شناسایی شل‌شدگی در اتصال فلنجی پیچ‌دار لوله تحت نیروهای تصادفی با استفاده از شبکه‌های عصبی کانولوشنی

This study investigates condition monitoring for detecting bolt loosening in flanged pipe connections critical components widely used in fluid-transfer industries. Since corrosion, cracks, and especially bolt loosening can lead to leakage and system failure, early detection is essential. In the experimental setup, a flanged pipe is subjected to random excitation using white-noise signals. Dynamic responses of the pipe are recorded under healthy conditions as well as faulty conditions involving half-loosened and fully loosened bolts. The collected sig-nals are then processed using a Haar filter bank at a selected decomposition level, employing both low-pass and high-pass filter coefficients. After filtering, the Smoothed Pseudo Wigner–Ville distribution is applied to obtain detailed time–frequency representations of the responses. These time–frequency images serve as input features for training convolutional neural networks (CNNs). The CNNs are trained, validated, and tested using the generated datasets to classify the condition of the flange connection. The proposed method demonstrates a classification accuracy exceeding 95%, highlighting its effectiveness. Finally, the performance of this approach is compared with other feature-extraction techniques to confirm its superiority.