Data-driven framework for failure prediction and classification in industrial machines

Abstract

In this work, an advanced deep learning-based hierarchical diagnosis framework for industrial machinery is presented. The proposed Hierarchical Fault Diagnosis Network (HFDN) model uses a layered architecture to decompose complex systems into manageable subsystems, enhancing accuracy and efficiency in fault detection, identification, and severity assessment. The core concept is to initially detect faults at a higher level and subsequently refine the diagnosis by progressively drilling down into lower levels i.e., fault identification and severity level assessment, thereby facilitating a more accurate and efficient fault detection process. The method integrates Empirical Mode Decomposition (EMD) for signal preprocessing with an attention-induced CNN utilizing quadratic neurons. EMD breaks down complex signals, while attention mechanisms highlight critical features for precise fault detection. By incorporating noise signals, the study addresses practical challenges in industrial fault diagnosis. To ensure a fair and comprehensive evaluation, the experimentation was carried out under varying Signal-to-Noise Ratio (SNR) conditions. Extensive testing across different fault classes and one non fault case demonstrates the model’s robustness, achieving an accuracy of 99.74% at 0 dB. This approach improves the fault diagnosis performance, enabling early and accurate fault detection, thereby optimizing maintenance strategies and reducing industrial downtime. Experiments were conducted on Case Western Reverse University rolling bearing dataset. The findings indicate that our model outperforms several established deep network models and achieves a high accuracy rate.