Adaptive Knowledge Transfer by Continual Weighted Updating of Filter Kernels for Few-shot Fault Diagnosis of Machines
Abstract : Deep learning (DL) based diagnosis models have to be trained by large quantities of monitoring data of machines. However, in real-case scenarios, machines operate under the normal condition in most of their life time while faults seldom happen. Therefore, though massive data are accessible, most are data of the normal condition while fault data are still extremely limited. In other words, fault diagnosis of real machines is actually a few-shot diagnosis problem. To deal with few-shot diagnosis, this paper proposes adaptive knowledge transfer with multi-classifier ensemble (AKTME) under the paradigm of continual machine learning (CML). In AKTME, knowledge learned by DL models is considered to be represented by the learnable filter kernels (FKs). The key of AKTME is a proposed continual weighted updating (CWU) technique of FKs. By CWU, shared FKs are distilled from multiple auxiliary tasks and adaptively transferred to the target task. Then by multi-classifier ensemble, AKTME is able to recognize faults with few fault data accessible. AKTME is applied on two few-shot diagnosis cases. Results verify that AKTME achieves higher diagnosis accuracies than recently proposed methods. Moreover, AKTME tends to improve the diagnosis accuracy as it pre-learns on more auxiliary tasks continually.
? Many of the existing intelligent data-driven fault diagnosis techniques in the literature struggle to adapt to different working conditions (such as different motor loads, different rotating shaft speeds, or varying amounts of environmental noise).
? In this paper, we propose a novel deep learning model for fault detection and diagnosis of time-series data that addresses many of the problems with existing intelligent data-driven fault diagnosis techniques.
? The proposed novel dual-path deep learning model obtains state-of-the-art classification performance, while also addressing many of the problems with existing intelligent data-driven fault diagnosis techniques such as being able to operate on raw time series data (with no need for complex feature engineering) and robustness to noise and changes in operating conditions
? Researchers therefore attempted to use domain adaptation (DA) to solve the problem. DA can be regarded as a special case of transfer learning, which aims to transfer shared knowledge across different but related domains.
? However, DA methods still suffer from some obstacles in solving the cross-domain fault diagnosis problem.
? For this motivation, this study introduces domain generalization (DG) into cross-domain fault diagnosis problems to remove the dependency on target domain data.
? By contrast, A6 performs worse than ADIG, which again proves the importance of the IN strategy in the generalization problem.
• We show that our proposed framework works directly on the raw temporal data obviating the need for manual feature extraction or noise removal.
• In this paper, a novel dual-path deep learning network, combining a recurrent neural network path and a deep convolutional network path (RNN-WDCNN) is proposed and applied to fault diagnosis for rolling element bearings.
• The deep convolutional path of the proposed model, therefore, consists of five convolutional stages for feature extraction and a final dimension reduction stage to compress the learned feature representations.
• With an inference time of 3.53 ms, this would allow the proposed RNN-WDCNN intelligent fault diagnosis system to process the vibration data online in real-time.
? The discrepancy between the source and target domains causes a domain shift through the different working conditions, where the performance of the diagnostic model degenerates when the model is trained by the source domain but is used as an inference engine in the target domain.
? Besides, an adaptive weight strategy achieves weight self-learning during multitask learning to improve performance.
? Comprehensive experiments based on two case studies are conducted to evaluate the performance of ADIG.
? To balance the multitask loss dynamically, a weight coefficient learner is proposed in this work to achieve adaptive weights.
? Modern industries are moving toward informatization and intelligentization in the fourth industrial revolution era, and modern machinery and equipment are widely used in various fields, such as construction, aviation, electric power, and metallurgy.
|