Deep Feature Reduction Algorithm And Application In Fault Diagnosis

The deep learning model can extract different low-level features and combine them to generate more abstract and expressive high-level features.However,the obtained deep feature sets are redundant,which will lead to problems such as timeconsuming training process and requiring many resources.In this paper,the feature reduction related technique in traditional machine learning is used to reduce the dimension of the feature set of high-dimensional image data,and the feature reduction technology is applied to the fault diagnosis of water supply pump in the industrial circulating water system.The main research work is as follows:To solve the redundancy problem of the feature set extracted by the deep learning model,a deep feature reduction algorithm based on fuzzy inconsistency measurement was proposed combined with the traditional machine learning feature dimension reduction method.The pre-trained model is used to extract the feature of target task data and obtain the deep feature set.The feature importance index of fuzzy inconsistencies was defined,and a deep feature reduction algorithm based on fuzzy inconsistencies was proposed to obtain the key deep feature set.The deep features of processed images are classified by a support vector machine(SVM)based on the key deep feature set.Firstly,the generality of the reduction algorithm is verified on the UCI data set,and then the effectiveness of the deep feature reduction algorithm is verified on the image deep feature set.Experimental results show that this method achieves a good reduction effect on the UCI data set and achieves good reduction effect and classification accuracy on the image deep feature set.To identify the working state of the water supply pump in industrial circulating water system timely and accurately,a fault diagnosis method of water supply pump based on deep transfer convolutional neural network(DTCNN)and support vector machine(DTCNN-SVM)is proposed.This method is combined with a deep feature reduction method based on fuzzy inconsistencies measurement to effectively remove the redundant features of convolutional neural networks.Firstly,the vibration signals strongly related to the working state are preprocessed by signal-image processing to realize the two-dimensional grayscale mapping of vibration time-series signals.On this basis,the vibration signal grayscale features are extracted by the DTCNN model which combined transfer learning and residual neural network,and the deep features are reduced based on fuzzy inconsistency measurement.Finally,support vector machine is used to establish the fault diagnosis model of the water supply pump.The vibration signal data collected from the pump platform of the blast furnace industrial circulating water system in a steel plant were validated.Experimental results show that the proposed method is effective in fault diagnosis.In summary,our paper first proposes a feature reduction algorithm based on fuzzy inconsistency measurement,which is used for image deep feature sets reduction.Then,combined with this method,transfer learning and convolutional neural network related technologies are used to solve the fault diagnosis task of water supply pump of an industrial circulating water system under a small number of samples.