A Comprehensive Approach to Gear Fault Classification Leveraging Vibration Data and Neural Networks

Detecting gear faults is crucial for maintaining the reliability and efficiency of machinery across industries such as manufacturing and transportation. Early and precise fault detection minimizes downtime, lowers maintenance costs, and prolongs equipment life. This study presents a comparative evaluation of deep learning models designed to classify gear damage levels using vibration data, with the goal of identifying the most effective architecture for this application. Among the tested models, a custom Multi-Layer Perceptron (MLP) demonstrated superior performance, achieving an accuracy of 97.08%, along with excellent precision, recall, and F1 scores. This highlights the MLP's ability to extract critical features from vibration signals while maintaining computational efficiency, making it a viable choice for real-time fault detection. Sequence-based models, such as BiLSTM and CNN_LSTM, also performed well, showcasing their capability to process temporal patterns inherent in vibration data. In contrast, the CNN model yielded lower performance, indicating that relying solely on spatial feature extraction may not be sufficient for sequential data analysis. The custom MLP model's effectiveness underscores its potential for practical deployment in predictive maintenance, offering a reliable and accurate solution for gear fault classification. By enabling real-time monitoring and fault diagnosis, this approach can significantly enhance operational efficiency and reduce maintenance expenses in industrial environments. The findings of this study demonstrate the promise of leveraging deep learning, particularly the custom MLP model, to address the challenges of gear fault detection and contribute to more sustainable and cost-effective machinery management practices.