【作者】 罗琦；

【导师】 刘亚俊； 毛卫东；

【作者基本信息】 华南理工大学 ， 工程硕士（机械工程）（专业学位）， 2022， 硕士

【摘要】 锤击法成桩的收锤标准是沉桩质量评估的重要指标,随外部因素的变化而变化。工程地质条件和桩土相互作用关系的复杂性导致了收锤标准的不确定性,仅仅以贯入度控制为锤击法成桩的收锤标准难以满足工程实际需求,如何结合液压锤的性能,在贯入度的基础上丰富收锤标准指标组合,多维度体现锤击成桩的质量状态,是液压锤锤击成桩质量监测技术中亟待解决的问题。因此,本文从沉桩施工收锤标准的影响因素出发,结合机器学习方法对锤击成桩过程参数建立基于沉桩状态预测的机器学习模型,实现对不同沉桩状态和桩端土层信息的智能识别,为桩工建筑行业的发展提供新的研究方向。本文首先从液压打桩锤锤击法成桩过程机理出发,通过多维组合指标表征收锤标准并对沉桩状态进行划分标定。通过核密度估计的统计分析方法对所采集的训练、测试数据进行独立同分布验证。通过皮尔逊相关系数和预测能量分数的融合分析方法对多维组合特征进行约简和分析得出影响沉桩质量的关键因素。算法实验结果表明,通过地勘信息的先验知识来改进BP神经网络的输出层结构能够较好识别沉桩过程桩端的土层信息,测试集识别率为94%;基于贝叶斯优化和Ada Boost的融合算法相较于传统机器学习方法对沉桩状态的识别具有更高的准确率。其次,本文通过LabVIEW软件搭建由计算机、声卡、声音传感器组成的声音信号采集系统对沉桩过程的声音信号进行实时采集。通过改进的高斯降噪算法对液压锤锤击桩头所产生振动的声音信号进行重构分解,经过信号预处理和线性预测倒谱系数算法进行了特征提取。通过算法实验的对比,验证了基于声学信号处理的沉桩质量监测方法的可行性,其中,基于高斯核函数的支持向量机算法对不同阶段沉桩声音信号具有较好的识别效果,其测试集准确率达到91%以上。针对沉桩各状态采样不均衡导致的数据偏态问题,本文对沉桩状态的预测模型进行了相应的优化和改进,提出了数据采样合成与传统机器学习的融合训练方法、深度学习与数据采样合成以及改进多分类Focal Loss损失函数的融合方法、基于类别权重分配的训练方法。针对收锤阶段和刚入设计持力层阶段样本的不均衡情况,采用召回率、精确度、F-measure、G-mean作为相应地算法评估指标。经算法实验对比,验证了上述方法对样本不均衡问题的改进作用。更多还原

【Abstract】 The ceasing driving standard of piling construction by hammering method is an important index of construction quality evaluation,which varies with the change of external factors.The complexity of geological conditions and pile-soil interaction leads to the uncertainty of the ceasing driving standard.It is difficult to meet the actual requirements of the project only by penetration as the standard.So as to fully stimulate the ultimate bearing capacity of PHC piles and reflect the quality state of PHC pile multi-dimensionally,it is an urgent problem to be solved in the construction quality evaluation how to combine the performance of hydraulic hammer and set the corresponding ceasing driving standard to enrich its indexes based on penetration.Therefore,based on the influencing factors of ceasing driving standard,machine learning method is used to realize the intelligent identification of different piling states and different soil categories combined with piling parameters which collected during the piling process.The research conducted in the paper is aimed to provide a new research direction for the development of pile construction industry.Firstly,based on the study of piling mechanism by hammering method,the piling states are divided and calibrated by multi-dimensional combination indexes to characterize the ceasing driving standard in this paper.The statistical analysis method based on Kernel Density Estimation(KDE)is used to split samples into training set and test set,which aims to ensure the independence and identical distribution of different data sets.Then,The key factors affecting the construction quality are obtained by the combination of Pearson correlation coefficient and Predictive Power Score(PPS).The algorithm experiments show that rectifying the output layer structure of BP neural network combined with geological distributed information can better identify the soil layer information during the piling process,and the recognition accuracy of the test set is 94%.And the combination of Bayesian optimization and Ada Boost has higher accuracy than the other machine learning methods.Secondly,LabVIEW software is used in this paper to build a sound signal acquisition system composed of a computer,a sound card and a sound sensor to collect the sound signal during piling process in real time.The improved Gaussian filtering noise reduction algorithm is proposed and used to reconstruct and decompose sound signal,and the features are extracted and calculated by signal preprocessing and Linear Predictive Cepstral Coefficient(LPCC)algorithm.Through algorithm experiments,the feasibility of the pile driving quality monitoring method based on acoustic signal processing is verified in the paper,in which,the Support Vector Machine(SVM)algorithm based on Gaussian kernel has a good recognition effect for sound signals collected from piling process in different stages,and the accuracy of the test set is more than 91%.Aiming at the problem of data skewness caused by uneven sampling in each state of piling process,the combination of synthetic minority oversampling techniques and traditional machine learning methods,the combination of deep learning methods,oversampling techniques and improved Focal Loss function,and the training method based on category weight distribution.In view of the imbalance of the samples collected after the pile end enters the design bearing stratum and the final stage,the indexes such as Recall,Precision,F-measure,G-mean are used corresponding algorithm evaluation indexes based on the confusion matrix.Through algorithm experiments,the improvement effect of the above methods on the sample imbalance problem is verified.更多还原