Denoising autoencoder anomaly detection for correlated data

Abstract

In the mining industry, primary digging units such as wheel loaders are critical components due to their position at the start of the process chain. Consequently, the cost of unexpected downtime is high: this motivates efforts to provide an early warning of faults using remote diagnostics.

Machines are equipped with sensors that measure machine health. Some sensors are highly correlated and a model based on machine learning techniques can leverage such relationships across sensors to detect within-group abnormalities.

Autoencoders are auto-associative artificial neural networks which are trained to compress and rebuild the original input with minimal loss. The information is stored in the lower dimensional hidden layers as an internal coding. This is susceptible to a phenomenon called spillover, where the error in a single input can propagate through the network, corrupting the coding and biasing the entire reconstructed data. A denoising autoencoder is a more robust variation on the traditional autoencoder, trained to remove noise and build an error-free reconstruction.

We created a denoising autoencoder to utilize the noise removal on corrupted inputs, and rebuild from working inputs. While this technique is novel to this problem it remained susceptible to spillover. We show our findings and discuss future anomaly detection techniques in correlated data.