Transfer learning for aluminium extrusion electricity consumption anomaly detection via deep neural networks

Effective anomaly detection can reduce the electricity consumption and carbon emissions in aluminium extrusion processes. The following two steps identify anomalies: electricity consumption forecasting and anomaly detection. Data-driven modelling is typical paradigm for building an accurate forecasting model. For a new extruding machine, there is insufficient extruded data for model training. The research objective of this work is to determine whether a forecasting model can be trained by transferring knowledge from a data-sufficient domain to a data-insufficient domain. A shared connected deep neural network is proposed for electricity consumption time-series anomaly forecasting. Anomalies are detected by the difference of predicted and measured values at a confidence interval. The experimental results show that the proposed approach can identify electricity anomaly events in real time. Furthermore, it is shown that transferring learning knowledge between domains significantly improves the forecasting results.