Due to the harsh working conditions and high cost of data acquisition in the actual environment of modern rolling mills, the resulting limited datasets issue leading in performance collapse of traditional deep learning (DL) methods has been plaguing researchers and needs to be urgently addressed. Hence, an improved single-sensor Deep Belief Network (IDBN) is first proposed to repetitively extract valuable information from hidden features and visible features of the previous improved Restricted Boltzmann Machine (IRBM) to alleviate this issue. Next, the multi-sensor IDBNs (MSIDBNs) are applied to obtain complementary and enriched health state features from different multi-sensor data to cope with limited datasets more effectively. Then, the Fast Fourier Transform (FFT) technique is adopted for the multi-sensor information to further enhance the effectiveness of feature extraction. Most importantly, the redefined pretraining and finetuning stages are designed for the MSIDBNs. Meanwhile, the optimal placement of multiple sensors is fully discussed to obtain the most efficient information about health content. Finally, two limited datasets are conducted to validate the superiority of the proposed MSIDBNs. Results show that the proposed MSIDBNs are capable of extracting valuable features from multi-sensor information and achieving more remarkable performance compared with the state-of-the-art (SOTA) methods under limited datasets.
Keywords: Health states diagnosis; Limited datasets; MSIDBNs; Multi-sensor information fusion; Rolling mill.
Copyright © 2022 ISA. Published by Elsevier Ltd. All rights reserved.