Data augmentation for deep learning-based detection of pump anomalous conditions

Abstract

Adopting effective asset maintenance approaches is critical in enhancing the longevity and cost-effectiveness of assets in civil infrastructure. Pumps are a crucial asset in many civil infrastructures such as wastewater treatment plants. Data-driven predictive maintenance (PdM) is an emerging asset maintenance method that diagnoses asset conditions proactively. However, the current PdM of pumping assets still requires extensive expert knowledge for finding robust feature extraction methods before applying machine learning methods. This is a significant barrier to the automation and robustness of the PdM of pumps. Deep learning-based algorithms offer the potential to address these issues by capturing data features in monitoring data and performing incremental learning of features without human interventions. To train an analytical model for accurate condition assessment, these methods require a great deal of training data, which is not often available due to time and cost limitations. This research aims to address the scarcity of training data by proposing a novel data augmentation method. The proposed approach consists of a signal-to-image data conversion method and multiple image augmentation methods. The LeNet-5 architecture was employed to produce the CNN model. The performance of the model was evaluated using a public data set. It was shown that the proposed augmentation method significantly enhances the validation accuracy and model generalisability.