A Fault Diagnosis Model for Bearing Using CNN Deep Learning Classification Method Based on 2D Spectrum Map Information

Harjito, Reyhan Nada (2025) A Fault Diagnosis Model for Bearing Using CNN Deep Learning Classification Method Based on 2D Spectrum Map Information. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5007211204-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (12MB) | Request a copy

Abstract

Kerusakan pada bearing dapat menyebabkan gangguan signifikan pada performa mesin, termasuk kegagalan besar yang berujung pada kerugian finansial dan risiko keselamatan. Diagnosis kerusakan bearing secara real-time menjadi tantangan penting, terutama dalam menangani sinyal getaran yang non-stasioner dan dipengaruhi oleh noise. Penelitian ini mengusulkan metode diagnosis kerusakan bearing berbasis Convolutional Neural Network (CNN) dengan arsitektur Residual Network (ResNet) menggunakan peta spektrum 2D yang dihasilkan melalui Short-Time Fourier Transform (STFT). Metode ini memanfaatkan kemampuan CNN untuk mengekstrak fitur dari data citra, meningkatkan akurasi deteksi kerusakan dibandingkan pendekatan sinyal 1D tradisional. Penelitian ini menggunakan pendekatan validasi model berupa 5-Fold Cross Validation untuk meningkatkan keandalan evaluasi dan mengurangi risiko overfitting. Dua dataset digunakan dalam penelitian ini, yaitu CWRU dan MFPT. Pengujian dilakukan pada dua dataset, yaitu CWRU dan MFPT. Dataset CWRU memiliki keunggulan dengan 10 kelas kerusakan, variasi jenis dan ukuran cacat, serta konfigurasi sinyal single-channel dan dual-channel (Drive End dan Fan End), termasuk variasi beban motor dari 0 hingga 3 HP. Sementara itu, dataset MFPT digunakan karena mencerminkan kondisi operasional nyata dengan noise alami, sehingga cocok untuk menguji kemampuan generalisasi model. Variasi dilakukan pada ukuran jendela data (window length) dan arsitektur CNN ResNet, termasuk ResNet-18, ResNet-34, ResNet-50, dan ResNet-101, untuk mengevaluasi performa terbaik dalam klasifikasi. Hasil dari proses pelatihan dan validasi menunjukkan bahwa memiliki nilai loss terkecil dan waktu pelatihan tercepat dibandingkan model lainnya, menjadikannya pilihan optimal untuk aplikasi diagnosis kerusakan bearing baik pada data eksperimen terkontrol maupun pada data kondisi nyata yang mengandung noise. Selain itu, ResNet-18 secara konsisten memberikan hasil terbaik, dengan nilai akurasi, precision, recall, dan F1-score mencapai 100% pada seluruh variasi data CWRU dan MFPT. Temuan ini menunjukkan bahwa model yang diusulkan efektif, efisien, dan andal dalam mendeteksi kerusakan bearing berdasarkan input peta spektrum dari sinyal getaran.
======================================================================================================================================
Bearing faults can significantly disrupt machine performance, potentially leading to major failures, financial losses, and safety risks. Real-time bearing fault diagnosis presents a critical challenge, particularly in handling non-stationary vibration signals affected by noise. This study proposes a bearing fault diagnosis method based on a Convolutional Neural Network (CNN) with a Residual Network (ResNet) architecture, utilizing 2D spectrum map generated through Short-Time Fourier Transform (STFT). The method leverages CNN’s ability to extract features from image-based data, enhancing fault detection accuracy compared to traditional 1D signal-based approaches. A 5-Fold Cross Validation technique is employed to improve model evaluation reliability and reduce the risk of overfitting. Two datasets are used in this study: the CWRU dataset and the MFPT dataset. The CWRU dataset offers advantages such as 10 fault classes, variations in defect type and size, single-channel and dual-channel signal configurations (Drive End and Fan End), and different motor load conditions ranging from 0 to 3 HP. In contrast, the MFPT dataset represents real operational conditions with natural noise, making it suitable for testing the model’s generalization capability. Variations in window length and CNN architecture including ResNet-18, ResNet-34, ResNet-50, and ResNet-101 are explored to evaluate the best performance for classification tasks. The training and validation results indicate that ResNet-18 achieves the lowest loss and fastest training time among all models, making it the optimal choice for bearing fault diagnosis in both controlled experimental data and real-world noisy data. Furthermore, ResNet-18 consistently delivers outstanding results, achieving 100% accuracy, precision, recall, and F1-score across all variations of the CWRU and MFPT datasets. These findings demonstrate that the proposed model is effective, efficient, and reliable in detecting bearing faults based on spectrogram inputs derived from vibration signals.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	bearing, diagnosis kerusakan, ResNet, CNN, peta spektrum 2D, fault diagnosis, 2D spectrum map
Subjects:	T Technology > TJ Mechanical engineering and machinery > TJ174 Maintenance and repair of machinery
Divisions:	Faculty of Industrial Technology and Systems Engineering (INDSYS) > Mechanical Engineering > 21201-(S1) Undergraduate Thesis
Depositing User:	Reyhan Nada Harjito
Date Deposited:	02 Aug 2025 07:34
Last Modified:	02 Aug 2025 07:34
URI:	http://repository.its.ac.id/id/eprint/124710

Actions (login required)

View Item