An Improved Wide-Kernel CNN for Classifying Multivariate Signals in Fault Diagnosis

Abstract

Deep Learning (DL) provides considerable opportunities for increased efficiency and performance in fault diagnosis. The ability of DL methods for automatic feature extraction can reduce the need for time-intensive feature construction and prior knowledge on complex signal processing. In this paper, we propose two models that are built on the Wide-Kernel Deep Convolutional Neural Network (WDCNN) framework to improve performance of classifying fault conditions using multivariate time series data, also with respect to limited and/or noisy training data. In our experiments, we use the renowned benchmark dataset from the Case Western Reserve University (CWRU) bearing experiment [1] to assess our models' performance, and to investigate their usability towards large-scale applications by simulating noisy industrial environments. Here, the proposed models show an exceptionally good performance without any preprocessing or data augmentation and outperform traditional Machine Learning applications as well as state-of-the-art DL models considerably, even in such complex multi-class classification tasks. We show that both models are also able to adapt well to noisy input data, which makes them suitable for condition-based maintenance contexts. Furthermore, we investigate and demonstrate explainability and transparency of the models which is particularly important in large-scale industrial applications.