Research On Prediction Remaining Useful Life Of Rolling Bearings Based On Deep Feature Representation And Domain Adaptation

Rolling bearing RUL prediction is a hot topic in current engineering field,an accurate prediction not only ensures the safety and reliability of the system,but also protect workers from accidents.It also increases revenue for enterprises.However,traditional feature extraction and machine learning based approaches are not able to automatically extract representative features.Although deep learning based RUL prediction solves such problems to some extent.Due to different distribution of different bearing’s aging data,the effectiveness and the accuracy of such models are impacted.To solve this problem,we propose a deep feature expression and field adapting based RUL prediction method.Our model uses an auto-encoder to extract features to overcome the issues that traditional feature extraction methods relies to much on domain knowledge and features extracted are not representative.The model implements multi-scale convolutional core and borrows the idea from confrontation to make auto-encoder as generator,conduct adversarial training on original marginal spectrum signal and it re-constructed signals.This approach is able to extract bearing’s deep sensitive features and is able to effectively increase the resolution and accuracy of reconstructed signal.To evaluated our model,we emulate a diagnose tested for collected data.The results shows the advancement of our model.This work provides fundamental research for the future work of bearing healthy status categorization and RUL prediction.Since simple time domain features and frequency domain features can not reflects the aging trend of bearing,our approach uses core PCA to create new healthy status indicators.Implementing core PCA on deep adaptive features extracted,we get kernel first principal component with good monotony,trend,and robustness.We then determine thresholds and the first prediction time to categorize bearing by healthy status and get data from aging process to train the RUL prediction model.To eliminate the issues that the distributions of vibrating data collected from bearing aging process are highly variant,we propose a filed adaptive bearing RUL prediction model.The model hires the adversarial relation between source domain and target domain and MAE as the adapting models.By feeding the deep adaptive features as training input,the model learns the invariant features and reduces the variance of feature distributions.Our model results the best prediction results in both cross-working condition and cross dataset scenarios.Which indicates promising application prospects.Our work studies the deep feature expression and field adaptive based bearing RUL prediction from theoretical and experimental perspectives.The results show that our model adaptively extract sensitive features from bearing and categorize them base on healthy status,which achieves more accurate RUL prediction in both cross-working condition and cross dataset scenarios.