@ARTICLE{Xiao_Wencong_Ball_2025,
 author={Xiao, Wencong and Cai, Gaipin},
 volume={vol. 41},
 number={No 1},
 pages={219–237},
 journal={Gospodarka Surowcami Mineralnymi - Mineral Resources Management},
 howpublished={online},
 year={2025},
 publisher={Komitet Zrównoważonej Gospodarki Surowcami Mineralnymi PAN},
 publisher={Instytut Gospodarki Surowcami Mineralnymi i Energią PAN},
 abstract={Accurately identifying the load status of the ball mill during the grinding process is conducive to improving the overall production efficiency and ensuring the safe operation of the entire grinding process. In this study, ball mill loads were classified into nine categories based on charge volume ratio (CVR) and material-to-ball volume ratio (MBVR). Different sensors are utilized to collect cylinder vibration and acoustic signals in the grinding process, respectively, and the raw data are converted into time-frequency images by continuous wavelet transform. In this paper, the ResNet18 model is improved from three aspects, namely, depthwise separable convolution (DSC), dropout layer, and Hardswish activation function, and an improved residual fusion network (IRF-Net) based on the merging of two time-frequency image signals is proposed for load recognition. In order to validate the performance of the proposed model, time-frequency images of the acquired data are analyzed, single and multiple signals are used as network inputs, respectively, compared with other classical models, and ablation experiments are performed on the different modules of the improvement. The results show that the improved residual fusion network achieves the best results in recognition with an accuracy of 98.33%, demonstrating good load recognition. The IRF-Net-based multi-signal time-frequency diagram identification method can be utilized to make a sound judgment on the load status of the mill.},
 title={Ball mill load identification method based on IRF-Net with multi-signal time-frequency images},
 type={Article},
 URL={http://ochroma.man.poznan.pl/Content/134577/219-237_Xiao-Cai.pdf},
 doi={10.24425/gsm.2025.153173},
 keywords={time-frequency image, residual networks, depthwise separable convolution, mill signals},
}