In modern industry, the Industrial Internet of Things (IIoT) has enabled system health analytics and monitoring through continuous data collection and networked connectivity. Bearing remaining useful life (RUL) prediction is one pivotal analytical task in preventing industrial system failures and optimizing maintenance schedules. Existing prediction methods using data face two critical engineering challenges: (1) performance degrades when deployed to unseen operational domains, and (2) imbalanced sensor data distributions cause biased predictions. Two challenges can even co-exist, which further limits the effectiveness of prediction methods using data. This study proposes a multi-task pointwise mutual information learning (MPML) based prediction model development method to tackle bearing RUL prediction under this compound challenge. MPML offers several key innovations through the following developments. First, an auxiliary task-assisted multi-task model learning scheme is devised to obtain task-wise generalizability for learning invariant latent features, and theoretical analysis is provided to explain the invariant feature learning mechanism. Furthermore, pointwise mutual information (PMI) modeling with a statistical explanation is proposed to impose adaptive biases, rectifying penalties for RUL misprediction in minority groups. Consequently, MPML addresses the unseen domain prediction through the multi-task learning scheme and effectively handles the data imbalance with the novel PMI-assisted loss. Extensive computational experiments are conducted to demonstrate the superiority of MPML, achieving a 33.12 square error compared to state-of-the-art methods.

In machine fault diagnosis, conventional data-driven models trained by empirical risk minimization (ERM) often fail to generalize across domains with distinct data distributions caused by various machine operating conditions. One major reason is that ERM primarily focuses on informativeness of data labels and lacks sufficient attention on invariance of data features. To enable invariance on top of informativeness, a learning framework, learning invariant features via in-label swapping for generalizing out-of-distribution (Lifeisgood), is proposed in this study. Lifeisgood is inspired by a simple intuition that invariance can be assessed by checking changes in loss due to swapping certain entries of features with the same labels. Lifeisgood also enjoys a theoretical guarantee on improving testing domain performance under certain conditions based on a swapping 0-1 loss proposed in this work. To circumvent the training difficulties associated with the swapping 0-1 loss, a swapping cross-entropy loss is derived as a surrogate and theoretical justifications for such a relaxation are also provided. As a result, Lifeisgood can be employed conveniently to develop data-driven fault diagnosis models. In the experiments, Lifeisgood outperformed the majority of state-of-the-art methods in terms of average accuracy and exceeded the second-best by 25% in terms of the frequency of beating the generic ERM. The code is available at: https://github.com/mozhenling/doge-lifeisgood.

Empirical risk minimization (ERM) is a celebrated induction principle for developing data-driven models. However, ERM has received both pros and cons for its capability on domain generalization (DG). To this end, this paper attempts to study the success and failure of ERM at supervised DG classification tasks, both theoretically and empirically, with causal perspectives. In the theoretical aspect, we first explore different properties of a causal metric termed information flow, followed with discussing relationships between the information flow and the mutual information in the proposed causal graph. Next, we analyze the roles of the transformed causal feature and the transformed spurious feature on modeling performances. It reveals that the interaction between the spurious influencer and the transformed causal feature is the key determining the failure or success of ERM on DG. In the empirical study, we first simulate various DG settings based on the MNIST, Fashion MNIST, and CIFAR10 datasets. Next, we verify developed theories by testing three different neural network configurations in designed experiments. In addition, experiments based on real-world datasets are conducted to further consolidate key points of the proposed theories. To extend application benefits of the theoretical discoveries, a new risk minimization framework with a novel feature intervention for regulating ERM is proposed. It achieves DG improvements over ERM on real-world datasets of image segmentation, image classification, and text classification.

Incorrect labels as well as the discrepancy between training and test domain data distributions can significantly affect the effectiveness of supervised data-driven models in machine fault diagnosis applications. Such a challenge can be characterized as the noisy label-domain generalization (NL-DG) problem. In this article, the extended invariant risk minimization (EIRM) is developed, which incorporates flat minima seeking to address the NL-DG challenge. The ability of handling NL-DG is realized by shifting the gradient penalty base from the dummy classifier to the entire model. EIRM is shown to be closely related to locating a flat minimum, which is crucial for label noise (LN) robustness and model generalization. Explorations on function smoothness and algorithm convergence are offered to understand EIRM from the theoretical aspect. An efficient implementation of EIRM is also developed to construct the fault diagnosis model. The EIRM-based fault diagnosis method is compared with strong benchmarks on multiple NL-DG tasks using actuator and gearbox fault datasets. Results indicate that the EIRM-based method on average is more effective than the benchmarks. The code is available at https://github.com/mozhenling/doge-eirm.

Industrial Internet of Things (IIoT) connects machines, and it is important to build intelligent models to prevent machine failures by identifying incipient faults. To develop intelligent fault diagnosis models, empirical risk minimization (ERM)-based modeling paradigm has been prevalently applied. However, during model training, ERM primarily focuses on instance-to-prototype (ItP) distances from a prototypical perspective, which may limit its effectiveness in analyzing data of diverse distributions. To improve the ERM model, we propose considering additional instance-to-instance distances (ItI) and prototype-to-prototype (PtP) distances, leading to a new modeling framework—distance-aware risk minimization (DARM). To gain awareness of extra types of distances, two novel losses are proposed based on reformulations of soft-max cross entropy. Theoretical explorations are conducted to justify the significance of collectively considering ItP, ItI, and PtP distances. Methodologically, DARM can jointly minimize three types of distance-aware losses to train neural networks for fault diagnosis in the same fashion as ERM. In a comprehensive computational study, DARM consistently outperformed ERM in domain generalization (DG) tasks based on various machine fault diagnosis data sets. In addition, DARM has superior performance over several recent DG methods. The code is available athttps://github.com/mozhenling/doge-darm.

As a classical and crucial component in industrial systems, the manipulators are widely employed in precision manufacturing scenarios because of their advantages of high stiffness, large load support capability, and high precision. During their service, it is inevitable that they encounter data imbalance scenarios due to the occasional and low-frequency failure behaviors. But in order to address these issues, the majority of the approaches already in use need the assistance of extra tools. Thus, a novel intelligent health state diagnosis model, named multiple neighbor homogeneous property-embedded graph auto-encoder (MNHP-GAE), is developed to get around this restriction and apply it to the manipulators. Its core is to realize the expansion and enrichment of the feature space by mining effective complementary information from homogeneous property samples without the assistance of data augmentation and other technologies. Specifically, the wavelet decomposition reconstruction and dynamic time warping are integrated to promote the quantification of the sample similarity and enable the construction of homogeneous property graph samples. Following that, a unique graph auto-encoder module with the multi-head attention mechanism is constructed to extract complementary information from homogeneous property nodes and match it for diagnostic tasks. Finally, through a multi-case experimental validation scenario constructed by a 3-PRR planar parallel manipulator experimental platform, the superior performances of the proposed MNHP-GAE model in highly unbalanced scenarios are fully demonstrated.

Machine learning has been widely applied to study AI-informed machinery fault diagnosis. This work proposes a sparsity-constrained invariant risk minimization (SCIRM) framework, which develops machine-learning models with better generalization capacities for environmental disturbances in machinery fault diagnosis. The SCIRM is built by innovating the optimization formulation of the recently proposed invariant risk minimization (IRM) and its variants through the integration of sparsity constraints. We prove that if a sparsity measure is differentiable, scale invariant, and semistrictly quasi-convex, the SCIRM can be guaranteed to solve the domain generalization problem based on a few predefined problem settings. We mathematically derive a family of such sparsity measures. A practical process of implementing the SCIRM for machinery fault diagnosis tasks is offered. We first verify our theoretical exploration of the SCIRM by using simulation data. We further compare SCIRM with a set of state-of-the-art methods by using real machinery fault data collected under a variety of working conditions. The computational results confirm that the machinery fault diagnosis model developed by the SCIRM offers a higher generalization capacity and performs better than the other benchmarks across the different testing datasets.

Lithium-ion rechargeable batteries are widely used in various electronic products and equipment due to their immense benefits in power supplying. The exact remaining useful life (RUL) prediction of lithium-ion batteries has shown excellent achievements in preventing severe economic and security consequences incurred in failing to provide necessary power levels. Recently, the nonlinear-drifted fractional Brownian motion made quite a splash in RUL prediction, since its first hitting time distribution can be approximated by weak convergence theorem and time-space transformation. However, the previous RUL prediction methods based on fractional Brownian motion only considered current state measurement. In this paper, a prediction framework based on nonlinear-drifted fractional Brownian motion with multiple hidden state variables is put forward to estimate RUL. Specifically, all the parameters of nonlinear function are defined as specific hidden state variables of lithium-ion battery degradation model, and all the state measurements are used to posteriorly estimate the distribution of the multiple hidden state variables by unscented particle filter algorithm. Four sets of lithium-ion battery degradation data provided by NASA Ames Research Center are used to validate the proposed prediction framework. According to comparison study with other methods, the proposed prediction framework demonstrates greater precision in the RUL prediction.