Distributed fault diagnosis with shared-basis and B-splines-based matched learning

In this paper, we present a so-called matched learning methodology for distributed fault diagnosis of dynamical network systems. The matched learning method is based on neural network approximation with special basis function (hidden neuron) configurations, such that the matching of the learned weights of the target node and its neighbors reflect the health conditions of the nodes (the target node and its neighbors). The advantage of this method is that it does not need to be pre-trained with healthy nodes' data. First, a least-squares-based solution framework for matched learning is introduced, and a shared-basis implementation of matched learning is presented. Then, in order to construct better basis functions for matched learning, based on the least-squares solution framework, we propose a B-splines-based implementation of the matched learning method. Experiments and discussions are given on both the shared-basis and the B-splines-based matched learning methods.