In recent years,with the rapid development of AI,the field of autonomous driving is already booming around the world.Autonomous driving is considered to be an inevitable trend in the future development of the automotive industry and will fundamentally change the way we travel in the future.Perception function is the key link of autonomous driving,and it is the guarantee of driving intelligence and safety.Accurate and real-time 3D object detection is the core function of autonomous vehicles to accurately perceive and understand the surrounding complex environment.It is also the basis of decision control processes such as path planning,motion prediction and emergency obstacle avoidance.3D object detection task should not only predict the category of the target,but also predict the size,distance,position,direction and other 3D information of the target.China's traffic road situation is very complex,to achieve high-level autonomous driving requires a variety of sensors to work together.The use of multi-sensor fusion sensing scheme can improve the vehicle's ability to interact with the real world.At present,the advanced multi-sensor fusion detection models are too complex and have poor scalability.Once one of the sensors is wrong,the whole system will not work.This limits the ability to deploy high-level autonomous driving application scenarios.Then,the visible light sensor has some disadvantages in night,rain,snow,fog,backlighting and other scenes,which reduces the safety of driving.To solve the above problems,this paper based on the advantages of the MEMS LiDAR and the infrared camera to design a separable fusion sensing system,which is simple,lightweight,easy to expand and easy to deploy,which realizes 3D object detection task.Set the LiDAR and the infrared camera as separate branch.The both can not only work independently but also work together,decoupling the interdependence of the LiDAR and the infrared camera.If a sensor fails,the other sensor will not be affected,which improves the deployment capability of the model.The model uses the Bird's Eye View(BEV)space as a unified representation of the two different modes.The advantage of BEV space is to simplify complex 3D space into 2D space and unify the coordinate system.It makes cross-camera fusion,multi-view camera merging and multi-mode fusion easier to achieve.The camera branch and the LiDAR branch unify the 2D space and the 3D space into the BEV space respectively,and solve the difference problem of the data structure representation and spatial coordinate system of the two different sensors.The camera branch chooses YOLOv5 algorithm as the feature extraction network.The YOLOv5 algorithm is widely used in the engineering field and is easy to deploy.Accurate depth estimation is the key for the camera branch to transform image features into BEV features.So,this paper improves the camera branch.It introduces the pointclouds Depth Supervision(DSV)module and the Camera Parameter Prior(CPP)module to enhance the depth estimation ability of the camera branch.The GPU accelerated kernel is designed in image-BEV view transformation to improve the speed of model detection.Although the camera branch performance is limited,when applied to the fusion branch,the fusion branch can significantly improve the performance of single-mode branch.The LiDAR branch can choose any SOTA pointclouds detection model.The fusion branch use a Gating Attention Fusion(GAF)mechanism to fuse BEV features from different branches,and then completes the 3D object detection task.If one of the sensors fails,the camera branch or the LiDAR branch can independently complete the 3D object detection task.This model has been successfully deployed to an embedded AI computing platform:MIIVII APEX AD10.The experimental results show that the proposed model is effective,easy to extend and easy to deploy.

With the continuous expansion of artificial intelligence in power system, the reliability, security and operation ability of power system have been greatly improved. With the massive collection of data for transmission and transformation operation and maintenance, in-depth learning, as a typical representative of data-driven in the system of artificial intelligence methods, is increasingly appearing in the application scenarios of intelligent state assessment of transmission and transformation equipment. Its excellent robustness can eliminate many kinds of interference in complex environment background, accurately identify and evaluate the target points under test, and greatly reduce the detection of transportation and inspection personnel. The workload of data analysis, processing, evaluation and other links can improve work efficiency and accuracy. This paper attempts to summarize the development process of in-depth learning, relevant landmark achievements and typical applications in the operation and inspection of power transmission and transformation equipment. This paper attempts to summarize the research progress of target recognition, target segmentation, image classification algorithms in depth learning in different spectral cameras of visible light, ultraviolet corona imaging and infrared thermal imaging.

本文分析了大豆油基植物绝缘油和棕榈油基植物绝缘油的相对介电常数、介质损耗因数、体积电阻率等参数,在不同温度下测试了两种植物油纸绝缘系统的频域介电谱.结果表明:两种不同油基的植物绝缘油相对介电常数不同;黏度低的植物绝缘油极化损耗更强;植物绝缘油的极性越强,体积电阻率越低,电导率越高;温度对植物油纸绝缘的频域介电谱曲线有较大的影响,温度升高使频域介电谱曲线向高频方向平移.

为了全面研究绝缘结构对植物油纸频域介电谱特性的影响,搭建了植物油纸频率响应实验平台,采用3种不同电极对植物绝缘油、植物油纸绝缘结构、纸板间带1层油隙及带2层油隙的植物油纸绝缘系统进行测试.结果 表明:绝缘结构对频域介电谱影响较大,在低频下植物油纸的介质损耗因数最低,植物绝缘油的介质损耗因数最高;在中高频下植物油纸的介质损耗因数最低,植物油纸与2层油隙组成的油纸绝缘系统介质损耗因数最高;3种测试电极测得同种试品的曲线存在一定差异,但整体趋势相同,可通过计算消除不同电极对测试结果产生的影响.

Natural ester, as a new environmentally green insulating oil, has been widely used in transformer. In an oil-immersed transformer, the normal aging, thermal failure, and discharge failure could easily lead to the decomposition of the oil-paper insulation system and produce different kinds of gases. Studying gas dissolution in natural ester and mineral oil could provide assistance in applying criteria to make a diagnosis of different kinds of faults in the transformer. In this paper, the molecular dynamics method was used to investigate the diffusion behavior of seven fault characteristic gases (including H-2, CO, CH4, C2H2, CO2, C2H4, C2H6) in natural ester and mineral oil. The simulation parameters of free volume, interaction energy, mean square displacement, and diffusion coefficient were compared between the natural ester and mineral oil. Meanwhile, the influence of temperature on the diffusion of gas molecules in two kinds of oils was also analyzed. Results showed that the free volume, the interaction energy, and the relative molecular mass of gas molecules were the factors influenced by the diffusion of gas molecules in natural ester and mineral oil. The order of the diffusion coefficients of gas molecules in natural ester was as follows: H-2 > CH4 > CO > C2H2 > C2H4 > CO2 > C2H6 and that in mineral oil was as follows: H-2 > CH4 > CO> C2H2 > C-2 H-4 > C2H6 > CO2. By comparing the diffusion behavior of gas molecules in natural ester and mineral oil, it was found that the smaller free volume and higher interaction energy of gas molecules in natural ester were the major reasons for the gas molecules to be more difficult to diffuse in natural ester. The rising temperature could enhance the free volume and reduce the interaction energy between gas molecules and oil. The diffusion coefficient of gas molecules increased exponentially with the follow of temperature. However, the temperature didn't affect the ordering of diffusion coefficient, free volume, and interaction energy of gas molecules in natural ester and mineral oil.

Abstract Recently with the rapid growth of electric power literatures, it is hard to artificially track and process hot electric scientific researches. In the past most, professionals use simple statistics to get high-frequency words, which is time-consuming and ignores the similarity between words. Moreover, different researchers have different requirements for prediction time span. In the paper we propose a prediction system for hot electric scientific researches and gives its implementation. It is based on our previous work and we improve it to suit indefinite prediction period. The proposed embedded RNN prediction model is flexible for heterogeneous time spans and can return prediction results rapidly and accurately. Our extensive experiments demonstrated that our approach has acceptable precision ratio as well as training time in comparison to SVM, RNN and linear regression algorithms. It also performs better when the embedded layers are multiple.

电力变压器状态参数预测分析可以为设备状态评估提供有力技术支撑。现有变压器状态参量预测模型主要基于单一或少数状态参量进行分析和判断,预测稳定性和科学性都有待提高。文中结合变压器设备大量状态信息、电网运行和环境气象数据,提出了一种考虑复杂关联关系深度挖掘的变压器状态参量预测方法,通过栅格长短时记忆网络提取各参量之间蕴含的内在规律和关联关系,用以修正状态参量的预测结果。将该方法用于某500kV变压器顶层油温趋势预测中,结果表明所提出的栅格长短时记忆网络能挖掘分析设备状态影响因素之间的关联关系,与未考虑关联性的预测方法及传统方法相比,通过参量序列间关联性提高了预测模型的稳定性,降低了预测误差。

油中溶解气体分析是变压器内绝缘故障诊断的重要方法之一.但误判案例分析表明,传统的基于深度信念网络(deep belief network,DBN)油中溶解气体故障诊断方法存在较多的局部放电、低温过热、低能电弧放电兼过热混淆等误判.为进一步提高故障诊断效果,提出一种面向变压器油中溶解气体分析的组合DBN故障诊断方法.该方法引入深度信念网络群识别故障类型及严重程度,根据第一层故障类型识别结果激活相应的二层DBN识别故障严重程度.研究不同输入下,网络层数、节点数对于组合DBN的油中溶解气体故障诊断准确率的影响,结果表明当输入为无编码比值加特征气体含量,网络层数选取为3时网络具有最高准确率;当网络节点数大于3,增加节点数无法显著提高网络识别准确率.组合DBN查准率及查全率均高于单一DBN,总体准确率由80.9％提高到90.1％.分析案例数据量对诊断结果的影响,查全率及查准率随数据量增加而增加,案例多的故障类型查准率高于案例少的故障类型.

油中溶解气体阈值是判断变压器异常的重要指标,其判断阈值来自于历史数据分类统计,而选取分类条件是计算阈值的前提.针对高维数据阈值计算分类条件难以确定的问题,提出一种基于分类水平的油中溶解气体分类条件选取方法.该方法通过SS-FCM计算数据子类中心点,进而求取分类集几何中心及分类水平,根据分类水平判断分类条件优劣并进行分类,结合威布尔分布模型计算差异化阈值.研究发现分类水平高的条件统计威布尔函数拟合度较高,并且计算得到的阈值误报率较小,相比于标准阈值具有更高的准确性,有效识别变压器潜在过热缺陷.

High-voltage test is critical to electrical equipments, and test apparatus used in high-voltage test and their test items are various and complex. In the paper we focus on the standardized management of test apparatus. Adapting to information-data-intelligent tendency, we construct a comprehensive management system for high-voltage electrical test apparatus to improve test accuracy and efficiency. We use two-dimensional bar code to build transceiver system. Through simple scanning, the instruments could easily change their status and location, and we can also get all its data stored in the database. The online report generation are achieved as well based embedded template, bookmark and electronic signature. The trial operation shows that our detection processing time reduces more than 33 percents, which increases detection accuracy and reliability.

Scientific and technical information database for electric power becomes indispensable with the development of big data and think tank in electric industry. Querying information database is time-consuming and cache management is an efficient solution. In the paper we focus on a multi-level cache management in order to improve search speed of sci-tech information. The proposed computation method of cached data value considers visiting volume, data size and user experience, which is accurate for choosing key words and caching device. Our caching management framework consists of four modules, and the used machine learning method could effectively predict the number of visits next period for key words. The experiments demonstrate that our proposed MLCM approach performs better in prediction accuracy rate, query time and user experience than traditional statistic, LRU, LFU algorithms.

The correlative change analysis of state parameters can provide powerful technical supports for safe, reliable, and high-efficient operation of the power transformers. However, the analysis methods are primarily based on a single or a few state parameters, and hence the potential failures can hardly be found and predicted. In this paper, a data-driven method of association rule mining for transformer state parameters has been proposed by combining the Apriori algorithm and probabilistic graphical model. In this method the disadvantage that whenever the frequent items are searched the whole data items have to be scanned cyclically has been overcame. This method is used in mining association rules of the numerical solutions of differential equations. The result indicates that association rules among the numerical solutions can be accurately mined. Finally, practical measured data of five 500 kV transformers is analyzed by the proposed method. The association rules of various state parameters have been excavated, and then the mined association rules are used in modifying the prediction results of single state parameters. The results indicate that the application of the mined association rules improves the accuracy of prediction. Therefore, the effectiveness and feasibility of the proposed method in association rule mining has been proved.

Dissolved gas analysis (DGA) of insulating oil can provide an important basis for transformer fault diagnosis.To improve diagnosis accuracy,this paper studies a transformer fault diagnosis method based on rectified linear units deep belief networks (ReLU-DBN).By analyzing relationship between the gases dissolved in transformer oil and fault types,non-code ratios of the gases are determined as characteristic parameters of ReLU-DBN model.ReLU-DBN adopts multi-layer and multi-dimension mapping to extract more detailed differences of fault types.In this process,the parameters of diagnosis model are pre-trained,and adjusted with back propagation algorithm with sample labels.Performances of the ReLU-DBN diagnosis model are analyzed with different characteristic parameters,different training dataset and sample dataset.Besides,influence of discharge and overheating multiple-fault on diagnosis model is studied.Diagnosis effect of the model with non-code ratios as characteristic parameters is better than those of the models with IEC ratios,Rogers ratios and Domenburg ratios.Compared with the results derived from support vector machine (SVM) and back propagation neural network (BPNN),the proposed ReLU-DBN method significantly improves accuracy for power transformer fault diagnosis.Diagnosis effect of the model considering multiple-fault is better than that of the model without multiple-fault.With increase of sample dataset,the diagnostic accuracy is improved.

Based on the detection and analysis of the abnormal phenomenon in UHV transformer partial discharge test, this paper eliminates the influence of external disturbance and the test system, analyzes the structure characteristics of transformer magnetic shielding and grounding, and carries out the discharge location. It first discovers the fracture of magnetic shielding ground line in UHV transformer, which causes the magnetic shielding floating potential discharge and thereby produce a large amount of acetylene gas. The defect is eliminated successfully, and then the transformer operates normally by replacing the magnetic shielding ground line. The analysis shows that the integrated use of partial discharge test, ultrasonic positioning method and oil chromatographic analysis can accurately diagnose defect types and position them.

传统的变压器局部放电模式识别算法由于需调整的参数多且难以确定最佳参数、学习速度慢等缺点,在实际工程应用中识别正确率低,识别速度慢.因此,提出了一种基于在线序列极限学习机(OS-ELM)算法的变压器局部放电模式识别方法,该算法是传统极限学习机(ELM)的在线学习改进算法,是一种新型的单隐含层前馈神经网络(SLFN).本文基于特高频检测法在真型变压器上进行局部放电实验,并获得大量实验数据.将本文所提方法与ELM、支持向量机(SVM)以及BP神经网络(BPNN)的模式识别效果和性能进行了比较分析.结果表明OS-ELM算法识别正确率比SVM和BPNN分别高出5.2％和23.2％;逐渐减小训练样本集大小,OS-ELM识别结果的波动明显小于SVM和BPNN,表现出更好的泛化能力;OS-ELM的训练时间仅为0.031 2 s,远远小于SVM和BPNN.因此,OS-ELM更适用于大数据量样本的工程应用.

变压器是电力系统的枢纽设备之一,基于油色谱的在线监测技术可以实时获取油中溶解气体含量的时间序列,从而实现对变压器的故障预警和辨识。在利用时间序列进行分析时,选择过多或者过少的数据均会导致分析结果不准确。因此,该文提出了基于相空间重构的油色谱分析数据最优长度选择方法。首先,利用李雅普诺夫指数验证时间序列是否具有混沌特性,对于具有混沌特性的时间序列数据,利用延迟坐标法重构其相空间,然后利用C-C法计算嵌入维数和延迟时间,在获得嵌入维数的基础上,通过分析嵌入时间序列的关联积分的收敛性,得到时间序列长度与嵌入维数之间的关系,从而可以得到数据的最优长度。最后,该文给出了计算变压器油色谱数据最优长度的流程,并利用获取到的现场的案例进行验证,验证结果表明:使用最优油色谱数据计算得到的结果与原来的计算结果一致,但是计算量和计算时间显著减少,计算效率显著提升。

Accurate evaluation, diagnosis and prediction of operation state of transmission line can provide technical support for safe, economic and efficient operation of power system. Traditional transmission line analysis and prediction model typically uses a single parameter, but transmission line operating state is also affected by meteorological conditions, operating conditions and other factors. Therefore, due to low quality of the measured data and random nature of environmental conditions, the traditional methods have great limitations in forecasting accuracy and timeliness. An association rule mining method based on Bayesian network is proposed in this paper, to mine association rules between parameters, more directly reflect correlation between the data and effectively improve computational efficiency. The mined association rules applied to forecast circuit state parameters can improve accuracy of predicted results. Finally, taking a 500 kV transmission line as an example, association rules are extracted, and the rules are used to predict load and temperature. The results show that the method can improve prediction accuracy, thus verifying validity and feasibility of the association rule mining methods.

In order to improve the safety of power equipment, the influence of thermal aging on the performance of insulating paper used for transformers was investigated, the same type insulation paper with different basis weight was used to simulate the thermal aging process. The paper folding endurance, the tensile strength and the fiber polymerization degree (DP) of the insulation paper were measured before and after the thermal aging. The result showed that thermal aging has deteriorated the paper's properties, the DP was decreased. The retention rate of the paper folding endurance was about 10% of the base paper, and the retention rate of the tensile strength was about 70% of the base paper. Compared to the tensile strength, the connection between the fibers has a greater impact on the paper folding endurance of the insulating paper. Therefore, paper folding endurance was a more realistic and efficient method to characterize the thermal aging degree of the insulating paper.

Big data analysis platform consists of a serious of application and technologies for gathering, cleaning, storing, and analysing, which contribute to manage large-scale information, get analysis result and make decision precisely. However, traditional data analysis platforms meet new challenges as both continued growth in both scale and complexity in real data, existing data analysis platforms rarely have comprehensive functions for users. In response to these challenges, the BDAP (Big Data Analysis Platform) is proposed. BDAP supports large-scale data storage, parallel ETL process, data mining, statistical analysis, text mining and power grid data processing which are based on Spark. This demo introduces two application scenarios, transformer power prediction and video classification are presented. Experimental results show that BDAP is efficient and effective for data mining.

在分布式计算和内存为王的时代,Spark作为基于内存计算的分布式框架技术得到了前所未有的关注与应用。着重研究BIRCH算法在Spark上并行化的设计和实现,经过理论性能分析得到并行化过程中时间消耗较多的Spark转化操作,同时根据并行化BIRCH算法的有向无环图DAG,减少shuffle和磁盘读写频率,以期达到性能优化。最后,将并行化后的BIRCH算法分别与单机的BIRCH算法和MLlib中的K-Means聚类算法做了性能对比实验。实验结果表明,通过Spark对BIRCH算法并行化,其聚类质量没有明显的损失,并且获得了比较理想的运行时间和加速比。