A novel approach to machining process fault detection using unsupervised learning

Abstract

The most common machining processes of turning, drilling, milling and grinding concern the removal of material from a workpiece using a cutting tool. The performance of machining processes depends on a number of key method parameters, including cutting tool, workpiece material, machine configuration, fixturing, cutting parameters and tool path trajectory. The large number of possible configurations can make it difficult to implement fault detection systems without having to train the system to a particular method or fault type. The research of this article applies a novel method to detect the changing state of a process over time in order to detect faulty machining conditions such as worn tools and cutting depth changes. Unlike studies in the previous literature in this domain, an unsupervised learning method is used, so that the method can be applied in production to unfamiliar processes or fault conditions. In the case presented, novelty detection is applied to a multivariate sensor feature data set obtained from a milling process. Sensor modalities include acoustic emission, vibration and spindle power and time and frequency domain features are employed. The Mahalanobis squared-distance is used to measure discordancy of each new data point, and values that exceed a principled novelty threshold are categorised as fault conditions.