Vol.9, No 1, 2011 pp. 21 - 32
UDC 007.52, 621.941, 519.863
METHODOLOGY OF DEVELOPING OPTIMAL BP-ANN MODEL FOR THE PREDICTION OF CUTTING FORCE IN TURNING USING EARLY STOPPING METHOD
Miloš Madić, Miroslav Radovanović
University of Niš, Faculty of Mechanical Engineering, A. Medvedeva 14, Niš, Serbia
E-mail: mirado@masfak.ni.ac.rs
Predictive modeling is essential for better understanding and optimization of the machining processes. This paper presents the modeling methodology for predicting the cutting force in turning AISI 1043 steel based on artificial neural networks (ANNs). Based on the previous theoretical and experimental studies, a comprehensive analysis of the ANN training and architectural parameters is carried out in order to develop an optimal ANN model of high predictive performance. In order to improve generalization capabilities of the ANN models, early stopping (ES) method is used in ANN training. The ANN models trained with backpropagation (BP) training algorithm are developed using experimental machining data. The optimal 3-2-1 BP- ANN model is selected based on multiple statistical criteria. It was found that the 3-2-1 ANN model has very good prediction performance in terms of agreement with experimental data.
Key words: Artificial Neural Networks, Modeling, Early Stopping Method, Prediction, Turning, Cutting Force
METODOLOGIJA KREIRANJA OPTIMALNOG BP-VNM MODELA ZA PREDIKCIJU GLAVNOG OTPORA REZANJA KOD STRUGANJA PRIMENOM METODE RANOG ZAUSTAVLJANJA
Predikciono modelovanje je od suštinske važnosti za bolje razumevanje i optimizaciju obradnih procesa. U ovom radu predstavljena je metodologija modelovanja za predikciju glavnog otpora rezanja kod struganja AISI 1043 čelika zasnovana na veštačkim neuronskim mrežama (VNM). Na osnovu prethodnih teorijskih i eksperimentalnih istraživanja, izvršena je sveobuhvata analiza VNM parametara strukture i treniranja, a sve u cilju kreiranja optimalnog VNM visokih predikcionih performansi. U cilju poboljšanja generalizacije modela VNM, primenjena je metoda ranog zaustavljanja (RZ) u procesu treniranja VNM. Treniranje modela VNM izvršeno je backpropagation (BP) algoritmom korišćenjem eksperimentalnih podataka. Optimalni 3-2-1 BP-VNM model je izabran na osnovu nekoliko statističkih kriterijuma. Utvrđeno je da 3-2-1 VNM model ima veoma dobre predikcione performanse u smislu slaganja sa eksperimentalnim podacima.
Ključne reči: veštačke neuronske mreže, modelovanje, metoda ranog zaustavljanja, predikcija, struganje, otpori rezanja