This thesis investigates the use of deep learning for the automatic identification of machine operations from multivariate time-series data emanating from sensors and actuators. Methods from deep learning and time-series analysis are reviewed with the aim of determining their suitability. A new approach is introduced to alleviate weaknesses in current approaches which include insufficient signal selection, requirement of large amount of training data or neglection of the physical nature of the system. It consists of: a preprocessing methodology based around stationarity tests, redundancy analysis and entropy measures; a deep learning algorithm classifying time series segments into operation categories; a process analytics framework dealing with operation length and frequency. The approach was applied successfully to several datasets from heavy machinery bulk handling systems.