| With the prosperity of Intelligent Machinery,the existing industrial systems incline to be complex and plant-scale.In the context of today’s industrial big data,how to timely judge the operation state and possible abnormalities of the system through intelligent fault diagnosis technology,so as to avoid risks in the early stage and ensure the safety of industrial equipment has become a research hotspot in relevant fields year by year.Traditional diagnosis methods need to establish complex mathematical models or rely on rich expert experience,and they often face the problems of low efficiency and difficult to complete in the implementation process.Therefore,the data-driven fault diagnosis method,which can complete the diagnosis only by processing a large number of historical data,has become the mainstream of the current research.Recently,machine learning algorithms such as support vector machine（SVM）quickly occupy an important proportion in data-driven fault diagnosis methods because of their advantages in automatic data extraction.However,these algorithms are not only limited by expert knowledge a in parameter selection,but also highlight the defect of insufficient generalization ability because the extracted features are only applicable to specific problems.The emergence of deep learning algorithms have well alleviated the above problems.Deep learning algorithms such as auto encoders（AE）and convolutional neural network（CNN）have gradually occupied a place in the field of fault diagnosis.However,most of the existing fault diagnosis methods are applied to the ideal situation of sufficient and balanced fault diagnosis data.In the environment of industrial big data,training data sets often face practical problems such as insufficient and uneven data and complex classification.At present,these existing models often fail to achieve the required diagnosis accuracy when dealing with data sets difficult to deal with.Based on this,this paper improves the existing AE based fault diagnosis models,and skillfully combines the improved algorithm with representative algorithms such as SVM and CNN.Experiments show that the finally proposed new fault diagnosis model not only improves the diagnosis accuracy,but also ensures the diagnosis efficiency.The main research work and innovations of this paper are as follows.1)Aiming at the situation of complex fault types and insufficient sample data,this paper proposes a fault diagnosis method based on improved stacked auto encoder（SAE）and SVM optimized by the differential evolution（DE）algorithm.Firstly,a small back propagation neural network（BPNN）with a Softmax classifier was added to the top of the SAE as an improvement,and then the improved SAE was used to extract the fault diagnosis data.Finally,the extracted feature data is input into SVM,and the whole fault diagnosis is completed by using the advantages of SVM,which has fast classification speed and does not need a large number of samples for training.In particular,when training SVM,the DE algorithm is used to automatically optimize the important parameters of SVM,so as to avoid the influence of expert experience on parameter selection.The data set selected in the simulation experiment has the characteristics of complex fault types and insufficient samples.The experimental results show that this model markedly enhances the accuracy of fault diagnosis without affecting the efficiency.2)The fault diagnosis method based on improved SAE and SVM optimized by the DE algorithm has achieved good results,but the comparative experimental results show that among the existing widely used fault diagnosis algorithms,the classification accuracy of CNN is obviously better than traditional machine learning such as SVM.However,the classification accuracy of CNN is also easily affected by the data set of non ideal conditions.Therefore,in order to give better play to the advantages of CNN in classification in the face of complex data problems,it is a feasible research direction to form a new fault diagnosis model by combining CNN with the front efficient data processing module.Based on this,a fault diagnosis method based on one-dimensional data enhancement and CNN is proposed in this paper.Firstly,the one-dimensional vibration signal data is input into the SAE optimized by BPNN with a Softmax classifier as the training data set,and then the output of the decoder of the SAE after training is combined with the original data.The combined data is the enhanced training data set.The enhanced data set is then directly converted into two-dimensional image data for training CNN.The trained CNN uses the two-dimensional images converted from the non enhanced test data set for performance test.In the experiments,the selected data set has the characteristics of complex fault types and the samples of each type are insufficient.The experimental data shows that the above-mentioned model can enhance the accuracy of fault diagnosis and also take into account the efficiency. |
