Elsevier

Measurement

Volume 111, December 2017, Pages 1-10
Measurement

A convolutional neural network based feature learning and fault diagnosis method for the condition monitoring of gearbox

https://doi.org/10.1016/j.measurement.2017.07.017Get rights and content

Highlights

  • A CNN based feature learning and fault diagnosis method for gearboxes is proposed.

  • The performance of feature learning of CNN with various data types is tested.

  • The selection of key parameters of CNN is discussed.

  • Feature learning with CNN provides better results than manual feature extraction.

  • CNN is more suitable to learn features from vibration signal in frequency domain.

Abstract

Feature extraction plays a vital role in intelligent fault diagnosis of mechanical system. Nevertheless, traditional feature extraction methods suffer from three problems, which are (1) the requirements of domain expertise and prior knowledge, (2) the sensitive to the changes of mechanical system and (3) the limitations of mining new features. It is attractive and meaningful to investigate an automatic feature extraction method, which can adaptively learn features from raw data and discover new fault-sensitive features. Deep learning has been widely used in image analysis and speech recognition with great success. The key advantage of this method lies into the ability of mining representative information and sensitive features from raw data. However, the application of deep learning in feature leaning for mechanical diagnosis is still few, and limited studies have been carried out to compare the effectiveness of feature leaning with various data types. This paper will focus on developing a convolutional neural network (CNN) to learn features directly from frequency data of vibration signals and testing the different performance of feature learning from raw data, frequency spectrum and combined time-frequency data. Manual features from time domain, frequency domain and wavelet domain as well as three common intelligent methods are used as comparisons. The effectiveness of the proposed method is validated through PHM 2009 gearbox challenge data and a planetary gearbox test rig. The results demonstrate that the proposed method is able to learn features adaptively from frequency data and achieve higher diagnosis accuracy than other comparative methods.

Introduction

Gearbox is a key component in mechanical transmission systems and widely used in wind turbines, machine tools, helicopters and other important mechanical equipments. A failure of a gearbox may cause unwanted downtime, production losses and even human casualties [1]. Therefore, it is necessary to detect the faults of gearboxes effectively to prevent accidents and guarantee an efficient operation of mechanical transmission systems.

While various methods have been developed for the fault diagnosis of gearboxes, intelligent fault diagnosis methods are widely accepted and used to tackle complicated mechanical diagnosis problems due to its adaptive learning mechanism, strong fault tolerant and high non-linear regression ability [2], [3]. Generally, intelligent fault diagnosis method includes three main steps [4]: (1) signal acquisition, (2) feature extraction, and (3) fault classification. In the step of signal acquisition, several signals including vibration, acoustic, electrical current, speed, temperature, etc., are acquired. Vibration signal is one of the most commonly used signals in mechanical fault diagnosis, since it usually represents the most intrinsic information of the health conditions of mechanical equipments. In the second step, fault-sensitive features are extracted and selected from raw signal through signal processing technology and dimension reduction strategies, such as Fourier spectral analysis, wavelet transformation and principal component analysis (PCA). In the step of fault classification, health conditions are diagnosed based on the extracted features through intelligent classification techniques, such as neural networks (NN), support vector machine (SVM) and self organizing map (SOM).

Whether the fault-sensitive features are extracted or not usually have a great influence on the performance of intelligent fault diagnosis methods. However, it is difficult to choose a suitable feature extraction method due to the highly requirement of extensive domain expertise and prior knowledge. At the same time, manual feature extraction usually depends on the existing features or evaluation criteria, which make it difficult to explore new useful features. Moreover, the feature extraction method is often sensitive to the changes in the physical characteristics of mechanical systems, which means a change in the components or fault conditions may dramatically alters the feature extraction method or its evaluation criteria. Lots of the efforts in intelligence fault diagnosis actually have been devoted into the design of a suitable feature extraction method for different diagnosis tasks. Therefore, it is urgent and meaningful to develop an automatic feature extraction method, which is able to learn features directly from raw signal and adaptive to the changes of mechanical systems.

Deep learning, also known as deep neural networks (DNN), attracts growing attention from researchers from various fields in recent years [5]. It employs a hierarchical structure with multiple neural layers and extracts information form input data through a layer by layer process. This “deep” layer structure allows it to learn the representations of complicated raw data with multiple levels of abstraction. Starting with the raw input, DNN automatically discover intricate structure in large datasets and learn useful features layer by layer [6]. Benefited from the feature learning characteristic, DNN has been widely used in visual recognition and language understanding, and this feature learning ability also becomes its key advantage [6], [7], [8], [9], [10]. It is obvious that the advantage of the feature learning ability of DNN just meet the requirements of an adaptive feature extraction method for mechanical fault diagnosis. There lies a great potential and a critical need to utilize DNN and its feature learning ability for fault diagnosis of mechanical systems.

As shown in Fig. 1, the applications of deep learning in mechanical fault diagnosis are divided into two stages, depending on whether the feature learning ability of DNN is used:

  • (1)

    During the past several years, almost all the researchers used DNN only as a classifier or a feature selection method for mechanical fault diagnosis in the similar way to traditional intelligent methods. Features are extracted by various signal analysis methods at first. Then these extracted features are used to train and test the DNN models [11], [12], [13], [14], [15], [16], [17], [18], [19], [20], [21], [22], [23], [24]. Li et al. [14] applied a deep belief networks (DBN) to diagnose gearboxes and bearings with statistic features in time, frequency and time-frequency domains; Chen et al. [16] extracted several time and frequency features and employed a convolutional neural network (CNN) to classify different health conditions of a gearbox. Verma et al. [19] developed a sparse auto-encoder (SAE) and extracted time, frequency and wavelet domain features as the model input to monitor air compressors. Shao et al. [24] composed an optimized DBN to enhance 18 time-domain features and diagnose faults of bearings. In this stage, although DNN models are applied in mechanical fault diagnosis, it is only used as a replacement of traditional intelligent methods. The key advantage of feature learning of DNN is not completely employed.

  • (2)

    To the best of our knowledge, it is only since 2015 that some researchers recognized the feature learning advantage of DNN and started to apply it to learn features from vibration data [10], [25], [26], [27], [28], [29], [30]. Chen et al. [10] developed a DBN to learn features from raw vibration signal for bearing fault classifications. Guo et al. [25] composed a hierarchical deep convolution neural network with two combined CNN to monitor health conditions of rolling bearings. One CNN is used to learn features from raw vibration data and recognize fault patterns. The other one is applied to evaluate the fault sizes of each fault pattern. Janssens et al. [26] explored CNN to diagnose bearing housings with raw frequency data. The vibration data of the bearing housings is preprocessed through fast fourier transformation (FFT) and inputted into CNN to detect faults. Zhao et al. [28] built a convolutional long short-term memory networks (C-LSTM), in which a CNN is used to extract local features from raw sensory data, and a LSTM is provided to predict tool wear. Sun et al. [29] demonstrated a SAE to extract features and monitor the health conditions of an induction motor. Lu et al. [30] applied a denoising auto-encoder (DAE) to learn features from raw vibration data and diagnose faulty bearings. Benefited from the feature learning ability of DNN, intelligent fault diagnosis becomes more automatic and effective.

CNN, as one of the main types of DNN models, has been applied with great success to learn features from raw data and becomes the dominant approach for almost all recognition and detection tasks in image and speech analysis [31], [32]. However, very few investigations have been conducted on the application of CNN in feature learning and fault diagnosis for a planetary gearbox or a gearbox with combined gear-bearing-shaft faults. At the same time, most of the studies of DNN based feature learning only focus on one type of raw data. The study of the different performance of feature learning from various types of data is still few. In this paper, we applied the CNN to learn features from raw vibration data in time domain, raw frequency spectrum of the data and their combination, and diagnose the health conditions of gearboxes. Manual features and three common intelligent methods, including fully-connected neural network (FNN), support vector machine (SVM) and random forest (RF), are used as comparisons. In addition, the selection of several key parameters of CNN is discussed, and the detection results of CNN with various configurations are tested in the experiments.

The rest of the paper is organized as follows. In Section 2, the typical architecture of CNN and its training method are briefly introduced. Section 3 details the CNN based diagnosis method and the selection of parameters of CNN model. In Section 4, the PHM 2009 challenge data and a planetary gearbox test rig are used to evaluate the effectiveness of the proposed method. Finally, the conclusions are drawn in Section 5.

Section snippets

Theoretical background of convolutional neural network

CNN is a type of deep learning models inspired by the visual system structure [6], [7]. Since the early 2000s, it has been successfully applied in the detection, segmentation and recognition of regions or patterns in images and speeches [33], [34]. CNN is now the dominant approach for most image detection and speech recognition tasks. In this section, we describe CNN in more detail.

Convolutional neural network based fault diagnosis method

Based on CNN, a novel intelligent fault diagnosis method is proposed to learn features from frequency spectrum of vibration data directly and detect faults of gearboxes. Although most of the applications of CNN in image recognition choose a 2D segment as input [31], [34], and many researchers used the same way to diagnose mechanical faults [16], [25], this study uses a 1D segment from raw data as the input, cooperated with 1D filter banks in the convolutional layers in the model. In our

Experiments and discussions

In this section, two fault diagnosis cases, a helical gearbox with combined faults of 2009 PHM data challenge and a planetary gearbox test rig, are used to validate the effectiveness of the proposed method.

Conclusion and future work

In this paper, a CNN model is presented to learn features from frequency data directly and detect faults of gearboxes. The CNN provides the abilities of feature extraction, feature selection and classification and forms an end to end machine learning system, which is able to process raw data as the input and provide the diagnosis result as the output. The effectiveness of the proposed method is validated using the PHM 2009 challenge data and the vibration data of a planetary gearbox. Raw

Acknowledgments

This research is supported by the National Natural Science Foundation of China (Grant No. 51475324) and the National Natural Science Foundation of China and Civil Aviation Administration of China jointly funded project (Grant No. U1533103).

References (52)

  • F. Jia

    Deep neural networks: a promising tool for fault characteristic mining and intelligent diagnosis of rotating machinery with massive data

    Mech. Syst. Signal Process.

    (2016)
  • W. Sun

    A sparse auto-encoder-based deep neural network approach for induction motor faults classification

    Measurement

    (2016)
  • C. Lu

    Fault diagnosis of rotary machinery components using a stacked denoising autoencoder-based health state identification

    Signal Process.

    (2017)
  • X. Luo

    A parallel matrix factorization based recommender by alternating stochastic gradient decent

    Eng. Appl. Artif. Intell.

    (2012)
  • B. Samanta

    Gear fault detection using artificial neural networks and support vector machines with genetic algorithms

    Mech. Syst. Signal Process.

    (2004)
  • N. Saravanan et al.

    Fault diagnosis of spur bevel gear box using artificial neural network (ann), and proximal support vector machine (psvm)

    Appl. Soft Comput.

    (2010)
  • Y. Lei

    An intelligent fault diagnosis method using unsupervised feature learning towards mechanical big data

    IEEE Trans. Industr. Electron.

    (2016)
  • G.E. Hinton et al.

    Reducing the dimensionality of data with neural networks

    Science

    (2006)
  • Y. LeCun et al.

    Deep learning

    Nature

    (2015)
  • Y. Bengio

    Learning deep architectures for ai. Foundations and trends®

    Mach. Learn.

    (2009)
  • Y. Bengio, A. Courville, Deep learning of representations, in Handbook on neural information processing, 2013,...
  • Y. Bengio et al.

    Representation learning: a review and new perspectives

    IEEE Trans. Pattern Anal. Mach. Intell.

    (2013)
  • Z. Chen, et al., Machine fault classification using deep belief network, in: 2016 IEEE International Instrumentation...
  • P. Kane et al.

    Application of psychoacoustics for gear fault diagnosis using artificial neural network

    J. Low Freq. Noise Vib. Active Control

    (2016)
  • C. Li, et al., Fault diagnosis for rotating machinery using vibration measurement deep statistical feature learning,...
  • A. Sharma et al.

    Feature extraction and fault severity classification in ball bearings

    J. Vib. Control

    (2014)
  • Cited by (0)

    View full text