Soft sensor for ball mill load based on multi-source data feature fusion

The real-time measurement of ball mill load (ML) in grinding process is difficult to realize, and the states of ML are identified mainly by the experience of the operator. Aiming at the problems, a new soft-sensor approach of ML based on the multi-source data feature fusion was proposed according to the relativity, the information complementation and redundancy among shell vibration, acoustic, electricity signal and ML. The approach consisted of five parts which were data filter, time/frequency transform, feature extraction, feature selection and soft sensor model. The shell vibration and acoustic signal in the time domain was transformed into the frequency domain using fast Fourier transform (FFT). The spectral signals were partitioned into three parts which were low, medium and high frequency bands according to the grinding mechanism. The kernel principal component analysis (KPCA) was used to extract the nonlinear feature of each part. The fused signals, which consisted of the frequency domain feature of vibration and acoustic signal, and the time domain feature of electricity signal, were selected as the input variables of the soft sensor model. The soft sensor model of ML was conducted based on the least square support vector machine (LSSVM). Experimental results show that the approach has better prediction accuracy for ML parameters than the PCA-LSSVM and the single sensor approaches.