MLDM: A Multi Learning Domain Model for Fault Identification of Centrifugal Fan

Abstract

Abstract In centrifugal fan fault diagnosis, the influence of multiple noise sources causes the collected fault signals to contain interference of different modes. Therefore, the difficulty of capturing fault characteristics greatly increases in this situation. In this article, a Multi Learning Domain Model (MLDM) scheme is proposed based on signal processing technology and artificial intelligence methods. The Haar Wavelet Convolution Extraction (HaarCE) module in the scheme is used to synchronously map features at different frequency scales, while the Lightweight Multi-Scale Feature Enhancement (LMFE) module enhances feature extraction at different time scales. Meanwhile, Upsampling is employed to enhance the expression of fault features. In addition, applying Information Rectification Learning (IRL) to feature maps extracted at each level allows for spatial representation and selection of extracted fault features, providing a basis for final decision-making. This scheme performs multi-level and multi-scale analysis on fault signals, jointly extracting time, frequency, and spatial information to improve the robustness and generalization ability of the model. Conduct experimental verification using data from the same type of centrifugal fan and rotor. The experimental results show that the recognition accuracy of the proposed model is about 93% for fan data at SNR=-4dB, which has certain competitiveness compared to other excellent models.