Skip to content

Machine Health Monitoring & Deep Learning

We briefly describe the advances in MHM, in particular the end-to-end intelligent fault diagnosis, under the lens of deep learning. More details can be found in recent reviews & monographs 123. A classical workflow in MHM 456 is typically composed of the following steps:

  • data acquisition,
  • feature extraction,
  • model construction,
  • decision making.

Data acquisition

A major body of work in MHM is based on vibration monitoring of components. This technique involves the analysis of vibration data coming from vital components of a rotating machine to detect features which reflect the operational state of machinery. The analysis leads to the identification of potential failures and their causes, and makes it possible to perform efficient preventive maintenance. The same signal analysis techniques can also be applied to signals of alternative modalities, such as acoustic emission, current-voltage, rotating speed etc, which may provide useful complementary information in certain applications 78. Datasets provided by dpmhm contain essentially vibration signals.

Recent studies910 reveal some important aspects that the data acquisition in view of fault diagnostic should follow. For example, as stated in 9:

Faults in bearings often manifest themselves at high frequencies, so the use of a high sampling rate – perhaps greater than 40 kHz – is recommended.

Some situations are frequently encountered in the data acquisition stage, e.g. sample imbalance (dominant healthy samples), domain imbalance (dominant samples from the source domain), inaccuracy in labelling and segmentation etc.

Feature extraction

In a classical workflow, feature extraction (or feature engineering) consists in transforming the raw signals (vibration/acoustic etc) into advanced forms using signal processing tools 11. Premilinary signal processings, e.g. signal separation and denoising, may be applied at this stage. Widely adopted feature extraction methods include: spectrogram, wavelet package transform etc. In place of raw signals, the handcrafted features can be taken as input to a representation learning network. See discussions of representation learning below and Signal Processing Techniques.

Model construction

Over the last decades a lot of statistical machine learning models have been successfully applied to MHM, including support vector machine (SVM), k-nearest neighbors (KNN), random forest etc 1213. Compared to more recent end-to-end models based on deep learning, these "shallow" methods generally achieve lower accuracy in MHM tasks but enjoy better explainability. Some of them are implemented as baseline methods in the dpmhm package.

In the following we outline mainly the deep learning based models. Our view on tasks of MHM can be summarized as:

Representation learning + Knowledge tranfer to downstream task

In the first stage, a NN for representation is trained by solving tasks adapted to the training dataset. In the second stage, the learned representation NN is transfered to some downstream tasks on the test dataset (e.g. by taking the penultimate layer of the trained classifier and adding supplementary layers then fine-tuning it). Such a paradigm is general enough to cover a wide range of application, from intelligent fault diagnostic to RUL prediction.

Architectures

Compared to classical "shallow" models, modern models based on deep learning can achieve higher accuracy while being more flexible and versatile, thus better suited for end-to-end system design. Depending on the specification of the dataset (e.g. unlabeled, unbalanced, limited) and the MHM task (diagnostic or prognostic 3), different network architectures have been proposed and some most popular ones are 214:

  • Convolutional Neural Network (CNN) 15
  • Auto-Encoder (AE) 16
  • Deep Belief Network (DBN) 17
  • Recurrent Neural Network (RNN)
  • Generative Adversarial Network (GAN)

Transfer learning

Transfer Learning (TL) is a machine learning methodology that focuses on applying knowledge gained while solving one problem to a different but related problem. It has received considerable attention in MHM and helps answer the question of whether a model trained on one type of machine/operating condition/fault can be adapted to another type 1819, which is central to the application in realistic scenarios.

Few-shot learning

Moreover, the real-case mechanical data has obviously unlabeled and unbalanced characteristics 20. Actually:

  • Compared to image datasets like ImageNet, vibration/acoustic/electrical signals are inherently more difficult to label, and manual labels can be inaccurate or incomplete.
  • Damage to rotating machinery is often irreversible. Unless initiated intentionally (which can be costly), anomalous data is collected less frequently than normal data. Moreover, different types of faults can also occur at different rates.

Few-shot learning (FSL) aims to improve the sample efficiency of learning, or in other words, to learn fast. This may be particularly relevant for MHM due to unbalanced data characteristics 2122. For example, diagnosis of a certain type of fault must be performed with a small amount of labeled data. FSL can be viewed as a specific form of TL in that intelligence gained on a generic context should be adaptable to a new context using only a small amount of information from the new context.

Representation learning

Self-supervised learning (SSL) refers to a machine learning paradigm to discover general representations from large-scale data without requiring human annotations, which is an expensive and time-consuming task. Such representations are useful for downstream learning tasks. Our view is that self-supervised representation learning is the key to transfer learning and few-shot learning in the context of MHM.

Classical representations of signals as time-frequency or time-scale features are manually designed and not adaptive to the signal class or to the ultimate task (e.g. diagnosis in MHM). However, in self-supervised learning, the representation is learned end-to-end directly from signals (or a classical form of representation) without the label information by solving some pre-tasks, which are generally discriminative or generative.

Decision making

A final step in the MHM workflow is to making decision with the outcomes of previous steps. This invovles often determining the health state of a machine given a test signal, which can be either diagnostic (a classification problem, i.e. to decide the type of fault) or pronostic (a prediction problem, i.e. to decide the RUL).

  1. Yaguo Lei, Bin Yang, Xinwei Jiang, Feng Jia, Naipeng Li, and Asoke K. Nandi. Applications of machine learning to machine fault diagnosis: A review and roadmap. Mechanical Systems and Signal Processing, 138:106587, April 2020. URL: https://www.sciencedirect.com/science/article/pii/S0888327019308088 (visited on 2023-02-23), doi:10.1016/j.ymssp.2019.106587. ↩

  2. Shen Zhang, Shibo Zhang, Bingnan Wang, and Thomas G. Habetler. Deep Learning Algorithms for Bearing Fault Diagnostics—A Comprehensive Review. IEEE Access, 8:29857–29881, 2020. Conference Name: IEEE Access. doi:10.1109/ACCESS.2020.2972859. ↩↩

  3. Behnoush Rezaeianjouybari and Yi Shang. Deep learning for prognostics and health management: State of the art, challenges, and opportunities. Measurement, 163:107929, October 2020. URL: https://www.sciencedirect.com/science/article/pii/S026322412030467X (visited on 2023-02-24), doi:10.1016/j.measurement.2020.107929. ↩↩

  4. Robert B. Randall and Jérôme Antoni. Rolling element bearing diagnostics—A tutorial. Mechanical Systems and Signal Processing, 25(2):485–520, February 2011. URL: https://www.sciencedirect.com/science/article/pii/S0888327010002530 (visited on 2023-02-24), doi:10.1016/j.ymssp.2010.07.017. ↩

  5. S. Joe Qin. Survey on data-driven industrial process monitoring and diagnosis. Annual Reviews in Control, 36(2):220–234, December 2012. Number: 2. URL: https://www.sciencedirect.com/science/article/pii/S1367578812000399 (visited on 2022-06-30), doi:10.1016/j.arcontrol.2012.09.004. ↩

  6. Yaguo Lei, Naipeng Li, Liang Guo, Ningbo Li, Tao Yan, and Jing Lin. Machinery health prognostics: A systematic review from data acquisition to RUL prediction. Mechanical systems and signal processing, 104:799–834, 2018. Publisher: Elsevier. ↩

  7. Oliver Mey, André Schneider, Olaf Enge-Rosenblatt, Dirk Mayer, Christian Schmidt, Samuel Klein, and Hans-Georg Herrmann. Condition Monitoring of Drive Trains by Data Fusion of Acoustic Emission and Vibration Sensors. Processes, 9(7):1108, July 2021. Number: 7 Publisher: Multidisciplinary Digital Publishing Institute. URL: https://www.mdpi.com/2227-9717/9/7/1108 (visited on 2022-02-28), doi:10.3390/pr9071108. ↩

  8. D. H. Pandya, S. H. Upadhyay, and S. P. Harsha. Fault diagnosis of rolling element bearing with intrinsic mode function of acoustic emission data using APF-KNN. Expert Systems with Applications, 40(10):4137–4145, August 2013. URL: https://www.sciencedirect.com/science/article/pii/S0957417413000468 (visited on 2023-02-28), doi:10.1016/j.eswa.2013.01.033. ↩

  9. Wade A. Smith and Robert B. Randall. Rolling element bearing diagnostics using the Case Western Reserve University data: A benchmark study. Mechanical Systems and Signal Processing, 64-65:100–131, December 2015. URL: https://www.sciencedirect.com/science/article/pii/S0888327015002034 (visited on 2022-02-19), doi:10.1016/j.ymssp.2015.04.021. ↩↩

  10. Dhiraj Neupane and Jongwon Seok. Bearing Fault Detection and Diagnosis Using Case Western Reserve University Dataset With Deep Learning Approaches: A Review. IEEE Access, 8:93155–93178, 2020. Conference Name: IEEE Access. doi:10.1109/ACCESS.2020.2990528. ↩

  11. Akhand Rai and S. H. Upadhyay. A review on signal processing techniques utilized in the fault diagnosis of rolling element bearings. Tribology International, 96:289–306, April 2016. URL: https://www.sciencedirect.com/science/article/pii/S0301679X15006052 (visited on 2023-02-24), doi:10.1016/j.triboint.2015.12.037. ↩

  12. Samir Khan and Takehisa Yairi. A review on the application of deep learning in system health management. Mechanical Systems and Signal Processing, 107:241–265, 2018. Publisher: Elsevier. ↩

  13. Ruonan Liu, Boyuan Yang, Enrico Zio, and Xuefeng Chen. Artificial intelligence for fault diagnosis of rotating machinery: A review. Mechanical Systems and Signal Processing, 108:33–47, 2018. Publisher: Elsevier. ↩

  14. Rui Zhao, Ruqiang Yan, Zhenghua Chen, Kezhi Mao, Peng Wang, and Robert X. Gao. Deep learning and its applications to machine health monitoring. Mechanical Systems and Signal Processing, 115:213–237, January 2019. URL: https://www.sciencedirect.com/science/article/pii/S0888327018303108 (visited on 2023-02-23), doi:10.1016/j.ymssp.2018.05.050. ↩

  15. Jinyang Jiao, Ming Zhao, Jing Lin, and Kaixuan Liang. A comprehensive review on convolutional neural network in machine fault diagnosis. Neurocomputing, 417:36–63, December 2020. URL: https://www.sciencedirect.com/science/article/pii/S092523122031225X (visited on 2022-10-14), doi:10.1016/j.neucom.2020.07.088. ↩

  16. Zheng Yang, Binbin Xu, Wei Luo, and Fei Chen. Autoencoder-based representation learning and its application in intelligent fault diagnosis: A review. Measurement, 189:110460, February 2022. URL: https://www.sciencedirect.com/science/article/pii/S0263224121013464 (visited on 2023-02-23), doi:10.1016/j.measurement.2021.110460. ↩

  17. Abdel-rahman Mohamed, George E. Dahl, and Geoffrey Hinton. Acoustic Modeling Using Deep Belief Networks. IEEE Transactions on Audio, Speech, and Language Processing, 20(1):14–22, January 2012. Number: 1. URL: http://ieeexplore.ieee.org/document/5704567/ (visited on 2020-01-08), doi:10.1109/TASL.2011.2109382. ↩

  18. Chuan Li, Shaohui Zhang, Yi Qin, and Edgar Estupinan. A systematic review of deep transfer learning for machinery fault diagnosis. Neurocomputing, 407:121–135, September 2020. URL: https://www.sciencedirect.com/science/article/pii/S0925231220306123 (visited on 2021-12-20), doi:10.1016/j.neucom.2020.04.045. ↩

  19. Huailiang Zheng, Rixin Wang, Yuantao Yang, Jiancheng Yin, Yongbo Li, Yuqing Li, and Minqiang Xu. Cross-Domain Fault Diagnosis Using Knowledge Transfer Strategy: A Review. IEEE Access, 7:129260–129290, 2019. Conference Name: IEEE Access. doi:10.1109/ACCESS.2019.2939876. ↩

  20. Tianci Zhang, Jinglong Chen, Fudong Li, Kaiyu Zhang, Haixin Lv, Shuilong He, and Enyong Xu. Intelligent fault diagnosis of machines with small & imbalanced data: A state-of-the-art review and possible extensions. ISA Transactions, 119:152–171, January 2022. URL: https://www.sciencedirect.com/science/article/pii/S0019057821001257 (visited on 2023-02-23), doi:10.1016/j.isatra.2021.02.042. ↩

  21. Ansi Zhang, Shaobo Li, Yuxin Cui, Wanli Yang, Rongzhi Dong, and Jianjun Hu. Limited Data Rolling Bearing Fault Diagnosis With Few-Shot Learning. IEEE Access, 7:110895–110904, 2019. Conference Name: IEEE Access. doi:10.1109/ACCESS.2019.2934233. ↩

  22. Chuanjiang Li, Shaobo Li, Ansi Zhang, Qiang He, Zihao Liao, and Jianjun Hu. Meta-learning for few-shot bearing fault diagnosis under complex working conditions. Neurocomputing, 439:197–211, 2021. Publisher: Elsevier. ↩