An incremental neural learning framework and its application to vehicle diagnostics

摘要: This paper presents a framework for incremental neural learning (INL) that allows base system to incrementally learn new knowledge from only data without forgetting the existing knowledge. Upon subsequent encounters of examples, INL utilizes prior direct its learning. A number critical issues are addressed including when make knowledge, how perform inference using both and newly learnt detect deal with aged systems. To validate proposed framework, we use backpropagation (BP) as learner multi-layer network intelligent system. has several advantages over algorithms: it can be applied broad range systems beyond BP trained networks; retains structures weights even during learning; committees generated by do not interact one another each sees same inputs error signals at time; this limited communication makes architecture attractive parallel implementation. We have two vehicle fault diagnostics problems: end-of-line test in auto assembly plants onboard misfire detection. These experimental results demonstrate capability successfully unbalanced noisy data. In order show general capabilities INL, also three machine benchmark sets. The showed good generalization comparison other well known algorithms.