Investigation On Decision-Making Of Shield Attitude Adjustment Based On Machine Learning

The measures of tending shield actual attitude to planned attitude are called shield attitude adjustment. The accuracy of shield attitude adjustment is important for shield excavation and tunnel construction precision. But because of the complicacy of factors affecting shield attitude adjustment and uncertainty of geological status, phenomena of attitude departure and snaky track appear frequently. The shield driver can avoid these disadvantages by evaluation of current excavation data and decision-making out of his experience. In an automatic system, the shield driver’s and other experience should be learned. The importation of machine learning methods is a alernative way for an automatic shield attitude adjustment system to learning experience and making decision.There are several problems need to be solved, when machine learning methods being imported to shield attitude adjustment decision-making. These problems are construction of sample set, attribute imbalance of different sample class, lack of negative samples in collecting sample stage, comprehensive decision involving attitude adjustment and excavation.For the construction of shield attitude adjustment sample set, shield attitude mechanical model involving prior knowledge should be established to obtain alternative sample features. And feature selection should be carried on considering importance and accessibility of alternative features. Class assignment of a sample comes from shield driver experience.There are three attributs being imbalant of different sample class:sample size, sample distribution and distance from samples center in festure space to the decision hyperplane. And this kind of imbalance makes the accuracy of support vector machine algorithm greatly decreased. In view of this situation, a sample class attribute support vector machine class penalty weight adjustment method based on the quantification of coefficient is proposed to keep high classification accuracy by tuning penalty coefficient accroding quantized attribute imbalance.In the process of accumulating samples, there will be quite a long period of time being lack of negative samples, even only positive samples. In order to make shield attitude adjustment decision in this period, the support vector data description method is introduced. For improving the training speed, a new method named core set approximation gauss support vector data desciption is proposed. A method is proposed for fast discrimination of the samples to be classified. Attitude adjustment decision function is given. Validation based on simulation and engineering data is carried out at the end of this chapter.In order to improve the ability of the shield to avoid risks, the scope of the learning is expanded to learning action strategy in different conditions. The strategy learning considers not only shield attitude adjustment, but also shield excavation. A method based on hierarchical reinforcement learning is proposed, and it will improve the initiative and intelligence level of shield attitude adjustment.