Enhanced Blade Fault Diagnosis Using Hybrid Deep Learning: A Comparative Analysis of Traditional Machine Learning and 1D Convolutional Transformer Architecture

Abstract

ABSTRACT Artificial intelligence offers a promising solution for the precise identification of faults in rotating machinery. The severe repercussions of turbomachinery blade failures, including fatalities and extensive damage, necessitate robust diagnostic tools. Early fault detection and diagnosis are vital and significant concerns for preventing these incidents and are particularly crucial in gas turbines and compressors to avoid costly downtime and maintain optimal plant performance. The financial consequences of unplanned downtime due to blade failures can be substantial, leading to loss of production, costly repairs, and potential legal liabilities. Effective fault diagnosis (FD) plays a key role in mitigating these financial liabilities by minimizing downtime and facilitating optimized maintenance planning. By investigating blade fault patterns and using appropriate diagnostic techniques, it becomes possible to predict potential failures and schedule maintenance proactively. This approach reduces operational failure and extends the lifespan of the equipment. Diagnosing blade failure is more challenging than bearing and gear faults, which exhibit standard fault characteristics observable in the time and frequency domain. Noise and complex design in multistage rotors can mask blade faults in vibration signals, necessitating automated feature extraction and expert diagnosis. This research investigates blade FD, comparing traditional machine learning approaches with a novel hybrid deep learning fused model based on a one‐dimensional (1D) convolutional transformer. Tested on an in‐house fabricated multistage rotor, the hybrid model demonstrated exceptional diagnostic accuracy, exceeding 93% for various fault scenarios. This represents a significant enhancement over existing traditional methods, which achieved 49.81%–86.75% accuracy, and also shows appreciable improvement compared with established artificial neural networks, which typically range from 88.43% to 90%. This enhanced performance was achieved with minimal human intervention and without complex signal processing. The Implementation of this approach within complex rotor systems offers a significant improvement in both the efficiency and reliability of blade FD.

Keywords

• 1D convolutional neural network
• automatic feature extraction
• blade faults
• fault classification
• hybrid deep learning
• traditional fault diagnosis