Study on Preprocessing Methods for Ultrasonic Signals from Internal Defects in Rolls.

Accurate detection of internal defects in rolls is crucial for industrial safety and product quality. Ultrasonic testing is a mainstream non-destructive method widely used for this purpose. However, in practice, ultrasonic echo signals often suffer from background clutter. When defects are located near the surface, weak defect echoes tend to couple with surface echoes, making signal extraction difficult and reducing the accuracy of subsequent feature extraction and classification. This paper proposes a novel ultrasonic signal preprocessing method aimed at improving the performance of subsequent defect identification models. The method first acquires ultrasonic signals from defect regions and background clutter reference signals from defect-free regions using a digital ultrasonic flaw detector. An improved median filter is then applied to remove spike interference and boundary outliers. On this basis, a multi-stage FIR (finite impulse response) filter is constructed, and particle swarm optimization is employed to adaptively optimize filter parameters, achieving an accurate estimation of background clutter. Finally, the clutter-suppressed defect signal is obtained through signal subtraction. Experimental results on a dataset of 5000 samples (2500 defective, 2500 non-defective) containing cylindrical artificial defects (diameter 8 mm, length 30 mm) demonstrate that using a CNN classifier with the same feature extraction and classification model, the signals preprocessed by the proposed method outperform traditional median filtering and wavelet denoising methods. The defect identification accuracy is improved by approximately 38 percentage points compared to median filtering and 20 percentage points compared to wavelet denoising, while also achieving a high recall rate, validating the effectiveness of the proposed method in enhancing roll internal defect detection.