Fault Diagnosis On Power Plant Fan Based On Deep Learning

Rotating stall is one of typical fault of centrifugal fans which is a universal auxiliary machine of the power plant.Therefore,it is valuable to apply fault diagnosis for the rotation stall of the fan to the project.With the development of data acquisition and storage technology,the amount of fault data that can be acquired has been increasing.Because of the difficulty of dealing with massive fault data with traditional fault diagnosis method,the fault diagnosis of centrifugal fan rotating stall based on deep learning is studied.The main research contents and results are as follows:Numerical simulation of centrifugal fans under normal operating conditions and varying degrees of rotational stall failure is conducted in Chapter 2.Due to the complex flow field and limited measurement methods of the fan,it is difficult to obtain the detailed information of the physical quantity inside the internal flow field of the fan.In this paper,the key position data information of the fan is obtained by the computational fluid dynamics method.After analyzing the meshing method of centrifugal fan,the flow field is simulated based on Realizable k-ε Model,Sliding Mesh Model and Throttle Valve Function.The results agree with experimental data well and the flow field distribution of different conditions and the pressure frequency characteristics of the fan outlet are analyzed in depth.Further,the cloud map of the velocity distribution of the section of the rotating region of the wind turbine is collected as a data source for the fault diagnosis of the rotor rotating stall.The network interest area was extracted based on the regional suggestion in Chapter 3.Due to the large amount of cloud pressure data and redundant information,a new regional recommendation network for extracting the region of interest was designed.The region of interest is the region that has a key impact on the fault classification effect.As a result,the region of interest is treated carefully to improve the accuracy of fault classification.Firstly,the region of interest is defined as the smallest rectangular region including the separation vortex and the stall cluster,and then the fan image set is established by cropping and labeling the cloud map of the fan velocity distribution.Finally,a new regional recommendation network SLENet is designed,which divides the segmented image block into two regions of interest and background regions to determine the region of interest.Compared with the traditional region of interest extraction method,its extraction recall rate and accuracy are higher.Rotational stall failure is diagnosed based on deep learning in Chapter 4.In this paper,the fault classification network structure and the corresponding pre-training loss function are designed for the fan image set established in the third chapter.Then,the overall scheme of fault diagnosis is designed in combination with the extraction of the region of interest in Chapter 3.Compared with the AlexNet fault diagnosis network,the results show that the fault diagnosis method adopted in this paper has a higher correct rate.