The multi-task hybrid model with a multi-branch regressor for flaw detection and sizing in ultrasonic testing

This article discusses the development of a multi-task hybrid neural network model with a multi-branch regression block structure for the simultaneous detection and quantitative assessment of defect sizes based on ultrasonic non-destructive testing data. The primary objective of the study is to improve the accuracy of determining defect geometric parameters through parallel feature processing using different activation functions within a single multi-task architecture. The study utilizes ultrasonic testing data for a welded joint made of austenitic stainless steel with artificial cracks. The methodology included expanding the previously developed CNN-GRU model for binary classification to a multi-task model, where the regression block is implemented as a multi-branch structure with parallel transformations and subsequent feature integration. Training was conducted with balanced loss functions to jointly optimize classification and regression problems. The results demonstrated the high efficiency of the proposed approach. The model demonstrated absolute classification accuracy and a low regression error: the average absolute error was 0.118 mm (5.3% of the average defect size). A comparison with a model of a similar architecture without a multi-branch structure confirmed that the proposed solution reduces the error by more than twofold and eliminates systematic prediction bias. The developed architecture may have practical implications for automated ultrasound diagnostic systems that require not only detection but also precise measurement of defect parameters.

1. Shi Y., Xu W., Zhang J., Li X. Automated Classification of Ultrasonic Signal via a Convolutional Neural Network. Applied Sciences. 2022;12(9). https://doi.org/10.3390/app12094179

2. Barshok K., Choi J.‑I., Lee J. Deep Learning‑Based Approach for Automatic Defect Detection in Complex Structures Using PAUT Data. Sensors. 2025;25(19). https://doi.org/10.3390/s25196128

3. Wang H., Fan Zh., Chen X., et al. Automated Classification of Pipeline Defects from Ultrasonic Phased Array Total Focusing Method Imaging. Energies. 2022;15(21). https://doi.org/10.3390/en15218272

4. Krolik A., Drelich R., Pakuła M., Mikołajewski D., Rojek I. Detection of Defects in Polyethylene and Polyamide Flat Panels Using Airborne Ultrasound‑Traditional and Machine Learning Approach. Applied Sciences. 2024;14(22). https://doi.org/10.3390/app142210638

5. Tunukovic V., McKnight Sh., Mohseni E., et al. A study of machine learning object detection performance for phased array ultrasonic testing of carbon fibre reinforced plastics. NDT & E International. 2024;144. https://doi.org/10.1016/j.ndteint.2024.103094

6. Soloviev A.N., Sobol B.V., Vasiliev P.V., Senichev A.V., Novikova A.I. Identification of defects in a coating wedge based on ultrasonic non-destructive testing methods and convolutional neural networks. PNRPU Mechanics Bulletin. 2023;(1):111–124. (In Russ.). https://doi.org/10.15593/perm.mech/2023.1.11

7. Sheehan P.S., Miorelli R., Robert S., Chapuis B., Chatillon S. Investigation of a Deep Learning Methodology for Automatic Detection and Characterization of Crack-Type Defects in Ultrasonic Non-Destructive Testing. In: 2025 ICU PADERBORN – 9th International Congress on Ultrasonics, 21–25 September 2025, Paderborn, Germany. 2025. P. 295–298. https://doi.org/10.5162/Ultrasonic2025/E4-a3

8. Shi S., Jin Sh., Zhang D., et al. Improving ultrasonic testing by using machine learning framework based on model interpretation strategy. Chinese Journal of Mechanical Engineering. 2023;36(1). https://doi.org/10.1186/s10033-023-00960-z

9. Pyle R.J., Bevan R.L.T., Hughes R.R., et al. Deep Learning for Ultrasonic Crack Characterization in NDE. IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control. 2021;68(5):1854–1865. https://doi.org/10.1109/TUFFC.2020.3045847

10. Fei Q., Cao J., Xu W., et al. A Deep Learning-Based Ultrasonic Diffraction Data Analysis Method for Accurate Automatic Crack Sizing. Applied Sciences. 2024;14(11). https://doi.org/10.3390/app14114619

11. Azad M.M., Jung J., Kim H.S., et al. An integrated multi-task transfer learning for damage detection, localization, and severity assessment of laminated composite plate. Composite Structures. 2025;371. https://doi.org/10.1016/j.compstruct.2025.119478

12. Dong X., Taylor Ch.J., Cootes T.F. Defect Classification and Detection Using a Multitask Deep One-Class CNN. IEEE Transactions on Automation Science and Engineering. 2022;19(3):1719–1730. https://doi.org/10.1109/TASE.2021.3109353

13. Kim K., Kim K.S., Park H.-J. Multi-branch deep fusion network-based automatic detection of weld defects using non-destructive ultrasonic test. IEEE Access. 2023;11:114489–114496. https://doi.org/10.1109/ACCESS.2023.3324717

14. Wang L., Qi Zh., Ding X., et al. Accurate detection and characterization of sub-millimeter cracks using nonlinear ultrasonics-informed parallel multi-branch convolutional neural network. Engineering Applications of Artificial Intelligence. 2026;166. https://doi.org/10.1016/j.engappai.2025.113638

15. Cao W., Sun X., Liu Zh., et al. The detection of PAUT pseudo defects in ultra-thick stainless-steel welds with a multimodal deep learning model. Measurement. 2025;241. https://doi.org/10.1016/j.measurement.2024.115662

16. Ivanov D.A. Review of an approach to processing ultrasonic non-destructive testing data using machine learning. Scientific and Technical Volga Region Bulletin. 2025;(6):29–33. (In Russ.).

17. Ivanov D.A., Kondratov D.V. The hybrid model for spatio-temporal processing of ultrasonic non-destructive testing data. Modern High Technologies. 2025;(12):69–77. (In Russ.). https://doi.org/10.17513/snt.40606

18. Virkkunen I., Koskinen T., Jessen-Juhler O., Rinta-aho J. Augmented Ultrasonic Data for Machine Learning. Journal of Nondestructive Evaluation. 2021;40(1). https://doi.org/10.1007/s10921-020-00739-5