Abstract
Mechanical equipment, operating in the oil and gas industry produces a wide range of sounds, often culminating in noise. A functional equipment creates a standard sound, while malfunctioning equipment creates an anomalous sound. In this paper, we propose an application of a contrastive learning algorithm to detect anomalous sound amidst industrial noise, which in turn helps in the identification of the malfunctioning equipment on an offshore rig.
The recent success of contrastive learning algorithms to detect novelty in visual representation has inspired our work. We hypothesise that for most of its lifetime, equipment operates normally. Instances of malfunction that produce anomalous sound and hazardous consequences are rare in their defined lifetime. To this end, we implement a contrastive learning algorithm to identify similarities between audio clips using 10-second length normal-audio clips collected from 6 different equipment for training purposes. It extracts local and global features from the augmented audio clips and presents them in latent space, where the loss function differentiates between normal and anomalous sound. Finally, we input normal audio and anomalous audio into the algorithm.
The performance of the proposed algorithm is measured using Receiver-Operating-Curve - Area-Under-Curve (ROC-AUC) (Koizumi et al., 2020; Nunes, 2021)[5 ]. It successfully differentiates normal and anomalous audio to achieve an average ROC-AUC of 0.77 on the scale (0-1). 26092 audio clips of normal equipment sound and 6065 audio clips of anomalous sound from Valve, pump, fan, slide-rail etc., are used to perform the proposed experiment. Diverse industrial anomalies produced by unbalanced-voltage change, clogging, leaking, contamination, loose belt, no grease etc., are present in the dataset. The proposed methodology of contrastive learning consists of a data-augmentation module followed by an encoder and a Neural Network (NN). Data augmentation transforms the input audio clip to ensure robust performance. The encoder extracts the features from audio clips and projects them in the space of contrastive loss function with the help of NN. Automatic failure detection using Artificial Intelligence is essential for Industry 4.0. Prompt decision by monitoring sound produced by mechanical equipment holds immense potential for asset maintenance in the petroleum industry.
This paper is one of the first instances wherein the application of contrastive learning for the maintenance of mechanical equipment has been demonstrated. The proposed method is a unique approach towards anomaly detection using acoustic-features and it significantly reduces human intervention in hazardous and hard-to-reach environments.
