Industrial Fault Detection Employing Meta Ensemble Model Based on Contact Sensor Ultrasonic Signal.

Area of Science:

• Mechanical Engineering
• Signal Processing
• Machine Learning

Background:

• Ultrasonic diagnostics is crucial for early industrial fault prediction.
• Contact microphones used in current methods are susceptible to noise contamination.
• Existing methods require robust feature extraction and selection for accurate fault classification.

Purpose of the Study:

• To develop a noise-resilient and accurate fault detection system using ultrasonic signals.
• To explore advanced feature extraction and dimensionality reduction techniques.
• To implement and evaluate a stacking classifier for enhanced fault classification performance.

Main Methods:

• Feature extraction in time and frequency domains from ultrasonic signals.
• Dimensionality reduction using Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), and t-distributed Stochastic Neighbor Embedding (t-SNE).
• Feature selection via Recursive Feature Elimination (RFE) and classification using k-Nearest Neighbor (KNN), Logistic Regression (LR), Decision Tree (DT), Gaussian Naive Bayes (GNB), and Support Vector Machine (SVM).
• Development of a stacking classifier with k-fold cross-validation for performance assessment.

Main Results:

• The proposed stacking classifier achieved approximately 5% higher accuracy across five cross-validation folds compared to individual models.
• The method demonstrated minimal variation in performance, indicating robustness.
• Real-time monitoring feasibility was confirmed with an execution time of 11 ms on a Cortex M4 microcontroller.

Conclusions:

• The developed stacking classifier offers a significant improvement in ultrasonic fault detection accuracy and reliability.
• The method effectively handles noisy data and reduces dimensionality for efficient processing.
• The system's low execution time makes it suitable for real-time industrial monitoring applications.