Industrial Pump Condition Monitoring with Audio Samples: a Low-Rank Linear Autoencoder Feature Extraction Approach

Condition monitoring of industrial pumps plays a crucial role in predictive maintenance across various industries. From the wide array of techniques used for this task, those based on vibration monitoring through different sensing approaches have gained popularity for the cost effectiveness in the deployment of sensors. In this paper, we focus on the examination of a plant-specific case involving an industrial pump. The use of audio signals captured during pump operation for fault detection is investigated, leveraging signal processing and machine learning techniques. Specifically, an Autoencoder-based approach to extract a linear latent representation of the audio data, facilitating the characterization of pump degradation is presented. Experimental results demonstrate the efficacy of the proposed approach in capturing temporal variations in pump sound signatures and thus, it can be potentially used for early fault detection. Future research directions include expanding the dataset to include samples from machines in various stages of their life cycle to enable comprehensive characterization of pump behavior.