Oil and gas pipeline leakage detection is a critical technology for ensuring energy transportation safety and protecting the ecological environment. However, practical applications face core challenges such as data scarcity, fluctuations in working conditions, and hardware limitations of edge devices. Existing methods struggle to balance detection accuracy, model lightweightness, and complexity. In response, this paper proposes a lightweight retrospective prototype-enhanced convolution fault diagnosis model, termed LiRetroFormer (Lightweight Retrospective Prototype-enhanced ConvFormer). The model employs multi-scale separable convolution (SMC) and broadcast self-attention (BSA) for feature extraction. By incorporating channel timing decoupling and a timing weight broadcast mechanism, LiRetroFormer rigorously controls the number of parameters while preserving multi-scale local features and global context capture capabilities. The retrospective prototype network (RPN) serves as the classification module, which dynamically optimizes the class prototype using a historical prototype buffer, thereby enhancing the characterization accuracy of leakage signal features. Experiments conducted on local and public datasets show that LiRetroFormer achieves high-precision identification of leakage conditions with aperture sizes from 0.4 mm (mm) to 10 mm. Compared with mainstream lightweight network MobileNet, LiRetroFormer obtains competitive accuracy with only one-tenth of the parameters. The proposed artificial intelligence-based model provides an efficient technical solution for lightweight pipeline leak detection in real industrial scenarios, showing good engineering application value.

The safe operation of the natural gas pipeline is very important for energy transmission. The simultaneous leakage of two points can easily lead to serious safety accidents. However, when the traditional positioning method faces this scenario, it is difficult to accurately separate the signal and estimate the time delay due to the aliasing of multi-source signals. There are problems of low positioning accuracy and weak anti-interference ability. In view of this, this paper proposes a AVMD-SOBI-GSCC two-point leakage location method : firstly, the dynamic characteristics of the leakage signal are restored by AVMD preprocessing technology, and the number of sources is determined by the leakage detection algorithm and the signal modal characteristics are extracted. The SOBI algorithm is used to realize the blind separation of multi-source aliasing signals and obtain the independent and effective leakage signals at the head and end of the pipeline. Then, the corresponding relationship between the leakage source and the sensor signal is established by similarity analysis. Finally, the GSCC algorithm is used to improve the accuracy of time delay estimation, and the accurate positioning of the two-point leakage position is completed by combining the propagation characteristics of natural gas in the pipeline. Experiments show that the average relative positioning error of this method is controlled within 2%, and the positioning success rate of this method is 98 %, which is 42 % lower than that of the traditional VMD-SOBI method, and 15 % higher than that of the EEMD-GSCC method. It provides a reliable technical solution for two-point leakage detection of pipelines and can be effectively applied to real-time monitoring system of natural gas pipelines.

The frequent occurrence of fires underscores the critical need for early fire detection, yet initial stages involve minute, irregular sources and sparse smoke, often obscured by complex backgrounds and lighting variations. Existing methods struggle with these faint signals, leading to high false alarm and missed detection rates. To address this, we propose FA-YOLOv10, an enhanced YOLOv10 framework employing a “two-stage network optimization and one-stage training strategy”. First, the C2f-MALA module suppresses background noise to reduce false alarms. Second, the PSAGMLCA module captures long-range dependencies to extract sparse edge features, minimizing missed detections. Complementing these, our training strategy integrates HSV-based preprocessing for consistent color feature separation under varying lighting, and loss function weight optimization to accelerate convergence on small targets. Experiments on the 6,650-image FireSmoke dataset demonstrate that FA-YOLOv10 outperforms state-of-the-art baselines. Specifically, compared with the YOLOv10n baseline (Table 11, Row 13), FA-YOLOv10 achieves a 1.7% increase in mAP0.5, a 3.8% improvement in precision, and a 0.4% improvement in recall. These quantitative gains—reflecting effective reductions in both false alarms and missed detections—validate FA-YOLOv10’s practical applicability in real-world fire warning scenarios.

To address the difficulty of building discrimination in complex regions and the insufficient extraction of boundary details in remote sensing imagery, this paper proposes a Boundary-Aware Deformable Convolution Feature Fusion Network (BA-DCFFNet) for building extraction. The proposed network adopts a dual-branch architecture composed of CNN and Transformer to capture local detailed information and global contextual dependencies, respectively. Firstly, an Enhanced Deformable Convolution Feature Fusion Block (EDCFFB) is introduced to achieve adaptive alignment and fusion of multi-level features, thereby improving the extraction of buildings with large-scale variations. Then, an Uncertainty-Aware Weighting Block (UAWB) is designed to emphasize hard regions, such as small-scale buildings, blurred boundaries, and shadow-occluded areas, so as to enhance the network’s discrimination capability in challenging scenes. In addition, a Boundary Feature Enhancement Block (BFEB) is introduced to further improve the accuracy of building extraction. Experimental results on different datasets show that the proposed method effectively improves the performance of building extraction.

To overcome the limitations of single-modal detection in pipeline leakage classification, this paper proposes a multimodal fusion framework for pipeline leakage detection, leveraging one-dimensional signals and Gramian Angular Difference Field (GADF) images. The framework enhances detection performance by leveraging the complementary nature of spatio-temporal features across modalities. First, we propose an improved GADF method based on wavelet threshold filtering. The improved GADF method suppresses noise while preserving impulsive features, thereby enhancing the clarity of high-frequency information in GADF images. Second, the spatial features extracted via the Swin Transformer and the temporal features obtained through the Gated Recurrent Unit (GRU) are fused in the spatio-temporal domain, leading to a significant improvement in fault classification accuracy. Furthermore, a Weighted Mutual Cross-Attention (WMCA) mechanism is introduced to generate joint features by weighted fusion of mutually attentive spatio-temporal and original features, effectively capturing cross-modal correlations. The proposed framework achieves an accuracy of 99.67% in the pipeline leakage detection tasks, significantly outperforming other multimodal methods. Furthermore, the average inference time per sample is merely 0.2225 ms, and inference efficiency fully satisfies the requirements of real-time detection in industrial applications.

In the industry, leak detection is a critical task to ensure the safe operation of oil and gas pipelines. Due to the low frequency and short duration of pipeline leakage in actual operation, and the small leakage is easily drowned by noise, these factors affect the accuracy and generalization of the leakage diagnosis model. This study proposes a new multimodal few-shot leak detection method based on spatio-temporal fusion. First, we combine VQ-VAE and short-time Fourier transform (STFT) to enhance one-dimensional leak signals. We then build a spatio-temporal bidirectional (STB) dataset by combining the enhanced signals with images converted from time-series data, improving data reliability. Next, we design a spatio-temporal fusion (STF) network, which uses the PKO algorithm to optimize key model parameters and integrates an intra-inter module (IIM) attention mechanism to fuse multi-source feature information, thereby boosting model performance. Finally, the proposed method is validated using pipeline data under different working conditions. The results show that this method achieves an accuracy of 98.67% in identifying various types of leak signals, comparative experiments demonstrate that this method outperforms other models, demonstrating strong robustness.

Natural gas pipeline leakage detection is crucial for ensuring the safe operation of energy transportation systems. However, reliable leakage-aperture identification remains challenging when only a limited number of samples are available. To address these problems, this study proposes a Hardness-Aware Weighted Prototypical Network (HAWPN) for few-shot natural gas pipeline leakage-aperture identification. HAWPN uses a one-dimensional residual attention encoder to extract discriminative embeddings from pipeline acoustic signals, a temperature-scaled similarity-weighted prototype strategy to improve prototype estimation under multi-support-sample settings, and hardness-aware prototype-distance regularization to enhance the separability of acoustically similar leakage-aperture classes. Experiments on a six-class natural gas pipeline leakage dataset showed that HAWPN achieved 98.61% ± 0.53% accuracy under the 6-way 1-shot setting, outperforming Vanilla ProtoNet by more than 10 percentage points. In the 6-way 5-shot setting, model performance approached saturation. These results demonstrate the effectiveness of HAWPN for few-shot leakage-aperture identification, particularly in the 1-shot scenario.

Accurate segmentation of oil spills in sea surface imagery is of critical practical importance for pollution detection, damage assessment, and emergency response. Nevertheless, marine scenes pose substantial challenges, including severe background clutter, pronounced class imbalance, multi-scale target variation, and indistinct object boundaries, which collectively limit the accuracy and robustness of existing segmentation models. To address these challenges, this study proposes an enhanced semantic segmentation framework based on Mask2Former, incorporating five dedicated modules. The Multi-Scale Feature Enhancement (MSFE) module strengthens multi-scale target representation through scale-adaptive weighting and residual scale augmentation. The Oceanic Context Aggregation (OCA) module enriches mask features by integrating directional strip pooling, a multi-scale context pyramid, and noise-aware gating mechanisms. The Class-Balanced Query Attention (CBQA) module alleviates class imbalance via class-aware query reweighting. The Boundary Refinement Module (BRM) and the Adaptive Boundary Upsampling (ABU) module jointly enhance mask predictions in boundary regions through residual refinement and boundary sharpening, respectively. Comprehensive comparative and ablation experiments conducted on the LADOS dataset demonstrate that the proposed method achieves superior performance over established baselines, including classical segmentation models such as U-Net, FCN, and DeepLabv3+, attaining an mIoU of 74.48%, fwIoU of 76.05%, mACC of 83.56%, and pACC of 86.38%. Particularly pronounced improvements are observed in underrepresented categories such as oil platforms and ships. These results substantiate both the individual effectiveness and the complementary nature of the proposed modules.

Traditional variational mode decomposition (VMD) struggles with parameter sensitivity, modal aliasing, and insufficient time-frequency resolution when processing nonstationary pipeline leakage signals, thus limiting its engineering applicability. Based on VMD, this article proposes an adaptive variational mode decomposition [PDE-AVMD, abbreviated as enhanced adaptive variational mode decomposition (EVMD) for simplicity]. First, the adaptive parameter optimization and dynamic convergence criteria are designed, and the iteration termination condition is automatically adjusted according to the signal complexity to realize the adaptive variational mode decomposition of the signal. Second, the partial differential equation (PDE) enhancement mechanism is introduced to construct a parameter evolution and constraint model based on the reaction-diffusion term, and the signal characteristics are adaptively adjusted to improve the convergence and stability and suppress the time-frequency leakage. Finally, new evaluation indexes modal reconstruction consistency index (MRCI), convergence stability index, and time-frequency resolution preservation index are proposed to establish a comprehensive evaluation system. Experiments show that EVMD is significantly superior to traditional algorithms in both self-collected pipeline data and open-source leakage diagnosis benchmark dataset. Compared with VMD, its mean square error is reduced by 34.9%, signal-to-noise ratio is increased by 4.93 dB, and MRCI is increased by 6.7%. It shows better robustness and accuracy in nonstationary, complex and real pipeline data, and provides a theoretically rigorous and engineering-feasible solution for intelligent fault diagnosis in complex industrial environments.

To address the issue of compromised detection accuracy due to noise interference in collected leakage signals during oil and gas pipeline monitoring, this study proposes a denoising method integrating Variational Mode Decomposition (VMD) and an Improved Red-Tailed Hawk Optimization algorithm(IRTH). The conventional Red-Tailed Hawk Optimization algorithm tends to converge to local optima when addressing complex structural optimization problems, making it challenging to locate global optimal solutions. To overcome this limitation, several improvements are implemented: the Halton sequence is introduced to achieve uniform population initialization, thereby enhancing initial population diversity; the simplex method is integrated during iterative updates to refine the positions of the worst-performing individuals and optimize overall population performance; finally, a tangent flight strategy is employed to expand the search space by simulating flight behavior, leveraging stochastic update rules to effectively prevent convergence to local optima and ensure practical applicability. Comparative experiments conducted on nine benchmark functions validate that the enhanced RTH algorithm exhibits significant improvements in both convergence speed and optimization accuracy. To mitigate the issues of over-decomposition and mode deficiency commonly encountered in VMD, the proposed enhanced algorithm is applied to optimize the parameter selection for VMD decomposition. The combined denoising approach is subsequently implemented in practical pipeline leakage signal processing, demonstrating effective noise reduction performance and offering robust technical support for improving the precision of oil and gas pipeline leak detection.

To address the limited recognition performance caused by insufficient samples in natural gas pipeline leak detection, this paper proposes a leak detection method that integrates continuous wavelet transform (CWT), efficient diffusion-model-based data augmentation, and a lightweight residual network (ResNet). First, CWT is employed to transform one-dimensional leak signals into two-dimensional time-frequency images, and a comparative analysis is conducted to determine the time-frequency representation method suitable for this task. Then, an efficient diffusion model is used to expand the training samples, thereby alleviating the negative impact of data scarcity under small-sample conditions. Finally, a lightweight ResNet classifier is constructed, in which the first convolutional layer is adjusted, channel widths are compressed, and an SE attention mechanism is introduced to improve model efficiency. Experimental results show that the proposed method can effectively improve leak recognition performance under small-sample conditions while maintaining a good balance between accuracy and computational efficiency, demonstrating promising potential for engineering applications.

The frequent occurrence of fires underscores the critical need for early fire detection, yet initial stages involve minute, irregular sources and sparse smoke, often obscured by complex backgrounds and lighting variations. Existing methods struggle with these faint signals, leading to high false alarm and missed detection rates. To address this, we propose FA-YOLO, an enhanced YOLOv10 framework employing a ”two-stage network optimization and one-stage training strategy”. First, the C2f-MALA module suppresses background noise to reduce false alarms. Second, the PSAGMLCA module captures long-range dependencies to extract sparse edge features, minimizing missed detections. Complementing these, our training strategy integrates HSV-based preprocessing for robust color feature separation under varying light and loss function weight optimization to accelerate convergence on small targets. Experiments on the 6,650-image FireSmoke dataset show FA-YOLO outperforms state-of-the-art baselines, achieving a 1.7% mAP increase, with 3.8% higher precision and 0.4% improved recall. These results validate FA-YOLO’s superior robustness and potential for practical deployment in complex fire environments.

Pipeline leakage detection is an important measure to ensure the national economy and public safety. However, since leakage signals are often contaminated by environmental noise, ensuring the accuracy of detection results is challenging. The paper presents a denoising method of Coarse-to-Fine decomposition based on Variational Mode Decomposition (VMD) in order to increase the detection precision. First, motivated by the bandwidth characteristics, a new iterative decomposition strategy of VMD is introduced to coarsely determine the optimal mode number. Then, the Whale Optimization Algorithm is employed to carefully search for the optimal balance parameter. After obtaining the optimal parameters, a series of modes are obtained from VMD. Finally, the denoised signal is reconstructed with the meaningful modes extracted through Kullback-Leibler divergence (KL). The effectiveness of the proposed method is validated using both simulated and actual leakage signal. The main innovation of this study lies in the proposed Coarse-to-Fine VMD strategy, which not only decouples the mode number (K) and balancing parameter (alpha), but also integrates KL divergence for adaptive mode selection, thereby ensuring superior robustness compared with conventional methods. The denoised results show that the SNR of the CTF-VMD-KL reaches 27.4047 dB, which is apparently better than that obtained by EMD-KL, VMD-KL, LMD-KL and CEEMD-KL for actual signal. The proposed method also surpasses popular methods in both denoising capability and signal integrity, establishing its superiority in pipeline leakage detection.

This paper presents an integrated UAV navigation system that combines 3D Gaussian Splatting (3DGS) for environment mapping, A* algorithm for global path planning, and Model Predictive Control (MPC) for trajectory tracking with obstacle avoidance. The 3DGS-generated Gaussian sphere map provides rich geometric information including opacity and scale parameters, which are utilized to construct a collision cost function embedded into the MPC optimization framework. Experimental results demonstrate that the proposed system achieves significant improvements in both path planning and trajectory tracking performance. The A* algorithm with Gaussian opacity-aware cost function reduces the integrated occupancy probability by 26.7% compared to traditional grid-based methods. The MPC controller with collision cost achieves a position RMSE of 0.030m, representing a 66.3% improvement over the baseline MPC without collision cost. The integrated system operates at an effective control frequency of 75Hz with an average single-step computation time of 13.3ms, meeting real-time requirements for UAV navigation in GNSS-denied environments.

In recent years, intelligent pipeline leakage detection technology has played a crucial role in ensuring pipeline safety and energy security. However, most existing methods assume balanced datasets, overlooking the inherent imbalance between normal and abnormal data in real-world scenarios. This limitation hampers effective feature extraction for anomaly detection. To address this challenge, we propose a novel multichannel and multi-branch one-dimensional convolutional neural network (MCB1DCNN). The model integrates a multi-channel convolution module and a multi-branch network structure to extract both global and local signal features. To mitigate the impact of data imbalance, we propose an adaptive weighted cross-entropy loss function. This function dynamically adjusts the loss weight of minority class samples based on the imbalance ratio. Furthermore, we construct a multi-channel acoustic signal dataset for oil and gas pipelines using the overlapping sample segmentation method. Variational mode decomposition (VMD) is applied to decompose acoustic signals into different frequency components, enabling comprehensive feature extraction. Ablation experiments analyze the impact of key model parameters. Experimental results show that MCB1DCNN outperforms several state-of-the-art methods in terms of accuracy, F1 score, false alarm rate, and missing alarm rate. These findings demonstrate its superior performance and practical applicability in real-world pipeline leakage detection.

During the actual industrial process,it is challenging to obtain samples of oil and gas pipeline leaks resulting in the scarcity of training data samples suitable for fault diagnosis. In order to tackle this issue, this paper suggests a method a Recursive Generalization self-attention generative adversarial network (RAGAN) aided by Wasserstein distance with gradient penalty. The initial step involves applying the short time Fourier transform to the acoustic signal of oil and gas pipeline leakage, treating it as real data. Subsequently, both the real data and random noise following a Gaussian distribution are fed into the generator. The output is utilised as a pseudo sample. The Wasserstein distance of the distribution of real data and fake samples is introduced as a loss term in the discriminator, and a gradient penalty is added. Finally, the network optimizes the parameters through back propagation until Nash equilibrium. PSNR and SSIM are used as sample reliability evaluation. The results show that the fake samples have high similarity with the real samples, which can be used to expand small sample data. Moreover, extending pseudo samples to small sample data sets can effectively improve the performance of fault diagnosis.

The safe operation of oil and gas pipelines is of vital importance for maintaining national energy security. Therefore, the implementation of efficient pipeline leakage detection is an important link to ensure the safe and stable operation of pipelines. In this paper, a pipeline leakage detection method based on an improved spiking residual network is proposed. First, a coding method is proposed to encode the original signal into a spiking sequence. The input oil and gas pipeline signals are encoded using short-time Fourier transform combined with spatial gating mechanism and LIF neurons. Second, wavelet convolution was introduced to improve the original spiking residual network. Finally, the improved spiking residual network is used to classify the pipeline signals after the coding process. The experimental results show that the classification accuracy of the model proposed in this paper reaches 100 % on the original signal data, and 95.62 % with the addition of 5 dB Gaussian white noise, which effectively shows that the method has high accuracy and strong robustness, and can effectively improve the oil and gas pipeline leakage detection effect.

Sewer systems are critical to smart city infrastructure, but conventional pipeline inspection methods cause high costs and inefficiency. This paper presents a real-time detection method for pipeline defects based on an improved you only look once version 5 (YOLOv5) algorithm. The proposed approach enhances the ability of the network to extract and fuse information by incorporating a selective kernel attention mechanism, a bidirectional cascade feature fusion structure, and an optimized loss function. Experimental results indicate that the proposed method can accurately identify and localize ten common types of defects. It achieves a mean average precision that is 4.5% higher than the original model and a frame rate of 69.9 frames per second, making it highly suitable for automated pipeline defect detection. Lastly, future research directions are outlined, including exploring lightweight architectures and adaptive mechanisms to improve the generalization of model to diverse defect types and environments.

Leak detection of oil and gas pipelines is essential for safe pipeline operation. Existing methods for detecting oil and gas pipeline leaks are inadequate and do not exploit the temporal characteristics of the leak signal. In this paper, a fusion recognition model of a long short-term memory network (LSTM) optimized by a one-dimensional convolutional neural network (1DCNN) and dung beetle optimization algorithm (DBO) is proposed. First, the one-dimensional signal is input into the 1DCNN. Then the adaptively extracted data features from the 1DCNN are input into the optimized DBO-LSTM for classification. The model is evaluated using the precision and duration metrics. In comparison to extant advanced models, the proposed model improves recognition precision and decreases detection time. The model proposed in this paper is able to extract pipeline data features more rapidly and precisely, thereby improving classification accuracy.