This thesis describes work aimed at developing practical methods for determining the best docking locations for an underwater remotely operated vehicle (ROV) when inspecting an offshore platform. ROVs are used extensively in the offshore oil and gas industry to conduct a large variety of intervention tasks such as visual inspection, operational monitoring, equipment installation and operation, debris recovery, and so on. However, they have found only limited use in the more difficult tasks such as the detailed inspection of complex weld geometries. These complex welds are, however, found extensively in the construction of the majority of offshore structures and platforms ('oil rigs'). Furthermore, there is a safety requirement to have them inspected regularly since failure of these welds can potentially lead to catastrophic failure of the structures, the majority of which are manned. A number of specialist ROV systems have been developed that are able to attach onto platform structures and use their manipulators to conduct inspection. However, due to the short reach of the manipulators and the complex geometry of the welds (often encumbered with protruding pipes and other fittings) the success of any inspection is crucially dependent on a good initial choice of ROV docking position. This thesis will describe the problems and current manual planning methods, and then detail the development of two new methods for automated optimisation of docking positions - firstly using neural networks, and secondly using more conventional numerical processing. This thesis will also review related work in the field, such as the development of neural networks and their applications in the general offshore environment and in the control of ROVs and robot manipulator arms, and other approaches to ROV docking. It will further describe the use of the system developed here for planning docking positions on example commercial ROV inspection work programmes.

No description available.

Novelty detection is a very useful tool to differentiate between known {normal} and unknown (novelty) sensory perceptions during robot exploration and inspection of environments. One way of doing this is to acquire a model of normality from frequently detected perceptions and then use this model of normality to detect perceptions which do not conform with the model. Since the novelty is unknown and occurs much less frequent than normal data; it is practically impossible to build models of abnormality. This thesis introduces dynamic novelty detection approaches which are used to model the dynamics of sensory perceptions in normal environments. In order to do this, spatia-temporal sensory data are modelled to predict future sensory perceptions, in an approach called the expectation model of normality. By calculating the prediction error using actual sensory observations, a novelty can be detected when there is a deviation from an estimated confidence level. In the first part of the thesis, an off-line radial basis function neural network based novelty filter is proposed. The filter is investigated in static and dynamic environments. In the static environment, the objects are assumed to be at fixed locations whereas in the dynamic environment the (normal) objects are moving continuously during the process of learning the expectation models of normality. Results show that the proposed filter is able to detect static and dynamic changes in the environment by predicting the normal future sensory data with high confidence. When something is changed in the environment the changes are predicted with lower confidence, and a novelty may be detected. In the second part of this thesis, an on-line dynamic novelty detection system is developed. The system is inspired by the Evolving-Connectionist-System network which grows dynamically when a new input perception or a predicted future perception is found novel. The validity of the system is verified in a number of robotic experiments. Experiments were carried out using different types of sensory systems to generate feature vectors for the novelty filter; 1) a fusion of visual features from a colour image, laser rangefinder readings and velocity information of the robot; and 2) a fusion of colour and 3D shape features using a Kinect sensor. The results are assessed using statistical tests and the proposed novelty filter is compared with well known Grow-When-Required neural network . Experimental results show advantages of the proposed novelty filter over the Grow-When-Required network. In the final part of this research, an on-line multi-channel novelty detection system is developed by combining two types of novelty filters. Both novelty filters are fed using different type feature extraction systems; these are a sensory-fusion system and a visual attention system. The results show that by integrating novelty outputs from both filters, a more robust novelty decision may be acquired.

No description available.

No description available.

No description available.

No description available.

No description available.

No description available.

No description available.