Echo state kernel recursive least squares algorithm for machine condition prediction

Kernel adaptive filter (KAF) has been widely utilized for time series prediction due to its online adaptation scheme, universal approximation capability and convexity. Nevertheless, KAF’s ability to handle temporal tasks is limited, because it is essentially a feed-forward neural network that lacks dynamic characteristics. Traditionally, a sliding widow that contains consecutive data points is constructed to deal with the temporal dependency between data points at neighboring time steps, but the restricted widow length may be incapable of capturing temporal patterns on a larger time scale. To manage this issue, a novel sequential learning approach called echo state KRLS (ES-KRLS) algorithm is proposed by incorporating a dynamic reservoir into kernel recursive least squares (KRLS) algorithm. The reservoir, consisting of a large number of sparsely interconnected hidden units, is treated as a temporal function that transforms the history of the time series into a high-dimensional reservoir state space. Subsequently, the spatial relationship between the reservoir state and the target output is effectively approximated by KRLS algorithm. With the utilization of the fixed reservoir, our novel method not only maintains the simplicity of the learning process but also leads to a significant improvement in the capability of modeling dynamic systems. Numerical results on benchmark tasks demonstrate the excellent performance of the novel method with respect to long-term prediction. Finally, an online prognostic method that combines ES-KRLS and a Bayesian technique is developed for tracking the health status of a degraded system and predicting remaining useful life (RUL). This prognostic method is applied to a turbofan engine degradation dataset to demonstrate its effectiveness.