Abstract:To address the challenges in small object detection tasks,such as the small size of target images,blurred target features,and difficulty in distinguishing targets from backgrounds,a method based on dual-stream contrastive feature learning and multi-scale image degradation augmentation is proposed.First,the input images of the contrastive learning model are subjected to multi-scale degradation augmentation,thus enhancing the model's ability to perceive and capture small targets.Second,contrastive learning representations are conducted in both spatial and frequency domains simultaneously to learn more discriminative target recognition features,thereby improving the model's ability to differentiate between targets and backgrounds.To verify the effectiveness of the proposed scheme,ablation experiments are designed,and the detection performance is compared with that of other advanced algorithms.Experimental results show that the proposed scheme achieves an improvement of 3.6％ in mean Average Precision (mAP) over the baseline algorithm on the MS COCO dataset,and an improvement of 7.7％ in mAP for small objects (mAP_S) compared to mainstream advanced algorithms.On the VisDrone2019 dataset,the proposed method achieves a 2.4％ increase in mAP compared to the baseline algorithm,demonstrating its superior overall performance over the baseline algorithm and other mainstream advanced algorithms.Visual analysis of detection results indicates a significant improvement in the rates of false negatives and false positives for small object detection.

Abstract:To address the issues of blurry texture of the repaired images and instable training process in existing image inpainting algorithms,this paper proposes a Generative Adversarial Network (GAN) based image inpainting approach leveraging the diffusion process.By incorporating the diffusion model into a dual-discriminator GAN,the generated images from the generator and real images undergo a forward diffusion process to obtain the inverted images and real images with Gaussian noise.These images are then fed into the discriminator to enhance the inpainting quality and improve the model training stability.Style loss and perceptual loss are introduced into the loss function to learn semantic feature differences,eliminate motion blur,and preserve more details and edge information in the inpainting results.Qualitative and quantitative analyses,along with ablation experiments,have been conducted on the datasets of CelebA and Places2.The evaluation and restoration outcomes show the superior performance of the proposed approach.Compared with current inpainting methods,the proposed approach achieves an average improvement of 1.26 dB in Peak Signal-To-Noise Ratio (PSNR) and 1.84% in Structural Similarity Index Measure (SSIM),while reducing the L1 error by an average of 25.7%.Furthermore,the changes in the loss function indicate that the image inpainting algorithm with diffusion process exhibits more stable training behavior.

Abstract:Deep Neural Networks (DNNs) exhibit vulnerability to specially designed adversarial examples and are prone to deception.Although current detection techniques can identify some malicious inputs,their protective capabilities remain insufficient when confronted with complex attacks.This paper proposes a novel unsupervised adversarial example detection method based on unlabeled data.The core idea is to transform the adversarial example detection problem into an anomaly detection problem through feature construction and fusion.To this end,five core components are designed,including image transformation,neural network classifier,heatmap generation,distance calculation,and anomaly detector.Firstly,the original images are transformed,and the images before and after the transformation are input into the neural network classifier.The prediction probability array and convolutional layer features are extracted to generate a heatmap.The detector is extended from focusing solely on the model's output layer to the input layer features,enhancing its ability to model and measure the disparities between adversarial and normal samples.Subsequently,the KL divergence of the probability arrays and the change distance of the heatmap focus points of the images before and after the transformation are calculated,and the distance features are then input into the anomaly detector to determine whether the example is adversarial.Experiments on the large-scale,high-quality image dataset ImageNet show that our detector achieves an average AUC (Area Under the ROC Curve) value of 0.77 against five different types of attacks,demonstrating robust detection performance.Compared with other cutting-edge unsupervised adversarial example detectors,our detector has a drastically enhanced TPR (True Positive Rate) while maintaining a comparable false alarm rate,indicating its significant advantage in detection capability.

Abstract:With the continuous intensification of global climate change and the rapid urbanization,urban waterlogging disasters caused by extreme rainfall events have become increasingly severe,posing a serious challenge for many cities around the world.Here,we propose a deep learning approach to predict urban waterlogging depth,which is based on Long Short-Term Memory (LSTM) and rainfall data from May to August 2021 measured by 75 national automatic meteorological observation stations in Zhejiang's Zhuji city and the water depth data of a typical waterlogging site.The relationship between rainfall and waterlogging depth constructed by LSTM provides the next 2-hour urban waterlogging depth forecast with an interval of 15 minutes.When compared with Random Forest (RF) and Artificial Neural Network (ANN) models,the proposed LSTM approach,using water depth and precipitation data over the past 4 hours to predict the next 2-hour waterlogging depth,demonstrates the best performance by lower root mean square error (<5.6 cm),higher correlation coefficient (>0.93) and Nash-Sutcliffe efficiency coefficient (>0.86).It can be concluded that the proposed deep learning approach is feasible and applicable for urban waterlogging depth prediction.

Abstract:Aiming at the distributed permutation flowshop scheduling problem with variable processing speed,a dual-population algorithm is proposed to optimize the makespan and the total energy consumption of the machine.First,an initialization method that mixes four strategies is used to generate a high-quality initial population.Second,specific evolution methods are designed according to the characteristics of the two populations,and the dynamic guide factor is introduced to adjust the evolution mode of the populations.Meanwhile,an energy-saving strategy for speed regulation is proposed to further optimize energy consumption.Finally,a dynamic population strategy is proposed to balance the resources of the two populations.Simulation results verify the effectiveness of each strategy,and show that the proposed dual population algorithm outperforms current multi-objective evolutionary algorithms.

Abstract:Path tracking is essential for unmanned driving.This article presents the design of a path tracking system for unmanned trucks,aiming to enhance accuracy and stability across various speeds.The system employs a Linear Quadratic Regulator (LQR) optimized through an improved Genetic Algorithm (GA).First,a two-degree-of-freedom dynamic model and a tracking error model of the vehicle are established based on natural coordinate system.Subsequently,an LQR controller is designed to eliminate steady-state errors and enhance tracking accuracy through feedforward control.Second,the genetic algorithm is enhanced to optimize the weight matrix of the LQR controller,resulting in improved accuracy and stability for path tracking.Finally,the control effectiveness of the designed LQR controller is simulated and verified across a range of operating conditions using the joint simulation platform of Matlab／Simulink and TruckSim.The results show that the GA-optimized LQR (Linear Quadratic Regulator) controller improves the tracking accuracy by about 68.5% and 49.4% at speeds of 30 km/h and 60 km/h,respectively,under the double lane change scenario;while under the U-turn scenario,the tracking accuracy is enhanced by approximately 12.0% and 25.5%,respectively.Specifically,it demonstrates higher stability,with position and heading errors controllable within 0.17 m and 0.11 rad,respectively,thereby validating the efficacy of the proposed tracking control scheme.

Abstract:To tackle the complexity and nonlinearity inherent in natural gas load sequences,this paper proposes a combined forecasting model that integrates Time2Vec,LSTM (Long Short-Term Memory),TCN (Temporal Convolutional Network),and attention mechanism.Initially,the Pearson correlation coefficient is used to conduct the correlation analysis to extract the meteorological features that exhibit strong relevance.Subsequently,the time vector embedding layer of Time2vec is introduced to convert the time series data into a continuous vector space,thus enhancing the model's computational efficiency in processing time series information.Then the temporal features extracted by Time2Vec,alongside the meteorological features selected using Pearson correlation coefficient,are fed into both the LSTM and TCN models for prediction,exploiting the long-term memory capability of LSTM and the local feature extraction capability of TCN.Finally,these two models are combined through attention mechanism,and assigned different weights according to the importance of the two to obtain the final prediction results.The experimental results show that the proposed Time2Vec-LSTM-TCN-Attention model outperforms other combined models in terms of adaptability and accuracy for natural gas load forecasting.

Abstract:The performance of power equipment fault detection models is affected by various factors including fault type,fault complexity,and image quality.Here,a fault detection model based on TrellisNet and attention mechanism is proposed for power equipment.First,Long Short-Term Memory (LSTM) is integrated with Convolutional Neural Network (CNN) to construct LSTM-CNN to obtain fault characteristics in images,which can effectively distinguish features of different fault types and reduce the influence of noise and other interference factors.In addition,the feature data obtained by LSTM-CNN are used as input,and by embedding the attention mechanism into TrellisNet,an AT-TrellisNet network with high resolution is constructed to detect the fault type of different power equipment.Finally,five common power equipment faults are selected for model validation.The experiment results show that compared with some existing detection models,the proposed model has higher detection accuracy,with a maximum of over 90％,which can meet the actual needs of power equipment fault detection.

Abstract:To address the increased load peak-to-trough ratio and user costs caused by disorderly charging and discharging of electric vehicle charging piles in residential communities,an optimized operation strategy is proposed for energy storage charging piles to achieve orderly charging and discharging.While reducing the peak-to-trough ratio,the strategy aims to minimize users' charging costs and maximize charging pile profits.A typical day is selected to establish an optimized charging and discharging scheduling model for the energy storage charging piles,which is solved by an Improved Multi-Objective Particle Swarm Optimization (IMOPSO) algorithm,and the charging and discharging power and time of the energy storage charging pile is adjusted in combination with the time-of-use electricity price.The MOPSO algorithm is improved by optimizing the inertial weights,learning factors and adaptively changing the position splitting factor.Experimental data results show that the algorithm can effectively improve the convergence speed,avoid falling into local optimum,and better handle multi-objective problem.In the energy storage scheduling model,it reduces the typical load peak-to-trough ratio by 55%,optimized by 36% compared to the original algorithm,rationally allocates charging piles to store power resources during low-demand period,effectively reduces charging costs by 20% to 30%,and improves charging pile profits,thus achieving a win-win situation for the power grid,users and charging piles.

Abstract:The Modular Multilevel Converter-based High Voltage Direct Current (MMC-HVDC) overhead line transmission scheme is susceptible to instantaneous DC faults,and utilizing the Energy Storage Units (ESUs) installed within each wind turbine to absorb unbalanced power during faults is an effective solution.However,existing literature often considers the wind farm as a single Wind Generator (WG),neglecting the differences in residual capacities among individual ESUs.This approach easily leads to overloading of ESUs with smaller residual capacities,while those with larger residual capacity still have unutilized reserve capacities,resulting in power imbalance during faults.To address these issues,this paper proposes a coordinated control strategy for DC fault ride-through based on optimized control of ESUs within WGs.The strategy adopts the variance of the State of Charge (SOC) as an indicator to quantitatively describe the differences in residual capacities of ESUs,and takes the maximum decline rate of SOC variance as the objective function.The residual unbalanced power after the conversion of non-fault pole converter station is optimally allocated to the ESUs within individual WGs,so as to reduce the differences of residual capacities while ensuring the power balance of the system during faults.A model is developed on the PSCAD/EMTDC simulation platform to compare the proposed optimized power allocation scheme with the traditional average allocation scheme.The results show that the optimized allocation scheme fully utilizes the power absorption capacities of ESUs,thereby improving the DC fault ride-through capability of the system.

Abstract:The new power system requires cooperative and optimized scheduling in the integration of Generation-Grid-Load-Storage (GGLS).At present,the scheduling automation system employs a relational database that relies on multiple associated tables for data query and storage,making it difficult to meet the rapid computation demands.Here,we propose a graph-based computing method for intra-day scheduling of coordinated GGLS.Firstly,the graph database is used to integrate the spatiotemporal data from power generation,grid,loads,and storage.Secondly,a comprehensive optimization model of intra-day scheduling of coordinated GGLS is formulated,taking into account various resources such as thermal power units,adjustable loads,and energy storage.Thirdly,a graph-based power flow calculation approach is proposed to quickly perform system security checks.Finally,based on the security check results,the system operating status is corrected until all operational constraints are satisfied.Through analysis of improved examples on the IEEE118-node and IEEE1354-node systems,it is verified that the proposed coordinated optimization strategy for GGLS can improve computational efficiency.

Abstract:To address the large-scale environmental mapping,lightweight robot swarms are employed to perceive the environment and multi-robot collaborative SLAM (Simultaneous Localization and Mapping) scheme has been developed to solve the problems of high individual cost,global error accumulation,excessive concentration of calculation and risk perplexed single robot SLAM schemes,which has strong robustness and stability.Here,we review the history of multi-robot collaborative SLAM,and introduce its fusion method and architecture.The current collaborative SLAM approaches are sorted out from the viewpoint of machine learning classification.The future development trends of multi-robot SLAM in directions of deep learning,semantic maps,and multi-robot VSLAM are projected.

Abstract:To investigate the response of N₂O emission in wheat field to the gradually elevated CO₂ concentration and the reduction of nitrogen fertilizer application,we conducted field experiments using Yangmai 22 as the experimental material.The experiments were carried out on a field equipped with an automatic CO₂ concentration control platform formed by Open-Top Chambers (OTC).Under the condition of ambient atmospheric CO₂ concentration (AC),a treatment of gradually elevated CO₂ concentration (EC) was set up,which increased by 40 μmol·mol^-1 annually from the winter wheat growing season of 2016-2017,reaching 120 μmol·mol^-1 higher than AC in the wheat growing season of 2018-2019.The nitrogen fertilizer treatments were set as a conventional level (N₁,25 g·m^-2) and a reduced level (N₂,15 g·m^-2).The N₂O fluxes were measured by a static opaque chamber-gas chromatograph method.The results indicated that throughout the winter wheat growing period,the changes of N₂O emission from winter wheat field were relatively consistent for all treatments with an overall fluctuating downward trend.No significant effect of EC treatment on N₂O emission was observed during the winter wheat growing seasons.While the N₂ treatment significantly reduced the cumulative N₂O emission by 45.2% compared with the N₁ treatment (P=0.004) under the AC,which was not significant under EC treatment.The impact of nitrogen reduction on N₂O emission was evident during winter wheat stage from booting to milking.Under the combined effect of gradually elevated CO₂ concentration and reduced nitrogen fertilizer application,the latter was the main factor affecting N₂O emission in winter wheat field.

Abstract:Gaseous ammonia (NH₃) is a key precursor in the formation of PM_2.5 in the atmosphere.To enhance the precision of identifying NH₃ sources through isotopic techniques and alleviate urban air PM_2.5 pollution,this paper reviews recent researches on the following aspects:methods for sampling and quantifying gaseous NH₃,techniques for determining or estimating its nitrogen isotope ratios,the nitrogen isotope compositions and fractionations of NH₃ in air and emitted from diverse sources,the isotopic compositions of NH₄⁺ in atmospheric particulate matter and rainfall,as well as the source apportionment of atmospheric NH₃.Several recommendations for future research are accordingly proposed.The minimum amount of samples should be determined before collecting air samples with passive method considering nitrogen isotope fractionation of NH₃.It is necessary to clarify the nitrogen isotope compositions of ammonia from biomass combustion,natural soils,oceans,sewage plants,vegetation,and other potential sources.The nitrogen isotopic variation mechanism for ammonia from diverse sources needs to be further investigated.Both mass concentrations and isotopic compositions of gaseous NH₃ and particulate NH₄⁺ should be measured simultaneously with higher temporal resolution to explore the mechanism of nitrogen isotope fractionation during the formation of particulate NH₄⁺.Additionally,exploring the isotopic fractionation of gaseous NH₃ under diverse meteorological conditions and air pollution status is recommended.