Research On Fault Diagnosis Of Rolling Bearings Under Variable Conditions Based On Improved AlexNet

Rolling bearing is the key component in rotating machinery,and its health is related to the stable operation of the entire mechanical equipment.It will effectively ensure human-machine safety and reduce economic losses,if the states of rolling bearing is monitored in real-time,and the failure locations and the damage degrees of rolling bearing are detected timely.However,in the actual operation of rolling bearing,there are often accompanied by different degrees of noise interference and different workloads.Meanwhile,traditional fault diagnosis methods are difficult to adaptively extract the favorable features of fault diagnosis of rolling bearing under complex and variable working conditions.Therefore,on the basis of summing up the previous research work,and aiming at above problems,based on AlexNet,the research of the thesis is as follows.1)Because rolling bearing in rotating machinery often works in the environment with variable loads and strong noise.Traditional fault diagnosis methods of rolling bearings are difficult to adaptively extract the favorable features under complex conditions,so a fault diagnosis method of rolling bearing with variable conditions based on improved AlexNet is proposed.Firstly,one-dimensional time-domain signals are translated into two-dimensional feature maps by using transverse insert samples method to satisfy the requirements of the improved AlexNet input.Secondly,the functional layers of AlexNet convolutional layers are improved and adjusted,and the profitable characteristics of state identification of rolling bearing could be automatically extracted via convolution and sampling operations of improved AlexNet from the twodimensional feature maps.Finally,the softmax cross entropy is considered as a loss function and Adam is used to realize the fault diagnosis of rolling bearing according to small batch iterative optimization method.The effectiveness of the method is verified by the vibration signals reflecting 12 kinds of fault states of rolling bearing under different working conditions.2)In reality,working conditions of rolling bearing are complex and variable easily,which faults can’t be diagnosed effectively.Therefore,a novel method based on improved AlexNet with Particle swarm optimization and bacterial foraging algorithm(PSO-BFA)is developed to diagnose the faults of rolling bearing.Firstly,AlexNet structure is simplified and the local normalization layer is added respectively after the first two pooling layers to reduce the training cost.And then,the softmax cross entropy of small batch samples is considered as a loss function,and Adam iterative optimization method is used to train the improved AlexNet with a small number of samples and a few iterations.The diagnosis accuracy of variable-load samples is designed as the fitness function of swarm intelligence algorithm which combines with the updating method of particle movement velocity in PSO and the bacterial movement direction in BFA is updated to find the structure parameters and other parameters of the improved AlexNet.Finally,according to the parameters found by PSO-BFA,the same training method is used to train the improved AlexNet with large samples and multiple iterations to realize multi-state fault diagnosis of rolling bearing under complex working conditions.The effectiveness of the method is verified by the vibration signals reflecting 16 kinds of fault states of rolling bearing under different working conditions.3)The intelligent fault diagnosis methods based on traditional machine learning are strongly dependent on the working conditions and samples of rolling bearing,which seriously affect the generalization performance.A fault diagnosis method for rolling bearing with variable working conditions based on improved AlexNet feature transfer is proposed.Firstly,the fault diagnosis model based on the improved AlexNet is trained with the rolling bearing fault samples under certain loads,and the parameters of previous layers of the network are saved after the training is completed;Secondly,the characteristics of the training samples and the test samples of different loads are extracted respectively according to the saved parameters,and the distance between the two types of sample features is measured by using the maximum mean difference(MMD).Finally,the weighted sum of the MMD and the softmax loss function of the training samples is taken as the objective function.Adaptively adjusting the fault diagnosis model of rolling bearing with variable working conditions to improve the accuracy of fault diagnosis.The effectiveness of the method is verified by the vibration signals reflecting 16 kinds of fault states of rolling bearing under different working conditions.4)Excessive noise and limited label samples would bring great difficulties to diagnose the faults of rolling bearing.A fault diagnosis method for rolling bearing with variable working conditions based on sparse denoising and improved AlexNet feature transfer is proposed.Firstly,the improved AlexNet is trained by using the samples reflecting the fault states of rolling bearing under certain load,and the first five layers of weight parameters are saved.Secondly,according to the training samples,the initial overcomplete dictionary is designed.The over-complete dictionarys are adaptively selected by using K-singular value decomposition(K-SVD)method.The signals are reconstructed with orthogonal matching pursuit(OMP)to achieve the purpose of denoising.Finally,in order to complete the fault diagnosis of rolling bearing at variable conditions,the weighted sum of the softmax cross-entropy value of training samples and MMD between training sample features with the test sample features is used as the loss function to on-line adjusting the weights of the after three layers of AlexNet.The effectiveness of the method is verified by the vibration signals reflecting 1 6 kinds of fault states of rolling bearing under different working conditions.