SUMMARY

Prognostics and health management (PHM) for complex industrial systems is required for cost reduction, maintenance schedules, and reducing system failures. Catastrophic failure usually causes significant damage and may cause injury or fatality, making early and accurate diagnostics of paramount importance. As important parts of industrial systems, the reliability and availability of critical components and machinery are essential to ensuring safe and continuous operation in modern industry. With the development of sensor and information technologies, multisensory systems are widely used in the modern industry, which use multiple measurements (e.g., vibratory, electrical, thermal, acoustic, and oil-based data, etc.) to adequately reflect the health condition of critical machinery. As a result, it is necessary to develop intelligent and accurate PHM methods for multisensory systems. However, the existing deep learning (DL)-based PHM methodologies still have some critical limitations: 1) the majority of the existing PHM methodologies were vibration-based methods and did not consider the integration of information across multiple measurements; 2) critical machines usually operate under normal condition while abnormal conditions rarely happen, leaving only a small or limited data of abnormal conditions. Also, modern industrial equipment generally operates in complex and varying conditions, which causes the cross-domain problem and substantially degrades the identification performance of DL-based models; 3) when new health conditions occur in testing samples, most DL-based approaches may misclassify samples into the existing types of health conditions defined from the training data, which will cause open-set misclassification problems in the DL model. To address these issues, a deep meta-learning framework for cross-domain multisensory systems is investigated, which uses information fusion strategy, meta-learning, and convolution neural network (CNN). First, to enhance the utilization of multisignal data in CNN-based models, a multisignals-to-RGB-image conversion method is proposed for feature-level information fusion, which uses principal component analysis (PCA) to fuse multisignal data into three-channel red-green-blue (RGB) images for further identification. Second, a novel method called information fusion-based meta-learning (IFML) for cross-domain few-shot problems is proposed, which can not only be generalized to identify abnormal conditions with limited labeled samples, but also be used for different cross-domain scenarios. Third, the deep meta-open generative adversarial network (DMO-GAN) is proposed to enhance model generalization and solve open-set misclassification problems. The effectiveness and generalization of the proposed framework are validated in several industrial datasets consisting of different multisensory systems.