Research On Fault Diagnosis Of Rolling Bearings Under Incomplete Data Conditions

Rolling bearing serves in complex operating conditions including transient speeds and time-varying loads,which inevitably causes fatigue,wear and even corrosion and other irreversible damage.As a result,bearings are generally the weakest parts of the entire mechanical system.If not monitored and evaluated properly,bearing failures eventuallys leads to production stagnation that causes economic losses,and even endangers the safety of operators.Therefore,effective and timely monitoring the rolling bearing operating conditions,are essential to improve the reliability of mechanical equipments.However,collecting rolling bearing data involves various factors such as complex and changing service environment,various fault types,application environment constraints,and realistic safety considerations,etc.Therefore,the collected bearing data sets are often incomplete,characterized by sparse sample numbers,missing sample labels and unbalanced sample numbers of various types.The existing fault diagnosis methods often rely on sufficient and balanced labeled data samples between classes.Those methods may suffer from loss of feature-learning and the generalization ability,and become less effective for the application of bearing fault diagnosis under incomplete data conditions.Therefore,targeting the incompleteness of bearing fault data sets and conducting corresponding fault diagnosis research are the core issues of rolling bearing health management assessment.In this paper,the key technologies involved in rolling bearing fault diagnosis under incomplete data conditions are discussed and studied,which are as follows:(1)To address the problem that deep learning methods require sufficient data samples,a model structure and training strategy that can achieve better recognition results with only a small number of samples are proposed.The model improves the learning and generalization ability of fault-related features by introducing an adaptive batch normalization module and a network attention-guided mechanism into the deep convolutional network.In order to further reduce the number of training parameters required by the network and prevent the network from overfitting,a training strategy based on deep parameter transfer learning is further adopted.The propsoed strategy initially pretrains the network using existing public data sets,and only some parameters near the output layer of the network need to be fine-tuned during the formal training process to ensure that the network obtains better generalization performance with only a small number of training parameters.The proposed model and learning strategy ensure that the model can be used to identify rolling bearing fault under a small amount of data samples.(2)To address the difficulty of collecting labeled data for industrial applications,a joint distribution adaptive cross-domain fault diagnosis method is proposed to establish an identification model of label-free real fault data through sufficient and high-quality laboratory artificially-damaged bearing data.The proposed method models the signal features of source and target domains separately by building a parallel deep domain adaptive network,and adopts a joint domain adaptation module to learn the distribution differences between cross-domain features,which improves the feature reuse capability and knowledge migration performance.The method fully considers the correlation between features obeying different distributions.By further limiting the correlation between parallel networks,it empowers the cross-domain diagnostic network with more powerful learning capability,which can fully cope with the distribution of fault features between naturally-damaged bearings and artificially-damaged bearings,and achieves good fault diagnosis recognition accuracy for real damaged bearings without labels.(3)To address the practical application of diagnosising bearings under unbalanced samples between classes and variability in data distribution,a method is proposed based on hybrid data augmentation(Re Mix)combined with an improved strong-weak supervised deep domain adversarial network model(Mi DAN),which generates new samples with fuzzy labels to improve the network’s focus on a few classes of samples.Then the strong-weak supervised module is used to separately learn key features learning are performed on the source and target domain data,and the learned pseudo-features are used as the input of the domain adversarial learning module to complete the adaptive learning of the feature distribution through dynamic adversarial training.The method is able to model and learn the distribution differences between features under data imbalance and improve the recognition accuracy of the model for cross-domain fault diagnosis tasks under data imbalance conditions.(4)The practical research on rolling bearing fault diagnosis is explored.In order to further enhance the practicality of the above method in fault diagnosis tasks under incomplete data conditions,the corresponding fault diagnosis system function module and overall framework are proposed for the previous fault diagnosis algorithm,and the practical research of rolling bearing fault diagnosis system is carried out.This rolling bearing fault diagnosis system integrates data analysis and intelligent fault diagnosis modules,and contains four basic modules: user management module,data acquisition module,data analysis and pre-processing module and intelligent fault diagnosis module.The system is conducive to promoting the protocolization process of rolling bearing fault diagnosis,and has positive significance for improving the information management level of bearing fault diagnosis.