Abstract:

An adaptive kernel learning (AKL) networks is proposed to model the industrial analyzer and meanwhile monitor its potential faults, which utilizes kernel function and geometric angle to build an adaptive network topology. Two forms of learning strategies for AKL networks are obtained and their corresponding recursive algorithms are developed, respectively. The proposed AKL algorithm applies to nonlinear MIMO modeling issues with controlled generalization ability. Numerical simulations on Tennessee Eastman (TE) process show that the proposed AKL networks can learn and monitor online the dynamics of the composition analyzer using relatively small samples, under stochastic and fault-existing environment.