| We still have problems generating tool paths for machining sculptured surfaces, especially in finding a gouge-free and collision-free tool position and orientation. The use of five axis machining, which has been proven to be an efficient technique to produce good surface quality for sculptured surfaces, also makes more difficult the problems of defining tool position and orientation. The fact that visual checking by a human operator is still a good way to detect an unacceptable tool position and orientation motivated this research.; The aim of this research is to build a detector that emulates a human vision brain system in its ability to detect tool gouging and collision. Fortunately, research in artificial intelligence has developed methods called Adaptive Resonance Theory 2 that mimic the human vision-brain system. Therefore, two major goals in this research are investigating ways to train an ART2 network, which include how to select good parameter values and how to design the training data, and understanding the capability of the trained ART2 network to handle new data. The performance level in this research is defined as the percentage of correct classification of the detector, i.e., whether the detector can correctly classify the position and orientation of the tool as good or bad.; We studied two ways to find good parameter values in training the ART2 network. The first method is a heuristic approach, which starts by selecting some arbitrary initial values for each parameter; then a parameter is selected as a moving parameter. We will then search for a value of the moving parameter that most improves the performance level. The best value of the moving parameter will be then fixed for the next experiment, which will select another parameter as the moving parameter. This method is a very time consuming. In the second method, called the continuous learning method, we start by selecting initial values for the parameters and find the performance level with those values. Then when we change the value of a parameter, we utilize the learning matrices of the trained network to ask the network to continue the learning process with a new value of a parameter. The second method works very well for training a network with a large set of training data since it can reduce the processing time. To understand the capability of a trained ART2 network in handling new data, there are four detectors built in this research: Detector One, Detector Two, Detector Three and Detector Sculptured Surfaces, which represent a series of detectors whose training sets range from a simple part to a more complex part. The finding is that to be able to handle more general data, the ART2 network needs to be trained with varied types of surfaces and combinations of the tool position and orientation. |
