塔式起重机不安全吊装行为的作用机理及智能识别研究

蒋伟光

摘要

塔式起重机（简称“塔机”）是工程建造活动中使用最为频繁的起重运输机械之一。然而，国内外塔机倒塌等安全事故频发的态势始终没有得到有效地遏制，造成了严重的人员伤亡及经济损失。大量的事故调查报告表明不安全吊装行为是塔机安全事故的导火索，如：2019 年4 月贵阳塔机工人斜吊使得塔身折断，同年甘肃庆阳塔机司机突然起制动造成上臂倾覆。依据海因里希法则，监控人员的不安全行为是减少安全事故的关键。因此，掌握不安全吊装行为对于塔机稳定性的作用机理，进而建立不安全吊装行为的智能识别模型成为塔机安全管理中的一个重要课题。

首先通过塔机可追溯事故案例分析，总结了施工活动中最典型的三种不安全吊装行为（倾斜起吊、突然制动、突然卸载）及事故机型（独立-尖头-上回转变幅式），并以该机型为例建立了塔机缩尺模型实验平台，以克服现场实验的危险性。在此基础上，从不安全吊装行为对于塔机失稳的作用机理、监控部位分析、识别模型建立及场景泛化方法上系统地开展研究。

在作用机理上，通过实验研究描述了倾斜起吊、突然制动及突然卸载的载荷激励，构建了塔机失稳的损伤指标，通过增量动力分析描述了在不安全吊装行为下各组件结构单元从弹性变形到塑性变形的过程。结果显示：倾斜起吊是对塔机稳定性影响最大的不安全吊装行为；在同一吊装载荷下，塔机底部标准节最能表征基础结构失效状态，两者皮尔逊相关性超过0.9，同时对不安全吊装行为作用下结构响应数据的分类最为敏感，可以作为监控设备布设的重点部位。

在识别方法上，基于塔机多姿态吊装实验建立了有标签的不安全吊装行为图片数据库，包含16200 条图片数据，覆盖塔机-90°至90°的回转姿态。提出了基于RGB通道的多传感器数据融合方法及不安全吊装行为智能识别模型。实验过程中，比较了时间序列归一化、格拉姆角场变换、连续小波变换三种特征变换方式及AlexNet、VggNet、ResNet、GoogleNet 及MobileNet 五种深度学习框架组合下的识别效益，得到了Rawdata-ResNet 最优的不安全吊装行为识别路径，实验室环境下精度达99.90%。

在场景应用上，提出了基于深度对抗迁移的不安全吊装行为智能识别模型的泛化思路，实现了工地现场的应用，避免了工地有标签不安全吊装行为监测数据少的问题。经实验分析，所提出的迁移模型能够较好地克服目标域数据集标签类型信息缺损时造成的“负迁移”问题，在工地现场数据验证集下的最优精度达到76.74 %。最后，基于Qt Designer 实现了不安全吊装行为智能识别的功能。

关键词：塔式起重机；不安全吊装行为；稳定性分析；智能识别；卷积神经网络；深度迁移学习

Abstract

Tower cranes are one of the most used machinery in construction activities. However,tower crane safety accidents still occur frequently at home and abroad, causing serious casualties and economic losses. A large number of accident investigation reports have shown: unsafe hoisting behaviors are the fuse of tower crane safety accidents. For example, in Guiyang and Gansu. China, tower cranes broke off due to unsafe hoisting behaviors. How to get the influence effect of unsafe hoisting behaviors on the stability of tower cranes, establish its intelligent recognition model, and realize the operation process control of hoisting operators. It has become an important topic for tower crane safety management.

This thesis summarizes the most typical unsafe hoisting behaviors (tilt hoisting, sudden braking, sudden unloading) and the accident type (independent & pointed tower crane) through accident cases. Taking the above type as an example, a tower crane experimental platform based on a scale model is established, to avoid the danger of on-site experiment. Then, this thesis systematically studies unsafe hoisting behaviors, including instability effects, monitoring parts seclection, intelligent models, and scene generalization application.

In the dynamic mechanism of unsafe hoisting behaviors. The load excitation effect of the above unsafe hoisting behaviors is summarized. Then the collapse level of the tower crane is constructed, and the elastoplastic changes of each component under different hoisting behaviors are discussed based on the incremental dynamic analysis method. The results show that tilt hoisting is most likely to cause tower crane instability. The Pearson correlation of the damage evolution between the bottom mast and the foundation is more than 0.9. And this mast is the most sensitive to the classification of different hoisting behaviors and can be used as the key part of the sensor layout.

In the intelligent recognition method of unsafe hoisting behaviors. This thesis establishes a labeled image feature data set of unsafe hoisting behaviors for the first time through experimental research, which contains a total of 16200 image data and covers the rotation attitude of - 90 ° to 90 °. An image data fusion method based on RGB channels is proposed, and an intelligent recognition model of the unsafe hoisting behavior is established. This thesis compares the recognition benefits of different feature extraction methods (Rawdata, GADF, CWT) and different deep learning frameworks (AlexNet, VggNet, ResNet, GoogleNet, MobileNet) in detail. The Rawdata-ResNet is the optimal recognition path of the unsafe hoisting behavior, and its recognition accuracy reaches 99.90%. 

In the construction site application of the recognition model. This thesis proposes a generalized idea of the intelligent recognition model based on deep transfer learning. It has realized the application transition from the scaled model to the full-scale model of the tower crane, avoiding the problem of less monitoring data of unsafe hoisting behaviors with labels on construction sites. According to the experimental data on-site, the proposed transfer model can overcome the "negative migration" problem caused by the lack of label type information in the target domain data set, and the optimal accuracy reaches 76.74% with the validation set. Finally, an offline functional module for intelligent recognition of the unsafe hoisting behavior is designed through Qt Designer.

Keywords: Tower crane; Unsafe hoisting behaviors; Stability analysis; Intelligent Recognition; Convolutional neural networks; Deep transfer learning