What does this research mean for the field?

The novel hybrid deep learning model achieves 90% to 92% accuracy in detecting minor blade faults in gas turbines, even in noisy environments. Novelty: ClaimNovelty.NOVEL_FINDING. Consensus alignment: ConsensusAlignment.CHALLENGES_CONSENSUS.

What question did this study set out to answer?

This research focuses on developing an advanced model for detecting minor blade faults in gas turbines using deep learning techniques.

March 10, 2026Open Access

A Novel Hybrid Deep Learning Approach Using Convolutional Neural Network–Transformer With Cross‐Attention Mechanism for an Effective Minor Blade Fault Detection

Puntos clave

This research focuses on developing an advanced model for detecting minor blade faults in gas turbines using deep learning techniques.
Developed an unsupervised hybrid deep learning model integrating convolutional neural network and transformer architecture.
Utilized a cross-attention mechanism and autoencoder structure for feature extraction and signal processing.
Conducted experiments on a multistage rotor system designed in the laboratory.
Achieved fault prediction accuracy between 90% and 92% at a signal-to-noise ratio of 15 dB.
Demonstrated capability to accurately identify minor blade faults in noisy conditions.
Showed improved efficiency and reliability in fault detection compared to traditional methods.

Resumen

Blade faults are widely regarded as one of the most frequent failures in gas turbines, and if left unchecked, they can result in catastrophic failure. To ensure the safe operation of these machines, it is essential to detect minor blade faults and monitor blade health regularly. In recent years, machine learning and artificial intelligence have been proposed as effective tools for machinery fault diagnosis. However, when features are manually extracted, some important and vital features may be overlooked, leading to unreliable results. This paper presents a novel unsupervised hybrid deep learning model based on convolutional neural network and transformer with cross‐attention and autoencoder structure to address these challenges in blade fault detection. The developed method has adaptive learning capabilities for automatic feature extraction and signal processing with minimal human intervention and was found to be more efficient in terms of the accuracy of fault prediction. The study was conducted on a multistage rotor system designed and developed within our laboratory, and the results revealed that the developed model can achieve an impressive fault prediction accuracy ranging from 90% to 92% at a signal‐to‐noise ratio of 15 dB, which is a challenging level of noise for various stages and fault combinations indicating that the model is capable of accurately identifying minor blade faults, even in the presence of significant noise interference. This approach revolutionizes blade fault detection in complex rotor systems, offering superior reliability and efficiency, even in the absence of labeled data and noisy environments.

Leer artículo completoexternamente

Me gusta

Guardar

Ver artículo completo