Research On Fault Diagnosis Method Of Rolling Bearing Based On Broad Convolutional Learning

The 14 th Five-Year Plan for the development of intelligent manufacturing proposes that equipment fault diagnosis is one of the core technologies of intelligent manufacturing.Rolling bearing is a key part of mechanical equipment,its high fault rate threatens the reliability and safety of mechanical equipment.Therefore,studying the fault diagnosis technology of rolling bearings is of great significance for the development of intelligent manufacturing.In order to meet the speed and accuracy of rolling bearing fault diagnosis technology,this paper proposes a rolling bearing fault diagnosis method based on broad convolutional learning system(BCLS);On this basis,focusing on the challenges of limited training data and inconsistent data distribution under varying operating conditions,fault diagnosis methods based on STFT and stacked broad convolutional learning system(SBCLS)and transfer manifold based broad convolutional learning system(TM-BCLS)are proposed,respectively.And the methods are verified through experiments.The main research contents are as follows:(1)Aiming at the poor performance of broad learning system(BLS)in extracting fault features of rolling bearings,combined with the feature extraction ability of convolutional neural network(CNN),a broad convolutional learning system(BCLS)fault diagnosis method based on CNN and BLS is proposed,achieving fast and accurate end-to-end pattern recognition.The BCLS method takes into account the strong feature extraction ability of CNN and the training algorithm of broad learning to ensure the training efficiency and accuracy.At the same time,the incremental BCLS(IBCLS)algorithm is proposed to improve the network expression ability and training efficiency by analyzing the increment of enhancement nodes,feature nodes,and input data.(2)The BCLS fault diagnosis method is verified by two datasets from CRWU experimental platform and rolling bearing experimental platform in the laboratory.The results show that compared to the BLS method,the BCLS method has a short training time of 1.97 s and 2.35 s,respectively,and high diagnosis accuracy of 99.38% and99.71%,respectively.In addition,the incremental learning function of IBCLS is also explored.Through experimental analysis on the two datasets mentioned above,the proposed incremental algorithm can quickly reconstruct the network structure by adding enhancement and feature nodes.At the same time,it can effectively deal with new dynamic data flows without retraining the network and updating network weights in real-time,thus improving the diagnosis accuracy.(3)Aiming at the problem that the poor generalization ability of BCLS due to the lack of training data,a diagnosis method based on STFT and SBCLS is proposed.STFT is used to conduct time-frequency analysis of vibration signals to obtain two-dimensional time-frequency images,which are input to BCLS.Then the residual learning is used to stack and connect BCLS with multiple BLS to form the SBCLS method,further improving network performance and achieving rolling bearing fault classification.The experimental results show that this method not only has the characteristics of small computation and short training time,but also maintains or improves the network classification accuracy,with an accuracy rate of over 99%,and effectively improves the generalization of the network to cope with the small amount of data.(4)For the problem that inconsistent data distribution under different working conditions affects the applicability of broad network,a fault diagnosis method based on TM-BCLS is proposed to achieve rolling bearing fault diagnosis under cross-working conditions.In TM-BCLS method,frequency domain signals of source domain and target domain are input into BCLS to initially extract features,transfer learning method JDA is used to reduce the distribution difference between source domain and target domain data,and manifold regularization is used to optimize the objective function of the network and improve the adaptive ability of the system.Using the CRWU variable load rolling bearing dataset for TM-BCLS transfer diagnosis experiments,where the source domain are labeled data and the target domain are unlabeled data.The results show that the proposed method has a better transfer effect in cross-domain learning tasks under varying loads.In the set single domain and multi domain transfer tasks,the average accuracy is 99.98% and 99.99%,respectively,which can effectively improve the applicability of the model under variable working conditions.