CF-HSACNN: A joint anti-noise learning framework for centrifugal fan state recognition. (October 2022)

• Record Type: Journal Article
• Authors: Fan, Zhixia; Xu, Xiaogang; Wang, Ruijun; Wang, Huijie
• Is Part Of: Measurement. Volume 202(2022)
• Publication Date: 2022-10
• Subjects: Centrifugal fan -- Fault diagnosis -- Anti-noise -- Self-attention mechanism -- Deep learning; Weights and measures -- Periodicals; Measurement -- Periodicals; Measurement; Weights and measures; Periodicals; 530.8
• DOI: 10.1016/j.measurement.2022.111902
• Languages: English
• ISSNs: 0263-2241

Abstract

Highlights: The proposed CF method focuses on separating noise from fault data from sensors to reveal discriminative features. This paper puts forward HSACNN, including SSAM, PSAM and HSAM. The attention mechanism of different angles is constructed, further improving the diagnostic effect of the model in the noisy environment. CF-HSACNN is proposed and its optimal hyperparameter settings are obtained through BO. A model interpretation method is explored. It is suitable for understanding the internal operating rules and parameter changes of the model under different conditions. Abstract: State recognition of centrifugal fans is a challenging task. Because the difference of the fault signals collected is small in a noisy environment, it is difficult to capture useful features, which leads to a decrease in diagnostic efficiency. With the purpose of solving this problem, paper proposes a method that complete-spectrum fast entropy discrimination matches hybrid self-attention convolutional neural network (CF-HASCNN). The method uses fixed-point iteration to perform signal reconstruction on the vector of the complete spectrum decomposition, and uses fuzzy entropy threshold discrimination to weaken the interference of noise on the signal. Then, the HASCNN captures subtle useful information in two aspects: space and single point. In addition, the operating rules of the model are studied to provide an explanation basis. The method which has proposed has stronger anti-interference ability, compared with other advanced diagnostic models.