PREDICTION OF PARTICLE SIZE DISTRIBUTION OF A BALL MILL USING IMPROVISED NEURAL NETWORK TECHNIQUE AND TIME SERIES

In the mining industry, it is important to minimize the wastage of raw materials while achieving the desired particle size distribution by grinding the original input mix. To date, the procedure is performed manually, and there is no such control mechanism for grinding that reduces wastage to achieve the desired output, resulting in the loss of material. This study aims to develop an autonomous system for predicting the desired states of breakage by analyzing the acoustic signatures of the materials being crushed. The signal envelope is detected from the time-series acoustic data, which changed gradually during grinding. We designed an autoregression model using the signal envelopes of different grinding stages to predict the desired particle size. In another scenario, the acoustic signatures are approximated to a Gaussian distribution, and the kernel density estimation function is applied to obtain the best-fitted observed data points with the help of local points. An improvised neural network technique is used to classify the unknown patterns of crushing at different breakage states, which validates the experimental results. The network is trained with the input patterns corresponding to the observed data points and the output of the autoregression model. The prediction accuracy of the proposed approach is approximately 97%.