Multi-information fusion-based belt condition monitoring in grinding process using the improved-Mahalanobis distance and convolutional neural networks

Due to the advantages of flexibility, high efficiency, and low processing heat, belt grinding has been widely applied in manufacturing industries. As belt wear will cause deterioration of the removal capacity, increasing the surface irregularity and adversely affecting the grinding quality, interests in belt condition monitoring have signficantly augmented in recent years, which not only secures the surface quality, but also helps to optimize the utilization of the belt's life cycle. A multi-information fusion-based belt condition monitoring method in grinding process using the improved-Mahalanobis distance and Convolutional Neural Networks (CNN) is proposed in this paper. Firstly, a time-domain mapping relationship between belt wear and material removal rate is put forward and a factor kt is derived to characterize the wear status. Furtherly, the evolution of abrasive grains degradation as well as the wear effect on grinding quality is analyzed. Secondly, a parallel multi-sensor integration grinding system including force, vibration, sound and acoustic emission sensors is established, based on which the single -factor and multi-factor sensitivity experiments are conducted to determine the optimal combination of characteristic signals. Finally, a multi-layer model including the grinding conditions classification and belt stages identification is established adopting the methods of improved-Mahalanobis distance and CNN. On one hand the model is not limited to a fixed condition and has a wider application scope, on the other hand avoids the impact of human experience on the features extraction and improves the model accuracy from the theoretical perspective. The experimental results show that the identification accuracy of the belt wear stage adopting the method in this paper is no less than 94 % for the 16 sampling conditions and more than 86 % for other grinding conditions. Furtherly, the contrast experiments indicate that the method in this paper is of a higher accuracy than the single-layer CNN model, which proves the effectiveness of the proposed method.