As the scale and complexity of complex intelligent systems represented by high-performance computing and embedded systems continue to grow, logs, as core operational data, have become critical for ensuring reliable system operation through automated anomaly detection. Traditional log anomaly detection methods driven by machine learning and deep learning primarily focus mostly on log sequence modeling but suffer from insufficient semantic understanding and limited generalization ability. Large Language Models (LLMs), with their superior semantic understanding and contextual reasoning capabilities, have effectively overcome this limitation. Since the advent of LLM technology, relevant research has emerged rapidly, achievements are scattered across multiple technical paths and lack a systematic review. This paper provides a comprehensive survey of log anomaly detection methods based on LLMs. We select 35 core papers and establish a unified technical classification framework, categorizing existing methods into five technical routes: prompt engineering, Retrieval-Augmented Generation (RAG), domain fine-tuning, Reinforcement Learning (RL), and large—small model collaboration. Our analysis reveals that Supervised Fine-Tuning (SFT) is currently the most widely adopted technical route, while the large—small model collaborative architecture, as an emerging paradigm, is shifting research focus from merely pursuing detection accuracy to balancing inference efficiency and industrial deployability. The current evaluation system remains heavily focused on detection performance metrics, with insufficient attention paid to efficiency overhead and interpretability. Finally, this paper identifies inference latency bottlenecks and data privacy challenges of LLMs when processing ultra-long and massive log streams, and provides insights into frontier directions such as lightweight deployment and online continual learning.

Traffic flow prediction is a key core technology in Intelligent Transportation Systems (ITS) and has significant value in improving the efficiency of urban traffic management. Time series prediction and machine learning models are widely used in traditional methods; however, spatio-temporal Graph Neural Networks (GNN) have become a research hotspot owing to their outstanding advantages in traffic flow representation learning. However, existing methods have significant limitations. First, they rely on static graph structures, making it difficult to model long-range spatial dependencies and regional differences. Second, they fail to capture the time delay effect between road segments, resulting in insufficient modeling of spatio-temporal dependencies. To address these issues, this study proposes a Space Dynamic Graph Convolutional (SDGC) traffic flow prediction model based on the Time-Lag Aware Attention (TLAA) mechanism. The encoder—decoder architecture is adopted, and the mechanism is combined with spatial dynamic graph convolution to mitigate the adverse effects of the time-delay phenomenon on prediction accuracy. Specifically, gated causal convolution incorporates attention mechanisms to explicitly model the asynchronous temporal dependencies of spatial nodes by learning delay parameters. The dynamic adaptive spatial correlation matrix can perceive real-time changes in traffic status and accurately depict the dynamic propagation process of traffic flow in the road network. Meanwhile, the introduction of spatio-temporal embedding enables the model to accurately depict the short-term fluctuations, long-term periodic patterns, and spatio-temporal heterogeneity of traffic flow under network topology constraints. Experiments on public benchmark datasets demonstrate that this method effectively addresses challenges related to time lag and dynamic behavior. Compared with existing state-of-the-art baseline models, it achieves significant reductions in key prediction error metrics Mean Absolute Error (MAE) and Root Mean Square Error (RMSE), ranging from 6.5% to 11.9%.

Unstructured road segmentation is a crucial component of environmental perception for autonomous driving systems, and it suffers from challenges such as the integrity of global topological modeling, preservation of boundary details, and trade-off between model efficiency and accuracy. To address these challenges, this paper proposes a lightweight axial context network AXON-Net. The network employs an encoder—decoder architecture and introduces a Channel And Spatial Attention Block (CASAB) in the encoder, which adaptively recalibrates feature weights by aggregating multidimensional statistical information to effectively suppress environmental noise, thereby enhancing feature discriminability in complex backgrounds. A Lightweight Partial Context (LightPCT) model is integrated at the bottleneck, utilizing a partial channel interaction strategy to reduce computational redundancy and efficiently capture long-range dependencies for restoring road topological connectivity. In addition, the decoder integrates Dual-Path Channel Fusion (DPCF) and Thin Structure Enhancer (TSE) modules for bridging the feature semantic gap and explicitly enhancing axial geometric features for the refined recovery of blurred road edges. Experimental results on unstructured road datasets constructed from the India Driving Dataset (IDD) and Off-Road Freespace Detection (ORFD) dataset show that AXON-Net achieves road Intersection over Union (IoU_Road) scores of 95.3% and 88.1%, respectively. With only 8.49×10³ parameters, it achieves a superior balance between segmentation accuracy and model efficiency. Ablation studies further validate the synergistic effectiveness of the proposed modules, demonstrating the potential application of the network for unstructured road perception tasks.

Video analytics extracts high-value information from video streams and plays a crucial role in applications such as intelligent transportation and public safety. Although traditional cloud-based video analytics offers powerful computational capabilities, uploading massive amounts of video data results in high bandwidth consumption and network latency. Edge computing reduces network latency by processing video data near cameras; however, it still faces two major challenges. First, frame-by-frame analysis leads to redundant inference, and existing frame reuse methods cannot fully exploit local similarities in historical frames. Second, uneven core workload arises because task allocation across big LITTLE lacks real-time load awareness. To address these issues, this study proposes Vable, an efficient video analytics system for big.LITTLE edge devices. It employs a multihistorical frame, block-level frame reuse mechanism, which partitions video frames into fine-grained blocks and employs a tree-based storage structure combined with Locality-Sensitive Hashing (LSH) for similarity matching, enabling efficient cross-frame computation reuse and significantly reduces redundant inference overhead. Additionally, Vable introduces a core workload-aware list-based Directed Acyclic Graph (DAG) partitioning algorithm that dynamically allocates analysis tasks by monitoring the real-time load of big LITTLE, thus balancing computation and communication overhead while avoiding latency increases caused by load imbalance. A prototype of Vable is implemented and evaluated using two real-world datasets. Experimental results demonstrate that Vable reduces the end-to-end latency by 59.23% in UA_V₁ compared to WiseFuse (the highest-latency baseline) and by 45.83% in UA_V₂ compared to Faastlane, effectively lowering the end-to-end latency while maintaining high throughput.

Image deraining aims to restore high-quality clean images from rain-degraded inputs and is a key technique for improving the robustness of outdoor vision systems. Existing Convolutional Neural Network (CNN)-based methods are limited by local receptive fields and thus struggle to effectively model long-range dependencies. Although some Transformer-based methods enhance global modeling capability, they show limited ability in handling multi-scale and directional rain streaks as well as recovering local high-frequency details. Moreover, existing frequency-domain enhancement methods often treat the frequency domain merely as an auxiliary representation, which lack fine-grained collaborative modeling of the complementary relationship between the spatial and frequency domains. To address these issues, this study proposes a Synergizing Frequency and Spatial Network (SynFSNet) for image deraining. Unlike existing methods that primarily use the frequency domain as a global enhancement tool, the proposed method uses the directional, density-related, and scale-varying characteristics of rain streaks in the frequency domain. Specifically, a Multi-Scale Fourier Fusion Module (MSFFM) is designed to progressively suppress rain streaks in a coarse-to-fine manner. Additionally, a Fourier Shaped Attention Module (FSAM) is introduced in which Rectangular Filter Attention (RFA) and Square Filter Attention (SFA) are used to enhance direction-sensitive and local-structure-sensitive frequency-domain representations, respectively. Furthermore, a dual-domain loss is employed to jointly constrain spatial structural restoration and frequency-domain consistency, thereby improving restoration quality in complex scenes. Experimental results show that SynFSNet achieves superior performance on multiple synthetic and real-world datasets, verifying the effectiveness of collaborative spatial—frequency domain modeling for complex image deraining tasks.

The video stream of maritime multifunctional navigation mark monitoring presents the distinctive scene characteristics of a dominant static sea surface background and sparse dynamic ship targets. Constrained by the low-bandwidth transmission of navigation mark-borne communication networks, compressed video streams experience severe quality degradation, including noise superposition, blocking artifacts, ringing artifacts, frame blurring, and inter-frame motion estimation bias. Existing general-purpose video quality enhancement methods are unable to jointly balance single-frame spatial detail recovery and multi-frame temporal information utilization, and therefore fall short of the requirements of the aforementioned scenario. To address this issue, a Content Adaptive Fusion-based Video Stream Quality Enhancement of Maritime Multifunctional Navigation Mark Monitoring (CAF-VSMM) method is proposed, in which a dual-branch fusion enhancement model is constructed. The single-frame enhancement branch, built on multiscale features and a spatio-temporal network, exploits the latent temporal information of a single-frame via virtual frames and integrates the pixel shuffle with a residual network to achieve multiscale spatial feature fusion, effectively suppressing compression artifacts and spatial degradation induced by the maritime environment. The multi-frame enhancement branch, based on a dynamic shifted window and inter-frame motion compensation, efficiently models long-range spatio-temporal dependencies and employs hierarchical offset estimation together with deformable convolution to achieve accurate inter-frame alignment in multiship scenes, thereby alleviating temporal blurring and motion misalignment. The Dynamic Feature Perception-based Content-Adaptive Fusion (DFP-CAF) module, based on dynamic feature perception, dynamically adjusts the feature weights of the two branches according to the video content, realizing an adaptive enhancement strategy that emphasizes detailed recovery for static backgrounds and temporal consistency for dynamic targets. Experimental results show that, on the self-built Port Cluster Waterway Monitoring Video Dataset (PC-WMVD), the Peak Signal to Noise Ratio (PSNR), Structural Similarity (SSIM), and Video Multimethod Evaluation Fusion (VMAF) of the proposed method reach 42.71 dB, 99.33%, and 77.14%, outperforming the second-best method by 3.90 dB, 1.19 percentage points, and 6.50 percentage points, respectively; on the public dataset JCT-VC, they reach 28.66 dB, 88.60%, and 55.40%, exceeding the second-best method by 1.63 dB, 5.42 percentage points, and 3.19 percentage points, respectively. The model has 37.52×10⁶ parameters and 51.74×10⁹ Floating-Point Operations Per Second (FLOPs) and runs at 41.52 frames per second, meeting the real-time requirements of waterway monitoring. The proposed method is tested in an intelligent monitoring system for the waterways of port clusters, where it significantly improves the identifiability of key targets, such as ships and navigation marks, and effectively supports downstream tasks, such as ship tracking, verifying its engineering practicability and application value.