Smith, Samuel M.

Autonomous Underwater Vehicles (AUV) collect a large volume of scientific data in every mission, using the onboard sensors, and store them in log files. The accessibility of these data is limited. Specific tools are required to extract the data to be processed on the user workstation with the installed analysis tools and scripts. The objective is to standardize and simplify the way data can be retrieved and processed from anywhere by anybody. The design of a server that manages the access to the data and to the applications that process them has been considered. Everything can then be done through the use of a single Java client executed on any Java compliant computer. Analysis tools are downloaded on the fly when needed and do not require any installation. New tools can be integrated into the application server in the form of plugins developed with an appropriate Java Library.

Recent advances in computer engineering make the computational approaches to controller design for high order systems practical. In this dissertation, a series of computational methods based on cell state space for the design and optimization of Takagi-Sugeno (TS) type Fuzzy Logic Controllers (FLCs) are presented. The approaches proposed in this research can be classified into two categories: feed forward design and feedback design. An Optimal Control Table (OCT) based on cell state space is used in all the feed forward design approaches. An FLC can be trained by Least Mean Square (LMS) algorithm with an OCT serving as the training set. For high order systems, due to physical memory limit, the cell resolution is generally low. A specially modified k-d tree representation of cell space is proposed to save the memory while keeping the cell resolution as high as possible. The control command for a point that is not a cell center is approximated by interpolating an OCT. All these commands can be used as training data to train an FLC. An iterative feedback design approach named Incremental Best Estimate Directed Search (IBEDS) is proposed to further optimize a training set. It is a kind of globally directed random search method. The general philosophy is that since the best possible performance of an FLC largely depends on the quality of the training set, if the training set is optimized, an FLC trained by the set would also be optimized. Based on IBEDS, two other feedback FLC design algorithms are also proposed. In one algorithm, subtractive clustering method is used to extract the structure of an FLC from an OCT. The coefficients of the FLC obtained are then optimized with IBEDS. The other algorithm applies IBEDS to three system models and finds the training set that has the worst performance for all the models. This training set is further optimized to improve robustness of a trained FLC.

The problem at hand is developing a controller design methodology that is generally applicable to autonomous systems with fairly accurate models. The controller design process has two parts: synthesis and analysis. Over the years, many synthesis and analysis methods have been proposed. An optimal method for all applications has not yet been found. Recent advances in computer technology have made computational methods more attractive and practical. The proposed method is an iterative computational method that automatically generates non-linear controllers with specified global performance. This dissertation describes this method which consists of using an analysis tool, continued propagation cell mapping (CPCM), as feedback to the synthesis tool, best estimate directed search (BEDS). Optimality in the design can be achieved with respect to time, energy, and/or robustness depending on the performance measure used. BEDS is based on a novel search concept: globally directing a random search. BEDS has the best of two approaches: gradient (or directed) search and random search. It possesses the convergence speed of a gradient search and the convergence robustness of a random search. The coefficients of the best controller at the time direct the search process until either a better controller is found or the search is terminated. CPCM is a modification of simple cell mapping (SCM). CPCM maintains the simplicity of SCM but provides accuracy near that of a point map (PM). CPCM evaluates the controller's complete and global performance efficiently and with easily tunable accuracy. This CPCM evaluation guarantees monotonic progress in the synthesis process. The method is successfully applied to the design of a TSK-type fuzzy logic (FL) controller and a Sliding Mode-type controller for the uncertain non-linear system of an inverted pendulum on a cart for large pole angles (+/-86 degrees). The resulting controller's performance compares favorably to other established methods designed with dynamic programing (DP) and genetic algorithms (GA). When CPCM is used as feedback to BEDS, the resulting design method quickly and automatically generates non-linear controllers with good global performance and without much a priori information about the desired control actions.

It is desirable to have robust high performance nonlinear control with a model-free design approach for the real time automatic control of practical industrial processes. The field has seen the application of Sliding Mode Controllers (SMCs). SMCs are nonlinear robust controllers, however most design approaches related to SMCs are model-based approaches. PID controllers and some Fuzzy Logic Controllers (FLCs) are model-free controllers, however their robustness is not integrated into their design parameters directly. This dissertation presents two new types of robust high performance nonlinear controllers with model-free design approaches. One introduces fuzzy logic to a model-free SMC which is a simple saturation function incorporating three design parameters. Due to the interpolative nature of fuzzy control, a TSK type FLC with the model-free SMCs as its rule's consequents will produce a controller with a nonlinear sliding curve and a nonlinear boundary layer. We call this controller a Fuzzy Sliding Controller (FSC). The other uses a new type of Variable Structure Controller (VSC), which intentionally switches from one controller to another controller during a step response. In conventional approaches to VSC, the control surface does not change its shape during a step response. The new type of VSC intentionally changes the shape of the control surface during the step response. This technique is analogous to that technique employed in image processing called "morphing" where a given image gradually changes over time to the image of a different entity. In order to avoid confusion with the conventional approach to a VSC, we use the term "Morphological" Controller (MC) for the VSC of the new type. The performance and robustness with respect to parameter variations, disturbances and slow sample rates of the proposed controllers are studied in detail with a DC motor and an Inverted Pendulum System. As a means to verify the proposed controllers in practical cases, we design the model-free SMC, the FSC and the MC for the highly nonlinear and uncertain dynamics of an Autonomous Underwater Vehicle, Ocean Voyager II. It is shown that the proposed controllers are high performance and high robustness controllers.