A fault location method based on ensemble complex spatio-temporal attention network for complex systems under fluctuating operating conditions

Abstract

Recently, traditional fault diagnosis methods achieve fault identification by establishing a sample set covering all fault degrees of different fault components in complex systems. However, the uncertainties including environmental stresses and own physical and chemical variations lead to an infinite number of fault degrees of fault components, and the fault identification of complex systems in practical engineering applications faces the challenge of fluctuating operating conditions due to variations of rotational speed and load. To address the above problem, a fault location method based on ensemble complex spatio-temporal attention network (ECSAN) is proposed in this paper, which can identify the critical fault components of complex systems by combining an ensemble learning mechanism with excellent basic estimators. A basic estimator consists of a feature extraction module, a feature enhancement module, and a classification module. In the first module, a complex spatio-temporal backbone network with strong generalization is developed to provide spatio-temporal features containing the inherent information of the sample data for complex systems. In the feature enhancement module, a lightweight complex attention layer is constructed to enhance the effective structural information of the features and reduce the interference of their redundant information. The classification module then adopts a Softmax layer to perform fault classification. Finally, an ensemble learning mechanism is designed to integrate the basic estimators. By constructing sample weights and introducing a knowledge transfer strategy, the generalization is further improved while saving training expenses. Two datasets from different experimental platforms are concerned to verify the effectiveness and superiority of this method under various operating conditions and fault degrees. The experimental results indicate that this method achieves 99.81% accuracy on a standard dataset of the mechanical system and 98.88% accuracy on a real dataset of the closed-loop control system of the water jet propulsion device, which is superior to comparison approaches.