摘要：Abstract Nowadays systems must adapt to rapidly changing environments and must show behaviour that evolve with time. On the other hand they are intensively monitored so that huge amounts of data are generally collected and available for further analysis. Data-driven diagnosis systems must hence face new challenges, in particular be supported by efficient classification algorithms that adapt the targeted model â°Ş or classifier â°Ş to the evolution of the system and be able to scale to big data. In this paper, a proposal which couples a dynamic clustering method with an on-line trend extraction algorithm that works incrementally on the incoming data is presented. The dynamic clustering method is based on micro-clusters that may drift, merge and split, hence following the evolution of the system. The trend extraction method applies to individual signals and generates a compact abstraction in the form of episodes. The episodes of all signals are then put together to feed the dynamic clustering method. This approach allows data reduction and makes it suitable for on-line data analysis and diagnosis in real-time and low-memory requirements. The proposed algorithm is tested successfully on a continuous stirred tank heater benchmark suffering faults with varying magnitude.
