Defect detection in solar panels remains constrained by the limitations of manual labeling and the inefficiency of traditional inspection methods, which often struggle with large, high-resolution imagery. This study presents a novel self-supervised learning approach using the Momentum Contrast (MoCo) framework to address these challenges without relying on annotated datasets. Leveraging MoCo’s robust feature extraction and K-Nearest Neighbors (KNN) clustering, our method achieves accurate defect identification, bypassing the dependency on labeled data. Evaluated on the ELPV dataset, our approach attained a notable 96.95% accuracy, demonstrating significant improvement over existing unsupervised methods like KDAD, SAOE, DRA, and BGAD-FAS, and even outperforming some supervised models such as Adapted VGG19, Adapted VGG16, and ShuffleNet. Additionally, our model achieved 99.44% accuracy on the EL dataset, underscoring its adaptability and robustness across different environments. This framework offers a scalable, automated solution for image-level defect detection, poised to enhance efficiency and reduce manual intervention in industrial applications.

Steel surface defect detection is critical for ensuring the reliability and safety of automotive manufacturing. However, existing methods often suffer from high computational cost, weak sensitivity to fine textures, and limited adaptability to diverse defect scales. To address these challenges, we propose WaveMamba-YOLO, a real-time detection framework that integrates frequency-domain enhancement with efficient state-space modeling. The architecture introduces three key modules: (1) CHDWT, which combines Haar wavelet decomposition and residual learning to preserve structural details during downsampling; (2) GLaM, a global-local-aware Mamba module that couples large-kernel convolution with state-space modeling to capture long-range dependencies at linear complexity; and (3) LWGA, a lightweight group attention mechanism that adaptively attends to micro-, regular-, medium-, and large-scale defects. Experiments on the Severstal Steel Defect and NEU-DET datasets demonstrate that WaveMamba-YOLO achieves superior performance, reaching 51.70% mAP@0.5 and 58.60% precision on Severstal and 77.70% mAP@0.5 on NEU-DET, consistently surpassing mainstream lightweight detectors. These results confirm the effectiveness of WaveMamba-YOLO in balancing detection accuracy and efficiency, highlighting its potential for real-time industrial inspection.

As object detection models grow in complexity, balancing computational efficiency and feature expressiveness becomes a critical challenge. To address this, we propose SAMF-YOLO, a novel model integrating three key components: SONet, BFAM, and FASFF-Head. The UniRepLKNet backbone, enhanced by the Star Operation, expands the feature space with high efficiency. FASFF-Head performs adaptive multi-scale feature fusion with minimal overhead, and the Bi-temporal Feature Aggregation Module (BFAM) strengthens the detection of small defects. Additionally, the Focaler-IoU loss improves bounding box regression for challenging object scales, and a self-supervised contrastive learning strategy enhances feature representation and model robustness without relying on labeled data. Experimental results demonstrate that SAMF-YOLO surpasses YOLOv11s with a 6.38% improvement in mAP@0.5 and a notable reduction in computational cost, confirming its superiority in accuracy, efficiency, and robustness. The code is released at https://github.com/Missing24ff/SAMF-YOLO.git.

Voltage prediction plays a crucial role in ensuring power quality and enhancing the efficiency of power grid operations. Driven by the need for accurate and reliable voltage prediction methods, this paper introduces a novel dual-agent transformer network based on prompt engineering. By integrating prompt embeddings derived from historical voltage trends into the input embeddings, the proposed method effectively captures temporal correlations within voltage data. Utilizing a dual-agent attention mechanism, which reduces computational complexity by 42% compared to standard transformers, while the prompt embeddings enable precise capture of temporal patterns. Extensive simulations on real-world datasets demonstrate that our approach outperforms existing state-of-the-art models, achieving a 0.5% improvement in prediction accuracy. This advancement provides a robust solution for power system scheduling and operation, underscoring the importance and practical value of our work in advancing voltage prediction techniques.

Industrial metal defect detection is difficult to run in real time on edge devices with limited computing power. To solve this problem, we propose a new lightweight detection network based on Mamba for industrial defect detection applications. Firstly, we designed a MobileMamba backbone network, which includes a Mamba network with temporal characteristics and a focus on global features. Moreover, we adopted a multi-scale approach to further stimulate the characteristics of Mamba and enhance the feature extraction ability of the model. We also add a contrastive auxiliary loss function on the basis of model training to compare the differences between foreground features and background features, as well as the differences between real labels and negative samples. We also defined the difficult negative samples and the easy negative samples, and used the negative samples with the top K% similarity as the difficult negative samples to make the model pay more attention to the difficult negative samples during training. Experiments on the NEU-DET, GC10-DET, and APDDD datasets show that the efficiency of our model is higher than that of the detection models at the same level, and our nano model can reason in real time on edge devices with an accuracy rate higher than that at the same level.

In complex driving environments, autonomous agents must interact with diverse road users who exhibit heterogeneous and often unpredictable behaviors. Traditional reinforcement learning (RL) methods struggle to maintain robust performance in the presence of adversarial or deceptive intent. We propose Adversarial Theory of Mind Reinforcement Learning (Adversarial ToM-RL), a novel framework that integrates cognitive modeling with adversarial training to improve agent resilience. Unlike prior adversarial RL that perturbs observations or dynamics, our method operates on belief-level perturbations within a partially observable Markov decision process (POMDP) to simulate deceptive intent in Theory-of-Mind reasoning. Empirical results in hybrid autonomous vehicle crossover scenarios demonstrate that Adversarial ToM-RL reduces collision rates by 38% compared to standard ToM-RL and improves success rates by 8.3%. Our method shows strong robustness against malicious behaviors such as deceptive yielding and late-blocking, maintaining low collision rates and stable performance in adversarial traffic. These findings highlight the critical role of adversarial cognitive modeling in ensuring robust decision-making for security-sensitive multi-agent systems. The framework is general, model-agnostic, and compatible with existing ToM-RL pipelines.