A Fleet-Wide Approach for Condition Monitoring of Similar Machines Using Time-Series Clustering

The application of machine learning to fault diagnosis allows automated condition monitoring of machines, leading to reduced maintenance costs and increased machine availability. Traditional approaches train a machine learning algorithm to identify specific faults or operational settings. Therefore, these approaches cannot always cope with a dynamic industrial environment. However, an industrial installation often contains multiple machines of the same type, which enables a fleet-based analysis. This type of analysis compares machines to tackle the challenges of a dynamic environment. In this paper a novel method is proposed for analyzing a fleet of machines operating under similar conditions in the same area by using inter-machine comparisons. The proposed methodology consists of two steps. First, the inter-machine difference is calculated using dynamic time warping by using the amount of warping as measure. This method allows comparing the measured signals even when fluctuations are present. Second, a clustering method uses the inter-machine similarity to identify groups of machines that operate in a similar manner. The generation of a fault usually causes a change in the raw signals and diagnostic features. As a result, the inter-machine difference between the faulty machine and the rest of the fleet will increase, leading to the creation of a separate group that contains the faulty machine. The methodology is evaluated and validated on phase current signals measured on a fleet of electrical drivetrains, where a phase unbalance fault is introduced in some of the drivetrains for a specific time duration.