Application of Kalman filtering to multivariate calibration and drift correction Original Research Article

This paper discusses a recursive digital filtering technique, the Kalman filter, which has potential applications for on-line process and environmental monitoring. Two different models are initially used with the Kalman filter algorithm for multivariate calibration of multicomponent spectral data sets obtained by diode array spectrophotometric measurements of synthetic transition metal mixture solutions. The predictive accuracies are compared with those obtained in previous work using direct multicomponent analysis (DMA) and partial least squares regression (PLS1). A model based on K-matrix regression in conjunction with the Kalman filter is generally found to produce improved predictive performances over DMA and a DMA-type Kalman filter model, but cannot match the performance of PLS1 when significant physical or chemical interference effects are present. A further modification of the model is applied to the determination and correction of linear and random baseline drift components in single- and multicomponent spectral data. Relative calibration and prediction errors obtained using this third model are found to be significantly lower than those achieved using Kalman filter models with no drift correction capability (all <1% when using a value of zero for q, the system noise variance).