A Self-Paced Robust Regression-based Industrial Virtual Sensor

Virtual sensors are widely used in industrial processes for predicting key quality variables that are often hard to measure timely. Nevertheless, data challenges like the widespread process noises usually make the data-driven virtual sensors weak and vulnerable in practical applications. In this article, a method termed as self-paced robust regression (SRR) is proposed for industrial virtual sensing. Specifically, a robust non-convex loss is defined to deliver a constant penalty for the samples with large errors. Furthermore, inspired by the self-paced based learning, SRR trains samples by assigning weights to alleviate the adverse impact of noisy samples, which progresses from easy to complex learning objectives. Accordingly, an efficient parameter learning algorithm is developed based on the difference of convex functions program and alternating optimization technique, whose convergence property is proved by theoretical analysis. Experiments conducted on an industrial case is provided to demonstrate the effectiveness and generation capability of the proposed virtual sensor.