PREDICTION OF LOCATOR PIN ERRORS IN FACING PROCESS USING NEURAL NETWORK AND SUPPORT VECTOR REGRESSION

• Main Author: Yusli Arbain Sugoro, Ramdhani
• Format: Theses
• Language: Indonesia
• Subjects: Teknik (Rekayasa, enjinering dan kegiatan berkaitan)
• Online Access: https://digilib.itb.ac.id/gdl/view/55518
• Institution: Institut Teknologi Bandung

Machining is a widely used process because it can achieve tight tolerances. The machining process results from many factors. One of the factors that is quite important is the positioning of the workpiece. The general principle of positioning the workpiece is the 3-2-1 principle. The position of the pins can cause the machining process workpiece to experience an error. The most common machining process is the process of flattening the workpiece (facing). This process is important because the leveled surface is usually used as a reference for measuring other features. The facing process is usually carried out on two opposite sides of the workpiece. One work at a surface may result in another feature error. Facing process error could be the result of a three-pin error in 3-2- 1 principle. The process of measuring pin error is usually difficult to carry out periodically on the production floor. Control of errors that occur is usually carried out using statistical quality control methods. The disadvantage of this method is that the root of the problem cannot be analyzed in detail. This study discusses how to identify pin errors based on the deviation measurement values of objects after facingg using neural network algorithms and supporting vector regression. The facing process is carried out on two opposite surfaces so that errors on the surface will affect the results on the other surface. The two processes separately modeled and become two cases. The first case is the prediction of the pin error value based on the deviation of the object at the fifteen measuring points. The second case is the prediction of the pin error value based on the deviation of the object at the sixteen measuring points and the surface error represented by the pin error in the first case. Machine learning models are created using the Keras program. The modeling process includes defining objectives, data preparation, parameter selection, model training, and model testing. The selection of parameters in the neural network algorithm is the selection of the number of hidden layers, the activation function used, and the learning rate value. The two hidden layer model with the ReLU activation function is the best parameter to predict the pin error value in both case 1 and case 2. Algorithms supporting vector regression have several parameters that can be adjusted, namely kernel, C, and ?. The Gaussian RBF kernel is the best kernel for case 1 and case 2. The performance of case 1 neural network model based on R2 score reaches 0.960. The case 2 artificial neural network model based on the R2 score reached 0.986. The support vector regression on case 1 based on R2 score reaches 0.9485. The support vector regression reaches 0.9921 of R2 score on case 2.