Real-Time Application Of Support Vector Machines For Chatter Detection Using Sound Signals In Boring Of Thin-Walled Aluminum Alloy 6063 Cylinders

Bilfaqih, Hilmy Al-May (2026) Real-Time Application Of Support Vector Machines For Chatter Detection Using Sound Signals In Boring Of Thin-Walled Aluminum Alloy 6063 Cylinders. Other thesis, Sepuluh Nopember Institute of Technology.

Text 5007211061_Undergraduate_Thesis.pdf Restricted to Repository staff only Download (10MB) | Request a copy

Abstract

The regenerative chatter phenomenon during the boring process of a thin-walled Aluminum 6063 decreases the quality of the specimen surface and shortens the boring insert life. Manual detection of Chatter based on the sound is subjective and slow. Therefore, this research aims to make real-time chatter detection using sound signals applied by machine learning. This methodology utilizes microphone-recorded acoustic signals, processed through spatial domain segmentation to synchronize local physical conditions with time- and frequency-domain features at 5 mm intervals. There are 157, 65, and 138 datasets of Non-Chatter, Transition, and Chatter. Due to the imbalance, the Transition data was augmented through overlapping segmentation by shifting to 1.5 mm and 3.5 mm with a similar interval at 5 mm. It achieves proportional data for Non-Chatter, Transition, and Chatter of 157, 141, and 138. Features were selected based on high correlation with Chatter and low inter-feature correlation. The selected features are spectral entropy, spectral spread, and energy, which showed label correlations of -0.91, -0.55, and 0.53, respectively. The inter-feature correlations were 0.61 (spectral entropy–spectral spread), -0.38 (spectral entropy–energy), and -0.312 (spectral spread–energy). The labeled features were partitioned to 60%, 20%, and 20% for training, validation, and testing. The stratified sampling is employed to randomize the data with balanced classes on every partition. The grid search generates 7, 35, 210, and 1050 parameter combinations from Linear, RBF, sigmoid, and polynomial kernels. The RBF and polynomial kernel satisfy the accuracy threshold of 90%. The RBF has the highest training accuracy at 98.08% compared to the polynomial kernel at 97.32%. The validation accuracy of the RBF kernel is 91.95%, similar to the accuracy of the polynomial kernel. The RBF kernel was selected because of the higher recall for Chatter of 89.29% compared to the polynomial kernel at 85.71%. This ensures that the system can detect critical conditions with higher sensitivity. The learning curves show that the RBF kernel generalization gaps range from 6.13% to 10.14% at the training samples of 80-261. The testing model of the RBF Kernel has an accuracy, recall, precision, and F1-score of 93.18%, 93.15%, 93.26%, and 93.15%. It shows a stable performance of the RBF kernel with the C and γ of 100 and 1. The model is reliable for the chatter detection system for industrial purposes.
=============================================================================================================================
Fenomena regenerative chatter selama proses boring pada material thin-walled Aluminum 6063 menurunkan kualitas permukaan spesimen dan memperpendek umur boring insert. Deteksi chatter secara manual berdasarkan suara bersifat subjektif dan lambat. Oleh karena itu, penelitian ini bertujuan untuk membuat deteksi chatter secara real-time menggunakan sinyal suara yang diaplikasikan dengan machine learning. Metodologi ini menggunakan sinyal akustik yang direkam oleh mikrofon, diproses melalui spatial domain segmentation untuk menyinkronkan kondisi fisik lokal dengan fitur time-domain dan frequency-domain pada interval setiap 5 mm. Terdapat 157, 65, dan 138 dataset untuk Non-Chatter, Transition, dan Chatter. Karena ketidakseimbangan data, data Transition ditambah (augmented) melalui overlapping segmentation dengan pergeseran sebesar 1,5 mm dan 3,5 mm dengan interval serupa pada 5 mm. Hal ini menghasilkan data yang proporsional untuk Non-Chatter, Transition, dan Chatter sebanyak 157, 141, dan 138. Fitur dipilih berdasarkan korelasi yang tinggi dengan Chatter dan korelasi antar-fitur (inter-feature correlation) yang rendah. Fitur yang dipilih adalah spectral entropy, spectral spread, dan energy, yang menunjukkan korelasi label masing-masing sebesar -0,91, -0,55, dan 0,53. Korelasi antar-fitur adalah 0,61 (spectral entropy–spectral spread), -0,38 (spectral entropy–energy), dan -0,312 (spectral spread–energy). Fitur yang telah diberi label dipartisi menjadi 60%, 20%, dan 20% untuk training, validation, dan testing. Stratified sampling digunakan untuk mengacak data dengan kelas yang seimbang pada setiap partisi. Grid search menghasilkan 7, 35, 210, dan 1050 kombinasi parameter dari kernel linear, RBF, sigmoid, dan polynomial. Kernel RBF dan polynomial memenuhi ambang batas (threshold) akurasi sebesar 90%. RBF mencapai akurasi training tertinggi sebesar 98,08% dibandingkan dengan kernel polynomial sebesar 97,32%. Akurasi validation dari kernel RBF adalah 91,95%, serupa dengan akurasi kernel polynomial. Kernel RBF dipilih karena nilai recall untuk Chatter yang lebih tinggi yaitu 89,29% dibandingkan dengan kernel polynomial sebesar 85,71%. Hal ini memastikan bahwa sistem dapat mendeteksi kondisi kritis secara lebih sensitif. Learning curves menunjukkan bahwa kernel RBF memiliki rentang generalization gap sebesar 6,12% hingga 10,14% pada sampel training 80-261. Model testing dari kernel RBF memiliki akurasi, recall, precision, dan F1-score masing-masing sebesar 93,18%, 93,15%, 93,26%, dan 93,15%. Ini menunjukkan performa yang stabil dari kernel RBF dengan nilai C dan γ sebesar 100 dan 1. Model ini andal untuk sistem deteksi chatter untuk keperluan industri.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Boring, Pearson Correlation Matrix, Chatter, Support Vector Machine, Pearson Correlation Matrix, Chatter, Support Vector Machine
Subjects:	T Technology > TJ Mechanical engineering and machinery > TJ1185 Metal-cutting--Problems, exercises, etc.
Divisions:	Faculty of Industrial Technology and Systems Engineering (INDSYS) > Mechanical Engineering > 21201-(S1) Undergraduate Thesis
Depositing User:	Hilmy Al-may Bilfaqih
Date Deposited:	04 Feb 2026 01:33
Last Modified:	04 Feb 2026 01:33
URI:	http://repository.its.ac.id/id/eprint/131969

Actions (login required)

View Item