Industrial data collected from similar processes under varying production specifications or monitoring configurations often exhibit structural heterogeneity, particularly in the form of varying feature dimensions. This presents a fundamental challenge for conventional fault detection models, which typically require fixed-length inputs and assume consistent feature spaces. As a result, practitioners are often forced to either discard valuable heterogeneous data or develop separate models for each product configuration, both of which compromise scalability and efficiency. To address this, we propose a novel heterogeneous AutoEncoder (HAE) framework that enables unified representation learning and fault detection on heterogeneous tabular data with arbitrary feature lengths. HAE adopts a Transformer-based encoder–decoder architecture guided by a binary padding indicator mask, which explicitly distinguishes between observed and padded attributes. To further improve discriminability and reduce intersource variation, a triplet-center loss is introduced to align latent representations and enhance interclass separability. We evaluate our method on two real-world datasets: a heavy-plate production process from the steel industry and a cross-project software defect prediction task. Experimental results show that HAE consistently outperforms a wide range of traditional, statistical, and generative imputation baselines or feature truncation method. Ablation studies further verify the effectiveness of each proposed component. HAE provides a scalable and generalizable solution for heterogeneous fault detection, without requiring strong assumptions on data missingness patterns or feature correspondences.

Imbalanced class distribution disrupts the training of a classifier, resulting in biases favoring majority classes. Data oversampling is a common strategy to tackle this issue. However, traditional methods may generate incorrect and unnecessary instances when facing complex data challenges, such as class overlap, small disjuncts, and noise samples. Therefore, there is a need for an oversampling method that can accurately characterize the data distribution. This paper introduces a novel deep generative oversampling approach for balancing the imbalanced tabular data by leveraging diffusion models and Generative Adversarial Networks (GANs). The model comprises a generator constructed from diffusion models and a discriminator with a Noise-Sensitive Auxiliary Classifier (NSAC) and is trained through an adversarial process. The synergy of these two models enhances stability and sample quality compared to GANs, with faster sampling speed and better conditional generating ability than diffusion models. In experimental validation across 22 real-world datasets, our method consistently outperforms six counterparts regarding Accuracy, F1-score, and MCC for binary and multi-class scenarios. Notably, our approach enhances classifier accuracy for minority classes while maintaining a high level for the majority class, a facet often compromised by other algorithms.

Industrial Internet of Things (IIoT) communication serves as the core for connecting industrial devices, systems, and platforms. In addition to real-time performance, reliability, and security, fairness in device access and data transmission has become increasingly important. This paper integrates Reconfigurable Intelligent Surfaces (RIS) with Non-Orthogonal Multiple Access (NOMA) technology in 6G communications, establishing a synergistic integration to jointly elevate spectral efficiency and fairness. Deploying Simultaneously Transmitting and Reflecting Reconfigurable Intelligent Surfaces (STAR-RIS), which provides$\mathbf{3 6 0}$-degree coverage, in factory environments helps improve signal coverage and reduce bit error rates. To tackle challenges such as beamforming coupling, dynamic decoding order, reflection path optimization, and phase adjustment -with particular attention to fairness for low-rate devices - a fairness-oriented optimization model is proposed. This model focuses on reflective channel beamforming, quality of service (QoS), and STAR-RIS phase shift matrices optimization. Pursuing the maximization of the minimum achievable rate under QoS constraints for this intricate non-convex problem, a two-layer iterative method is employed. The outer layer handles adaptive SIC decoding order updates, nesting an inner layer that tackles the coupled optimization of the base station beamforming and STAR-RIS coefficients. Simulation verification reveals: 1) The proposed algorithm significantly improves system fairness. 2) The integration of STAR-RIS and NOMA provides higher gain in device communication. 3) Optimizing RIS phase shifts and beamforming effectively enhances communication rates. 4) The proposed approach exhibits superior performance relative to other schemes.

Recently, active reconfigurable intelligent surfaces (RIS) have attracted much attention due to its ability to amplify signals. The combination of active RIS and unmanned aerial vehicle (UAV) can improve the performance of communication systems. For the multi-objective optimization problem of active RIS assisted UAV communication systems, this work proposes a new low-complexity scheme to maximize system energy efficiency (EE). Unlike previous work, we focus on the impact of UAV hovering altitude on the system's EE and separately investigate the effects of the amplification coefficient and phase shift of active RIS on the system's EE. We also adopt non-orthogonal multiple access (NOMA) technology to improve spectral efficiency, further boosting the system EE. To address challenges such as non-convexity caused by UAV altitude updates and the coupling between active RIS's amplification coefficients and phase shifts, this work proposes a block coordinate descent (BCD) iterative optimization framework that decomposes the problem into three subproblems. By solving each subproblem, the optimal UAV altitude and the amplification coefficients and phase shifts of the active RIS are derived to maximize system EE. Numerical results show that: 1) Compared with other benchmark schemes, the proposed active RIS-NOMA scheme achieves significant improvements in system EE. 2) UAV demonstrates great communication potential in active RIS-NOMA systems, especially when direct links are blocked.

Class imbalance, which is common in real-world classification tasks, often leads to biased models favoring majority classes. Data oversampling is a widely used strategy to address this issue. However, traditional oversampling methods often generate incorrect or redundant instances when class overlap occurs, increasing decision boundary complexity. To this end, we propose a novel Generative Oversampling approach to addressing Class Imbalance and Overlap (GOIO) in the classification of tabular data. GOIO combines a Metric-Learning-based Variational Autoencoder (MLVAE) and a Conditional Latent Diffusion Model (CLDM) to handle class imbalance and overlap effectively. The MLVAE employs a triplet-center loss to the adverse effects of class overlap by transforming the data distribution into a more separable latent feature space. Following this, the CLDM is trained with class-center feature prompting and classifier-free guidance strategy to capture class-specific latent distributions accurately. Minority class samples are synthesized in the latent space using the CLDM and then reconstructed into the data space via the MLVAE decoder. Comprehensive experiments on 18 real-world and five synthetic datasets demonstrate that GOIO outperforms the state-of-the-art oversampling methods in F1-score, MCC, and Accuracy. Ablation studies further validate the effectiveness of the proposed contributions in addressing class imbalance and overlap.

With the prosperity of the intelligent surveillance, multiple cameras have been applied to localize pedestrians more accurately. However, previous methods rely on laborious annotations of pedestrians in every frame and camera view. Therefore, we propose in this paper an Unsupervised Multi-view Pedestrian Detection approach (UMPD) to learn an annotation-free detector via vision-language models and 2D-3D cross-modal mapping: 1) Firstly, Semantic-aware Iterative Segmentation (SIS) is proposed to extract unsupervised representations of multi-view images, which are converted into 2D masks as pseudo labels, via our proposed iterative PCA and zero-shot semantic classes from vision-language models; 2) Secondly, we propose Geometry-aware Volume-based Detector (GVD) to end-to-end encode multi-view 2D images into a 3D volume to predict voxel-wise density and color via 2D-to-3D geometric projection, trained by 3D-to-2D rendering losses with SIS pseudo labels; 3) Thirdly, for better detection results, i.e., the 3D density projected on Birds-Eye-View, we propose Vertical-aware BEV Regularization (VBR) to constrain pedestrians to be vertical like the natural poses. Extensive experiments on popular multi-view pedestrian detection benchmarks Wildtrack, Terrace, and MultiviewX, show that our proposed UMPD, as the first fully-unsupervised method to our best knowledge, performs competitively to the previous state-of-the-art supervised methods. Code is available at https://github.com/lmy98129/UMPD.

Image recognition techniques have become the mainstream solution for indoor unmanned aerial vehicle (UAV) localization and navigation due to the absence of global positioning system. However, unlike autonomous vehicles that enjoy the dividends of the "big model" era, UAVs fail to deploy such models due to the hardware limitations. To this end, this paper proposes a compact all multilayer perceptron (MLP) deep neural network structure that offers a paradigm for theory-guided prompt structural compression of large-scale MLP models. First, we propose a gradient-based sensitivity analysis (GB-SA) method. Unlike existing SA methods, GB-SA obtains nodes’ sensitivity indices with gradient information, openning the possibility of efficient SA for large models. We start by integrating GB-SA with MLP and then extend the mode to MLP-Mixer, which is a promising all-MLP deep neural network. By deeply combining GB-SA and MLP-Mixer, SA-MLP-Mixer emerges as a compact model without reducing the model precision. Finally, we evaluate the effectiveness of the proposed model on the benchmark. The experimental results show that SA-MLP-Mixer has an airborne level model scale and accurate localization capability.

As an extended task of crowd counting, fine-grained crowd counting aims to estimate the number of people in each semantic category instead of the whole in an image, and faces challenges including 1) inter-category crowd appearance similarity, 2) intra-category crowd appearance variations, and 3) frequent scene changes. In this paper, we propose a new fine-grained crowd counting approach named DSR to tackle these challenges by modeling Discriminative Semantic Relationship, which consists of two key components: Word Vector Module (WVM) and Adaptive Kernel Module (AKM). The WVM introduces more explicit semantic relationship information to better distinguish people of different semantic groups with similar appearance. The AKM dynamically adjusts kernel weights according to the features from different crowd appearance and scenes. The proposed DSR achieves superior results over state-of-the-art on the standard dataset. Our approach can serve as a new solid baseline and facilitate future research for the task of fine-grained crowd counting.