Research Of Fault Diagnosis Algorithm Based On Cost Sensitive Deep Convolutional Neural Network

In modern industrial manufacturing,mechanical equipments have become increasingly larger and more complex.While significantly increasing the productivity,those equipments are becoming much more difficult to maintain.In reality,even a minor failure may cause serious accidents and lead to economic losses or casualties due to the accumulation of time,so we hope that faults can be diagnosed in its early stage to prevent accidents.However,this will inevitably require a lot of time to complete this work if only human skill is used,this long period of downtime also means a lot of economic losses.In order to realize rapid and effective detection and maintenance of mechanical equipments,intelligent fault diagnosis has drawn more and more attention and it has become an important research topic in modern industry 4.0.Intelligent fault diagnosis not only improves the speed and accuracy of fault diagnosis but also reduces the maintenance cost of those equipments.With the help of machine learning especially deep learning,fault diagnosis algorithms are becoming more and more effective and intelligent in terms of accuracy and stability,however,the current fault diagnosis algorithms still have a lot to optimize.For example,most fault diagnosis algorithms rely on the use of expert experience to complete the task of feature extraction,the training speed and diagnostic effect is difficult to balance.Besides most studies do not consider the different costs of the misclassification errors and then the evaluation of the algorithm is too optimistic.In face of those problems mentioned above,this paper try to solve those problems by using convolutional neural network to process raw vibration signals and combine it with cost-sensitive learning.The convolutional neural network(CNN)has many advantages such as fewer parameters,can make full use of local features and has stronger ability to extract features,based on those characters it has achieved great success in the field of image recognition.Fault diagnosis can be treated as a classification problem from the perspective of machine learning,characters extracted from the original mechanical vibration signal used for classification is vital to the result of the final classification.Based on the excellent characteristics of CNN This paper uses it as the fundamental network and lets it deal with the original fault signal directly,which not only eliminates the dependence on expert experience but also make the algorithm more intelligent.As about the design of network architecture this paper use smaller kernels(3 X 3),deeper network(7 convolutional layers)and reduce the number of fully connected layer(only 2 FC layers),which can improve feature extraction ability effectively while reduces the amount of parameters of the entire network.Besides a new method is used in this paper to convert raw vibration signal to two dimensional image,which combines data resampling and dislocate time series.This method can make use of period information of vibration signal and improve the simplicity of signal processing.The problem of data imbalance exists in many practical applications,this is especially obvious in the field of fault diagnosis,but most studies did not pay enough attention to it.Since different categories of faults are unbalanced in data,models will be skewed after training and categories with less data often have higher error rates in data-driven diagnostic techniques,however,these categories are more important than others and with much greater misclassification costs.In order to solve the problem of data imbalance in fault diagnosis,this paper applies the method of cost-sensitive learning and combines it with CNN,and expands the evaluation index to multi-classification imbalance problem.Experimental results on popular fault diagnosis datasets show that the method of this paper has achieved better classification result than traditional ones.