Decision Tree-Based Classification for Planetary Gearboxes' Condition Monitoring with the Use of Vibration Data in Multidimensional Symptom Space

Piotr Lipinski, Edyta Brzychczy, Radoslaw ZimrozSensors 2020 · DOI · OpenReview

TL;DR

This paper introduces a novel approach to monitoring the health of planetary gearboxes by using decision trees to analyze multidimensional vibration data. By leveraging spectral analysis of vibration signals and treating multiple spectral components as features, the method achieves a remarkable 99.74% classification accuracy, even under challenging non-stationary operating conditions, offering a more robust and interpretable alternative to traditional methods.

What Problem Does This Paper Address?

Monitoring the condition of rotating machinery, particularly complex planetary gearboxes, is crucial for ensuring reliability in critical applications like wind turbines and mining equipment. Traditional methods often rely on advanced signal processing to extract features and then use AI for classification. However, these methods can struggle with non-stationary operating conditions (where the signal characteristics change over time) and often reduce complex data into a single, aggregated feature, losing valuable information. This loss of information can lead to misclassifications, especially when the gearbox operates under varying loads. The paper aims to improve diagnostic accuracy and interpretability by utilizing the full multidimensional spectral information from vibration data and applying decision trees.

Methods

The core of this research lies in two main areas: feature extraction and classification.

Feature Extraction: Instead of aggregating spectral components into a single feature, the authors extract 15 spectral components related to the planetary mesh frequency and its harmonics from segmented vibration signals. This segmentation is necessary because spectral analysis typically requires locally stationary signals, and the gearbox operates under non-stationary conditions. The segmentation length is chosen based on the kinematics of the system (e.g., a digging cycle). This approach preserves the multidimensional nature of the diagnostic data.

Classification: The paper explores the use of decision trees (DTs) for classifying the gearbox condition into "good" or "bad." Specifically, they investigate:

  • CART (Classification and Regression Trees): Using both Gini index and entropy as splitting criteria.
  • Random Tree (RT): An ensemble method that builds multiple decision trees.

These DT-based classifiers are compared against K-Nearest Neighbors (KNN) and meta-classifiers like Random Forest (RF) and AdaBoost (ADA). A key aspect of the DT approach is its ability to perform feature selection implicitly, identifying which spectral components are most informative for classification. The authors also implement stopping criteria to prevent overfitting and create a more generalized and interpretable decision tree.

Datasets

The study utilizes a dataset comprising vibration signals from two planetary gearboxes of the same type, used in a bucket wheel excavator. The dataset consists of 2183 input vectors, each described by 15 features and a class label (good or bad condition). The data was collected under varying load conditions, including "no load" and "small load" scenarios, which are known to be challenging for traditional methods. The dataset is split into training (65%) and testing (35%) sets, and also evaluated using a 10-fold cross-validation approach.

Key Insights and Ideas

  • Multidimensionality is Key: Aggregating spectral features into a single dimension significantly degrades classification performance, especially under non-stationary conditions. Exploiting the multidimensional symptom space (mDSS) is crucial.
  • Decision Trees Offer Interpretability: Unlike methods like PCA that transform features, decision trees select and use original spectral components. This preserves the physical link to the machine's energy and allows for easier interpretation of diagnostic rules.
  • Feature Importance Revealed: Decision trees not only classify but also highlight which spectral components are most discriminative, providing valuable insights for implementing condition monitoring systems.
  • Simplicity Can Be Powerful: The study demonstrates that relatively simple decision tree models can achieve performance comparable to or even exceeding more complex ensemble methods like Random Forest and AdaBoost, especially when properly generalized.

Key Results

  • The proposed decision tree-based classification method achieved a 99.74% classification accuracy on the test dataset with a generalized CART-E model.
  • This represents a significant improvement (approximately 19% better) compared to previous work using aggregated 1D features on the same data.
  • The CART-E classifier, using only four key features (pp10, pp2, pp7, and pp11), outperformed other tested algorithms, including CART-G, RT, KNN, RF, and ADA, in terms of accuracy and simplicity.
  • The generalized decision tree model provides a set of "if-then" rules that are easily implementable in real-time monitoring systems.

Limitations and Future Work

  • Damage Specificity: The current approach is primarily effective for detecting wear-type (distributed) damage. It is acknowledged that it may not be suitable for localized damage types.
  • Machine and Damage Type Dependence: The trained models are specific to the type of machine (planetary gearboxes in excavators) and the type of damage (wear). Retraining would be necessary for different machines or damage mechanisms.
  • Future Work: The authors suggest that further research could explore other types of machines and damage, as well as investigate more advanced feature selection techniques within the decision tree framework. The potential for real-time implementation is highlighted as a promising direction.

Summary updated

Abstract

Monitoring the condition of rotating machinery, especially planetary gearboxes, is a challenging problem. In most of the available approaches, diagnostic procedures are related to advanced signal pre-processing/feature extraction methods or advanced data (features) analysis by using artificial intelligence. In this paper, the second approach is explored, so an application of decision trees for the classification of spectral-based 15D vectors of diagnostic data is proposed. The novelty of this paper is that by a combination of spectral analysis and the application of decision trees to a set of spectral features, we are able to take advantage of the multidimensionality of diagnostic data and classify/recognize the gearbox condition almost faultlessly even in non-stationary operating conditions. The diagnostics of time-varying systems are a complicated issue due to time-varying probability densities estimated for features. Using multidimensional data instead of an aggregated 1D feature, it is possible to improve the efficiency of diagnostics. It can be underlined that in comparison to previous work related to the same data, where the aggregated 1D variable was used, the efficiency of the proposed approach is around 99% (ca. 19% better). We tested several algorithms: classification and regression trees with the Gini index and entropy, as well as the random tree. We compare the obtained results with the K-nearest neighbors classification algorithm and meta-classifiers, namely: random forest and AdaBoost. As a result, we created the decision tree model with 99.74% classification accuracy on the test dataset.
BibTeX
@article{lipinski2020decision,
  author = {Piotr Lipinski and Edyta Brzychczy and Radoslaw Zimroz},
  title = {{Decision Tree-Based Classification for Planetary Gearboxes' Condition Monitoring with the Use of Vibration Data in Multidimensional Symptom Space}},
  journal = {Sensors},
  year = {2020},
  doi = {10.3390/s20215979},
}
Figure