AI-Based RUL Prediction for Rolling Element Bearings | #sciencefather #database #scientistaward #DeepLearning #ExplainableAI
One of the most critical steps in any data-driven prognostics model is the accurate preprocessing of raw sensor data to enhance signal quality and extract meaningful features. In this study, vibration data from rolling element bearings are collected under accelerated degradation conditions. Due to the presence of noise and overlapping frequency bands caused by multiple failure modes (e.g., inner race, outer race, cage defects), traditional filtering techniques often fall short in preserving fault-specific signatures.
Degradation Zone Classification
To model the degradation behavior of bearings, it is essential to categorize their health status into different zones (e.g., healthy, early degradation, severe degradation). In this work, the extracted features are input to a multi-class Support Vector Machine (SVM), trained to classify these zones. The SVM uses a nonlinear kernel function to handle the complex and high-dimensional nature of bearing data. Cross-validation is applied to tune hyperparameters and prevent overfitting. This step provides a high-level understanding of the bearing's current condition and helps localize the prediction window for RUL estimation, increasing the efficiency of the subsequent deep learning models.
Remaining Useful Life (RUL) Prediction
Once the degradation zone is known, RUL estimation becomes the next focus. Long Short-Term Memory (LSTM) networks are used due to their ability to learn long-term dependencies in sequential data. LSTM models capture the temporal progression of degradation and are ideal for tracking how features evolve over time toward failure. A sliding window technique is used during the training phase to feed sequences of past data into the LSTM for time-series forecasting. The network predicts the time remaining before the bearing crosses a predefined failure threshold. Performance metrics such as Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE) are used to evaluate the model’s accuracy. To further enhance performance, the study also explores the use of CNN-LSTM and RNN-based models, which combine spatial and temporal feature learning. These hybrid deep learning models are trained and tested using the real-world bearing dataset collected during the experiments.
Explainable AI for Interpretability
One of the major limitations of deep learning models in industrial applications is their lack of transparency, often referred to as the "black-box" problem. To mitigate this and build user trust, Explainable AI (XAI) techniques are integrated into the proposed methodology.
-
SHAP (SHapley Additive exPlanations) is used to interpret feature importance and visualize how individual features contribute to RUL predictions.
-
LIME (Local Interpretable Model-agnostic Explanations) is applied to provide local explanations for specific predictions, aiding operators and engineers in understanding model decisions.
-
These techniques allow stakeholders to validate whether the model is focusing on relevant fault indicators or being influenced by noise or irrelevant data.
The inclusion of XAI makes the system more transparent, trustworthy, and suitable for safety-critical applications in industries like aerospace, railways, wind turbines, and manufacturing.
Experimental Setup and Validation
An experimental test rig was developed using SKF Extra-light series deep groove ball bearings. Accelerated degradation was induced under controlled conditions. Vibration signals were collected using high-sensitivity accelerometers, with each test generating run-to-failure data captured at one-second intervals. Six tests were conducted, with each test producing data under different loading and speed conditions. The system was validated using four successful datasets, with the remaining used for testing generalizability and robustness.
Contributions and Key Findings
-
Developed a complete end-to-end pipeline for RUL prediction in bearings using VMD, feature extraction, ML classification, deep learning, and XAI.
-
Demonstrated that VMD improves fault isolation in vibration signals compared to EMD and other traditional methods.
-
Proposed a hybrid model combining SVM for degradation zone identification and LSTM for time-series RUL forecasting.
-
Incorporated SHAP and LIME for model interpretability, enhancing usability in industrial environments.
-
Achieved high accuracy in experimental validation with low RMSE and MAE, demonstrating real-world applicability.
Comments
Post a Comment