An adaptive constrained clustering approach for real-time fault detection of industrial systems

Thanks to the pervasive deployment of sensors in Industry 4.0, data-driven methods are recently playing an important role in the fault diagnosis and prognosis of industrial systems. In this paper, a novel Adaptive Constrained Clustering algorithm is defined to support real-time fault detection of an industrial machine, by clustering the incoming monitoring data into two clusters over time, representing the nominal and non-nominal work conditions, respectively. To this aim, the proposed algorithm relies on a two-stage procedure: micro-clustering and constrained macro-clustering. The former stage is responsible for grouping the batches of work-cycle data into micro-clusters, while the data stream continuously arrives from the data acquisition system. Then, after condensing the micro-clusters into vectors of cluster features, and leveraging on additional knowledge on the nominal and non-nominal working conditions (i.e., constraints on some samples), the second stage aims at offline grouping the micro-clusters features into macro-clusters. Experimental results on a real-world industrial case study show that the proposed real time framework achieves the same results of offline baseline methods (e.g., Constrained K-means) with a higher responsiveness and processing speed; in comparison to stream baseline methods (e.g., Stream K-means), the proposed approach obtains more accurate and easily interpretable results.