导航

ACTA AERONAUTICAET ASTRONAUTICA SINICA ›› 2022, Vol. 43 ›› Issue (8): 625429.doi: 10.7527/S1000-6893.2021.25429

• Special Topic: Application of Fault Diagnosis Technology in Aerospace Field • Previous Articles     Next Articles

Gearbox fault diagnosis with small training samples: An improved deep forest based method

SHAO Yiwei1,3, CHEN Jiayu1,2, LIN Cuiying1, WAN Cheng3, GE Hongjuan1,2, SHI Zhilong4   

  1. 1. College of Civil Aviation, Nanjing University of Aeronautics and Astronautics, Nanjing 211106, China;
    2. Civil Aviation Key Laboratory of Aircraft Health Monitoring and Intelligent Maintenance, Nanjing University of Aeronautics and Astronautics, Nanjing 211106, China;
    3. College of Electronic and Information Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing 211106, China;
    4. Hangzhou Branch, Aircraft Maintenance and Engineering Coorporation, Hangzhou 310000, China
  • Received:2021-02-28 Revised:2021-04-30 Online:2022-08-15 Published:2021-04-29
  • Supported by:
    the Fundamental Research Funds for the Central University (1007-XZA20017);National Natural Science Foundation of China (U1933115)

Abstract: The gearbox is an important transmission part of complex aero-equipment, and fault diagnosis of gearbox plays a vital role in ensuring reliable and continued airworthiness. With the continuous development of artificial intelligence technology, deep learning-based methods have become a research hotspot in this field. However, deep neural networks have strict requirements on hyperparameter settings and training data volume, and are difficult to meet the requirements of fast, accurate and stable diagnosis in the actual industry. To solve these problems, a diagnostic method is proposed based on the improved deep forest to realize the efficient diagnosis of the multiple and mixed faults of gearbox with small training samples. For the contradiction between the long characteristic of the single-sample data in the rotating machinery vibration signal and the high cost of data processing in the deep forest model, a deep forest model is designed based on principal component analysis (PCA) feature extraction to solve redundant calculation of data in the original model. The improved deep forest enhances data transmission and processing capability in multi-grained scanning and cascade forest. While ensuring diversity of data, it enhances the representativeness of the internal features in the model, so as to improve the operating efficiency and diagnostic performance of the algorithm. Experiments on gearbox fault diagnosis under small training samples are carried out by controlling the variable of the ratio of total datasets to training samples. The results show that average diagnostic accuracy of the proposed method reaches 97.3% and 82.8% under the condition of 50% and 10% training-total dataset ratio, respectively, which verifies the effectiveness of the proposed method. The diagnostic performance of the proposed method is found to outperform the existing intelligent diagnostic methods for gearbox.

Key words: deep learning, fault diagnosis, small training samples, deep neural network, principal component analysis

CLC Number: