Convolutional LSTM Neural Network Autoencoder Based Fault Detection in Manufacturing Predictive Maintenance

The smart manufacturing has revolutionised the intelligent predictive maintenance by integrating IoT technologies with big data analytics, artificial intelligence, cloud computing and other evolving technologies. An effective predictive maintenance demands not only measuring equipment, but the underlying ecosystem that starts with data acquisition from sensors and propagates all the way to visualisation on engineer friendly dashboards. For process monitoring and performance optimization in a smart factory, it is important to recognise time series events like equipment peaks, changeovers and failures. In this article, a model proposed is a deep convolutional LSTM autoencoder architecture using an autoencoder approach to classify real world machine and sensor data to condition based label. The proposed model outperformed baseline architectures. A window size of 45 was used to determine that the model produced a RMSE of 58.45, an MAE of 22.48, and a sMAPE of 0.869, most of which represents significant improvements of up to 37% over existing methods. Having a window size 90, it remained on top with an RMSE score of 72.16 and MAE of 29.64 and sMAPE of 0.847. These results show that it processed a real world manufacturing data and correctly estimated RUL and its complete predictive maintenance.