YOLO电气工程图纸符号目标检测数据集建筑空间中的目标检测从标注到部署 数据集基本信息目标类别 [‘3wayswitch’, ‘AA’, ‘ACpower supply’, ‘AIR CIRCUIT BREAKER WITH SYN’, ‘AIR CIRCUIT BREAKER WITH SYN. FACILITY’, ‘AIR CIRCUIT BREAKER WITHOUT SYN’, ‘AIR CIRCUIT BREAKER WITHOUT SYN. FACILITY’, ‘AMMERTER’, ‘AMMETER SELECTOR SWITCH’, ‘AUTO MANUAL SELECTOR SWITCH’, ‘Air circuit breaker’, ‘Ammeter’, ‘Announciator’, ‘BRANCH CIRCUIT MINIMUM SIZE PER CKT BKR’, ‘BRANCH CIRCUIT MINIMUM SIZE PER CKT BKR - REMAIN’, ‘BUS BAR’, ‘Battery’, ‘Brake Cluthes’, ‘Busbar’, ‘CABLE TERMINATION’, ‘CT CURRRENT TRANSFORMER’, ‘Capacitive Voltage transformer’, ‘Circuit Breaker’, ‘Connection point’, ‘Connectors’, ‘Counter’, ‘Coupling capacitor’, ‘Current transformer’, ‘DIESEL GENERATOR’, ‘DISTRIBUTION TRANSFORMER’, ‘Disconnector Isolator’, ‘Disconnector Switch’, ‘Distribution Panel Meter’, ‘ELT’, ‘ELTR’, ‘EMERGENCY STOP PUSH BUTTON’, ‘EXISTING POWER -WHIP- CONNECTION TO CUBICLES - REMOVE’, ‘Earth ground’, ‘FIELD WIRING’, ‘Field isolator terminal’, ‘Generator’, ‘HOME RUN TO PANEL MINIMUM SIZE PER CKT BKR’, ‘HOME RUN TO PANEL MINIMUM SIZE PER CKT BKR - REMAIN’, ‘INDICATION LAMP’, ‘LED -Light Emitting Diode’, ‘LOW VOLTAGE WIRING’, ‘Light Dependent Resistor’, ‘MASTER TRIP RELAY’, ‘MCB MINIATURE CIRCUIT BREAKER’, ‘MCCB MOULDED CASE CIRCUIT BREAKER’, ‘MULTIFICTIN RELAY WITH IEC’, ‘MULTIFUNCTION METER’, ‘MUTIFUNCTION RELAY WITH IEC 61850’, ‘Motor terminal’, ‘PLC terminal’, ‘PT POTENTIAL TRANSFORMER’, ‘Potential transformer’, ‘Potentiometer’, ‘RCD Circuit Breaker’, ‘RESTRICTED EARTH FAULT RELAY’, ‘Resistor’, ‘SANDWICH TYPE BUS DUCT’, ‘SWITCH LEG - LETTER DENOTES CONNECTION ZONE’, ‘Single Phase energy meter’, ‘Switch disconnector fuse’, ‘Synchronous generator’, ‘TEST TERMINAL BLOCK’, ‘THREE-CONDUCTORS’, ‘TNC SELECTTOR SWITCH TRIP NEUTRAL CLOSE’, ‘TRIP CKT SUPERVISION RELAY’, ‘Transistor’, ‘Transistor -NPN-’, ‘VOLTMETER’, ‘VOLTMETER SELECTOR SWITCH’, ‘Wave trap’, ‘auxillarycontacts’, ‘branch’, ‘cabletray’, ‘camswitches’, ‘capacitor’, ‘circuit breaker with overload protection’, ‘circuit_breaker’, ‘connectionpoint’, ‘contactors’, ‘control suppply isolator’, ‘dc power supply’, ‘diacsandtriacs’, ‘diode’, ‘disconnector’, ‘door indicator’, ‘doublepole switch’, ‘doublepoleswitch’, ‘drawout circuit breaker’, ‘emergency switch’, ‘flow sensor’, ‘footswitch’, ‘fuse’, ‘heater’, ‘home and building’, ‘hydralic motors’, ‘hydraulic misc’, ‘hydraulic valve’, ‘hydrauliccylinder’, ‘hydraulicpump’, ‘inductor’, ‘isolation check’, ‘layout’, ‘levelsensor’, ‘light sensor’, ‘limit switch’, ‘lubrication shed’, ‘mcc drive cell terminal’, ‘meter’, ‘motion sensor’, ‘motor’, ‘octo coupler’, ‘overload trip’, ‘pandid’, ‘panel’, ‘photo electric sensor’, ‘pilot light’, ‘pneumatic misc’, ‘pneumatic motor’, ‘pneumatic valve’, ‘pneumaticcylinder’, ‘pneumaticpump’, ‘power supply’, ‘prefabricated circuit’, ‘pressure sensor’, ‘proximity sensor’, ‘pull buttons’, ‘push button’, ‘relay’, ‘resistor’, ‘selector switch’, ‘single pole switch’, ‘socket’, ‘solenoid’, ‘surge protector’, ‘switch’, ‘temperature sensor and controller’, ‘terminal block’, ‘thermistor’, ‘timer’, ‘transformer’, ‘vendor cable’, ‘voltmeter’]中文类别[‘三路开关’, ‘AA电池’, ‘交流电源’, ‘带同步装置的空气断路器’, ‘带同步设施的空气断路器’, ‘不带同步装置的空气断路器’, ‘不带同步设施的空气断路器’, ‘安培表’, ‘安培表选择开关’, ‘自动手动选择开关’, ‘空气断路器’, ‘电流表’, ‘指示器’, ‘每回路断路器分支电路最小尺寸’, ‘每回路断路器分支电路最小尺寸-剩余’, ‘母线’, ‘电池’, ‘制动离合器’, ‘母线排’, ‘电缆终端’, ‘电流互感器’, ‘电容式电压互感器’, ‘断路器’, ‘连接点’, ‘连接器’, ‘计数器’, ‘耦合电容器’, ‘电流互感器’, ‘柴油发电机’, ‘配电变压器’, ‘隔离开关’, ‘隔离开关’, ‘配电盘仪表’, ‘紧急定位发射器’, ‘紧急定位发射接收器’, ‘紧急停止按钮’, ‘现有电源-鞭状-连接至柜体-移除’, ‘接地’, ‘现场布线’, ‘现场隔离端子’, ‘发电机’, ‘至配电盘主回路每断路器最小尺寸’, ‘至配电盘主回路每断路器最小尺寸-剩余’, ‘指示灯’, ‘发光二极管’, ‘低压布线’, ‘光敏电阻’, ‘主跳闸继电器’, ‘微型断路器’, ‘塑壳断路器’, ‘符合IEC标准的多功能继电器’, ‘多功能电表’, ‘符合IEC 61850标准的多功能继电器’, ‘电机端子’, ‘PLC端子’, ‘电压互感器’, ‘电压互感器’, ‘电位器’, ‘漏电断路器’, ‘限制性接地故障继电器’, ‘电阻器’, ‘夹层式母线槽’, ‘开关支路-字母表示连接区域’, ‘单相电能表’, ‘开关熔断器组合’, ‘同步发电机’, ‘测试端子排’, ‘三导体’, ‘TNC选择开关跳闸中性闭合’, ‘跳闸回路监视继电器’, ‘晶体管’, ‘NPN型晶体管’, ‘电压表’, ‘电压表选择开关’, ‘阻波器’, ‘辅助触点’, ‘分支’, ‘电缆桥架’, ‘凸轮开关’, ‘电容器’, ‘带过载保护的断路器’, ‘断路器’, ‘连接点’, ‘接触器’, ‘控制电源隔离器’, ‘直流电源’, ‘双向二极管和双向晶闸管’, ‘二极管’, ‘隔离开关’, ‘门指示器’, ‘双极开关’, ‘双极开关’, ‘抽屉式断路器’, ‘紧急开关’, ‘流量传感器’, ‘脚踏开关’, ‘保险丝’, ‘加热器’, ‘住宅与建筑’, ‘液压马达’, ‘液压杂项’, ‘液压阀’, ‘液压缸’, ‘液压泵’, ‘电感器’, ‘隔离检查’, ‘布局’, ‘液位传感器’, ‘光传感器’, ‘限位开关’, ‘润滑棚’, ‘MCC驱动单元端子’, ‘仪表’, ‘运动传感器’, ‘电动机’, ‘八通道耦合器’, ‘过载跳闸’, ‘PANDID设备’, ‘配电盘’, ‘光电传感器’, ‘指示灯’, ‘气动杂项’, ‘气动马达’, ‘气动阀’, ‘气动缸’, ‘气动泵’, ‘电源’, ‘预制电路’, ‘压力传感器’, ‘接近传感器’, ‘拉钮’, ‘按钮’, ‘继电器’, ‘电阻器’, ‘选择开关’, ‘单极开关’, ‘插座’, ‘电磁铁’, ‘浪涌保护器’, ‘开关’, ‘温度传感器与控制器’, ‘接线端子排’, ‘热敏电阻’, ‘定时器’, ‘变压器’, ‘供应商电缆’, ‘电压表’]训练集1838 张验证集338 张测试集336 张总计2512 张 data.yaml 配置信息该数据集提供了data.yaml文件内容如下train:../train/imagesval:../valid/imagestest:../test/imagesnc:147names:[3wayswitch,AA,ACpower supply,AIR CIRCUIT BREAKER WITH SYN,AIR CIRCUIT BREAKER WITH SYN. FACILITY,AIR CIRCUIT BREAKER WITHOUT SYN,AIR CIRCUIT BREAKER WITHOUT SYN. FACILITY,AMMERTER,AMMETER SELECTOR SWITCH,AUTO MANUAL SELECTOR SWITCH,Air circuit breaker,Ammeter,Announciator,BRANCH CIRCUIT MINIMUM SIZE PER CKT BKR,BRANCH CIRCUIT MINIMUM SIZE PER CKT BKR - REMAIN,BUS BAR,Battery,Brake Cluthes,Busbar,CABLE TERMINATION,CT CURRRENT TRANSFORMER,Capacitive Voltage transformer,Circuit Breaker,Connection point,Connectors,Counter,Coupling capacitor,Current transformer,DIESEL GENERATOR,DISTRIBUTION TRANSFORMER,Disconnector Isolator,Disconnector Switch,Distribution Panel Meter,ELT,ELTR,EMERGENCY STOP PUSH BUTTON,EXISTING POWER -WHIP- CONNECTION TO CUBICLES - REMOVE,Earth ground,FIELD WIRING,Field isolator terminal,Generator,HOME RUN TO PANEL MINIMUM SIZE PER CKT BKR,HOME RUN TO PANEL MINIMUM SIZE PER CKT BKR - REMAIN,INDICATION LAMP,LED -Light Emitting Diode,LOW VOLTAGE WIRING,Light Dependent Resistor,MASTER TRIP RELAY,MCB MINIATURE CIRCUIT BREAKER,MCCB MOULDED CASE CIRCUIT BREAKER,MULTIFICTIN RELAY WITH IEC,MULTIFUNCTION METER,MUTIFUNCTION RELAY WITH IEC 61850,Motor terminal,PLC terminal,PT POTENTIAL TRANSFORMER,Potential transformer,Potentiometer,RCD Circuit Breaker,RESTRICTED EARTH FAULT RELAY,Resistor,SANDWICH TYPE BUS DUCT,SWITCH LEG - LETTER DENOTES CONNECTION ZONE,Single Phase energy meter,Switch disconnector fuse,Synchronous generator,TEST TERMINAL BLOCK,THREE-CONDUCTORS,TNC SELECTTOR SWITCH TRIP NEUTRAL CLOSE,TRIP CKT SUPERVISION RELAY,Transistor,Transistor -NPN-,VOLTMETER,VOLTMETER SELECTOR SWITCH,Wave trap,auxillarycontacts,branch,cabletray,camswitches,capacitor,circuit breaker with overload protection,circuit_breaker,connectionpoint,contactors,control suppply isolator,dc power supply,diacsandtriacs,diode,disconnector,door indicator,doublepole switch,doublepoleswitch,drawout circuit breaker,emergency switch,flow sensor,footswitch,fuse,heater,home and building,hydralic motors,hydraulic misc,hydraulic valve,hydrauliccylinder,hydraulicpump,inductor,isolation check,layout,levelsensor,light sensor,limit switch,lubrication shed,mcc drive cell terminal,meter,motion sensor,motor,octo coupler,overload trip,pandid,panel,photo electric sensor,pilot light,pneumatic misc,pneumatic motor,pneumatic valve,pneumaticcylinder,pneumaticpump,power supply,prefabricated circuit,pressure sensor,proximity sensor,pull buttons,push button,relay,resistor,selector switch,single pole switch,socket,solenoid,surge protector,switch,temperature sensor and controller,terminal block,thermistor,timer,transformer,vendor cable,voltmeter]️ 标注可视化数据集下载数据集下载⬇️⬇️⬇️ 点击下载数据集详细介绍YOLO电气工程图纸符号目标检测数据集是一个专为电气工程领域设计的图像识别数据集其核心目标是支持对电气工程图纸中各类符号的自动识别与定位。从名称和类别数量来看该数据集适用于电力系统设计、自动化控制、建筑电气施工等场景能够为智能图纸解析、CAD软件辅助识别、设备布局分析等应用提供基础数据支持。在实际工程中准确识别图纸中的符号有助于提高设计效率、降低人工错误率并为后续的自动化处理和信息提取奠定基础。本数据集包含2512张高质量电气工程图纸图像覆盖147种不同的电气符号类别涵盖开关、继电器、变压器、电缆、接线端子、指示灯等多种常见电气元件。这些类别均来源于实际工程标准具有明确的行业定义和应用场景确保了数据集的专业性和实用性。由于电气符号种类繁多且细节复杂该数据集在标注时需兼顾精确性和一致性以满足深度学习模型训练的需求。在标注规范方面该数据集采用了严格的矩形边界框标注方式每个符号均被精确标记确保模型能够学习到清晰的视觉特征。同时标注过程中遵循统一的行业标准避免了因人为差异导致的标注误差提升了模型的泛化能力和可靠性。这种高质量的标注体系使得该数据集在训练目标检测模型时具备较高的参考价值尤其适合用于YOLO系列算法的优化与部署。该数据集的典型应用方向包括电气图纸的智能识别系统、自动化设计工具的辅助模块、以及工业设备的故障诊断与维护支持。对于企业而言可以基于该数据集开发定制化的电气图纸分析工具提升设计与施工效率。建议在实际应用中结合具体业务流程进行模型微调并考虑与其他工程数据源进行融合以增强系统的实用性和准确性。YOLO26 目标检测算法原理如果说 YOLOv8 统一了多任务框架那 YOLO26 做的就是极致简化推理流程——去掉 NMS、去掉 DFL追求端到端的干净架构。端到端无 NMS 推理传统 YOLO 模型在推理后需要执行非极大值抑制NMS来去除冗余检测框这增加了后处理延迟和部署复杂度。YOLO26 默认采用一对一one-to-one检测头直接输出每张图像最多 300 个检测结果(N, 300, 6)完全省去 NMS 步骤推理流水线大幅简化。在 CPU 上的 ONNX 推理速度相比 YOLO11n 提升高达 43%。无 DFL 回归YOLO26 移除了分布焦点损失Distribution Focal Loss, DFL结构检测头更加轻量。边界框回归不再受限于预设的离散区间既降低了模型复杂度又保持了同等甚至更优的回归精度。Progressive Loss 与 STAL训练过程中通过渐进损失Progressive Loss将监督信号逐步从辅助头部转移到推理阶段的主头部。小目标感知标签分配STAL策略特别提升小目标的正样本标签覆盖率改善小物体检测性能。MuSGD 混合优化器创新性地将大语言模型常用的 Muon 优化器与 SGD 结合实现更稳定高效的训练收敛。相比纯 SGDMuSGD 在大 batch 训练时收敛更快且不易震荡。模型尺度与 COCO 性能YOLO26 提供 N/S/M/L/X 五种尺度模型mAP50-95CPU ONNX(ms)T4 TensorRT(ms)参数量YOLO26n40.938.91.72.4MYOLO26s48.687.22.59.5MYOLO26m53.1220.04.720.4MYOLO26l55.0286.26.224.8MYOLO26x57.5525.811.855.7MNano 版本参数量仅 2.4MCPU 推理速度比 YOLO11n 快 43%适合边缘设备和小数据集X 版本 mAP 达 57.5适用于对准确率要求极高的场景。双头架构设计YOLO26 检测模型内置两个检测头——一对一默认无 NMS输出 300 个检测框和一对多需 NMS输出 8400 个候选框。训练时一对多头作为辅助监督信号加速收敛推理时默认使用一对一头实现端到端输出。YOLO26 支持全部七大视觉任务检测、实例分割、语义分割、深度估计、分类、姿态估计和定向边界框检测OBB一个框架覆盖从 2D 检测到 3D 感知的完整需求。关键训练参数配置详解别一上来就改 lr00.002 或 warmup_epochs5。YOLO26 的默认参数是经过大批量实验验证的只在明确知道要改什么的时候才动。参数默认值本项目设置含义model-yolo26n.pt预训练权重n/s/m/l/x 五种尺度epochs100150训练总轮数imgsz640416输入图像尺寸batch1664批次大小根据显存调整lr00.010.02初始学习率lrf0.010.01最终学习率因子lr0 × lrfmomentum0.9370.937SGD 动量weight_decay0.00050.0005权重衰减L2 正则化warmup_epochs3.03.0预热轮数warmup_momentum0.80.8预热期间动量初始值box7.57.5边界框回归损失权重cls0.50.5分类损失权重dfl1.51.5DFL 损失权重YOLO26 可置 0patience100-1早停轮数cos_lrFalseTrue余弦学习率衰减close_mosaic1010最后 N 轮关闭 Mosaic 增强optimizerautoSGD优化器选择ampTrueTrue自动混合精度训练freezeNone0冻结前 N 层骨干网络参数dropout0.00.1分类头 Dropout 率防过拟合三个损失权重的分工box负责《框在哪》cls负责《框里是什么》dfl在 YOLO26 中可设 0 因为默认检测头已去掉 DFL。大部分场景保持默认比例即可除非有明确的偏好需求。学习率调参经验训练初期 loss 不下降 → 调大lr0或增加warmup_epochs训练后期 loss 震荡 → 降低lr0增大weight_decay验证集 loss 先降后升过拟合→ 增大dropout、weight_decay或减小epochs小数据集100张→lr00.001,weight_decay5e-4,dropout0.2大数据集1000张→lr00.01,weight_decay5e-4,batch32训练可视化与TensorBoardYOLO26 默认开启 TensorBoard 日志记录训练过程中可实时查看 loss 曲线和指标变化。在命令行执行tensorboard --logdir runs/detect/train即可打开 Web 看板每轮训练完成后自动刷新。训练结束后 results.png 生成的六合一曲线图是最直观的诊断工具——一眼就能看出是否过拟合、是否收敛、哪个 loss 在震荡。断点续训机制如果训练意外中断断电、OOM 等不用从头开始。YOLO 每轮保存last.pt恢复训练时指定modellast.pt并设置resumeTrue学习率和优化器状态都会从断点恢复训练曲线也和连续训练一样平滑。模型评估指标解读如果训出来的 mAP50 很高但 mAP50-95 很低说明模型对定位宽容度不够——框的位置稍微偏一点就判定为错误通常跟标注质量有关。训练完成后YOLO26 训练器自动输出以下指标mAP50mAP0.5IoU 阈值为 0.5 时的平均精度均值反映模型粗粒度定位能力。mAP50 ≥ 0.90 表示定位效果很好。mAP50-95mAP0.5:0.95IoU 从 0.5 到 0.95 共 10 个阈值上取平均。更严苛反映精确定位能力是更具代表性的综合评价指标。Precision精确率检测结果中正确检测的比例。高 Precision 意味着虚警少。Recall召回率真实目标中被成功检测到的比例。高 Recall 意味着漏检少。F1 ScorePrecision 和 Recall 的调和平均数 2×P×R/(PR)综合衡量模型平衡性。Confusion Matrix混淆矩阵N×N 矩阵对角线为正确分类非对角线为误分类可定位类别间混淆情况。训练曲线分析train/box_loss单调下降 → 边界框回归正常收敛val/box_loss先降后升 → 过拟合风险应增大数据增强或减小模型metrics/mAP50(B)增长放缓趋于平稳 → 模型接近收敛上限train 持续降、val 持续升 → 严重过拟合建议减小模型或增大 dropout指标使用场景安全生产场景如安全帽检测→ 优先保证 Recall宁可多报不能漏报质检场景如缺陷检测→ 优先保证 Precision降低误检率减少人工复核量通用检测 → 均衡优化关注 mAP50-95 和 F1 ScorePR 曲线实战解读每张PR_curve.png里每个类别有一条不同颜色的曲线。不用盯着每个类看——先找出最差的那条最靠左下角的它代表你模型的弱项。加大那个类别的数据量或检查标注质量通常比调任何参数都见效。YOLO 系列算法演进史YOLO 的每次大版本更新都解决了一个具体痛点v3 的多尺度预测解决了小目标问题、v5 的自动锚框省去了手动计算、v26 的端到端推理彻底干掉了 NMS。YOLOv1 (2015)将检测问题建模为单一回归问题直接在输出层预测边界框和类别概率。速度极快但定位精度较差。YOLOv2/YOLOv3 (2016-2018)引入 anchor box 机制、特征金字塔FPN和多尺度训练。YOLOv3 的 Darknet-53 骨干网络成为经典架构至今仍被广泛使用。YOLOv5 (2020)Ultralytics 在 PyTorch 上的实现带来了完善的训练框架、自动锚框聚类和数据增强 pipeline大幅降低了使用门槛。YOLOv8 (2023)统一了检测、分割、分类和关键点任务的框架引入 C2f 模块和无锚框anchor-free检测头。YOLOv9 (2024)提出 GELAN通用高效层聚合网络和 PGI可编程梯度信息同等参数量下精度提升显著。YOLOv10/YOLOv11 (2024)YOLOv10 首次提出无 NMS 端到端检测。YOLOv11 进一步优化骨干网络和训练策略在速度和精度之间取得新平衡。YOLO26 (2026)集历代之大成核心创新包括① 端到端无 NMS 一对一检测头 ② 无 DFL 的简化回归头 ③ Progressive Loss 渐进式监督转移 ④ STAL 小目标感知标签分配 ⑤ MuSGD 混合优化器。关键设计演进总结Anchor-based → Anchor-freev8 起C3 → C2f → C3k2骨干模块持续优化多任务统一框架v8 起NMS 后处理 → 无 NMS 端到端v10/v26单一损失函数 → 多任务联合损失Box Cls DFLv26 可去 DFLYOLO26 训练步骤详解YOLO26 的训练步骤已经被封装得很简洁了——装好 ultralytics 包准备 data.yaml一条命令就能启动。环境准备pipinstallultralytics yolo checks python-cimport torch; print(torch.cuda.is_available())命令行训练yolo detect traindatadata.yamlmodelyolo26n.ptepochs150imgsz416batch64device0lr00.02patience-1cos_lrTrueclose_mosaic10Python API 训练fromultralyticsimportYOLO modelYOLO(yolo26n.pt)resultsmodel.train(datadata.yaml,epochs150,imgsz416,batch64,device0,lr00.02,patience-1,cos_lrTrue,close_mosaic10,augmentTrue,ampTrue,workers8,)训练输出文件runs/detect/train/weights/best.pt验证集上 mAP 最高的权重runs/detect/train/weights/last.pt最后一轮保存的权重runs/detect/train/results.csv每轮 loss 和指标数据runs/detect/train/confusion_matrix.png混淆矩阵图runs/detect/train/results.png训练曲线图runs/detect/train/val_batch*_pred.jpg验证集预测效果图单张推理测试modelYOLO(runs/detect/train/weights/best.pt)resultsmodel(test_image.jpg)results[0].show()多 GPU 分布式训练yolo detect traindatadata.yamlmodelyolo26n.ptdevice0,1epochs100恢复中断的训练yolo detect trainmodelpath/to/last.ptdatadata.yamlresumeTrue部署与应用建议大规模数据集2512 张训练出的模型泛化能力通常已经不错。部署的重点在于效率和稳定性而不是继续压榨精度。fromultralyticsimportYOLO modelYOLO(runs/detect/train/weights/best.pt)# 推荐 TensorRT FP16精度几乎无损、速度翻倍model.export(formatengine,device0,halfTrue)# 备选 ONNX兼容性好model.export(formatonnx,opset17)运维建议生产环境部署后建议建立模型版本管理和 A/B 测试机制。新模型上线前在 10% 流量下跑 24 小时对比核心指标漏检率、误检率、P99 延迟和老模型无显著劣化再全量切换。