Fault diagnosis of gas turbine generator bearings using enhanced valuable sample strategy and convolutional neural network

Abstract

Abstract Gas turbine bearings operate continuously under complex and harsh conditions such as high temperatures, high pressures and high speeds. Bearing fault monitoring data often exhibits anomalies, noise, missing values, and strong coupling and non-linearity due to real-world random factors. In addition, the traditional convolutional neural network (CNN) is still limited by the scarcity of labeled samples in real-world conditions and cannot fully extract fault features. To address the complexities of strongly coupled fault data, inconsistent data quality, and the shortage of authentic labeled samples, this paper presents a novel method for bearing fault diagnosis. This method utilizes an enhanced valuable sample strategy in conjunction with CNNs, integrating data enhancement and active learning (AL) for preprocessing to create valuable training sets for the networks. At the outset, random dropout and scaling jitter operations are applied to the original data, effectively removing anomalies and gaps in the fault signals. This process simultaneously emulates variations in the frequency spectrum and amplitude of vibration signals under real operating conditions, thereby enhancing the quality of gas turbine bearing fault data. Subsequently, the method incorporates AL techniques, iteratively selecting unlabeled data samples with the highest information value, and annotating them using support vector machines. This strategy creates a valuable training dataset that further enhances the fault diagnosis capabilities of the model. In the subsequent phase, a three-dimensional CNN is employed to extract fault feature information efficiently from the fault data, culminating in precise fault classification. In order to validate the effectiveness and superiority of the algorithm, validation and comparison analyses were carried out using the Case Western Reserve University and XJTU open datasets. To evaluate the practical effectiveness, a gas turbine main bearing dataset collected in a real environment, named the BaiChuan dataset, is used for the engineering validation of the proposed method.