Pipeline Detection, Control And Navigation Of Pipeline Robot

The pipelines, as an important approach to deliver water, coal gas, and oil, are widelyapplied to many fields, such as oil, chemical engineering, construction, nature gas, andnuclear industry. These bring many conveniences and enormous economic benefits forhuman life, agriculture, and industry. Meanwhile, pipeline robots are tremendouslydemanded for pipeline inspect and service. Pipeline robots integrate intelligent mobile carrier,various sensors, tools, and nondestructive testing technology. They detect, clean up, andservice pipelines which are complex environments and difficult for human. With thedevelopment of machine learning, pattern recognition, and statistical theory, research onintelligence behavior of robot has been upgraded to a more important position. Robots withautonomous learning ability have gradually become a new trend in robot development.Aiming to improve the intelligence of pipeline robots, the thesis has investigated two corequestions of pipeline robots: first, automatic understanding of pipeline images. To cope withthe real time and class imbalance which is frequently overlooked in pipeline detection, theproposed methods improve the real time and recognition rate of pipeline inspection viamutli-classifiers ensemble and fenture selection；second, intelligent control and autonomousnavigation of mobile robot in unstructured pipeline environment. Fuzzy control, imitationlearning, and multi-module optimization method are introduced into the localization, motioncontrol, and autonomous navigation of pint-sized mobile robot, which improve the intelligentrobot’s capacity of motion and autonomous navigation in unstructured complicatedenvironment.The main contributions of the thesis are summarized as:1、The real-time problem and recognition rate of automatic pipeline inspection based onvision is investigated. Rich visual information may result in a difficult problem in real-time detection. The cascade of boosting multi-classifiers is introduced into pipeline detection. Formulticlass problems, the classes corresponding to the maximal Bhattacharyya distance,which is a separability criterion of classes, are first classified in a cascade structure. Themost of non-object images which are easily recognized by simple classifiers are discarded inthe first stages with short processing time, while more complex defect classes are graduallyclassified by several cascade stages. The complexity analyses of the cascade classifiers aredone based on Gaussian maximum likelihood. An automatic hierarchical image segmentationmethod is proposed and the needed geometrical features are extracted from pipeline images.The whole course can be considered as a process of coarse-to-fine segmentation, in whichfalse alarms of the segmented image are progressively eliminated.2、To deal with unbalanced data distribution in anomaly inspection, this thesis proposesan ensemble classification method and feature selection of unbalanced data. The costsensitive function and asymmetric misclassifcation costs of classes are specially designed torelieve class imbalance in feature selection based on Tabu search. The final decisions of theensemble classification are integrated based on the weighted sum rule, the nearest-neighborrule, and the nearest consensus rule, respectively. The weighting coefficient is automaticallyobtained via Tabu search. And either oversampling the minority class or undersampling themajority class in data level are employed to relieve unbalanced data problems. Therecognition rate of pipeline anomaly in unbalanced data is improved. To cope with thediversity of feature distribution, firstly, the semi-supervised K-means clustering method isused to pre-process the features, such that unbalanced data problems are relieved; then, C4.5decision tree is used to improve the class dominance and forced assignment. The approachhas achieved good performance in difficult classification with noise and local optimum, suchas automatic anomaly detection and so on.3、The tracked mobile robot is a nonholonomic dynamical system with intrinsicnonlinearity and its turning motion is accomplished by overcoming friction-force between the tracks and the ground. The thesis proposes a novel path following control for trackedrobot based on hierarchical fuzzy structure. The whole structure composes of three–level lowdimensional sub-controllers: fuzzy velocity control, fuzzy steering and fuzzy supervisorycontrol. Consequently, the numbers of control rules are largely reduced due to thehierarchical structure and the complexity of designing the controller is decreased to a greatextent. These are important to a pint-sized robot. The structure applied to pint-sized trackedrobot has the apparent advantages of easy implementation. All input and output havephysical meanings such that the design of fuzzy control becomes easy and linguistic rulescan easily be obtained by prior knowledge.4、Learning-based automatous navigation of modularized mobile robot is addressed. Wediscuss to apply imitation learning to decision and path planning of modularized mobilerobot. By learning expert demonstration, we extend the Maximum Margin Planning (MMP)to develop a function mapping multi-module perceptual data to costs. These algorithmschoose the joint cost function (or reward function) between multi-modules, such that therobot’s planned behavior mimics an expert’s demonstration as closely as possible and findparameters for each module that optimize the performance of the overall system. The methodcan improve the robustness of automatous navigation and reduce the intervence of humanbeing. And by learning the dynamic perceptual data, local autonomous obstacle-avoidingbehavior of pipeline robot are implemented.