Leak detection in compressed air systems using unsupervised anomaly detection techniques

Abstract

Critical components of mobile mining machinery, such as brake and lubrication, are typically powered by compressed air. The compressed air system is subject to leaks in either pipelines or the air-actuated components; and these leaks can cause accelerated wear or unexpected machine shutdowns. Given the size of these machines, fault-finding can be timeconsuming. Remote diagnostics is possible by analysing the machine’s air accumulator pressure with respect to component activation times. However, since every component draws air directly from the accumulator, the resulting drops in pressure all superimpose over the accumulator’s charge and discharge cycles. The result is a highly dynamic trend, making visual diagnostic difficult for anything but major leaks. In this paper, we apply unsupervised anomaly detection techniques to detect developing air leaks. Our method uses machine learning to associate patterns in pressure drop from the accumulator with the activation of each air-powered component. We first apply a wavelet transform to the accumulator pressure trend to make patterns apparent in the time-frequency domain. We then use the Random Forest algorithm’s feature importance to select the most informative wavelet scales. Finally, we trial two anomaly detection methods over the selected inputs: the first uses a clustering approach (LOF), while the second uses a neural-network approach (autoencoder).Once the learning phase (using historical data) is complete, we test the system on an intermittent leak which occurs only when a particular component is activated. We find that both systems perform well, and the LOF trades accuracy for speed with respect to the autoencoder.