Haitao Li, Junjie Chen, Jingli Yang, Qingyao Ai, Wei Jia, Youfeng Liu, Kai Lin, Yueyue Wu, Guozhi Yuan, Yiran Hu, Wuyue Wang, Yiqun Liu, Minlie Huang. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Due to the high-temperature, high-humidity, and high-salt-fog conditions of the marine environment in which ship gearboxes operate, their surfaces are highly susceptible to corrosion. When corrosion reaches a certain level, it may affect the normal functioning of internal gears and even lead to unpredictable damage to the entire system. This paper focuses on the gearbox surface as a critical inspection area and proposes an ultrasonic corrosion defect detection method based on a teacher-student network architecture. The teacher-student network is trained using non-corroded signals along with various simulated transformed versions of these signals. By relying solely on normal samples for training, the model leverages the student network's inability to replicate anomalies for defect detection. The detection threshold is determined using a quantile-based method, which is robust against extreme outliers and offers greater flexibility in controlling the false alarm rate. Experimental results demonstrate that, compared to existing unsupervised detection methods, the proposed approach achieves significantly higher accuracy and effectively identifies corrosion defects on gearbox surfaces, thereby enhancing the reliability and safety of marine propulsion systems.

In modern industry, the operational status of rotating machinery has a significant impact on both production safety and economic efficiency. Deep learning-based intelligent diagnosis methods have become effective tools for fault detection. However, under complex operating conditions, single-modality diagnostic approaches often lack robustness and generalization ability. As a result, multimodal information fusion has emerged as a popular research direction. To address this issue, a multimodal cross-fusion network (MCFN) integrating attention mechanisms and a cross-guided fusion strategy is proposed. First, an improved bi-directional long short-term memory network is employed to extract temporal features from one-dimensional modality data, and a residual neural network (ResNet) captures spatial features from two-dimensional modality data. Then, a self-attention mechanism is introduced to enhance the resolution of spatiotemporal features. Finally, a cross-guided fusion module is proposed to effectively integrate multimodal information via feature interaction modulation and normalization. During model optimization, a focal loss and label smoothing strategies are introduced to improve generalization on imbalanced data. Experiments on the PU dataset involving 13 classes of bearing faults demonstrate that MCFN achieves an average testing accuracy of 97.18%, significantly outperforming comparison algorithms.

The industrial Internet of Things (IIoT) significantly enhances mechanical fault diagnosis. However, IIoT-based intelligent diagnostic models struggle with sample insufficiency and high energy consumption due to collection costs and limited computing resources. Therefore, this article proposes an energy-efficient mechanical fault diagnosis method based on the neural-dynamics-inspired metric SpikingFormer (MSF) to achieve accurate fault recognition under insufficient samples. The design and construction of a data acquisition system based on the aircraft engine platform and the ship water jet propulsion platform effectively support the operation of the developed diagnostic algorithm. Specifically, an event-driven multiscale mask spiking self-attention (MMSSA) mechanism is designed to focus critical spatiotemporal features from different scales under low computational complexity. Meanwhile, a rate encoding metric classifier (REMC) is constructed to bridge spiking learning and prototype representation, thereby accurately classifying fault under insufficient samples. Finally, a customized backpropagation strategy based on neural dynamics is developed to enable the MSF to learn effectively and be stable. The superiority of the MSF in energy consumption and diagnostic accuracy is verified through comparison with six authoritative methods across standard, laboratory-acquired, and real-world datasets. The results showed that the parameter count of MSF is$7.04\times $and$20.46\times $less than the strong baseline method, respectively, and the diagnostic accuracy on the two real datasets is 4.52% and 6.91% higher than the latest method, respectively.

To alleviate data distribution under different operating conditions, domain generalization (DG) has been applied in mechanical diagnosis. Still, its effectiveness is limited when unknown fault states appear in the target domain. Consequently, open set DG (OSDG) has emerged to identify unknown classes in unknown domains. However, data collection costs and safety concerns have resulted in a significant class imbalance in OSDG. This imbalance causes the decision boundary to be skewed toward abundant positive classes, ultimately leading to misclassifying unknown states and increasing security risks. Currently, there is a lack of methods to simultaneously address domain shift and class shift in an imbalanced unknown domain. To tackle this issue, this article proposes a multisource domain-class gradient coordination meta-learning (MDGCML) framework, which can learn the generalized boundaries of all tasks by coordinating gradients between interdomains and interclasses. Based on the MDGCML, a joint learning paradigm involving the sharing of parameters between open-set classifiers and closed-set classifiers is constructed to enable quick adaption of the model to unknown domains. The superior performance of the proposed framework has been verified on two datasets.

The dynamic fault environment, incremental data accumulation, and specific needs in train transmission systems make continual learning essential for fault diagnosis. Recent advancements in continual learning have improved diagnostic adaptability, but current methods face challenges: 1) Complex architectures to prevent catastrophic forgetting increase training difficulty and computational costs, hindering deployment. 2) Lack of new class samples leads to delayed model evolution due to long sample accumulation periods. This article introduces a lightweight continual learning method based on model expansion. A hash space metric mechanism using lightweight convolution is developed to reduce computational costs while maintaining accuracy. Additionally, a joint strategy of knowledge enhancement and compression improves model evolution by refining knowledge subsets. Compared with the state-of-the-art method, the proposed method reduces parameters by 2.59 times, FLOPs by 2.67 times, and inference time by 2.89 times, and leads by 2.5% and 0.23% in incremental accuracy and incremental forgetting rate, respectively.

Aircraft engines operate under more demanding and unique environments, which require the inner components to be able to withstand extreme conditions. Intershaft bearings serve as the critical part of power transmission. Therefore, their accurate and reliable fault diagnosis is of paramount importance to ensure secure and dependable functioning of the engine. In this field, scarcity of labeled fault data owing to high collection costs is a common challenge. To address this, this article proposes a semi-supervised cross-domain diagnostic method for aircraft engine intershaft bearings, utilizing a virtual domain-driven approach to achieve high accuracy with limited labeled data. Specifically, a dynamics-based simulation model is developed to generate source domain data, reducing the dependency of deep learning models on experimental platforms and lowering platform construction costs. Additionally, a hyperbolic geometric metric learning strategy is designed to capture hierarchical features in high-dimensional data, which handles the correlation between different fault types and enhancing classification accuracy. Furthermore, a domain-class adversarial decoupling mechanism is developed to mitigate the domain bias, enabling the precise representation of fault modes and maximizing the utility of unlabeled virtual domain data. Using datasets from both real-world aircraft engine scenarios and public resource experiments validate the proposed method, illustrating its superior performance compared to state-of-the-art techniques on public domain benchmark datasets.

With the rapid development of industrial intelligence, the domain adaptation technology depending on the availability of the target domain has gradually become a reliable solution to address the performance degradation of fault diagnosis due to the variation of operating conditions for rotating machinery. However, the fault diagnosis scenarios under unseen operating conditions are commonly confronted in practical engineering. In response to the above limitations, a novel multi-source domain generalization network (MDGN) is proposed in this paper to enhance the fault diagnosis performance of rotating machinery under unseen operating conditions by decoupling the domain-relevant feature information for obtaining domain-invariant diagnosis knowledge. First, a temporal convolutional autoencoder (TCAE) is developed as a feature extraction module to fully capture the effective spatiotemporal information of the sample data. Then, a domain classification module is designed for adversarial transfer learning to facilitate the acquisition of domain-invariant features. Finally, domain divergence loss and boundary margin loss are constructed for MDGN to supervise the feature distribution alignment among domains and feature distinction among classes. The mechanical comprehensive diagnosis simulation platform (MCDSP) bearing dataset collected in our laboratory is employed to evaluate the domain generalization performance of the proposed method. The experimental results demonstrate that the method achieves an average fault diagnosis accuracy of 94.50%, which is better than the comparison algorithms.

To solve the problem of nonlinear factors in the fault detection process of complex systems, this article proposes a fault detection model based on vector quantization sparse autoencoder. First, a feature extraction model, which consists of a self-normalizing convolutional autoencoder module, a vector quantization module, a gradient module, and a loss module, is developed. The first module employs self-normalizing convolutional layers with good stability and generalization ability to extract the nonlinear structural features of complex systems. A nearest neighbor search strategy is implemented in the vector quantization module to further mine the nonlinear information. The gradient module adopts a straight-through estimation technique to improve the training efficiency. Sparse constraints are introduced into the loss module to obtain the essential features and enhance interpretability. Thereafter, a construction rule based on local Mahalanobis distance and K nearest neighbors is designed to calculate K Mahalanobis neighbor metrics that depend on the sparse features obtained by the feature extraction model. A comprehensive statistic for fault detection is constructed to accurately track the operating status of complex systems by combining the loss metric and the K Mahalanobis neighbor metric. Finally, the threshold of the fault detection statistics is determined by modeling the generalized extreme value distribution. Three case studies, a numerical simulation, the Tennessee Eastman benchmark process, and a typical circuit system, are adopted to demonstrate the effectiveness and merits of the proposed fault detection model.

The content-based publish/subscribe systems are attracting more and more attention in Internet applications due to their intrinsic time, space, and synchronization decoupling properties. With the increase in system scale, the efficiency of event matching becomes more critical for system performance. However, most existing methods suffer significant performance degradation when the system has large volumes of subscriptions. This paper presents DOCO (DOuble COmbination event matching algorithm) to improve the efficiency of event matching in content-based publish/subscribe systems. Via assembling the attributes in the attribute space by pairs, a novel index structure is built up for classification of the subscriptions. On the arrival of an event, the event matching process is only carried out on some related units of the index structure, thus, the number of subscriptions that involved in the event matching process is reduced. A series of experiments are designed to verify the performance of the proposed algorithm, and a comparison with other event matching algorithms is also carried out. The experimental results show that DOCO can improve the efficiency of event matching in content-based publish/subscribe systems.

Traffic matrix is a vital performance parameter for network management and optimization, thus it is in great need to achieve the traffic matrix accurately. Network tomography is a commonly adopted framework to estimate traffic matrix based on link loads in real networks. Since the model of network tomography always behaves the ill-posed characteristic, which means the traffic matrix estimation under network tomography framework is still a major challenge. To address this problem, a novel approach named ATME is presented. ATME can reduce the reconstruction errors of traffic matrix by using the criteria of TM's sparsity on each time slot. Besides, a prediction method based on grey predictive model is used to update the approximate value of the negative entries achieved by orthogonal match pursuit algorithm. Experimental results demonstrate that ATME is adaptive for initial value of sparsity, and can also obtain a higher accuracy on traffic matrix estimation.

Following the successful deployment of the Earth Observing System Data and Information System (EOSDIS), NASA's Earth Science Enterprise (ESE) is developing a Strategy for the Evolution of ESE Data Systems (SEEDS). SEEDS seeks to coordinate best practices in data system technologies throughout NASA's Earth Science Enterprise, while deferring decisions on implementation to individual projects within the Enterprise. To meet these potentially contradictory goals of technology coordination and implementation independence, SEEDS would employ processes based on community moderated input, review, and consensus. This paper outlines a possible SEEDS standards process, as proposed by the authors and refined in a series of NASA-sponsored public workshops.