Data Driven Quality Control of Cast Iron Based on Adaptive PSO-LWLR Thermal Analysis Model

Rapid detection of the chemical compositions is of great significance to stabilize the quality of cast iron production, which is usually achieved by thermal analysis in front of the furnace. However, the traditional thermal analyzer usually adopts a fixed parametric mathematical model. Under complex and changeable working conditions (such as fluctuation of furnace burden, changes in process conditions and sample cup batches, etc.), the fixed model often causes large deviations in chemical composition detection, and its parameters are not easy to adjust, which greatly limits the popularization and application of thermal analyzers in front of the furnace. To this end, this study is devoted to exploring a more adaptive thermal analysis model. The temperature eigenvalues of cooling curves are taken as input parameters, local weighted linear regression (LWLR) algorithm is used for local fitting, and particle swarm optimization (PSO) algorithm is used to optimize the bandwidth parameter of LWLR. Finally, a PSO-LWLR hybrid model is formed. In a large foundry, the thermal analysis temperature data and corresponding chemical composition data of more than 200 furnaces of hot metal were collected from actual production. By using the above PSO-LWLR model, and comparing with the multiple linear regression (MLR) model and BP artificial neural network (BP-ANN) model, it is proved that the proposed model can achieve higher prediction accuracy. Unlike the global model with fixed parameters, PSO-LWLR is a locally optimized Just-In-Time (JIT) learning model. It can build different optimized local models for different query samples, so it has stronger adaptability and robustness under time-varying working conditions.