Unsupervised Anomalous Sound Detection of Noisy Real-World Production Facility Using Deep Autoencoder and Spectrogram

Abstract In this pilot study, we develop an unsupervised deep learning methodology to automatically detect anomalous sounds collected by an acoustic sensor carried by an inspection robot, where all the audio data are collected near hot oil pumps in a noisy real-world oil and gas production facility. Approximately 800 time-series audio clips, each with a 10-second duration, are used for training and evaluating the model. The audio data comprise clips of a hot oil pump working normally and a second hot oil pump working abnormally with a faulty motor bearing. We start by training a deep autoencoder architecture using normal data, and the training data are generated by computing a spectrogram from each signal and extracting features. Such deep autoencoder can take spectrograms of the original time-series audio signals as inputs and act as a classifier to discriminate between normal and abnormal sounds. Specifically, the reconstruction error generated by the autoencoder model is calculated to measure the degree of abnormality of the sound event. Using the test dataset that contains abnormal cases, the developed autoencoder achieved accurate prediction with a high recall of 1.0. The developed methodology has a broader impact and may be readily extended to detect audio anomalies for other types of rotating equipment in the process train such as cooling fans, and compressors. The developed model can potentially serve as a compute engine in a cloud-based software platform to enable autonomous inspection and predictive maintenance for oil and gas production facilities.