Information Theory Filters for Wavelet Packet Coefficient Selection with Application to Corrosion Type Identification from Acoustic Emission Signals

Gert Van Dijck, Marc M. Van HulleSensors 2011 · DOI

TL;DR

This paper tackles the challenge of accurately identifying different types of corrosion in industrial settings by analyzing acoustic emission signals. They propose using information theory-based methods, specifically mutual information, to select the most informative features from wavelet packet decompositions, leading to significantly improved corrosion classification accuracy compared to traditional methods.

What Problem Does This Paper Address?

Corrosion in chemical process installations can lead to catastrophic failures, plant shutdowns, and environmental hazards. Acoustic emission (AE) signals generated by structural changes during corrosion offer a promising avenue for monitoring and predicting corrosion types. However, effectively extracting and selecting the most relevant features from these complex AE signals is a significant challenge. Existing methods, like the Local Discriminant Basis (LDB) algorithm, often overlook the statistical dependencies between extracted features (wavelet coefficients), leading to suboptimal classification accuracy. This paper aims to address this by developing and evaluating information theory-based feature selection methods that account for these dependencies.

Methods

The core of this research lies in Wavelet Packet Decomposition (WPD) for feature extraction. AE signals are decomposed into a library of wavelet packet coefficients, which represent different time-frequency characteristics of the signal. The challenge then becomes selecting the most discriminative coefficients for classifying corrosion types.

The paper compares several feature selection approaches:

  • Local Discriminant Basis (LDB) algorithm: A standard method that selects basis functions based on maximizing a discriminant measure for each coefficient individually. The paper highlights its limitation in ignoring dependencies between coefficients.
  • Information Theory Filter Approaches: These methods use mutual information to quantify the relationship between features and the corrosion class, explicitly accounting for dependencies. Three specific criteria are explored:
  • MI Parzen: A density-based estimator of mutual information using Parzen windows.
  • MI knn: A distance-based estimator of mutual information using k-nearest neighbors.
  • Relevance-Redundancy (rel-red): A criterion that selects features highly relevant to the class while penalizing redundancy with already selected features.

A Sequential Forward Search (SFS) strategy is employed with these mutual information criteria to iteratively build a feature set.

Datasets

The study utilizes a dataset of acoustic emission signals collected from experiments simulating different corrosion types. The dataset includes:

  • 197 time series of "absence of corrosion."
  • 194 time series of uniform corrosion.
  • 214 time series of pitting.
  • 205 time series of stress corrosion cracking (SCC).

These signals were generated from experiments involving two types of steel (carbon steel and stainless steel) under various corrosive media and temperature conditions. The data was collected from a bypass of a chemical process plant, ensuring representative environmental conditions.

Key Insights and Ideas

  • Dependencies Matter: The paper strongly emphasizes that ignoring statistical dependencies between wavelet coefficients significantly degrades classification accuracy. This is a key insight that differentiates their work from prior methods like LDB.
  • Mutual Information as a Powerful Tool: Mutual information proves to be a robust criterion for feature selection because it inherently captures complex relationships and dependencies between variables.
  • Relevance-Redundancy Strikes a Balance: The relevance-redundancy criterion emerges as particularly effective, balancing the need for informative features with the avoidance of redundant ones, leading to the highest accuracies.
  • Feature Selection Can Outperform Using All Data: Surprisingly, using a carefully selected subset of features often leads to better classification performance than using the entire raw signal. This is attributed to the "curse of dimensionality" and the presence of noisy samples in the raw data.
  • Harmful Corrosions are Identifiable: The most dangerous corrosion types (pitting and SCC) can be identified with very high sensitivity, which is crucial for industrial safety.

Key Results

  • Accuracy Improvements: The proposed information theory-based feature selection methods significantly outperform the LDB algorithm and the "no feature selection" baseline.
  • The highest accuracy achieved was 86.4% using the normalized mutual information criterion (relevance-redundancy) with a naive Bayes classifier.
  • This is a substantial improvement over the LDB algorithm, which achieved a maximum of 71.9% accuracy with a k-nearest neighbor classifier.
  • Even the raw acoustic signals yielded a disappointing accuracy of 70.7%.
  • Classifier Performance: The study evaluated four classifiers (k-NN, decision tree, Gaussian Mixture Model, naive Bayes). The relevance-redundancy approach consistently yielded the best results across most classifiers.
  • Confusion Matrix Insights: The confusion matrix revealed that while pitting and SCC were identified with high sensitivity (93.0% and 96.6% respectively), absence of corrosion and uniform corrosion were more easily confused, likely due to their similar continuous emission signatures.

Limitations and Future Work

  • Steel Type as an Additional Variable: The authors note that the steel type is also indicative of corrosion and could be used as an additional discrete input variable for machine learning algorithms.
  • Chemical Composition: Exploring the use of chemical composition as continuous input variables is suggested, though this would require a larger dataset with diverse steel types.
  • Hybrid Filter-Wrapper Approaches: The paper mentions that a hybrid filter-wrapper approach is in preparation, which could potentially further enhance classification accuracy.
  • Stopping Rules for Feature Selection: While the study stopped feature selection at 50 features, they suggest that practical applications could benefit from automated stopping rules based on validation set performance.

Summary updated

BibTeX
@article{dijck2011information,
  author = {Gert Van Dijck and Marc M. Van Hulle},
  title = {{Information Theory Filters for Wavelet Packet Coefficient Selection with Application to Corrosion Type Identification from Acoustic Emission Signals}},
  journal = {Sensors},
  year = {2011},
  doi = {10.3390/s110605695},
}
Figure