【摘要】 起重机运行状态实时检测是工业安全生产的重要保障之一。针对起重机部件提出基于多任务学习Mask R-CNN的分割与关键点网络结构,该结构由Mask R-CNN定位框与区域语义分割网络、DeepLabCut关键点检测网络构成;在吊钩桥式起重机中采集数据进行测试,利用贪婪线性搜索算法与贝叶斯优化算法,搜索得到此模型的最优超参数组合为:学习率0.005,批数2,学习率策略为余弦衰减。该模型测试误差为2.46个像素点,测试AP可达95%,像素点误差反映到实际误差在5 cm以内,满足实际检测需求,可拓展应用于无人化、自动化起重机运行状态监测。更多还原

【Abstract】 Real time detection of crane operation status is one of the important guarantees for industrial safety production. A segmentation and key point network structure based on multi task learning mask R-CNN is proposed R-CNN crane key parts positioning frame and regional semantic segmentation network, crane key point detection network based on DeepLabCut; in a hook crane to collect data for testing, using greedy linear search algorithm and Bayesian optimization algorithm, the optimal super parameter combination of this model is: learning rate 0.005, batch size 2, learning rate strategy is cosine decay. The test error of the model is 2.46 pixels, and the test AP is up to 95%. The pixel error can be reflected within 5 cm of the actual error, which can meet the actual detection needs, and can be extended to the unmanned and automatic crane operation state monitoring.更多还原