Multimodal remote sensing image classification improves models’ capacity to recognize complex land-cover patterns by integrating data from heterogeneous sensors such as hyperspectral image (HSI) and light detection and ranging (LiDAR). However, many existing classification models ignore aleatoric and epistemic uncertainties introduced during data acquisition and labeling. As a result, they become less robust to noise and more vulnerable to spurious correlations, which ultimately weakens their ability to generalize to unseen data. To mitigate these issues, multimodal joint distribution modeling is reformulated as a flow matching optimization problem that learns a distribution evolution process under an unknown target distribution. A reinforcement learning-driven flow matching (RL-FM) framework is proposed for multimodal remote sensing image classification. Specifically, feature distributions of remote sensing images from different modalities are first modeled using variational autoencoders, and a multimodal mixture distribution is then constructed via Gaussian mixture strategy to serve as an initial distribution for flow matching. To perform flow matching optimization when target distribution is unknown, label information is further exploited to guide the transformation of initial distribution toward target distribution. At the same time, the distribution evolution process of flow matching is formulated as a Markov decision process (MDP), enabling the model to learn an evolution path from initial distribution to target distribution by maximizing the expected cumulative reward. RL-FM jointly accounts for immediate classification loss and long-term generalization performance, thereby alleviating suboptimal convergence caused by myopic gradient updates. Furthermore, by incorporating counterfactual causal inference into policy optimization, a counterfactual proximal policy optimization (CPPO) is designed. CPPO can strengthen the model capacity to capture the causal relationship between action and reward, thus improving generalization in scenarios with limited labeled samples. Experimental results on multiple benchmark datasets demonstrate that the proposed RL-FM achieves state-of-the-art performance on multimodal remote sensing image classification tasks. The code is available at: https://github.com/zwdmw/RL-FM

This paper studies the problem of high false alarm rates in the nonlinear no-Gaussian quality-oriented fault diagnosis framework. Compared with existing methods, advanced data-driven (ADD) leverages information entropy to handle nonlinear non-Gaussian data, thereby enhancing the accuracy and robustness of fault detection while avoiding the assumption of Gaussian distributions. The preprocessing-postprocessing strategy and quality-oriented fault detection high-order indices are designed to ensure the correct correlation between faults and quality indicators. In addition, the pseudo-samples fault location method mitigates the smear effect in kernel function to identify faulty reasons. In the end, the effectiveness of ADD is verified through experimental results obtained from both simulation and industrial scenarios.

While pretrained foundation models (FMs) exhibit remarkable generalization across diverse tasks, their full fine-tuning under limited labeled data remains prohibitively expensive and prone to performance degradation or overfitting. To mitigate these challenges in the context of multimodal remote sensing image classification, we propose multimodal prompt tuning with modified knowledge distillation (MPT-MKD), a parameter-efficient learning framework enabling accurate class predictions while updating only a small fraction of model parameters. Prompt tuning is adapted to multimodal remote sensing by injecting modality-specific learnable prompts (spectral and light detection and ranging (LiDAR) prompts) into a student FM, whereby cross-modal interactions are captured through self-attention. In addition, a knowledge distillation (KD) scheme is designed that leverages both limited labeled samples and abundant unlabeled data, where a teacher FM, with parameters frozen except for a learnable classifier, generates soft pseudo-labels for the student FM. To mitigate potential inaccuracies in the logits predicted by the teacher FM, a similarity- and class-consistency-based selection strategy is employed to identify a more reliable subset of unlabeled samples and to reconstruct modified soft targets, thereby alleviating the adverse effects of noisy distillation signals. Experimental results on three real hyperspectral image (HSI) and LiDAR datasets demonstrate the effectiveness of the proposed method.

In complex decision-making tasks, agents often suffer from memory interference, inefficient exploration under sparse rewards, and limited adaptability in dynamic environments. While existing deep reinforcement learning (DRL) methods often suffer from performance limitations when handling long-horizon tasks due to memory interference, insufficient exploration, and poor adaptability. Inspired by multi-timescale memory mechanisms in neuroscience, this article proposes a hierarchical memory-based DRL (HM-DRL) architecture that integrates three complementary memory layers: a perceptual memory layer for real-time detection of environmental changes, an episodic memory layer for structured event sequence storage, and an abstract memory layer that constructs causal graph and supports backward reasoning. Furthermore, HM-DRL incorporates a dynamic gating mechanism and a compound reward function to facilitate effective integration of the policy network and memory outputs, thereby enhancing policy optimization and crisis response. Within the designed open-ended multi-task simulated survival environment, the HM-DRL architecture demonstrates substantial improvements in mitigating memory interference during long-horizon tasks, enhancing adaptability to sudden crises and dynamic environments, and improving learning efficiency under sparse reward scenarios. This article offers an effective paradigm for addressing core memory management challenges in long-horizon decision-making and establishes a foundation for developing agents with causal reasoning capabilities.

Visual reinforcement learning (VRL) aims to extract effective visual information from high-dimensional observations to optimize decision-making policies. While existing VRL methods have achieved significant progress in various control tasks through data augmentation and auxiliary tasks, agents remain susceptible to distractions from redundant information and irrelevant factors, leading to overfitting and degraded generalization in unseen environments. To address this challenge, we propose a similarity-guided state attention (SSA) for VRL. In this method, a similarity guidance module (SGM) is designed to leverage the similarity between state embeddings from the original and augmented observations to guide the encoder to focus on task-relevant regions in the original observation, thereby producing a corresponding state attention map. Meanwhile, the state attention map of augmented observation is obtained by decoding its state embedding. Furthermore, cosine similarity is introduced to measure the global similarity between state attention maps of original and augmented observations, which is incorporated into self-supervised learning objective together with binary cross-entropy loss to encourage alignment of state attention maps in representation space. The combination of SGM and cosine similarity-based alignment of state attention maps facilitates self-supervised learning to obtain more robust state representations for downstream reinforcement learning, thereby enabling the agent to learn optimal policies and improve its generalization ability in unseen environments. Experimental results on the DeepMind Control Generalization Benchmark (DMControl-GB) demonstrate that SSA achieves superior robustness and generalization performance compared with representative VRL baselines.

Objects in complex environmental conditions such as low light and smoke occlusion are difficult to be detected accurately, and multimodal fusion of visible and infrared images with complementary physical properties provides a solution. However, different modalities often exhibit significant modal heterogeneity due to differences in imaging mechanisms, which can easily induce intermodal information interference. We decomposed inter modal information interference into task coupling interference, distribution difference interference, and noise superposition interference. To address the above issues, the variational decoupled bottleneck adaptation network (VDBAN) has been proposed. First, we decoupled the class distribution and spatial distribution of the two modalities, which in turn is targeted to captured task-related distribution information from multimodal data. In addition, we also customized differentiated fusion parameters and strategies for object recognition and localization tasks, adapting to the feature learning preferences of different tasks. Second, we designed the global mutual information adaptation mechanism and the structural mutual information adaptation mechanism, which promoted cross-modal knowledge sharing by enhancing the shared information between modalities. Finally, we constructed a variational fusion information bottleneck to compress the information flow in the multimodal feature fusion process to make the model focus on the task related information, and filter out the task-independent noise information. Numerous experimental results have demonstrated that VDBAN shows state-of-the-art detection performance in multimodal object detection task.

Eliminating the effect of input saturation is greatly significant in maintaining system performance. This article proposes a performance-based adaptive anti-saturation control scheme for a class of uncertain nonlinear systems with input saturation limit. Unlike the compensation-based anti-saturation schemes that do not take saturation occurrence into account, the proposed method adopts a monitoring modulation approach to proactively prevent the occurrence of input saturation. Specifically, an adaptive controller is first designed based on prescribed performances and shifting functions. Then, a monitoring module is employed to supervise the behavior of the control signal, and a modulation module is intended to implement the controller reconfiguration by updating the shifting functions. It is shown that by using the proposed scheme, the control signal never violates the saturation limit, the Zeno phenomenon is avoided, and the tracking error satisfies a modified prescribed performance.

Meta-transfer learning introduces domain adaptation techniques into meta-learning frameworks, aiming to adapt to scenarios where training/testing tasks have both task distribution differences and domain distribution differences. However, existing meta-transfer learning methods tend to follow a metric-based meta-learning paradigm. Facing challenges such as data imbalance, task heterogeneity, and knowledge sharing in cross-domain few-shot hyperspectral image (HSI) classification. To address the above challenges, we proposed a balanced meta-adaptation network (BMAN) by customizing a meta-transfer learning framework for the cross-domain few-shot HSI classification task. First, to cope with the data imbalance challenge, we broke the traditional optimization mechanism in which the source domain and target domain are independent of each other. An inner optimization mechanism based on meta-generalized empirical risk minimization was constructed to capture the target domain-oriented few-shot classification knowledge. Second, to address the challenge of task heterogeneity, a balanced meta-sampling mechanism was designed to align the two domain task distributions, and an outer layer optimization mechanism capable of capturing unbiased knowledge was constructed to correct the feature learning pReferences of the model. Finally, to cope with the challenge of knowledge sharing, based on the idea of “seeking common ground while reserving differences,” we allowed the source domain to retain domain-specific knowledge while learning sharable knowledge, thus promoting cross-domain knowledge transfer. Experiments have shown that BMAN has achieved state-of-the-art performance in cross-domain few-shot HSI classification tasks.

In recent decades, increasingly severe precipitation-driven flood events have been observed across many regions, posing substantial challenges to infrastructure safety and risk management. Reliable estimation of design precipitation associated with rare events is essential for improving flood preparedness and engineering design. In the Chinese mainland, the Pearson Type III (PE3) distribution has been recommended as a unified model for estimating precipitation extremes since 2006. However, this assumption neglects pronounced spatial heterogeneity and may lead to biased estimates of design precipitation, affecting engineering design and risk assessment. In this study, we evaluate 7 commonly used extreme-value distributions for estimating century-scale precipitation extremes in the Chinese mainland, using high-resolution gridded precipitation data from 1961 to 2022. Our results create a composite Regionally Optimal Distribution framework considering 3 algorithms (ROD-3) involving Generalized Extreme Value (GEV), Weibull (WEI), and PE3. The ROD-3 framework provides the optimal fit for over 92% of grids in the Chinese mainland, with GEV best for 54%, followed by WEI and PE3. Compared to PE3, ROD-3 significantly improves goodness-of-fit in 65% of regions and reduces estimation bias for the 100-year return period design precipitation in 22% of regions. In addition, a decision-tree classifier identified a kurtosis threshold (∼0.71) as a simple criterion for distribution selection. Further validation using global land precipitation data reveals a consistent threshold, supporting the framework's transferability across diverse climatic regions. Overall, the framework provides an operational tool for region-specific design standards and risk-informed infrastructure planning, improving the reliability of design precipitation estimation.

Hyperspectral image (HSI) classification models face dual challenges in open-set domain generalization: limited generalization ability due to unseen-domain shifts, and the need for unknown class recognition that breaks the closed-set assumption of traditional models. To address these challenges, we propose the Markov meta-Mamba network (M3Net), which provides a meta-reinforcement learning-based solution for open-set domain generalization of HSI classification model. Specifically, a meta-task construction mechanism is proposed, treating source-domain background pixels as virtual unknown classes to simulate open-set HSI classification tasks during training, thereby providing task support for meta-reinforcement learning. Then, the open-set HSI classification task is reconstructed as a Markov decision process. By leveraging reinforcement learning's multi-step temporal credit assignment, non-causal factor sensitivity is suppressed, improving the model's cross-domain generalization performance. Finally, the theoretical linkage between Mamba and meta-learning is established, demonstrating that Mamba inherently operates as a meta-learner when processing task sequences. Building on this, a Mamba-based meta-task embedding framework is designed, where shared meta-parameters and task-specific parameters are jointly optimized to achieve cross-task knowledge induction across open-set HSI classification tasks, thereby enhancing the model's generalization capability for unseen open-set tasks. Experiments on three cross-domain hyperspectral image datasets show that M3Net has achieved the most competitive performance in the open-set domain generalization.

In order to address the issue of overly conservative offline reinforcement learning (RL) methods that limit the generalization of policy in the out-of-distribution (OOD) region, this article designs a surrogate target for OOD value function based on dataset distance and proposes a novel generalized Q-learning mechanism with distance regularization (GQDR). In theory, we not only prove the convergence of GQDR, but also ensure that the difference between the Q-value learned by GQDR and its true value is bounded. Furthermore, an offline generalized actor-critic method with distance regularization (OGACDR) is proposed by combining GQDR with actor-critic learning framework. Two implementations of OGACDR, OGACDR-EXP and OGACDR-SQR, are introduced according to exponential (EXP) and open-square (SQR) distance weight functions, and it has been theoretically proved that OGACDR provides a safe policy improvement. Experimental results on Gym-MuJoCo continuous control tasks show that OGACDR can not only alleviate the overestimation and overconservatism of Q-value function, but also outperform conservative offline RL baselines.

Hypersonic flight vehicles (HFVs) exhibit complex nonlinear dynamics and time-varying uncertainties, placing stringent demands on the adaptive learning capabilities of flight control systems. Meanwhile, the limited onboard communication and computational resources further require efficient resource utilization. To address these issues, this article proposes a dynamic event-triggered control (DETC) scheme based on actor-critic framework for HFVs, achieving synergistic optimization of control performance and resource conservation. First, an optimal event-triggered controller is developed within actor-critic framework to regulate the frequency of control input updates. A static event-triggered condition is formulated, and the stability of the closed-loop system is analyzed, with Zeno behavior eliminated, ensuring both stability and feasibility. Building on this, a dynamic variable is introduced, leading to a dynamic event-triggered condition that adaptively adjusts the trigger threshold based on real-time system states. Unlike static event-triggered mechanisms with fixed thresholds, this mechanism further optimizes the frequency of control input updates, achieving more efficient resource utilization while maintaining the desired control performance. Moreover, rigorous theoretical analysis demonstrates that the closed-loop system is stable and Zeno behavior is eliminated under the DETC scheme. Finally, the effectiveness of our proposed DETC scheme is verified through a simulation example.

Implicit Q-learning, a prominent approach within the implicit value learning (IVL) paradigm, mitigates distribution shift in offline reinforcement learning (offline RL) by avoiding out-of-distribution action queries. However, current IVL methods lack stepwise training analysis in value learning and commonly assume the behavior policy to be optimal, both of which overlook more realistic suboptimal scenarios. This work presents cautious implicit value learning (CIVL), a novel offline RL method that redefines the advantage weight to identify a previously unrecognized constraint in IVL training. By establishing a formal connection between the identified constraint and the target V-network, CIVL provides theoretical motivation for introducing the target V-network. In policy learning, CIVL introduces a negative-action penalty (NAP) that penalizes actions with negative advantage values, thereby better handling suboptimal actions. Since the NAP can be shown to stochastically induce an approximate gradient penalty on the Actor, CIVL introduces a floating term that synergistically increases the probability of the gradient penalty occurring. Finally, CIVL further incorporates trajectory-position weighting to provide temporally fine-grained control over the V-network's bias. Benchmark experiments demonstrate that CIVL outperforms existing IVL methods in terms of normalized score and training stability.

This study proposes an adaptive critic network-based event-triggered control (ETC) scheme for hypersonic flight vehicles with input constraints. First, a nonquadratic performance index is formulated for input-constrained systems to overcome the limitations of traditional quadratic performance indexes. Next, an event-triggered condition is designed to trigger controller updates based on a fixed threshold, communicating only when necessary, effectively saving computational and communication resources. Subsequently, a critic-only network is implemented to approximate solutions to the Hamilton-Jacobi-Bellman equation (HJBE). Thereafter, system stability is analyzed using the Lyapunov method, demonstrating that the ETC scheme not only guarantees asymptotic stability of the closed-loop system but also avoids Zeno behavior. Lastly, a simulation example confirms the efficacy of the proposed ETC scheme.

The goal of emotion recognition in conversation (ERC) is to identify the emotion expressed in each utterance in a conversation. However, some previous approaches have not sufficiently considered the effect of information about the relative positions of each speaker’s utterance and the target utterance in a conversation on emotion analysis and have ignored the differences in the emotions expressed in the target utterance under different subthemes. We introduce a dynamic extraction of subdialogs (DESDs) approach to emotion recognition to address these issues. The method extracts subdialogs using information about the relative positions of each speaker’s utterance and the target utterance. The emotional dynamics in the dialog are then captured more accurately by considering each speaker’s contribution to the emotional expression. Additionally, we extract the subdialog topic information to capture the impact of different subthemes on the emotion of the target utterance. Through experiments conducted on four different datasets, we validate the effectiveness of the network. Our code is available at https://anonymous.4open.science/r/DESD-1FDF

Due to the lack of prior knowledge about unknown classes during training, existing methods for cross-domain open-set image recognition typically rely on threshold-based solutions. However, such approaches often struggle to capture the complex boundary relationships between known and unknown classes, which can lead to negative transfer effects caused by feature confusion between the two. To address this issue, this paper proposes a graph isomorphic distillation diffusion model (GIDDM) that aims to learn the boundary relationships between known and unknown classes from a closed-set classifier that models predictive uncertainty. First, a diffusion classifier is designed to quantify model predictive uncertainty through a Monte Carlo sampling strategy performed on the noise distribution during the reverse denoising process. The uncertainty distribution is modeled, and the cumulative distribution function is used to compute the probability of a sample belonging to an unknown class. Second, an open-set recognition framework is constructed, treating the closed-set diffusion classifier as a teacher classifier, and guiding the student classifier to learn the complex boundary relationships between known and unknown classes through knowledge distillation. Third, the knowledge distillation process is further formalized as a graph isomorphic optimization problem, where the predictive manifolds of the student and teacher classifiers are constrained to be consistent, thereby enhancing knowledge transfer between the classifiers. Finally, the entire process is integrated into a unified open-set adversarial domain adaptation framework, reconstructing the traditional optimization objectives of closed-set adversarial domain adaptation to ensure sufficient separation between known and unknown classes while aligning the distributions of known classes in both the source and target domains. Experiments conducted on multiple hyperspectral image (HSI) datasets demonstrate that the proposed method achieves state-of-the-art performance on cross-domain open-set image recognition tasks. The code demo can be accessed on the following website: https://github.com/wzr78998/GIDDM.

In industrial Internet of Things (IIoT), sensor failures, network interruptions, and communication delays often lead to structured missing data, significantly hindering the timely detection of incipient faults. This article addresses the problem of quality-related incipient fault detection and location (IFDL) in IIoT under incomplete data. Compared with existing methods that rely on fully observed data, IFDL integrates a neighborhood probability guided imputer with a multiscale graph regularized autoencoder (AE), preserving the temporal continuity and manifold structures that characterize incipient faults. With the restored data and learned features, generalized singular value decomposition (GSVD)-based correlated subspace modeling isolates the quality-related variations embedded in the process to ensure accurate identification of incipient faults. Furthermore, a confidence-weighted trajectory deviation index, combining contextual and Bayesian uncertainty, identifies the variables contributing to process faults. The effectiveness of the proposed method is demonstrated through both simulation studies and an industrial case.

To address the challenge of balancing reconstruction performance and inference speed in the existing infrared image super-resolution algorithms, this paper introduces a novel Large Kernel Reparameterization Attention mechanism. Based on this, we propose the reparameterizable large kernel attention network for infrared image super-resolution. During training, a multi-branch large kernel network is employed to fully extract information, while at inference time, it is equivalently transformed into a single-branch large kernel network, achieving a trade-off between processing performance and inference speed. Compared to state-of-the-art methods, our approach improves the average PSNR on a self-constructed infrared dataset by 0.0008 dB. Additionally, on the RK3588 Neural Processing Unit, it requires only 37ms to perform 4 $$\times$$ super-resolution on 320 $$\times$$ 180 images.

With the wide application of airdrop technology in rescue activities in civil and aerospace fields, the importance of accurate airdrop is increasing. This work comprehensively analyzes the interactive mechanisms among multiple models affecting airdrops, including wind field distribution, drag force effect, and the parachute opening process. By integrating key parameters across various dimensions of these models, a multidimensional parameter dynamic evolution (MPDE) target prediction method for aerial delivery parachutes in radar-detected wind fields is proposed, and the Runge–Kutta method is applied to dynamically solve for the final landing point of the target. In order to verify the performance of the method, this work carries out field airdrop experiments based on the radar-measured meteorological data. To evaluate the impact of model input errors on prediction methods, this work analyzes the influence mechanism of the wind field detection error on the airdrop prediction method via the Relative Gain Array (RGA) and verifies the analytical results using the numerical simulation method. The experimental results indicate that the optimized MPDE method exhibits higher accuracy than the widely used linear airdrop target prediction method, with the accuracy improved by 52.03%. Additionally, under wind field detection errors, the linear prediction method demonstrates stronger robustness. The airdrop error shows a trigonometric relationship with the angle between the synthetic wind direction and the heading, and the phase of the function will shift according to the difference in errors. The sensitivity of the MPDE method to wind field errors is positively correlated with the size of its object parachute area.