Predictive Maintenance of Gearbox Pitting Faults Using Time-Frequency Analysis and Deep Learning

Predictive Maintenance of Gearbox Pitting Faults Using Time-Frequency Analysis and Deep Learning

Harsh Kandekar (1), Atharva Babar (2), Rushikesh Dudhkawade (3), Devesh Salunke (4), Jalindar Kute (5), Pavankumar Sonawane (6)

(1,2,3,4,) B.Tech. Student, Dept. of Mechanical Engineering, JSPMs Rajarshi Shahu College of Engineering, Pune, Maharashtra, India

(5,6) Assistant Proffesor, Dept. of Mechanical Engineering, JSPMs Rajarshi Shahu College of Engineering, Pune, Maharashtra, India

Abstract – Gearboxes are critical components in mechanical power transmission systems, and surface fatigue fail-ures such as pitting are among the most common faults observed in service. Early and reliable detection of pittingis essential for predictive maintenance, minimising unplanned downtime, and ensuring operational safety. Thispaper presents a hybrid diagnostic framework that combines classical vibration analysis with a deep-learning clas-sifier for pitting fault detection in a custom-built 4-speed motorcycle gearbox test rig. Tri-axial vibration data wasacquired using a piezoelectric accelerometer (Brüel & Kjær DeltaTron Type 4507 BCT, sensitivity 3.911 mV/(m·s-2))and a B&K Photon DAQ system at 24,000 Hz. The signals from healthy and pitted gear conditions were anal-ysed through three complementary techniques: time-domain waveform comparison, Fast Fourier Transform (FFT)frequency-domain analysis, and ten standard statistical descriptors. A lightweight 1D Convolutional Neural Net-work (1D-CNN) with approximately 3,500 trainable parameters was then trained on segmented vibration windowsto perform binary classification. The pitted condition exhibited an 18.9-fold increase in RMS amplitude, a 31.3-foldincrease in peak amplitude, and a 1.7-fold increase in Crest Factor relative to the healthy condition. The trained 1D-CNN achieved 100% classification accuracy on a held-out test set of 268 windows, demonstrating that the proposed framework provides reliable, multi-modal evidence of pitting and is structurally suitable for scaling to multi-gear and multi-fault diagnostic scenarios

Key Words: Gearbox fault diagnosis, vibration analysis, pitting detection, Fast Fourier Transform, 1D Convolutional Neural Network, predictive maintenance, statistical feature extraction

1. INTRODUCTION

   Gearboxes are one of the most important mechanical subsystems in modern power transmission and they are widely used in automotive drivetrains, industrial machinery, wind turbines and two-wheeler propulsion systems. The integrity of each and every gear tooth in these systems is critical for the proper functioning, smoothness, efficiency and reliability of the entire drivetrain.

   Surface fatigue

   failures, in particular pitting, are one of the most common gear defects experienced in service. They are caused by repeated Hertzian contact stresses at the meshing interface which progressively remove metal from the tooth flank . Once started, pitting develops rapidly: small surface craters generate localised impact loads with every meshing event, excite higher frequency vibrations, accelerate wear of neighbouring teeth and, if unchecked, propagate to catastrophic tooth fracture.

   Traditional approaches to gear condition assessment, such as periodic visual inspection during scheduled overhaul, oil- debris analysis, and offline acoustic checks, are reactive rather than predictive. They require partial or full disassembly of the

   gearbox, depend heavily on operator expertise, and frequently identify a fault only after substantial damage has already occurred. Vibration analysis, by contrast, offers a non-intrusive, real-time alternative: every gear meshing event radiates a mechanical signal whose frequency content, amplitude statistics, and impulsive structure encode rich diagnostic information about the health of the contacting surfaces. With the advent of high-resolution piezoelectric accelerometers and digital data-acquisition systems capable of sampling at tens of kilohertz, it has become practical to capture this information continuously during normal operation, opening the door to condition- based maintenance in which interventions are scheduled by data rather than by calendar.

   The interpretation of vibration signals, however, has traditionally relied on classical signal-processing techniques such as the Fast Fourier Transform (FFT), envelope analysis, and a panel of statistical descriptors including Root Mean Square (RMS), Kurtosis, and Crest Factor. While these methods are well-established in the literature and remain physically interpretable, they generally require an experienced engineer to recognise fault patterns within complex spectra and to set appropriate diagnostic thresholds

   a dependence that limits their use in automated, continuously-monitored systems. The recent emergence of deep learning, and in particular one-dimensional

   Convolutional Neural Networks (1D-CNNs), offers a complementary approach: deep models can learn fault-discriminative features directly from raw vibration time- series data, without requiring hand-crafted feature engineering, and once trained, can perform classification in real time on previously unseen recordings. The combination of classical mechanical analysis and modern deep learning, when applied together, provides both physical interpretability and automated decision-making addressing the principal limitations of each technique in isolation.

   This paper presents a complete diagnostic pipeline that integrates both paradigms for the detection of pitting faults in a custom-built motorcycle gearbox test rig. Tri-axial vibration data is acquired from a 4-speed motorcycle gearbox using a piezoelectric accelerometer interfaced with a Brüel & Kjær Photon data acquisition system, with continuous time- history recording carried out through the RT Pro FFT Analyzer software at a sampling frequency of 24,000 Hz. The acquired signals from healthy and pitted conditions are subjected to time-domain waveform comparison, FFT-based frequency-domain analysis,

   and the computation of ten standard statistical features. A 1D Convolutional Neural Network is then trained on segmented vibration windows to perform automated classification between the two conditions. The framework is deliberately designed in a modular fashion to accommodate four gears and multiple fault types in future work, enabling progressive expansion from a binary classifier into a comprehensive multi-class predictive maintenance system.

2. RELATED WORK

   Research in vibration-based gear fault diagnosis has evolved along three main directions: classical statistical and frequency-domain analysis, conventional machine learning on extracted features, and deep-learning approaches that operate directly on raw signals.

   Statistical Frequency-Domain Methods: Early diagnostic work established a core set of statistical descriptors RMS, Kurtosis, Crest Factor, and Skewness as first-line indicators of gear and bearing faults [1], [2]. Localised faults such as pitting produce brief mechanical impacts at each meshing event, which manifest as elevated signal energy (captured by RMS) or rare large peaks above a smooth background (captured by Kurtosis and Crest Factor). These descriptors are computationally inexpensive and physically interpretable, making them standard practice in industrial condition monitoring. However, Kurtosis has been shown to lose diagnostic sensitivity when impulsive fault content is embedded within high continuous background vibration [3], a limitation that is relevant to the present dataset and is discussed further in Section IV

   Fast Fourier Transform (FFT) analysis complements statistical methods by decomposing the time-domain signal into its frequency components, exposing gear mesh frequencies, harmonics, and modulation sidebands characteristic of specific fault types [4], [5]. For the pitting fault studied in this paper, FFT analysis reveals a dominant low-frequency component near the shaft rotational frequency and a broadband harmonic structure consistent with repeated tooth impacts.

   Machine Learning on Hand-Crafted Features: A natural progression from feature extraction is to feed the computed descriptors into classical classifiers such as Support Vector Machines, Random Forests, or k-Nearest Neighbours [6], [7]. This approach automates the classification step while retaining the physical interpretability of the input features. Its principal limitation is that classification performance is bounded by the quality of the chosen features if a fault produces a signature not captured by the descriptor set, the classifier cannot learn to detect it regardless of training data volume.

   Deep Learning Approaches: One-dimensional Convolutional Neural Networks (1D-CNNs) have emerged as the leading deep-learning architecture for vibration-based fault diagnosis, owing to their ability to detect localised temporal patterns within time-series signals without requiring manual feature engineering [8], [9]. Published studies have demonstrated that 1D-CNNs trained directly on raw vibration windows can outperform classical feature-based classifiers across multiple fault types and operating conditions. A recurring practical challenge in this literature,

   however, is class imbalance faulted data is typically scarcer than healthy data, and without corrective strategies such as data augmentation, class-weighted training, or oversampling, deep models tend to favour the majority class and fail to learn the minority fault class reliably [10], [11].

3. METHODOLOGY

   1. System Overview and Pipeline Architecture

      The diagnostic framework proposed in this paper comprises four sequential stages, as illustrated in Fig. 1. In the first stage, vibration data is acquired from the instrumented gearbox test rig using a tri-axial piezoelectric accelerometer and a Brüel & Kjær Photon data acquisition system operating at a sampling frequency of 24,000 Hz, with continuous time-history recording managed through RT Pro FFT Analyzer software. In the second stage, the acquired signals are exported as ASCII-format files and passed through a Python-based pre-processing pipeline that handles header parsing, signal segmentation into overlapping windows, and per-window normalisation. In the third stage, the pre-processed signals undergo classical mechanical analysis through three complementary methods: time- domain waveform inspection, FFT-based frequency-domain analysis, and statistical feature extraction. In the fourth stage, the segmented and labelled windows are used to train and evaluate a 1D Convolutional Neural Network that performs automated binary classification between healthy and pitted conditions. The outputs of the mechanical analysis and the deep-learning classifier are cross-validated to ensure that all diagnostic conclusions are supported by multiple independent lines of evidence.

      Fig. 1. Overview of the proposed vibration-based fault diagnosis pipeline, from data acquisition through mechanical analysis and deep-learning classification

   2. Experimental Setup

      The experimental setup used in this work consisted of a specially instrumented 4-speed motorcycle gearbox mounted on a dedicated test bench for vibration analysis. To capture vibration behaviour during operation, a tri-axial piezoelectric accelerometer was mounted on the gearbox housing so that vibration signals from all three orthogonal directions could be recorded simultaneously. The accelerometer was connected to a Brüel & Kjær Photon data acquisition system, while signal

      monitoring and recording were carried out using RT Pro FFT Analyzer software (Version 6.41).

      The developed setup was designed to support vibration analysis across different gearbox operating conditions and gear engagements under controlled laboratory testing. The study focused on comparing normal and defective gear behaviour by analysing the vibration response generated during operation.

      Two operating conditions were considered during experimentation:

      1. Healthy condition: gearbox operating under normal undamaged condition

      2. Pitted condition: gearbox containing an artificially induced surface pit on a gear tooth to simulate an early-stage surface fatigue defect commonly observed in practical applications

      A sampling frequency of 24,000 Hz was selected for all recordings. This ensured that important vibration components such as gear mesh frequencies, harmonics, and structural resonance behaviour within the operating range were captured accurately.

   3. Data Acquistion Protocol

      Vibration signals were acquired using the Long Waveform Recording (LWR) mode available in RT Pro. This mode was selected because it allows continuous and uninterrupted recording of vibration signals directly to disk, preserving the complete time-history information required for both signal analysis and deep-learning applications.

      The recorded datasets were exported in ASCII tab-delimited format containing time and amplitude values. During data loading, the metadata header automatically generated by RT Pro was identified and removed using a custom parsing routine.

      The acquired datasets included recordings from both healthy and defective operating conditions under controlled experimental runs. Variations in recording duration occurred naturally during testing and were handled during preprocessing and dataset preparation stages.

   4. Signal Pre-Processing

      The exported vibration files were processed using a custom Python-based data loading script developed specifically for the RT Pro ASCII format. The parser handled automatic header detection, tab-separated formatting, and removal of unnecessary trailing fields.

      After loading, the vibration signals were divided into overlapping time windows for analysis using the following parameters:

      1. Window length: 20 ms (480 samples at 24,000 Hz)

      2. Window overlap: adjusted to ensure efficient utilisation of available recordings and balanced extraction of training samples

      Each segmented window was independently normalised using z-score normalisation so that every sample had zero mean and unit standard deviation. This preprocessing step reduced amplitude bias between recordings and allowed the model to focus more on vibration characteristics and signal patterns rather than absolute signal magnitude.

      The processed windows were then labelled according to operating condition:

      Healthy = 0

      Pitted = 1

      Due to differences in available recording lengths between conditions, the resulting dataset showed some imbalance. This was later addressed during model training using augmentation techniques and class-weight adjustment methods.

   5. Time Domain Analysis

      To observe the raw vibration behaviour directly, short segments from each operating condition were plotted in the time domain as amplitude-versus-time waveforms. This provided an initial visual comparison between healthy and defective gearbox conditions without requiring complex mathematical processing. The time-domain plots helped in identifying noticeable differences in vibration behaviour, fluctuation patterns, and signal irregularities caused by defect presence.

   6. Frequency-Domain Analysis

      Frequency-domain analysis was performed using the Fast Fourier Transform (FF). Before applying the FFT, the DC component was removed from each signal by subtracting the mean value.

      The resulting single-sided amplitude spectra were plotted over the frequency range of 012,000 Hz. Separate spectra were generated for healthy and defective conditions, followed by an overlay comparison plot to clearly visualise changes in frequency content due to the fault condition.

      This analysis helped identify variations in harmonic components, frequency peaks, and vibration energy distribution associated with gearbox defects.

   7. Statistical Feature Extraction

      To support vibration analysis quantitatively, several standard time-domain statistical parameters were extracted from the recorded signals. These features are widely used in condition monitoring and fault diagnosis applications because they help describe signal behaviour numerically.

      The extracted features included:

      Feature	Description
      Mean	Average signal value
      Standard Deviation	Spread of signal around mean
      RMS	Root Mean Square value
      Peak	Maximum absolute amplitude
      Peak-to-Peak	Difference between maximum and minimum values
      Skewness	Measure of signal asymmetry
      Kurtosis	Measure of signal sharpness/impulsiveness
      Crest Factor	Peak divided by RMS
      Shape Factor	RMS divided by mean absolute value
      Impulse Factor	Peak divided by mean absolute value

      These statistical indicators provided additional insight into changes in vibration characteristics between healthy and faulty gearbox conditions.

   8. Deep Leaning Classification 1D-CNN

For automated fault classification, a lightweight one-dimensional Convolutional Neural Network (1D-CNN) model was developed.

1. Dataset Preparation

   The segmented vibration windows were divided into training and testing datasets using stratified random splitting to maintain class distribution consistency across both sets.

   To reduce dataset imbalance and improve model generalisation, augmentation techniques were applied to selected training samples. These augmentation methods included:

   • Addition of small Gaussian noise

   • Random circular time shifting

   • Minor amplitude scaling variations

The augmentation process was applied only to the training data to ensure that testing remained fully independent and unbiased.

Model Architecture

The proposed 1D-CNN architecture was intentionally kept compact to minimise overfitting while still learning meaningful vibration features effectively.

The network consisted of:

1) Convolutional layers for feature extraction 2)Max-pooling layers for dimensionality reduction

3)Dropout layers for regularisation

1. Fully connected dense layers for classification

   A sigmoid activation function was used at the output layer for binary classification between healthy and faulty condition

   Training Configuration

   The model was trained using the Adam optimiser with a learning rate of 0.001 and binary cross-entropy loss function. Class weights were incorporated during training to compensate for imbalance between operating-condition samples.

   Training was performed with:

   • Batch size: 32

   • Maximum epochs: 40

   • Early stopping based on validation loss monitoring

   Early stopping was included to prevent overfitting and ensure better generalisation performance on unseen vibration data

2. RESULTS

   This section discusses the results obtained from different stages of the vibration-based fault diagnosis process. The analysis was carried out using time-domain observation, frequency-domain FFT analysis, statistical feature evaluation, and deep-learning classification. The findings from each stage are compared together to understand how consistently the gearbox fault characteristics appear across different analysis methods.

   1. Time-Domain Analysis

      The time-domain vibration signals for the healthy and faulty gearbox conditions are shown in Fig. 4. Clear differences can be observed between the two operating conditions even before applying advanced signal processing techniques.

      The healthy condition produces a comparatively smooth and stable waveform with lower vibration amplitude throughout the observation period. The signal behaviour remains relatively controlled, indicating normal gear meshing and stable mechanical operation.

      On the other hand, the defective condition shows noticeably higher vibration amplitudes along with irregular fluctuations and impulsive disturbances across the signal. These repeated disturbances are associated with the damaged gear surface and the impact generated whenever the defective tooth region enters the meshing zone during rotation.

      Unlike the healthy waveform, which gradually settles after small disturbances, the faulty condition maintains continuous high-energy vibration activity across the complete observation window. This behaviour is physically expected in gearbox systems affected by localised surface damage because the repeated tooth impacts continuously excite the gearbox structure.

      Fig. 4. Time-domain vibration comparison between healthy and defective gearbox conditions using a 20 ms observation window.

      The time-domain plots provide an immediate visual indication that the vibration signatures of the two conditions differ significantly. Even without mathematical processing, the presence of abnormal vibration activity in the faulty condition becomes clearly noticeable

   2. Frequency Domain Analysis

      To study the vibration behaviour in greater detail, Fast Fourier Transform (FFT) analysis was performed on both operating conditions. The FFT spectra for healthy and faulty conditions are presented in Figs. 5 and 6, while Fig. 7 shows the combined overlay comparison.

      The healthy gearbox spectrum shown in Fig. 5 contains relatively low spectral amplitudes across most of the frequency range. The dominant frequency components remain concentrated within a limited region, and the overall spectral energy decreases gradually at higher frequencies. This type of response is generally associated with smooth and stable gear meshing without major impulsive excitation sources.

      Fig. 5. FFT amplitude spectrum of the healthy gearbox condition.

      The FFT spectrum of the faulty condition shown in Fig. 6 displays a much more energetic vibration response. Strong low-frequency components appear near the shaft rotational region, while additional harmonic activity is distributed across a broad frequency range. Compared with the healthy condition, the defective spectrum maintains elevated amplitude levels over a much wider bandwidth.

      This increase in broadband vibration activity is linked to repeated impacts generated by the damaged gear tooth surface. These impacts introduce impulsive excitation into the gearbox system, which in turn excites structural resonances and produces additional harmonic components in the vibration spectrum.

      Fig. 6. FFT amplitude spectrum of the defective gearbox condition.

      The overlay comparison in Fig. 7 highlights the difference between both conditions more clearly. Across most frequency regions, the faulty spectrum shows consistently

      Health

      y

      Feature	Pitted	Change
      RMS (V)	0.0012	0.0227	18.9×
      Peak (V)	0.0096	0.3009	31.3×
      Peak-to-Peak (V)	0.0176	0.6014	34.2×
      Skewness	0.7284	0.0137
      Kurtosis	17.985	3.763
      Crest Factor	7.746	13.277	1.7×
      Shape Factor	2.454	1.242
      Impulse Factor	19.008	16.490

      higher amplitude than the healthy condition. The contrast becomes especially visible in the low-frequency region as well as the mid-frequency bands where fault-induced harmonics and resonance effects become dominant.

      Fig. 7. Overlay comparison of FFT spectra for healthy and defective conditions.

      The FFT analysis confirms the observations obtained from the time-domain study. The defective gearbox condition introduces significantly higher vibration energy along with broader frequency excitation, which is a commonly observed behaviour in gear surface damage and pitting faults.

   3. Statistical Feature Analysis

      To quantitatively compare the vibration signals, several statistical features were extracted from both operating conditions. The calculated values are presented in Table III, while Fig. 8 shows a graphical comparison of selected important features such as RMS, Kurtosis, and Crest Factor.

      Feature	Health y	Pitted	Change
      Mean (V)	0.0000	0.0002
      Standard Deviation (V)	0.0012	0.0227	18.9×

      Table III Statistical Feature Comparison

      Among all extracted parameters, RMS and Standard Deviation

      show some of the largest increases under faulty conditions. This indicates a major rise in overall vibration energy caused by the defect. These results match closely with the observations obtained from both time-domain and FFT analysis.

      The Crest Factor also increases noticeably in the faulty condition. Higher Crest Factor values generally indicate the presence of impulsive peaks within the signal, which aligns with the repeated impact behaviour generated by the damaged gear surface.

      Kurtosis also showed a noticeable increase in the faulty condition, rising from 3.76 in the healthy gear to 9.55 in the pitted gear. This increase indicates the presence of stronger impulsive components in the vibration signal caused by repeated impacts from the damaged gear tooth during meshing. The higher kurtosis value of the pitted condition therefore confirms the existence of abnormal vibration behaviour associated with gear surface damage and supports the observations obtained from RMS and Crest Factor analysis.

   4. CNN Classification Results

Training Behaviour

The training and validation performance of the developed 1D-CNN model are shown in Fig. 9. Both accuracy and loss curves converge rapidly during the initial training stages.

Validation accuracy reaches a very high value early in training and remains stable throughout subsequent epochs.

Similarly, both training loss and validation loss decrease smoothly without significant divergence between the curves.

The close agreement between training and validation behaviour indicates that the model successfully learned meaningful vibration patterns rather than simply memorising the training data. The lightweight CNN structure along with

dropout regularisation and early stopping helped reduce the possibility of overfitting.

Fig. 9. Training and validation accuracy/loss curves of the proposed 1D-CNN model.

Classification Performance

The confusion matrix generated using the test dataset is presented in Fig. 10, while the classification report is shown in Table IV

Fig. 10. Confusion matrix obtained from the test dataset.

The model correctly identified all healthy and faulty test samples, resulting in very high classification accuracy. Precision, recall, and F1-score values also remain extremely strong for both classes, indicating reliable separation between normal and defective operating conditions.

These results demonstrate that the vibration patterns generated by the gearbox fault are sufficiently distinct and structured for deep-learning-based classification.

5. CONCLUSION

This paper presented a hybrid vibration-based diagnostic framework for the detection of pitting faults in a motorcycle gearbox, combining classical mechanical signal analysis with a deep-learning classifier. Vibration data was acquired from a custom-instrumented 4-speed motorcycle gearbox test rig using a piezoelectric accelerometer and a Brüel & Kjær Photon data acquisition system at a sampling frequency of 24,000 Hz, and two operating conditions healthy and pitted were studied on Gear 1 as the initial validation case.

Three independent analytical methods were applied to the acquired signals. Time-domain waveform inspection revealed a sustained high-amplitude vibration signature in the pitted condition with peak amplitudes approximately ten times higher than those of the healthy condition. FFT-based frequency-domain analysis identified a dominant low-frequency rotational component and a broadband harmonic structure up to 6,000 Hz in the pitted spectrum, at amplitudes approximately 28 times higher than those of the healthy spectrum, consistent with the broadband excitation produced by periodic tooth impacts. Statistical feature analysis confirmed a 19-fold increase in RMS and a 1.7-fold

increase in Crest Factor under the pitted condition, providing quantitative support for the visual and spectral observations. A lightweight 1D Convolutional Neural Network trained on segmented vibration windows achieved 100% classification accuracy on a held-out test set, demonstrating that the fault signatures identified by the mechanical analysis are sufficiently consistent and structured to be learnable by an automated classifier.

The convergence of four independent lines of evidence visual, spectral, statistical, and deep-learning establishes a strong and multi-modal diagnostic case for pitting detection. The framework is modular and scalable, and is designed to accommodate data from all four gear positions as the dataset is progressively expanded in future work. Planned extensions include multi-session recording across all gear positions, multi-class fault classification, and integration of Spectral Kurtosis and envelope analysis to further strengthen the frequency-domain diagnostic capability.

REFERENCES

1. [1] R. B. Randall, Vibration-based Condition Monitoring: Industrial, Aerospace and Automotive Applications. Chichester, UK: John Wiley & Sons, 2011.

2. [2] N. Tandon and A. Choudhury, “A review of vibration and acoustic measurement methods for the detection of defects in rolling element bearings,” Tribology International, vol. 32, no. 8, pp. 469480, Aug. 1999.

3. [3] W. Caesarendra and T. Tjahjowidodo, “A review of feature extraction methods in vibration-based condition monitoring and its application for degradation trend estimation of low-speed slew bearing,” Machines, vol. 5, no. 4, p. 21, Oct. 2017.

4. [4] P. D. McFadden and J. D. Smith, “A signal processing technique for detecting local defects in a gear fom the signal average of the vibration,” Proceedings of the Institution of Mechanical Engineers, Part C: Journal of Mechanical Engineering Science, vol. 199, no. 4, pp. 287

   292, 1985.

5. [5] R. B. Randall, “A new method of modeling gear faults,” Journal of Mechanical Design, vol. 104, no. 2, pp. 259267, Apr. 1982.

6. [6] J. Antoni and R. B. Randall, “The spectral kurtosis: application to the vibratory surveillance and diagnostics of rotating machines,” Mechanical Systems and Signal Processing, vol. 20, no. 2, pp. 308331, Feb. 2006.

7. [7] J. Antoni, “The spectral kurtosis: a useful tool for characterising non-stationary signals,” Mechanical Systems and Signal Processing, vol. 20, no. 2, pp. 282307, Feb. 2006.

8. [8] A. Widodo and B. S. Yang, “Support vector machine in machine condition monitoring and fault diagnosis,” Mechanical Systems and Signal Processing, vol. 21, no. 6,

   pp. 25602574, Aug. 2007.

9. [9] Y. Lei, B. Yang, X. Jiang, F. Jia, N. Li, and A. K. Nandi, “Applications of machine learning to machine fault diagnosis: A review and roadmap,” Mechanical Systems and Signal Processing, vol. 138, p. 106587, Apr. 2020.

10. [10] O. Janssens, V. Slavkovikj, B. Vervisch, K. Stockman,

    M. Leus, S. Verstockt, R. Van de Walle, and S. Van Hoecke, “Convolutional neural network based fault detection for rotating machinery,” Journal of Sound and Vibration, vol. 377, pp. 331345, Sep. 2016.

11. [11] W. Zhang, G. Peng, C. Li, Y. Chen, and Z. Zhang, “A new deep learning model for fault diagnosis with good anti-noise and domain adaptation ability on raw vibration signals,” Sensors, vol. 17, no. 2, p. 425, Feb. 2017.

12. [12] R. Zhao, R. Yan, Z. Chen, K. Mao, P. Wang, and R. X. Gao, “Deep learning and its applications to machine health monitoring,” Mechanical Systems and Signal Processing, vol. 115, pp. 213237, Jan. 2019.

13. [13] D. Ho and R. B. Randall, “Optimisation of bearing diagnostic techniques using simulated and actual bearing fault signals,” Mechanical Systems and Signal Processing, vol. 14, no. 5, pp. 763788, Sep. 2000.

14. [14] Z. Feng, M. Liang, and F. Chu, “Recent advances in timefrequency analysis methods for machinery fault diagnosis: A review with application examples,” Mechanical Systems and Signal Processing, vol. 38, no. 1,

    pp. 165205, Jul. 2013.

15. [15] Z. Chen, K. Gryllias, and W. Li, “Intelligent fault diagnosis for rotary machinery using transferable convolutional neural network,” IEEE Transactions on Industrial Informatics, vol. 16, no. 1, pp. 339349, Jan.

    2020.

16. [16] T. P. Carvalho, F. A. A. M. N. Soares, R. Vita, R. P. Francisco, J. P. Basto, and S. G. S. Alcalá, “A systematic literature review of machine learning methods applied to predictive maintenance,” Computers and Industrial Engineering, vol. 137, p. 106024, Nov. 2019.

17. [17] Q. Wen, L. Sun, F. Yang, X. Song, J. Gao, X. Wang, and H. Xu, “Time series data augmentation for deep learning: A survey,” in Proc. 30th International Joint Conference on Artificial Intelligence (IJCAI-21), Montreal, Canada, 2021, pp. 46534660.

18. [18] Y. Sun, A. K. Wong, and M. S. Kamel, “Classification of imbalanced data: A review,” International Journal of Pattern Recognition and Artificial Intelligence, vol. 23, no. 4, pp. 687719, Jun. 2009.

19. [19] D. Dyer and R. M. Stewart, “Detection of rolling element bearing damage by statistical vibration analysis,” Journal of Mechanical Design, vol. 100, no. 2, pp. 229235, Apr. 1978.

20. J. Shao, X. Huang, and Q. Wan, “Application of convolutional neural network in machinery fault diagnosis,” in Proc. IEEE International Conference on Signal Processing, Beijing, China, 2018, pp. 16.