Anotace:
Accurate detection of mill conditions during the cement grinding process directly impacts the quality of particle size distribution and energy consumption per ton of cement. This paper proposes a mill condition classification method based on three-axis wireless vibration sensing and deep feature learning to address issues such as distortion of mill condition characterization caused by power grid disturbances in traditional electrical power methods and sound transmission attenuation in mill sound methods. First, three-axis wireless vibration sensors were installed on the mill shell to collect three-dimensional vibration signals. After filtering and outlier removal, the Fast Fourier Transform (FFT) was applied to generate frequency-domain energy distribution images, creating a vibration spectrum dataset with physical interpretability. Next, a deep dilated separable convolution and multi-head attention fusion network model is proposed. In this model, dilated convolution captures multi-scale frequency domain features using adjustable dilation rates, and the multi-head attention mechanism dynamically adjusts the weight distribution of key frequency bands, enabling adaptive extraction of global frequency domain correlations and local resonance features. Experimental results show that the use of three-dimensional vibration signals to characterize mill conditions improves accuracy by 10% compared to one-dimensional signals. Classification accuracy increased by 6.7% compared to traditional convolutional neural networks, and by 7.6% and 5.5% compared to linear models and machine learning methods, respectively.