A semi-supervised fault diagnosis method for ventilation fan using multi-head attention-enhanced generative adversarial network

Ventilation fans are widely used in industry factories and commercial buildings. After experiencing a long-time operation, fans are prone to abnormalities resulting in system performance degradation, energy waste, and even safety issues. Recent works have shown that the machine learning-based techniques outperform most of the traditional vibration signal-based diagnostic method. However, the insufficient number of fault training samples has become the main obstacle for the supervised fault diagnosis. Therefore, in order to improve the fault diagnosis performance with imbalanced training dataset, this paper reports a multi-head self-attention enhanced semi-supervised generative adversarial network (MSA-SGAN) method for ventilation fans. The original unsupervised GAN was improved to a semi-supervised GAN (SGAN) and thus the ability of multi-classification could be achieved. In addition, the SGAN was enhanced by integrating multi-head self-attention (MSA), which allows for increased emphasis on relevant and significant features. An experimental system was established, and different types of fan fault were simulated. Using the experimental data, the fault diagnosis model based on the proposed MSA-SGAN was trained and validated. Results showed that the proposed fault diagnosis method exhibited an excellent performance including overall accuracy, recall, and precision as compared to the other traditional methods. In the case of the imbalanced dataset, the proposed method shows superior performance compared to other traditional supervised and semi-supervised methods.