Sparse Sample Regression Based Just-In-Time Modeling (SSR-JIT): Beyond Locally Weighted Approach

Abstract In the present work, a new method for just-in-time (JIT) modeling is proposed. To develop virtual sensors or soft-sensors that can cope with changes in process characteristics as well as nonlinearity, JIT modeling such as locally weighted regression (LWR) and locally weighted partial least squares (LW-PLS) has been investigated and successfully used in various industries. The conventional JIT modeling methods predict output variables by constructing a local model by using past samples located in the neighborhood around the new target sample (query) every time when the output prediction is required; the modeling samples are selected or weighted according to the similarity between the samples and the query. The similarity is usually determined on the basis of the distance from the query. However, the use of distance does not assure the high prediction accuracy. To overcome this limitation of the conventional JIT methods, the proposed method selects past samples that are useful for constructing an accurate local model by using elastic net, which builds a sparse regression model to estimate the query, and uses the derived regression coefficients to evaluate the similarity for conducting LW-PLS. This sparse sample regression based just-in-time modeling (SSR-JIT) has a potential for surpassing the conventional distance-based JIT modeling. In fact, it was demonstrated that SSR-JIT outperformed LW-PLS in the prediction accuracy through two case studies with real industrial data.