Abstract: The soft sensor model for mill load (ML) is difficulty to adapte the time-vary characters of the mineral process, which needs to be update online accrrording the changes the of opearatal condition. Aiming at the problems, based on the vibration specrum of the mill shell, a on-line soft sensor approach is proposed to measure the ML parameters, such as materail to ball volume ratio, pulp densityi and charge volume ration inside the mill. The appraocho are realized by the interagion of recursive principal component analysis (RPCA) and on-line least square support vector regression (LSSVR). At first, to the training samples, spectral principal components (PCs) at low, medium and high frequency bands of the shell vibration spectrum were extracted through PCA. Then, the spectral PCs of serial combination with different bands were used to construct ML parameters models based on LSSVR. At last, when a new sample was given, after predict with the older models, the inputs and regression parameters of the soft sensor models are updated by RPCA and on-line LSSVR algorithm respecitively. Therefore, the integration of the RPCA and on-line LSSVR makes the on-line soft sensor of ML parameters soft sensor practical. A case study shows that the proposed approach has higher accuracy and better predictive performance than the other normal approaches.