AGMA 91FTM11-1991 pdf download.Initial Design of Gears Using Artificial Neural Net by: T. Jeong and T. P. Kicher, Case Western Reserve;  and R. J. Zab, Joy Technology.
The artificial neural net is composed of highly interconncted layers which attempt to achieve human neuron-like performance [3]. It is designed to emulate the human neural activities, exhibiting abilities such as learning, generalization, and abstraction [4]， using mathematical implementations. A typical model of the artifcial neural net is illustrated in figure 4. The modeled net has three layers; input, hidden (or middle), and output layers. This model is extremely simple, compared to the hundred trillion connections of the human neural system [2]. The terms shown in parenthesis in fgure 4 are the anatomic terms of the human neural system.
In figure 4，each node in one layer receives multiple signals from the nodes in the previous layer. The strength of each signal is determined by the value of the connecting weight between paired nodes. The signals conveyed to the node are summed and averaged (or mathematically evaluated) to decide whether this node will activate or not. If the node activates, the sigoal generated will be transmitted to the nodes in the next layer. The artificial neural net is not functional without existing knowledge, just as a human engineer can not perform a task without pre-existing knowledge of the field. The net must be trained with known knowledge patterns that consist of input and the corresponding target output. The knowledge patterns are fed through the net so that the connecting weights can be learmned and memorized. Once all the connecting weights are established, the net will produce the proper output when the same or similar input pattern is seen. Accordingly, the quality of the knowledge patterns used for training influences the quality of the estimated outputs. The net is said to be successfully trained if the estimated outputs match the target outputs within a certain level of error. Because the training knowledge patterns may not be perfect, there is always the chance that an errant estimation may appear. In comparison, it can also be said that the performance of the human engineer will be inaccurate if incorrect knowledge was used in training. ARTIFICIAL NEURAL NET ALGORITHMS ; Many artificial neural net algorithms have been developed and implemented. Although there are some structural variations, the basic idea is equivalent in terms of implementing a human neural system. Each algorithm has its own characteristics and applicable regime. After the nature of initial gear design was investigated, two algorithms, namely LVQ (Learning Vector Quantization) and GDR (Generalized Delta Rule), were selected to emulate two steps of initial gear design. LVQ is also known as the pattem recognition or classification method, which classifies available knowledge patterns in a pattern space [5]. Each patterm must have its own class label (or class I.D.). LVQ forms clusters which include identically labeled patterns while remembering their weight centers. When a new input patterm without a class label, not encountered previously, is seen, LVQ locates the cluster weight center which is closest to the new input pattern and sends the class label of the selected cluster as the output. In other words, LVQ simply tells where the new input patterm belongs.AGMA 91FTM11 pdf download.