Adaptive local fusion systems for novelty detection and diagnostics in condition monitoring

This paper describes the application of Kohonen Self Organising Maps in a dynamic machine condition monitoring application which learns fault conditions over time. The authors describe the implementation of a novelty detection and adaptive diagnostic system which forms a modular component of a larger on-line condition monitoring system. NEURAL-MAINE, a project under the European Union's EUREKA programme, aims to advance on-line condition monitoring applications by the use of neural networks, data fusion and multiple sensor technology. NEURAL-MAINE aims to implement a system on two levels: at the first,, Local Fusion Systems (LFS) are used to model individual machine components, such as high pressure turbines; at the second, a larger 'Overseer' system sits above the LFS and takes in their input, as well as plant operating parameters, in order to give a global view of the condition of the plant. This paper describes the implementation and testing of the two major components of the Local Fusion System, namely the novelty detection and the adaptive diagnostic systems. Novelty detection works by using a Kohonen-based neural network(1) to learn the normal operating state of the component that is being modeled, with new values being passed into the network to see if it is similar to the learned normal states of operation. If the new fused sensor values are unlike the values which the Kohonen neural network represents, this is nagged as novel, and the values are passed to the adaptive diagnostic system. The adaptive diagnostic system takes as its input the values that the novelty detection system identified as novel and passes these into the Kohonen-based diagnostic networks. If the diagnostic networks recognize the values then a local diagnosis is given and passed to the Overseer system. If the diagnostic networks do not recognize the values, then these patterns are dynamically learned as the new fault condition occurs, and the diagnostic networks are updated. This system allows fault knowledge for a specific machine to be dynamically built, which can then potentially be applied on other machines of phe same type.