Using Artificial Neural Networks to Predict Grinds for the Small Block Chevrolet Camshaft in Aftermarket Applications

During the first design stages in the production of a new after-market camshaft for racing applications, custom auto parts companies rely on experienced technicians to predict the best shape of the camshaft’s lobes. With nothing other than their eye to guide them, numerous iterations of the shape of the lobes were required before the proper combination was achieved. Using the engine rpm as the independent variable, a very good initial estimate of the lobe parameters was achieved by teaching an Artificial Neural Network (ANN) the camshaft characteristics. Introduction Artificial Neural Networks are modeled loosely on the way the human brain computes information. Because of the analogies of an ANN and the human brain, some of the vocabulary used when describing an ANN is similar to that used to describe the way the brain processes information. Early applications of ANN were in the field of artificial intelligence. ANNs are now being used to model the complex physics necessary to model heat transfer equipment, see Yang et. al (2000). After-market camshaft design is also complex: only one independent variable is provided to predict 12 dependent variables. Because of the relative ease with which an ANN can be used to learn and then predict complex mathematical and physical relationships, it was chosen to model aftermarket camshaft design. The current work was completed as an undergraduate research requirement for an internal combustion engines class at Christian Brothers University. Required Input for Grinding Custom Camshafts Traditionally, custom camshafts that are used in racing cars are machined through a trialand-error process by the custom camshaft manufacturer. Even the best initial prediction, by a thirty year racing enthusiast, usually needs to be refined in order to provide a good starting point to grind the new camshaft from stock for a particular application. In racing terminology, one refers to the grind of a particular camshaft. Camshaft blanks are machined metal rods that have a larger diameter at the axial locations where the lobes are to be ground. The grind of a camshaft is a set of variables input into a Computer Numerically Controlled (CNC) grinder that determine the camshaft’s specific shape. The CNC grinder then grinds the lobes from the camshaft blank. The input variables to the CNC are: lift, duration, lobe separation, and the degree at which the intake and exhaust valves are opened and closed with respect to either top dead center or bottom dead center. Camshafts are designed to produce peak engine power at a particular engine rpm ( although low and mid range torque is also an important consideration). The commercially available grinds are designed for a set of specific engine rpm. If the modifications produced by the race team necessitate peak power at an rpm that is different from the ones associated with the commercially available camshafts, a custom camshaft must be ground. For a given engine, the camshaft rpm at peak engine horsepower is usually the only camshaft variable specified by a race team to the machinist. For this reason the ANN was developed with camshaft rpm as the only independent variable, and all the other camshaft variables as dependant variables. The most widely produced series of after-market camshafts is for the small block Chevrolet “street-strip” application. These camshafts are used for engines produced between the years 1957 to 1986. The Lunati camshafts were chosen for the data base because they have been in a leader of quality after-market camshafts since the early 1960’s. Description of the Neural Network A neural network consists of a number of layers, and a number of nodes assigned to each layer. The first layer is the input layer, with the number of nodes in this layer corresponding to the number of independent variables. In this study since engine rpm is the only independent variable, there is only one node in the first layer. The last layer of the network is the output. It has a number of nodes corresponding to the number of independent variables. Between the first and last layer are hidden layers. The number of nodes in each of these layers generally vary from layer to layer. It is outside the scope of this paper to detail the theory of neural networks. A good history and a more detailed description of neural network theory can be found in Haykin (1999). A fully-connected feedforward network with a sigmoid activation function. Initial weights and biases used between nodes were generated with a standard random number generator. A backward-error propagation algorithm is used to teach the ANN the data. The network consisted of an input layer with 1 node, an output layer with 12 nodes, and 3 hidden layers with 6, 14, and 16 nodes respectively. A relaxation factor of 0.5 was used. After 150,000 iterations the calculated errors between the calculated dependent variables and the values from the data file were low enough to demonstrate that the ANN successfully learned the network. Initially data for 10 camshafts ( 2000 < rpm <6000) was used to teach the ANN.