg, 2 kg, 3 kg, and 5 kg. The research includes applying time series data augmentations like Laplace noise, Gaussian noise, uniform noise, magnitude warping, and channel permutation. Furthermore, new conveyor belt tension signals were generated using a Time VAE model. This study investigates the influence of time series augmentation methods on the accuracy of deep learning models. A CNN-LSTM deep learning model capable of classifying conveyor belt signal data has been developed. The biggest positive impact on the classification accuracy was the addition of Laplace noise, which improved the baseline (no augmentation) accuracy by 4.07 % to 98.64 % for 0.8 second signal data.