Abstract
As one of research and practice hotspots in the field of intelligent manufacturing, the machine learning approach is applied to diagnose and predict equipment fault for running state data. Despite deep learning approach overcomes the problem that the traditional machine learning approaches for fault diagnosis is difficult to characterize the complex mapping between the massive fault data, the exponentially grown and newly generated data is learned repeatedly, and these approaches cannot incrementally correct the model to adapt the situation that the states and properties of equipment are changed over time, resulting in the increase of time costs and the decrease of diagnosis accuracy of model. In this paper, a dynamic deep learning algorithm based on incremental compensation is proposed. Firstly, the feature modes of the newly generated data are extracted by using deep learning algorithm; it is then compared with the fault modes extracted from the historical data. Next, a similarity computing model is presented to dynamically adjust the weights of incrementally merged modes. Finally, the SVM algorithm is employed to classify the weighted modes by supervised way, and the BP algorithm utilized to fine tune the model, in order to complete the dynamic and compensatory adjustment of deep learning with original modes and incremental modes. The experimental results of bearing running data demonstrate that the proposed approach could significantly improve the accuracy of diagnosis and save the time cost, contributing to meet the varied needs of the real-time equipment fault diagnosis.
Highlights
Under the background of the Industry 4.0, it is gradually crucial to figure out how to extract the feature information of fault from the equipment running state data and make an effective analysis to achieve the fault diagnosis and prediction, which has become a hotspot in the field of intelligent manufacturing
The proposed ICDDL approach is used for incremental learning by comparing with BP, Support Vector Machine (SVM), AE, and denoising autoencoder (DAE), respectively
With the variance degree of importance for feature modes which changed over time being taken into account, it can be seen that the fault diagnosed accuracy of proposed ICDDL model has a great improvement to a certain extent because the incremental feature modes are weighted by dynamic compensation
Summary
Under the background of the Industry 4.0, it is gradually crucial to figure out how to extract the feature information of fault from the equipment running state data and make an effective analysis to achieve the fault diagnosis and prediction, which has become a hotspot in the field of intelligent manufacturing. With the rapid development of industrial Internet of things, the deep learning approaches have new problems: Because of the exponential growth in the size of newly generated equipment state data, it is clearly unreasonable to rely on existing fault modes for matching. A dynamic deep learning approach is proposed to dynamically adjust the weights of the feature according to the difference between the new feature modes and the existing feature modes, and effectively complete the incremental learning of the newly generated state data. The study implements the dynamic learning and incremental learning of equipment fault modes, which could extract feature and diagnose fault for the newly generated data from the real-time operation of equipment, and solves the problem that the new state of equipment caused by abrasion is not.
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
