Convolutional Neural Network Based Faults Classification Algorithm And Its Application

With the deep and extensive application of information and sensing technology,more and more data has been accumulated in industrial production process,which contains a lot of important information about production status and safety.With the increase of data scale,and the complexity degree of the mapping relationship between data,traditional data-driven methods based information extraction becomes inadequate.In recent years,the Convolutional Neural Network(CNN)has emerged in the field of deep learning,which can realize multi-level and multidimensional complex feature extraction.It has also gained a lot of successful applications in the field of image processing.Image fault diagnosis based on CNN model is also one of the current research hotspots.However,most of the data generated by industrial processes are numerical data.The research on numerical fault data in the industry based on convolutional neural networks is still in the preliminary exploration stage.Therefore,this paper is based on this background.The research in this thesis has been funded by the Natural Science Foundation of Zhejiang Province.The main research work and achievements are as follows:(1)In order to give full play to the advantages of CNN in the field of image processing,a CNN classification algorithm based on radar chart representation(Radar-chart CNN,Radar-CNN)is proposed,which visualizes numerical data with radar chart.The numerical data is converted into two-dimensional image data,in which the information between the numerical data is reflected by the characteristics of the edge and shape of the radar char.By constructing the commonly used convolutional neural network model and setting the network model parameters,CNN can fully extract the image information of the radar chart to reflect the complex information contained in the numerical data.(2)In order to further study the influence of the feature scale and sequence of numerical data on the performance of radar-chart models,a method based on Rank and Sequential Forward Selection(SFS)is proposed to pre-process the numerical data.The features selected after preprocessing are not redundant.The method reduced the time to draw the radar chart.At the same time,the representative features selected will enable the radar chart to reflect more accurate information and obtain the best classification accuracy.(3)In the process of direct modeling and classification of industrial numerical data,the convolutional neural network model has problems such as insufficient use of features and poor performance of model classification.An improved CNN based on adaptive convolution kernel(ACK-ICNN)is proposed.In order to increase the reuse rate of features,a multi-scale convolution kernel model structure is constructed,which is realized by fusion processing different features of convolution kernel extraction and enhances the adaptability of the model;in order to further improve the performance of the algorithm,the grid search algorithm is used to adaptively select the optimal convolution kernel size in CNN,which can extract the optimal features.(4)The above three improved CNN-based algorithms are applied to the TE chemical process to establish the fault classification model.The proposed algorithms are compared with the typical data-driven methods of extreme learning machine,one nearest neighbor and support vector machine.The experimental results show that compared with the above three methods,the three algorithms proposed in this paper can effectively improve classification accuracy at various faults.