A Deep Autoencoder-Based Convolution Neural Network Framework for Bearing Fault Classification in Induction Motors.

Area of Science:

• Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Machine fault diagnosis is crucial for industrial condition monitoring.
• Data-driven methods offer advanced fault classification but require quality features and labeled data.
• Motor current signals are accessible but often noisy, complicating feature extraction.

Purpose of the Study:

• To propose a novel deep auto-encoder (DAE) and convolutional neural network (CNN) based model for bearing fault classification in induction motors.
• To leverage non-invasive motor current signals for fault detection.
• To address challenges posed by noisy industrial data and limited labeled datasets.

Main Methods:

• Utilized a deep auto-encoder (DAE) to estimate the system's nonlinear function from normal operating data, generating residual signals.
• Employed a convolutional neural network (CNN) to classify bearing fault types based on these residual signals.
• Implemented a semi-supervised learning approach for enhanced fault classification.

Main Results:

• Achieved a classification accuracy exceeding 99% for bearing faults.
• Demonstrated that the DAE significantly improves classification accuracy, especially with small amounts of labeled data.
• The combined DAE-CNN model's performance is comparable to existing methods on the same dataset.

Conclusions:

• The proposed DAE-CNN model is an effective approach for bearing fault classification using motor current signals.
• The method shows promise for robust condition monitoring in industrial settings.
• The semi-supervised approach enhances accuracy and data efficiency in fault diagnosis.