Transferable common feature space mining for fault diagnosis with imbalanced data. (July 2021)

• Record Type: Journal Article
• Authors: Lu, Na; Yin, Tao
• Is Part Of: Mechanical systems and signal processing. Volume 156(2021)
• Publication Date: 2021-07
• Subjects: Fault diagnosis -- Transfer learning -- Imbalanced data -- Deep feature -- Classification; Structural dynamics -- Periodicals; Vibration -- Periodicals; Constructions -- Dynamique -- Périodiques; Vibration -- Périodiques; Structural dynamics; Vibration; Periodicals; 621
• DOI: 10.1016/j.ymssp.2021.107645
• Languages: English
• ISSNs: 0888-3270

Abstract

Highlights: To fully explore the health data, fault diagnosis is divided into two stages. A weakly supervised domain adaptive Autoencoder is developed to mine common feature. A unique feature branch is combined to form a dual channel module of relation computation. Few shot training strategy is employed to balance the training progress. Abstract: Many deep transfer learning methods for fault diagnosis have been proposed in this decade. Some of the existing methods focus on addressing the problem of fault data scarcity and fault knowledge transfer across different domains with different number of samples. There is still much room to improve considering the best performance so far on imbalanced and transfer fault diagnosis. The existing researches apply synthetic data generation, weighted sample or cost and transfer learning techniques to solve the problem. However, the synthetic samples might not follow the true fault data distribution or exploit excessively over the available small data which could lead to model bias or overfitting. In addition, the value of the abundant normal condition data has not been well explored which may carry essential information for fault discrimination. To address these problems, a novel two stage transferable common feature space mining method for fault diagnosis is developed which is termed as Common Feature and Compare Net (CFCNet). The fault diagnosis task has been divided into two stages, common feature learning and fault category diagnosis. In the first stage, CFCNet trains a weakly supervised domain adaptive convolutional Autoencoder to learn the common features underlying multi-domain data, which makes efficient use of all the available data and is termed as Common Feature Net. In the second stage, the trained Common Feature Net and a Unique Feature Net is combined to construct a dual-channel feature extraction and comparison architecture. CFCNet could mine both the transferable common features and unique features of different faults. Based on a feature concatenation and similarity computation structure, CFCNet enables an efficient similarity estimation mechanism for fault diagnosis. Training strategy of few shot learning is adopted to train CFCNet which can balance the training progress instead of the imbalanced data. Extensive experiments have verified the superior performance of the proposed method.