| As equipment becomes more complex and large in scale,fault will cause huge economic losses and even casualties.Fault detection technology is an effective method to improve the safety of the system.It can determine the status of the equipment and judge the occurrence of the fault.Traditional methods usually extract features manually and then perform pattern recognition.Feature extraction relies on expert experience,and the generalization ability is very low for specific data.At present,deep learning is widely used in the field of machine vision,which solves the problems of classification,target detection and recognition of visual objects to a large extent,and the recognition results even outperform humans.It is necessary to introduce deep learning technology in mechanical vibration fault detection to realize detection of fault objects.Based on the results of deep learning in the field of vision,this thesis applies it to the vibration fault detection of rotating machinery in non-visual fields.The details are as follows:1.Mechanical vibration fault detection based on one-dimensional convolutional neural network(1D-CNN)under different working conditions is studied.For the time series data of mechanical vibration,1D-CNN is selected for feature extraction,a mechanical vibration fault detection method based on 1D-CNN is proposed.Firstly,1D-CNN with simple structure is constructed.The linear operation of one-dimensional arrays can reduce the amount of calculation and calculation time.Secondly,in view of the inconsistent data distribution when working conditions change,a two-layer 1D-CNN model is proposed for cross-condition fault detection.By introducing the minimization training of the category label loss in the first layer network,in order to realize the accurate classification of the source domain data.By introducing the domain confrontation training to maximize the loss of the domain label in the second layer network,the data of the unknown operating conditions are trimmed to make the distribution consistent with the data of the known operating conditions.Finally,the simulation verifies that this method has the advantages of small computation and high recognition rate.2.Mechanical vibration fault detection based on recurrence plot(RP)and convolutional neural network(CNN)is studied.On the basis of the 1D-CNN,considering that the 1D-CNN lacks nonlinear analysis of the signal and cannot visualize the fault type,a mechanical vibration fault detection method based on the RP and the CNN is proposed.Firstly,a mechanical vibration fault database is constructed,and the vibration signal is converted into a RP using the method of RP.The characteristics of different working conditions are analyzed on the two-dimensional image,and the characteristic difference between different states of the signal is displayed.Secondly,the CNN simulates the multi-layer architecture of the human brain to extract the topological attributes of the image,and build a shallow CNN model to classify the fault features of the RP.Finally,the simulation verifies that the method has the advantages of being able to visualize fault types and high recognition rate.3.Mechanical vibration fault detection based on transfer learning is studied.It is a common phenomenon that the data of mechanical fault are lacking of labels or even can not be obtained.Aiming at the problem of insufficient network model training caused by too little data,a mechanical vibration fault detection method based on transfer learning is proposed.Firstly,the ability and focus of extracting RP features for different models are different.The models pre-trained on classic data set are reused and merged,and the target domain data set is used to fine-tune the parameters of the model to establish a network model for small sample data sets.Secondly,the features extracted by the network model are not all valid information.The attention mechanism is used to filter the valid feature maps in the convolutional layer and the valid information in a single feature map.Finally,the simulation verified the effectiveness of the method. |
