Research On State Recognition Method Of Centrifugal Fan Based On Deep Learning

Centrifugal fan is indispensable mechanical equipment in thermal power plants.It mainly undertakes the roles of blower,induced draft fan,primary fan,sealed fan and powder discharger,and consumes about 1.5% to 3.0% of the power generation of the power plant.In the actual operation of the thermal power plant,the centrifugal fan has poor operating conditions and a high fault rate.Vibration signals are often used as the medium for fan fault detection,and different identification models are built according to different diagnostic tasks.Due to the harsh working conditions and the common influence of various components,the collected vibration signals contain multi-frequency noise and various oscillation mode characteristics.This will increase the difficulty of fault analysis.With the development of artificial intelligence,deep learning models have made outstanding achievements in the field of fault diagnosis.Therefore,in this paper,a deep learning method based on the vibration signal drive of fan is studied.The main research contents include: exploring the influence of feature mapping rules on model performance based on deep learning models,innovating the construction of functional modules to adapt to specific application scenarios,and using feature extraction knowledge to improve the learning ability of discriminative features.First,based on the influence of feature mapping rules,a one-dimensional adaptive densely connected network(1-DADense Net)is established for time-series vibration signals.Through the strong local perception ability and dense connection,it can fully learn the fault information contained in the vibration signal.A better fault state classification effect is achieved.Second,build functional modules based on different application scenarios.Aiming at the problems of low diagnostic accuracy and waste of computing resources in the multi-mode vibration fault diagnosis of fan signals,a lightweight multi-scale multi-attention feature fusion network(LMS-MAFFNet)is proposed.The designed multi-scale and multi-attention functional modules make the model expression biased towards features with larger response values from the two dimensions of space and channel,thereby improving the recognition efficiency of the network.Furthermore,a new model interpretation method is created.Study the dynamic response of its internal structure and parameters,which is suitable for explaining the model operation mode in various environments.Finally,the feature extraction knowledge matches the discriminative feature learning mechanism.In a strong noise environment,due to the complex nonlinearity of the fan signal,the model cannot be biased towards the useful features,which affects the accuracy of fault diagnosis.A complete ensemble empirical mode decomposition with adaptive noise(CEEMDAN)and fast independent component analysis(Fast-ICA)match hybrid attention convolutional neural network(CF-HSACNN)method is proposed.For the noisy signal,the feature extraction method is used to filter the noise and highlight the fault signal.The coding mechanism in the innovative model improves the weight response principle,captures discrete fault information,reduces the probability of feature activation failure and error.and further amplifies the difference between fault features.