MACHINE LEARNING FOR ROTATING MACHINERY HEALTH MONITORING A CASE STUDY ON BEARING FAULT CLASSIFICATION

ABSTRACT

This paper addresses the need for effective health monitoring of rotating machinery bearings. We propose an intelligent diagnostic method based on machine learning and multi-feature fusion. By extracting a comprehensive set of 20 time-domain and time-series features from vibration signals, we evaluated eight different classifiers. Logistic Regression was identified as the optimal model, achieving a high classification accuracy of 97.5% on an independent test set for three health states (healthy, degrading, faulty). The results, validated through statistical significance testing, demonstrate the method's potential for reliable predictive maintenance in industrial applications.

Keywords: Bearing Health Monitoring; Machine Learning; Feature Engineering; Logistic Regression; Fault Diagnosis.

Introduction

Rolling bearings are critical components in rotating machinery, with bearing degradation accounting for a significant percentage of motor failures. Traditional maintenance strategies are often inefficient, leading to unplanned downtime and economic losses. Data-driven methods, particularly machine learning, offer a promising approach for Prognostics and Health Management (PHM) by learning degradation patterns from sensor data. However, challenges remain in effectively fusing multi-modal features and ensuring model interpretability. This study proposes a novel approach that combines multi-modal feature engineering—integrating time-domain statistics with time-series dynamic features—and applies Logistic Regression for accurate and interpretable bearing health state classification.

Feature Engineering and Methodology

The experimental data was sourced from the PRONOSTIA platform, using the complete life-cycle vibration data of seven bearings under constant conditions (1800 rpm, 4000N load). A total of 263 valid samples were extracted, each labeled into one of three health states: Healthy, Degrading, or Faulty. The data was split into training (70%), validation (15%), and test (15%) sets using stratified sampling.

Our core contribution lies in the feature engineering scheme. A set of 20 features was constructed for each sample, categorized as:

Basic Time-Domain Features (8): RMS, Peak, Crest Factor, Impulse Factor, Shape Factor, Margin Factor, Kurtosis, Skewness.

Time-Series Dynamic Features (12): Lag features (1, 2, 3, 5-step), rolling statistics (mean, std for windows 3, 5, 10), difference features (first and second order), and cumulative features (Cumulative RMS, Time Ratio). These features capture short-term dependencies, local trends, and long-term degradation effects.

Eight machine learning models were trained and evaluated. Hyperparameter tuning was performed using grid search on the validation set. The final model performance was assessed on the held-out test set.

Results and Discussion

The performance of all models on the test set is summarized in Table 1. Logistic Regression achieved the highest accuracy, outperforming more complex ensemble methods.

Table 1.

Model Performance Comparison Table

The confusion matrix for the best Logistic Regression model is shown in Figure 1. The model perfectly classified all Healthy and Faulty samples, with only one Degrading sample misclassified as Faulty, resulting in an overall accuracy of 39/40.

Figure 1. Confusion Matrix for the Logistic Regression Model on the Test Set

A binomial test confirmed that this accuracy is statistically significant (p < 0.001) compared to a random guess baseline of 33.3%. The 95% confidence interval for the accuracy is [92.7%, 100%].

Analysis of the model's coefficients provided insights into feature importance, as shown in Figure 2. The cumulative RMS feature was identified as the most critical, highlighting the importance of long-term energy accumulation in assessing degradation. Rolling statistics and lag features also ranked highly, confirming the value of our time-series feature engineering.

Figure 2. Top 10 Most Important Features Based on Logistic Regression Coefficients

Conclusion

This paper presented an intelligent bearing health diagnosis system based on multi-feature fusion and Logistic Regression. The proposed method achieves a high classification accuracy of 97.5%, demonstrating its effectiveness. The key findings are:

Logistic Regression is highly suitable for this task, offering a superior balance of accuracy, speed, and interpretability.

The novel combination of time-domain and time-series features, particularly cumulative RMS, is critical for capturing the degradation process.

The model's performance is statistically robust, providing a reliable foundation for predictive maintenance applications.

Future work will focus on validating the approach on datasets with varying operating conditions and exploring advanced architectures like 1D-CNNs for end-to-end feature learning.

References:

1. Ren L, Cui J, Sun Y, Cheng X. Multi-bearing remaining useful life collaborative prediction: A deep learning approach. Journal of Manufacturing Systems. 2017;43:248-256.
2. Li X, Elasha F, Shanbr S, Mba D. Remaining useful life prediction of rolling element bearings using supervised machine learning. Energies. 2019;12(14):2705.
3. Ren L, Sun Y, Wang H, Zhang L. Prediction of bearing remaining useful life with deep convolution neural network. IEEE Access. 2018;6:13041-13049.