A novel multi-source sensing data fusion driven method for detecting rolling mill health states under imbalanced and limited datasets

Monitoring and maintaining the health states of the rolling mill is a constant concern of the steel industry. Therefore, in this paper, multi-source sensors are mounted on the rolling mill to collect various data. Meanwhile, for better health states monitoring with multi-source sensing data, a new deep learning (DL) method based on the improved one-dimension Convolutional Neural Network (I1DCNN) and the improved two-dimension Convolutional Neural Network (I2DCNN) is proposed. First, I2DCNN is fed the 2D kurtogram images generated from the vibration signals by fast kurtogram, while I1DCNN is fed the acoustic signals. Meanwhile, Group Normalization (GN) is embedded to improve the robustness. More importantly, Global averaging pooling (GAP) replaces the traditional fully connected layer to improve model spatial feature extraction. Then, the overfitting problem is mitigated by introducing the dropout layer. Finally, the imbalanced and limited datasets are conducted to test and evaluate the proposed method. Experimental results suggest that the proposed method can achieve efficient and accurate health states monitoring with multi-source sensing data, compared to the other states of the art DL methods.

Automated summary

This paper proposes a new deep learning method for monitoring the health states of a rolling mill using multi-source sensing data. The method feeds 2D kurtogram images generated from vibration signals and acoustic signals to the improved one-dimensional Convolutional Neural Network (I1DCNN) and improved two-dimensional Convolutional Neural Network (I2DCNN) respectively, while introducing Group Normalization (GN) and Global averaging pooling (GAP) for better performance. Experimental results demonstrate that the proposed method achieves efficient and accurate health states monitoring compared to other state-of-the-art deep learning methods.