Performance Degradation Prediction Using LSTM with Optimized Parameters

Yawei Hu, Ran Wei, Yang Yang, Xuanlin Li, Zhifu Huang, Yongbin Liu, Changbo He, Huitian Lu2022 · DOI

TL;DR

This paper introduces an improved Long Short-Term Memory (LSTM) model, optimized with an enhanced Particle Swarm Optimization (IPSO) algorithm, for predicting the performance degradation of rolling bearings. By intelligently fusing multi-domain vibration features and optimizing LSTM hyperparameters, the proposed IPSO-LSTM method achieves higher predictive accuracy and faster convergence compared to traditional LSTM, PSO-LSTM, ELM, and SVR approaches.

What Problem Does This Paper Address?

Predicting the degradation of mechanical components, particularly rolling bearings, is crucial for preventing catastrophic failures in industrial machinery. Traditional methods often struggle with the complexity and dynamic nature of bearing degradation, requiring difficult manual feature engineering or suffering from slow convergence and local optima in optimization. Existing deep learning models like LSTM, while powerful for time-series data, are sensitive to hyperparameter choices, which can significantly impact their predictive accuracy. This paper aims to address these challenges by developing a robust and accurate degradation prediction model that automates hyperparameter optimization and effectively captures complex degradation patterns.

Methods

The core of this research lies in the integration of three key components:

  • Feature Extraction and Fusion: Multi-domain vibration signals (time and frequency domains) are extracted from bearings. To overcome the limitations of single-domain features and reduce redundancy, the Kernel Joint Approximate Diagonalization of Eigen-matrices (KJADE) method is employed for feature fusion. This process maps data to a high-dimensional feature space, enabling better nonlinear processing and creating a more effective, lower-dimensional feature set.
  • Degradation Index Calculation: A performance degradation index is calculated using the between-class and within-class scatter matrix (SS). This index quantifies the dispersion and aggregation of health and monitoring samples, providing a clear measure of degradation.
  • Optimized LSTM Model: A Long Short-Term Memory (LSTM) network is used as the backbone for predicting degradation trends due to its ability to handle sequential data. To overcome the challenges of manual hyperparameter tuning (number of hidden nodes in each layer, batch size), an Improved Particle Swarm Optimization (IPSO) algorithm is introduced. IPSO modifies the traditional PSO by adaptively adjusting learning factors and inertia weight, and incorporating an adaptive adjustment factor for particle position updates. This enhanced optimization strategy aims to achieve faster convergence and avoid local optima, leading to a more accurate LSTM model.

Datasets

The study utilizes two datasets for training and evaluation:

  • University of Cincinnati's Intelligent Maintenance System (IMS) Center Dataset: This dataset contains full-life vibration signals of bearings, including examples of rolling, inner, and outer faults. It was used for initial validation and comparison with other methods.
  • Lab Experiments Dataset: This dataset was generated from controlled experiments using four HRB6305 bearings under accelerated damage conditions. It was used to further validate the proposed method, particularly for rolling element faults.

Key Insights and Ideas

  • Feature Fusion is Key: Relying on single-domain features for bearing degradation is insufficient. Fusing features from both time and frequency domains using methods like KJADE is crucial for capturing the complex dynamics of degradation.
  • Hyperparameter Optimization Matters: The performance of LSTM models is highly sensitive to hyperparameter choices. Manual tuning is inefficient and prone to suboptimal results.
  • IPSO Outperforms Standard PSO: The proposed IPSO algorithm demonstrates superior convergence speed and a reduced tendency to get stuck in local optima compared to the standard PSO, leading to better hyperparameter selection for the LSTM.
  • Holistic Approach Yields Results: The combination of advanced feature fusion (KJADE), a robust degradation index (SS), and an optimized deep learning model (IPSO-LSTM) creates a powerful system for accurate performance degradation prediction.

Key Results

The proposed IPSO-LSTM method consistently outperformed traditional LSTM, PSO-LSTM, ELM, and SVR across various fault types (outer, roller, inner) and datasets.

  • Reduced RMSE: The Root Mean Square Error (RMSE) values were significantly lower for the IPSO-LSTM method compared to the other models. For instance, in Case 1, RMSE values for outer bearings were 0.012 (IPSO-LSTM) vs. 0.042 (LSTM) and 0.025 (PSO-LSTM). Similar improvements were observed for roller and inner bearings.
  • Accurate Degradation Trend Tracking: Visualizations (Figures 6-12, 15, 17) show that the IPSO-LSTM predictions closely track the true degradation trends, especially in the later stages of bearing life, where other methods often deviate.
  • Faster Convergence: Figure 9 and Figure 16 illustrate that the IPSO algorithm converges faster than the standard PSO, reaching optimal parameters in fewer iterations.

Limitations and Future Work

  • Generalizability: While the method shows promise across different fault types, further validation on a wider variety of bearing types and operating conditions would be beneficial to confirm its generalizability.
  • Data Availability: The study mentions that models and codes are not shared at this time as they form part of an ongoing study. Making these available would facilitate reproducibility and further research.
  • Real-time Application: The paper focuses on offline analysis. Future work could explore adapting this method for real-time monitoring and prediction in industrial settings.
  • Other Mechanical Systems: The authors suggest that the method can be tailored and applied to other mechanical systems for online health and prognosis management, indicating a broad potential for future research.

Summary updated

BibTeX
@misc{hu2022performance,
  author = {Yawei Hu and Ran Wei and Yang Yang and Xuanlin Li and Zhifu Huang and Yongbin Liu and Changbo He and Huitian Lu},
  title = {{Performance Degradation Prediction Using LSTM with Optimized Parameters}},
  year = {2022},
  doi = {10.3390/s22062407},
}

Loading related content...

Figure