The rapid growth of digital textual content has intensified the need for automatic text summarization sys- tems that are both effective and reliable. While extractive summarization methods are interpretable and preserve factual content, they often suffer from redundancy and limited coherence. In contrast, abstractive approaches based on large pretrained transformer models improve fluency and readability but are prone to factual inconsistencies and hallucination. To address these limitations, this paper proposes a semantic- aware hybrid text summarization framework that integrates supervised extractive sentence scoring with constrained abstractive generation. The proposed approach employs an interpretable sentence importance model based on lexical, positional, and semantic features, learned using a Gradient Boosting Regressor. Semantic redundancy among candidate sentences is explicitly controlled using Word Mover’s Distance, enabling improved content diversity without increasing training complexity. The selected sentences are subsequently refined using a transformer-based abstractive model to enhance coherence and linguistic quality while preserving factual consistency. The framework is evaluated on a benchmark news summa- rization dataset using ROUGE-1, ROUGE-2, and ROUGE-L metrics. Experimental results demonstrate consistent improvements over a purely extractive baseline and competitive performance compared to repre- sentative extractive, abstractive, and hybrid approaches. Ablation studies further confirm the contribution of supervised sentence scoring, semantic redundancy control, and abstractive refinement to performance stability. Overall, the results indicate that combining interpretable sentence selection with semantic simi- larity modeling within a hybrid architecture provides a balanced and practical solution for automatic text summarization.
Existing pseudonym schemes in Vehicular Ad-Hoc Networks (VANETs) incur significant communication and storage overhead. Moreover, while RSU-dependent schemes are constrained by the cost and scala- bility of infrastructure deployment, RSU-independent schemes frequently rely on centralized components that introduce both privacy vulnerabilities and performance bottlenecks. To address these issues, this paper proposes a fog-computing-based pseudonym-management architecture that utilizes distributed fog nodes to execute pseudonym generation, distribution, storage, and accountability. Within this framework, a vehicle entering a fog domain receives a pool of short-term pseudonyms from its nearest fog node for intra-domain communications. This architecture significantly reduces communication latency, alleviates the burden on the centralized certificate authority, and enhances system scalability. In addition, based on the proposed architecture, a vehicle-centric RSU-independent pseudonym-change mechanism is introduced. The mech- anism dynamically determines the optimal timing for pseudonym changes by quantifying driving-behavior similarity and incorporating real-time vehicle states such as speed, direction, and relative position. Theo- retical analysis shows that the proposed scheme provides effective privacy protection. Experimental results demonstrate that, compared with traditional RSU-based and centralized schemes, the proposed scheme sig- nificantly enhances anonymity while substantially reducing the risk of being tracked, offering a practical and scalable solution for privacy-preserving VANETs.
Anxiety is a prevalent mental health problem among university students, creating the need for scalable and reliable tools for screening, severity classification, and decision support. This pilot study aimed to design, develop, and preliminarily evaluate a generative artificial intelligence chatbot for screening anxiety severity among Engineering students at the Universidad Nacional de San Cristóbal de Huamanga, Peru, during the 2025-II academic semester, using the Hamilton Anxiety Rating Scale (HAM-A). A pilot sample of 50 students was selected through cluster sampling. The chatbot, developed with Flutter and Dart and integrated with a generative language model API, automated HAM-A administration, score calculation, severity classification, result storage, and personalized feedback generation. The system was preliminarily validated by comparing chatbot-generated severity classifications with those assigned by a clinical mental health expert. Agreement was assessed using Cohen’s Kappa coefficient and a confusion matrix, while classification performance was analyzed using class-specific and multiclass metrics. The results showed 46 exact matches out of 50 cases, with an observed agreement of 92.0% and Cohen’s Kappa of κ = 0.8086, indicating almost perfect concordance in this pilot sample. Overall accuracy was 0.9200, with a Weighted F1-score of 0.9032, Macro F1-score of 0.7560, and Balanced Accuracy of 0.7197. The chatbot showed strong performance in moderate anxiety classification and robust identification of severe anxiety cases, with no false negatives in the severe category. However, the mild category showed lower recall, indicating a tendency to classify some mild cases into higher severity levels. These findings suggest that the chatbot has preliminary potential as a first-level screening and decision- support tool for university mental health programs. Nevertheless, due to the pilot nature of the study and the limited sample size, the results should be interpreted cautiously and validated in larger, more diverse, and preferably multicenter samples. The chatbot should be considered a complementary support system, not a substitute for formal clinical diagnosis or professional psychological evaluation.
Accurate runtime prediction is essential for efficient HPC job scheduling, yet users chronically overesti- mate their jobs’ requirements. We analyze 7.3 million completed jobs from the NREL Eagle supercomputer and find that the problem is far worse than previously reported: median time-limit utilization is just 6.7%, with users consuming a median of 10.6 minutes against 4-hour requests. We train ensemble models (Ran- dom Forest, Gradient Boosting, HistGradientBoosting) and an MLP neural network enriched with user behavioral features—historical runtimes, utilization habits, submission frequency—and temporal context. Our central finding concerns evaluation methodology: under the random train/test splits common in prior work, Random Forest reaches R2 = 0.602 (MAE = 0.99 h), but under a realistic temporal split (train
The rapid growth of lightweight Internet of Things (IoT) applications has intensified the need for effi- cient task scheduling mechanisms in fog computing environments, where delay sensitivity and resource constraints are critical concerns. To address these challenges, this paper proposes MPTS, a Multi-Queue Priority-Based Task Scheduling algorithm designed to minimize service delay while ensuring fair resource allocation for heterogeneous and delay-sensitive IoT workloads. The proposed algorithm classifies incom- ing tasks into short and long jobs based on burst time and schedules them using multiple priority queues with a dynamic time-frame mechanism, effectively mitigating starvation and improving response time. The performance of MPTS is evaluated using a Cooja-based simulation environment implemented on Con- tiki OS, considering realistic fog–IoT network settings. The proposed approach is compared against two benchmark scheduling schemes: Greedy Knapsack-based Scheduling (GKS) and Delay and Performance Optimization in Fog Computing (DPOFC). Simulation results demonstrate that MPTS achieves approx- imately 18–24% reduction in average end-to-end service delay and 47–50% lower network usage, while maintaining comparable energy consumption across varying numbers of IoT devices. These results confirm that MPTS significantly enhances Quality of Service (QoS) by jointly optimizing delay, network utilization, and energy consumption, making it well suited for delay-sensitive and resource-constrained fog-enabled IoT applications.
The Routing Protocol for Low-Power and Lossy Networks (RPL) remains the de facto IPv6 routing standard for constrained Internet of Things (IoT) deployments. However, its default objective functions, Objective Function Zero (OF0) and the Minimum Rank with Hysteresis Objective Function (MRHOF), rely on isolated single-metric evaluations or rigid metric combinations. Consequently, in dense deployments characterized by node mobility and high traffic volumes, these conventional mechanisms often fail to optimize parent selection. Although multi-metric objective functions—incorporating fuzzy logic and entropy-based designs—have demonstrated that evaluating link quality, residual energy, and node stability simultaneously improves routing performance, many existing schemes rely on pre-deployed static weights or neglect application-specific context during parent selection. To bridge this gap, this paper proposes an enhanced objective function framework that introduces the concept of graph-theoretical critical nodes, offering greater adaptability in defining routing topologies. By identifying nodes whose high placement within the Destination-Oriented Directed Acyclic Graph (DODAG) would be detrimental to overall network performance, we dynamically regulate their topological hierarchy. Specifically, we model the network as an undirected graph and apply a hybrid genetic algorithm to solve the 3-Component Critical Node Problem (3C-CNP). The resulting critical node set serves as an input for two proposed parent selection algorithms designed to demote or exclude critical nodes, thereby minimizing their adverse impact on routing stability. Extensive Contiki-NG simulations demonstrate that Variant 2 reduces parent churn by a factor of 3 to 4, lowers energy consumption by 35\% to 40\%, and significantly extends network lifetime under high traffic stress compared to standard OF0, achieving a superior balance between stability and efficiency.
Assistants based on large language models accelerate scholarly reading, but they hallucinate, which is unacceptable in academic work where every claim must be traceable to a source. We present CITAE, an open-source web application that integrates the complete scientific-literature workflow—discovery, read- ing, verification, organisation and citation—into a single system, and makes AI grounding visible to the user rather than hidden inside the model. Built as a React frontend over a layered Node.js/Express backend and a PostgreSQL database, its reading assistant ranks the user’s papers and highlights with Okapi BM25 (k1 = 1.5, b = 0.75) and injects the top-ranked, numbered passages into the prompt; the answer’s inline [n] markers are then bound back to the passages they cite and rendered as trust cards that expose the lit- eral supporting text and its source. A second, single-document assistant additionally verifies every quoted passage against the source and discards any that cannot be matched, so fabricated evidence never reaches the reader. A provider-agnostic client routes every model call through each user’s own keys, failing over across Groq, Google Gemini and OpenRouter, so the AI features run at no cost under a bring-your-own-key scheme and degrade gracefully. Around this core, CITAE unifies multi-source discovery over Crossref, Se- mantic Scholar, OpenAlex and arXiv with claim verification, paper comparison, literature reports, spaced review and citation in seven styles. We report an engineering evaluation on the production code: in-memory BM25 retrieval takes 0.5 ms over 100 and 6.4 ms over 1,000 passages; the deterministic verification guard attains precision, recall and F1 of 1.000 on a labelled 50-quotation probe set, rejecting all adversarial “genuine-prefix + fabricated-tail” strings; and an 86-case test suite covers the deterministic core. Offered bilingually with a Spanish-first default, free and open source, CITAE is an equity-oriented alternative to today’s fragmented, mostly proprietary and English-first landscape.
Cyberbullying detection remains challenging due to diverse linguistic patterns used by offenders. The in- ductive biases of deep learning architectures add to this challenge. Single-model approaches often capture only part of abusive language. This results in distinct but partially overlapping error spaces, limiting their effectiveness in real-world scenarios. To address this issue, we propose BBCS-Net. This is a modular hybrid framework that leverages diverse inductive biases. It combines Convolutional Neural Networks (CNNs), Bidirectional Long Short-Term Memory (BiLSTM) networks, and a margin-based Support Vec- tor Machine (SVM) classifier. In the proposed framework, a Bidirectional Encoder Representations from Transformers (BERT) model is first fine-tuned on a publicly available fine-grained cyberbullying dataset consisting of approximately 47,000 samples across six classes. This generates contextualized word repre- sentations. CNN and BiLSTM models are then trained independently. These models learn complementary feature representations. CNNs capture localized contextual cues. BiLSTMs capture long-range sequential dependencies. Penultimate-layer features from both models are fused at the feature level and classified using an SVM. The proposed approach achieves an average accuracy and F1 Score of about 98.87% us- ing nested cross-validation for model selection. On a held-out test set, it achieves 94.75% accuracy and a macro-F1 score of 94.75%. Error-space analysis shows that BBCS-Net recovers many model-specific errors. It corrects about 70.37% of CNN-only errors and 57.69% of BiLSTM-only errors. This highlights the effectiveness of feature-level hybridization in overcoming inductive-bias-driven failure modes. Latency analysis confirms the framework’s practical feasibility. It achieves a single-sample inference latency of 17.75 ms. During batch processing, the ONNX runtime reduces this to about 4.18 ms per sample. This supports near real-time cyberbullying moderation.
Text sentiment analysis of the social media text faces challenges posed by unstructured data and labori- ous human labeling for intent-driven, hierarchical classification. This work compares conventional ML models (SVM, Naïve Bayes, Logistic Regression) with contextual DL models (DistilBERT) in terms of their performance on Sentiment140 dataset (1.6 million tweets) where a balanced 300,000 tweets were selected (200,000 training, 50,000 validation and 50,000 test). As a solution to the bottleneck of human labeling for detailed topic classification, a zero-shot classification pipeline that uses Natural Language Inference (NLI) for classification of 1,000 tweets into a two-layer taxonomy of 30 parent topics and 330 subtopics has been created without any human-labeled samples. SVM is able to achieve 81.52% accuracy, while DistilBERT scores 84.44% on 50,000 tweet test set and 83.0% on a small 1,000 tweets sample.
Credit card fraud (CCF) poses a growing threat to the global economy, with payment card fraud losses reaching approximately USD 33.8 billion in 2023. While deep learning approaches such as CNNs, autoencoders, and Transformer-based models have improved detection accuracy, their “black box” nature restricts adoption in high-stakes financial domains where transparent and auditable reasoning is essential for trust and regulatory compliance. To address this challenge, we propose MTLFraudDetect, a multi-task framework that combines fraud detection with complementary model-analysis techniques. The model combines fraud classification with two auxiliary tasks—data reconstruction and amount prediction—while employing curriculum learning and PCGrad to stabilize multi-task optimization. For model analysis, the framework combines reconstruction-based anomaly-sensitivity signals with SHAP-based decision attribution. Reconstruction errors indicate transaction-level and feature-wise deviations from patterns learned by the reconstruction branch, whereas SHAP examines how the classifier’s shared latent representation and transaction amount contribute to its final prediction. Under an explicitly separated development-test protocol with training-fold-only preprocessing and out-of-fold hyperparameter selection, MTLFraudDetect achieves competitive F1-score and AUC-PR performance while providing complementary evidence at the levels of global anomaly detection, feature-wise error localization, and post-hoc decision attribution. These results do not by themselves establish formal explanatory faithfulness or complete interpretability.
Calculating motion features frame by frame and organizing them into a 3D matrix is a typical CNN-based solution for human action recognition (HAR). With the widespread use of consumer electronics, reducing computational costs and enabling efficient edge-side action recognition have become a research hotspot. In this paper, we extract action key frames via a well-designed algorithm to reduce computational overhead, so that the proposed method can be deployed on mobile electronic devices. Then we construct local and global motion features from these key frames and feed them into a cascade neural network for action recognition. The primary contributions include three aspects. First, the strategic adoption of key frames is introduced to greatly reduce the number of input parameters. The number of key frames can be adjusted to adapt to the temporal scales of different actions. Second, multiple origin points are adopted to construct motion matrices with larger dimensions than those constructed using a single origin point. Thus, deeper neural networks can be employed to achieve higher recognition accuracy. Third, a cascade neural network is proposed for action prediction, which leverages global and local information to achieve better efficiency and accuracy. Experimental results on UTKinect-Action3D, Florence-3D and our self-built HanYue-3D datasets demonstrate that our method achieves accuracy and efficiency competitive with state-of-the-art (SOTA) approaches. Moreover, the flexibility of the proposed method enables users to readily balance effectiveness and efficiency, making it well-suited for resource-constrained mobile devices.
We present a comprehensive parameterized complexity framework for the cryptanalysis of classical ciphersystems composed of multiple cryptographic transformation layers. Our main theoretical contributionsare threefold. First, we prove that Cipher-Decode is fixed-parameter tractable (FPT) when parameterizedby the structural complexity vector (k, ℓmax, |Σ|), yielding an algorithm with running time O∗(|Σ|k·ℓmax ).Second, we establish that the problem is para-NP-hard when parameterized solely by the number of layersk, and W[1]-hard via a complete formal reduction from k-Clique, providing a tight complexity dichotomy.Third, we design a branch-and-bound algorithm with statistical pruning based on Kullback-Leibler divergenceand the Index of Coincidence, reducing the effective search space by two to six orders of magnituderelative to exhaustive search. Experimental evaluation on 47 benchmark instances spanning Vigenère,substitution, ADFGVX, and multi-layer composite ciphers confirms the theoretical scaling predictions(R2 = 0.9998) and demonstrates speedups of 115× to over 2000× against naive baselines, with directcomparisons against simulated annealing and beam search. To the best of our knowledge, this is the firstrigorous FPT analysis of multi-layer classical cipher cryptanalysis.
Considering the landscape information of the optimization problem represents a promising research direction, as landscape features can provide a metaheuristic algorithm with useful information for understanding the search state, allowing it to adjust the search strategy accordingly. As a successful branch of metaheuristic algorithms for solving single-objective bound-constrained problems, however, the original design of differential evolution (DE) does not sufficiently consider information about the landscape of the solution space. To address this, an effective DE based on reinforcement learning (RL) that dynamically selects a suitable mutation operator via landscape information during the convergence process is proposed in this study, which consists of two phases: offline and online. In the “offline” phase, the proposed algorithm uses an RL-based algorithm to construct a learning model to understand the relationships between landscape characteristics and search operators. Then, in the “online” phase, the learning model constructed from the offline phase and a lightweight fitness landscape analysis (FLA) method are used by the DE to dynamically determine the suitable mutation operator based on the search state encountered by DE during the convergence process, which can reduce the costs for the FLA in every iteration. To understand the performance of the proposed algorithm, the CEC2021 and CEC2022 benchmark functions are used to evaluate its search performance against different DE-based algorithms for solving single-objective optimization problems. Simulation results show that the proposed algorithm outperforms other state-of-the-art DE-based algorithms and other DE algorithms based on FLA and RL in most cases.
Traditional cryptographic hash functions and existing neural hashing schemes suffer from a fundamental vulnerability: their deterministic nature renders them susceptible to rainbow table attacks and precomputation threats. To address this, this paper presents the Temporal-Stochastic Neural Cipher (TSNC), a novel framework that bridges dynamic security with deterministic verification. Unlike static architectures, TSNC eliminates the computationally expensive training phase by leveraging account creation timestamps and passwords to generate cryptographically secure seeds. The framework establishes a robust encryption pipeline through three logically cascaded modules: (1) timestamp-driven dynamic dictionary construction that ensures temporal uniqueness for character embeddings; (2) sequential feature extraction utilizing recurrent transformations to capture inter-character dependencies; and (3) a stochastic neural network with input-dependent dynamic topology, where the network depth and width adaptively evolve based on intermediate features. Theoretical analysis confirms the method's computational irreversibility and intractability against reverse engineering. Comprehensive simulations validate TSNC's security properties, demonstrating strict collision resistance (zero collisions in $10^7$ trials), a profound avalanche effect with output feature vectors exhibiting near-zero cosine similarity under minimal input perturbations, and temporal orthogonality. Furthermore, the algorithm incorporates a constant-time execution mechanism, effectively mitigating timing side-channel attacks while maintaining high computational efficiency. To promote reproducibility and further research, the source code of the proposed algorithm is publicly available at https://github.com/yongqianxiao/TSNC.
Accurately differentiating between fungal and bacterial infections is critical for treatment but remains challenging due to slow manual diagnostics and class imbalance in medical datasets. This paper introduces OD-Abs, a one-class outlier detection classifier based on an autoencoder architecture to distinguish infection types in microscopic images. By training solely on the reference fungal class, the model learns to identify bacterial instances as outliers. Our approach yields statistically significant improvements compared to other one-class approaches, achieving an F1-score of 0.967, an AUC of 0.980, and a specificity of 0.992. Additionally, we explore the explainability of our method by analysing reconstruction error maps from both a technical and biological perspective and conclude that the model’s decision-making aligns with biologically relevant morphological features. This proof of concept study establishes the OD-Abs autoencoder model as a highly robust tool for handling the task of classifying images of bacteria and fungi from abscesses.
Convolutional Neural Networks (CNNs) have achieved significant performance in various computer visiontasks, but at the cost of enormous computing resources. To alleviate this dilemma, this paper proposes alightweight and efficient multiscale attention fusion module (SCEMA) by combining spatial, and channelreconstruction convolution with an efficient attention mechanism and adds the module to the YOLOV8 networkstructure named YOLO-SCEMA. SCEMA adopts a parallel processing strategy, with the left branchperforming feature refinement operations through spatial and channel reconstruction units to reduce redundantcalculations, and the right branch effectively integrating features of different scales through featuregrouping and cross-spatial learning, enhancing the model’s understanding of multiscale image contentand improving its performance in handling complex image structures. The experimental results on the opensource datasets ExDark,VisDrone2019 and FYP show that YOLO-SCEMA has increased the mAP(50) scoreby 7.37%,3.24% and 1.5% , and compared to the YOLOv8 benchmark while reducing the parameter andcomputational complexity by 36.9% and 8.6%, respectively. Compared with the latest YOLO series, YOLOSCEMAperforms better in detection accuracy and parameter quantity.
Genetic Programming (GP) can be used to evolve human-interpretable Behavior Tree controllers for realtimestrategy games. Current Behavior Trees-based GP approaches compute fitness only at the end ofeach episode, which does not allow learning from finer-grained tactical decisions during gameplay. In thiswork, we integrate tabular Q-learning within the BT controller to gate actions at terminal nodes based onlearned action-values, and collect additional rewards at each tick of the game that are used to augmentthe final fitness signal. Since the Q-table is reset at the start of each episode, our approach allows thelearned information to affect fitness, while keeping the learned values from one generation separate fromsubsequent generations (a form of Baldwinian learning). Applied to plain-terrain MicroRTS micromanagementchallenge (population of 100 BTs, 2000 generations, against a deterministic rush opponent), ourapproach achieves a maximum fitness of 26.44 compared to GP’s max fitness of 22.0, a relative improvementof 20.2%. Mean best-fitness was also increased by 17.8% compared to the GP-only baseline. Theseresults are reported from a single representative run per configuration; multi-seed replication is identifiedas a priority for future work. The additional per-episode RL reward signal is strongly correlated with elitefitness (r = 0.89), confirming that it provides informative guidance for the evaluation. Unit coordinationexhibited by the hybrid agents is also found to be more structured, with defined front-line and support roles.The agent maintains its interpretable Behavior Tree structure.
Lithium-ion batteries have been widely utilized in modern society due to their excellent performance, making the accurate prediction of their remaining useful life (RUL) of paramount importance. Current RUL prediction methods are primarily categorized into battery discharge model-based approaches and data-driven approaches. However, the charging and discharging processes of batteries are typically accompanied by nonlinear variations in internal structural parameters, which pose significant challenges for model-based prediction methods. To address this issue, this paper proposes a predictive model based on Long Short-Term Memory (LSTM) networks optimized by an Improved Dream Optimization Algorithm (IDOA). First, to overcome the limitations of the original Dream Optimization Algorithm (DOA)—namely, the tendencies to become trapped in local optima and slow convergence—an optimal point set is employed for population initialization, and an adaptive population reduction mechanism is introduced to accelerate convergence. Second, the IDOA is utilized to optimize the hyperparameters of the LSTM network, thereby enhancing its predictive capability. Finally, simulation experiments are conducted using the CALCE and NASA battery datasets via the leave-one-out cross-validation method. The performance is evaluated using the average Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) calculated over ten independent runs. Comparative analysis with the original DOA and other baseline models demonstrates that the proposed IDOA-LSTM model can accurately predict the RUL of lithium-ion batteries with higher precision. Specifically, for the four batteries in the CALCE dataset, the RMSE values are 0.0201, 0.0198, 0.0213, and 0.0186, while the MAE values are 0.0181, 0.0174, 0.0163, and 0.0154, respectively. The results indicate that the average prediction accuracy of the proposed model is improved by 53.8% compared to the unoptimized model.
The explosive expansion of Low Earth Orbit (LEO) satellite mega-constellations presents critical challenges in operational risk management, real-time task allocation, and dynamic resource management due to their inherent time-varying topology and systemic uncertainties. To overcome the limitations of traditional optimization methods, this paper proposes a novel, hybrid artificial intelligence framework integrating K-means Clustering, Fuzzy Logic (FL), and Deep Reinforcement Learning for Task Allocation (DRL-TA). First, an adaptive K-means clustering mechanism periodically segments the constellation based on real-time channel quality and traffic load, effectively reducing network state dimensionality and overhead for the DRL agent. Second, a Fuzzy Inference System is employed to model non-deterministic operational elements (e.g., temperature, orbital deviation) and predict the real-time Safety Risk Score 〖Fuzzy〗_risk. This interpretable risk score is integrated as a penalty term within the DRL agent's reward function. Finally, the DRL-TA algorithm learns the optimal policy for computationally offloading and resource allocation by jointly analyzing the clustered network state and the fuzzy-predicted risk. Validated on a simulated 1,000-satellite LEO constellation over 10,000 training episodes, the integrated DRL-TA framework demonstrates significant performance gains: achieving a 25% reduction in average task completion delay and a 15% improvement in overall task success rate compared to conventional load-balancing and pure DRL baseline methods. The DRL policy exhibited stable convergence within 8,500 episodes, with a final average episodic return accuracy exceeding 97% of the theoretical maximum. This demonstrates the framework's efficacy in creating a reliable, high-performance, and risk-aware LEO edge computing environment.
In power system intelligent fault detection, real-time monitoring is critical due to grid complexity. To address the challenges posed by the complexity of continuous action selection and gradient estimation errors in transmission line fault monitoring, we developed a Deep Deterministic Policy Gradient (DDPG) algorithm enhanced by the Stochastic Variance Reduction Gradient (SVRG) method, termed SVRG-DDPG. This algorithm leverages the SVRG technique to mitigate the gradient estimation errors typically encountered in the DDPG algorithm. Utilizing the SVRG-DDPG framework, we further developed a transmission line fault monitoring model that directly employs real-time voltage sensor data from actual transmission line environments, encompassing a range of state information from normal operation to fault conditions. To achieve real-time monitoring and optimize system performance, we also propose a transmission line fault monitoring system based on a lightweight edge computing architecture. Using real-time voltage sensor data, the SVRG-DDPG-based fault monitoring model achieves a residual error within 30 kV accuracy. To enable real-time fault diagnosis in resource-constrained edge environments, we propose a lightweight edge-cloud collaborative architecture that dynamically allocates computational resources based on fault severity and sensor data volume. The framework is validated using high-fidelity simulation data from a power grid (covering 8 fault types under 12 operational conditions, e.g., humidity >90%, load fluctuations ±40%), which aligns with the dynamic resource demands of edge devices in practical systems. Finally, our method achieves 92% accuracy in fault diagnosis with 42% lower latency compared to baselines, leveraging SVRG-enhanced DDPG for adaptive edge-cloud synchronization. Experimental results on real-world 5G-V2X data validate its suitability for low-latency transmission line monitoring.
