A Comprehensive Approach for Gearbox Fault

Detection and Diagnosis Using Sequential Neural

Networks
1st Parvathy Sobha 2nd Midhun Xavier 3rd Praneeth Chandran
Department of Engineering Sciences and Department of Computer Science, Department of Civil, Environmental and
Mathematics Electrical and Space Engineering Natural Resource Engineering
Luleå tekniska universitet Luleå tekniska universitet Luleå tekniska universitet
Luleå, Sweden Luleå, Sweden Luleå,Sweden
[email protected] [email protected] [email protected]

Abstract—Gearbox faults can lead to significant damage and Existing methods for gearbox fault detection include visual
downtime in industrial machinery, resulting in substantial losses inspection, acoustic analysis [4], and vibration analysis [3].
for manufacturers. Detection of faults in gears in the incipient Visual inspection involves physically inspecting the gearbox
state is essential to ensure safe and reliable operation of for signs of wear, damage, or other faults. This method requires
industrial machineries. In recent years, there has been an significant expertise and can only detect faults that are visible
increasing interest in using machine learning algorithms to to the naked eye. However, recent studies have proposed
automate gearbox fault detection. This paper proposes a machine combining visual inspection with machine learning techniques
learning approach for identifying different categories of faults in
to improve the accuracy and efficiency of fault detection [5].
a gearbox based on vibration signals. The approach involves
preprocessing the signals, extracting features, and training a
Acoustic analysis involves measuring and analyzing the
machine learning model. The proposed method was evaluated on sound generated by the gearbox to identify potential faults.
a dataset of vibration signals collected from a two-stage gearbox This method is limited in its ability to distinguish between
under different operational conditions. The research is focused different types of faults and is susceptible to environmental
on developing a sequential neural network-based method for noise. Nevertheless, studies have shown that acoustic analysis
gearbox fault detection. The results showed that the developed can be effective in detecting certain types of gearbox faults [6].
method achieved high training and validation accuracies and Vibration analysis is currently the most widely used method
relatively low training and validation losses, indicating the for gearbox fault detection. Vibration monitoring is a technique
model's ability to accurately detect and classify faults in used to detect faults in gears by analyzing the vibrations they
gearboxes. The testing accuracies were also high, demonstrating produce. Healthy and faulty gears generate different vibration
the model's ability to generalize well to new data. The practical signals, which can indicate the presence of faults such as gear
implications of the research are significant for improving the tooth wear, gear tooth cracks, and pitting. Various techniques
reliability and maintenance of gearboxes in various industrial such as time domain, frequency domain, and time-frequency
applications. The developed method has the potential to reduce domain techniques are used to analyze these vibration signals.
downtime, maintenance costs, and improve safety and efficiency. The frequency domain technique involves using the Fast
Fourier Transform (FFT) of the time domain signal to evaluate
Keywords—gearbox fault detection, vibration analysis,
sequential neural network, maintenance the condition based on the signal's frequency content.
However, vibration signals from gears are considered non-
stationary and non-periodic, making it challenging to detect
I. INTRODUCTION gear faults using conventional FFT analysis [7]. To overcome
The problem of gearbox fault detection is a critical issue in this limitation, sophisticated signal processing method, such as
industrial machinery maintenance. Gearboxes are essential wavelet analysis, can be used for feature extraction from the
components in many machines, and faults in gearboxes can noisy gear signal.Vibration analysis requires advanced signal
lead to significant damage, downtime, and financial losses for processing and machine learning techniques for accurate and
manufacturers [1]. Identifying gearbox faults is crucial to reliable results [8].
prevent catastrophic failures, minimize downtime, and reduce
maintenance costs. However, gearboxes are subjected to a The objective of this paper is to propose a method for
range of wear and tear [2]. If left undetected, faults in analyzing vibration signals to monitor the health of
gearboxes can escalate rapidly, leading to complete machine industrial systems by analyzing the gear box. The proposed
breakdown, loss of productivity, and increased maintenance methodology involves data collection, feature extraction,
costs. Early detection and repair of gearbox faults can prevent and training a machine learning model to predict the state of
these issues and ensure optimal machine performance [3]. the gear condition. The effectiveness of the proposed
approach is evaluated on a dataset of vibration signals

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE

 collected from a two-stage gearbox under different ability to distinguish between the different fault categories. A
operational conditions. The paper presents the results of the higher ranking score indicates that the feature is more
study, demonstrating the accuracy and reliability of the important in distinguishing between fault categories. Different
proposed approach in identifying different fault categories, feature ranking methods such as ANOVA, ReliefF, Extra
including normal working condition, surface wear, chipped, Trees Classifier, and Correlation-based feature selection etc.,
cracked, and tooth missing. This paper provides a useful
can be used to rank the features. Feature selection, on the other
framework for gearbox fault detection that can improve the
reliability and safety of industrial machinery. hand, involves the selection of the top-ranked features from
the feature ranking step. The selected features should have
The paper is structured as follows, in Section II, the high discriminatory power and low redundancy. The goal of
framework developed for this study is described. Section III feature selection is to reduce the dimensionality of the feature
discusses the case study and the application of the space, thereby improving the performance and interpretability
framework to the case. In Section IV, the modeling results
of the fault detection model.
and discussion are presented. Finally, Section V concludes
of the paper. B. Machine Learning Model Development
This step involves developing a machine learning model to
II. METHOD detect faults based on the extracted features. The following are
Gearbox fault detection using machine learning involves the the actions involved in this step:
use of algorithms to automatically identify the presence of 1) Model selection: involves selecting the appropriate
faults in gearbox systems from vibration signals. The general machine learning algorithm for the problem at hand. There are
framework for gearbox fault detection using machine learning several types of machine learning models that can be used for
involves the two key steps which are briefed explained in this gearbox fault detection, including, supervised learning models
section. - which learn from labeled data and can be used to classify the
data into different fault categories, Decision Trees, Random
A. Data Collection and Feature Extraction:
Forests, Support Vector Machines (SVMs) [13]–[15], and
This step involves collecting the vibration signals from the Neural Networks [16] are examples; unsupervised learning
gearbox system using sensors and processing the collected models - these models learn from unlabeled data and can be
data for analysis. The following are the actions involved in
used to identify patterns and anomalies in the data, K-Means
this step:
Clustering, DBSCAN clustering [17]–[19], Principal
1) Data collection and preprocessing: data collection
Component Analysis (PCA) [20], and Autoencoders [21] are
involves the collection of the vibration signals from the
examples; semi-supervised learning models - these models
gearbox system using sensors and are most commonly
learn from a combination of labeled and unlabeled data and
mounted on the gearbox housing or on the shaft.
can be used when only a small amount of labeled data is
Preprocessing involves cleaning and preparation of the
available, self-training [22] and co-training models [23] are
collected data for analysis. The data may be filtered,
examples; transfer learning models - these models learn from a
normalized, and segmented to remove noise and artifacts.
pre-trained model on a related task and can be used when there
2) Feature extraction: involves the extraction of relevant
is a limited amount of labeled data available for the specific
features from the vibration signals. The extracted features
application, fine-tuning [24] and domain adaptation [25] are
should capture the underlying fault patterns in the data. There
examples.
are several types of features that can be extracted from
The choice of model for gearbox fault detection depends
vibration signals for gearbox fault detection such as, time on several factors, including the amount and quality of labeled
domain features [9]- describing the behavior of the vibration data, the complexity of the underlying fault patterns, and the
signal over time, mean, standard deviation, kurtosis, and desired performance metrics. A common approach to model
skewness are examples; frequency domain features [10]- selection is to evaluate multiple models using a validation set
describing the behavior of the vibration signal in the frequency and select the one with the best performance. Alternatively, an
domain, spectral density, power spectrum, and Fourier ensemble of multiple models can be used to improve the
coefficients are examples; time-frequency domain features accuracy and robustness of the fault detection system.
[11]- describing the behavior of the vibration signal in both the 2) Model training: involves training the selected machine
time and frequency domains wavelet transform, short-time learning algorithm using the labeled data. The labeled data
Fourier transform, and Gabor transform are examples and consists of the extracted features and the corresponding fault
finally statistical features [12]- describing the statistical labels.
properties of the vibration signal include entropy, correlation, 3) Model validation: involves evaluating the performance
and covariance are examples. of the trained model on a validation set. The purpose of model
3) Feature ranking and feature selection: involves the validation is to estimate the performance of the model on an
identification and selection of the most relevant features for independent dataset that was not used for training the model.
the fault detection model. Feature ranking involves the 4) Model testing: involves evaluating the performance of
computation of a ranking score for each feature, based on its the trained model on new, unseen data. The purpose of model
testing is to assess the generalization capability of the model, B. Machine Learning Model Development
i.e., how well it can predict the outcome for new data that it Sequential neural networks are employed for this study as they
has not seen before. can learn from large amounts of data and have been shown to
achieve state-of-the-art results in various applications. Also
III. CASE STUDY
they can effectively capture complex patterns and
In this case study a machine learning-based model is dependencies in data. This makes them a suitable choice for
developed for analyzing vibration signals to monitor the health the task of gearbox fault detection, where large amounts of
of gear box in the industrial setup based on the method defined vibration signal data are available for training.
in section II. Gear boxes are prone to defects and failures, The sequential neural network model has four hidden
which may lead to various undesirable consequences. layers that are stacked one after the other. The layers are fully
Therefore, real-time monitoring techniques are essential to connected or dense layers. The input layer has 512 units with a
mitigate the risk of failures, reduce machinery downtime, and rectified linear unit (ReLU) activation function and uses L2
increase productivity by performing efficient maintenance regularization with a strength of 0.001. The dropout layer with
strategies. Vibration signals are widely used for fault detection a rate of 0.2 is used to prevent overfitting by randomly setting
in industrial/manufacturing machinery, making them the focus 20% of the input units to 0 at each update during training time.
of this case study. The following layers are also dense layers with ReLU
A. Data Collection and Feature Extraction: activation functions and have 256, 128, and 64 units,
respectively. All of these layers use L2 regularization with a
The dataset [26] used in this study consists of vibration strength of 0.001 and dropout with a rate of 0.2. The output
signals collected from a two-stage gearbox under different layer has 5 units (one for each gear state) with a softmax
operational conditions captured using a portable data activation function to produce a probability distribution over
acquisition system of National Instruments (NI, 356A01). The the classes. The hyper parameter learning rate was optimized
vibration data used for the present study focuses on the using the Adam optimizer, and an early stopping criterion was
planetary gearbox. Each vibration signal was recorded in three used to prevent overfitting or underfitting. The model was
directions (x, y, z), for a period of five minutes, and with a evaluated on the loss function
sampling frequency of 10 kHz. The first operating condition ‘sparse_categorical_crossentropy’ since the labels used in the
included a motor speed of 1500 rpm with a load of 10Nm. The study were integers.
second operating condition in which the data were recorded The study utilized a data set consisting of 50000 x 60
had a motor speed of 2700 rpm with a load of 25Nm. The samples, where each gear state including faults had a size of
features used in the study were extracted from time domain, 10000 x 60 samples for training the model. The data set is
frequency domain, wavelet analysis and time-frequency normalized prior to being fed into the model. The resulting
domain. The features used for this study includes peak-to- data set was then split into training and testing sets in an 80:20
peak, kurtosis, crest factor, skewness, standard deviation, ratio. The model was evaluated using 5-fold cross-validation
variance, energy spectrum, velocity spectrum, Fast Fourier on the training set, which contained 40000 samples. Finally,
Transform (FFT), Scale Averaged Wavelet Power (SAWP), 20 percent of the data set was reserved for the final testing of
Laplace Wavelet Kurtosis, M6A, M8A. The features were the model. A learning rate schedule function lr_schedule() is
extracted for the vibrations signal collected across the three used to adjust the learning rate during training. This function
directions individually. Five classes of gear state were will keep the learning rate constant for the first 10 epochs, and
monitored for this study: - Normal/healthy, surface wear, then decrease it exponentially with a decay rate of 0.1. The
cracked tooth, chipped and tooth missing. Each of the state code also sets up an early stopping criterion using the
had 10,000 signals for which the above mentioned features EarlyStopping callback from Keras. This will monitor the
were estimated. validation loss during training and stop training if the
This study employes extra tree classifier to identify the validation loss has not improved in the past 5 epochs. This
most important features. Here, a total of 78 features were also helps prevent overfitting and ensures that the model does
extracted from the vibration signals, with 26 features extracted not continue training if it has already reached its optimal
from each of the x, y, and z directions. These features were performance on the validation set. When using K-fold cross
then fed into the extra tree classifier, which analyzed the validation, this code will be applied within each fold of the
features and ranked them according to their importance. After cross validation loop. Each fold will use a different subset of
the analysis, the top 60 features were selected based on their the data for training and validation, but will use the same
feature ranking score and were used as input for the modeling initial learning rate, optimizer, learning rate schedule, and
stage. By selecting the top features, the model can focus on the early stopping criteria.
most important information in the dataset, reducing the impact
of irrelevant or redundant features, and improving the overall IV. RESULTS AND DISCUSSION
performance of the model. The developed model based on sequential neural network and
feature selection achieved promising results for the task of
gearbox fault detection. The model was trained and tested on
two different RPM (revolutions per minute) values, 1500 and
2700 RPM. Fig. 1. represents the results of a 5-fold cross-
validation study conducted on these two operating conditions
(RPMs) of Gear.

Fig. 1. Five-fold cross-validation results for 2700 RPM (left) and 1500 RPM (right)

For the 2700 RPM operating condition, the training accuracy

ranged from 93% to 94% from fold 1 to fold 5. The training
loss ranged from 0.3348 to 0.3416. The validation accuracy
ranged from 90% to 93%. The validation loss ranged from
0.2969 to 0.365 and the testing accuracy ranged from 90% to
91%. The testing loss ranged from 0.3536 to 0.3647. For the
1500 RPM operating condition, the training accuracy ranged
from 95% to 96% from fold 1 to fold 5. The training loss
ranged from 0.3631 to 0.3861 and the validation accuracy
ranged from 91% to 96%. The validation loss ranged from
0.2967 to 0.4289 and the testing accuracy ranged from 91% to
92%. The testing loss ranged from 0.408 to 0.4184.

Fig. 2 shows the training and validation loss during each

epoch for fold 1 for 1500 RPM. The training loss starts from a
relatively high value of 1.2 and decreases over time, reaching
a final value of 0.4. On the other hand, the validation loss Table. I. shows the final model results after the cross
starts from a value of 0.9 and decreases over time, reaching a validation. The training accuracy and loss indicate how well
final value of 0.5. This indicates that the model is learning to the model was able to fit the training data and the model
generalize well to new data, as the training and validation loss achieved higher accuracy and lower loss on the 1500 RPM
both decrease over time. Fig. 2 suggest that the model is training data (92.55% accuracy and 0.3861 loss) compared to
performing well on the testing set, with testing accuracy the 2700 RPM training data (91.47% accuracy and 0.3146
ranging from 90% to 91%. Additionally, the testing loss is loss). The validation accuracy and loss indicate how well the
generally lower than the training and validation loss, model was able to generalize to new data and the model
indicating that the model is not overfitting to the training data. achieved higher validation accuracy on the 1500 RPM data
(96.39% accuracy and 0.2967 loss) compared to the 2700
Fig. 2. Training and validation loss in each epoch for fold 1
RPM data (93.89% accuracy and 0.2848 loss). This suggests
that the model was better at generalizing to new data at the
lower RPM. The testing accuracy and loss indicate how well
the model performed on a completely new and unseen dataset.
The model achieved higher testing accuracy on the 1500 RPM
data (91.34% accuracy and 0.407 loss) compared to the 2700
RPM data (90.61% accuracy and 0.3535 loss). This suggests
that the model was better at predicting the outcome of new
data at the lower RPM.

TABLE I. TRAINING AND TESTING PERFORMANCE

 Rpm 1500 2700
Training accuracy 93% 91%
Training loss 0,3861 0,3146
Validation accuracy 96% 94%
Validation loss 0,2967 0,2848
Testing accuracy 91% 91%
Testing loss 0,407 0,3535

The results of the developed method for fault detection have

demonstrated its effectiveness in diagnosing faults in
gearboxes. The high training, validation accuracies, and
relatively low training and validation losses, indicate that the
model can accurately detect and classify faults in the gearbox.
The testing accuracies, although slightly lower than the
training and validation accuracies, are still relatively high and
demonstrate that the model can generalize well to new data.
The practical implications of these findings are significant for
improving the reliability and maintenance of gearboxes in
various industrial applications. By accurately detecting faults
in gearboxes, maintenance personnel can proactively address
the underlying issues before they become more severe and
potentially lead to catastrophic failures. This can help to
reduce downtime and maintenance costs and improve the
overall reliability and safety of machinery. Additionally, the
ability to diagnose faults in gearboxes using a machine
learning approach can potentially reduce the need for manual
inspection and diagnosis, saving time and resources. Overall,
the developed method for fault detection has the potential to
improve the reliability and maintenance of gearboxes in
various industrial applications, leading to improved safety,
reduced costs, and increased efficiency.
V. CONCLUSION
In conclusion, this paper presents the development of a
comprehensive method for fault detection in gearboxes using
vibration analysis. The proposed framework includes several
stages, such as data preprocessing, feature extraction, and fault
diagnosis, which aim to identify and diagnose faults accurately
and effectively. The proposed method was applied to a case
study involving a gearbox data, and the results demonstrated
the effectiveness of the developed framework in detecting and
diagnosing faults in gearboxes.
The developed sequential neural network with feature
selection has demonstrated promising results in the detection of
faults in gearboxes. The model achieved high training and
validation accuracies and relatively low training and validation
losses, indicating that it can accurately detect and classify faults
in gearboxes. The model also performed well on the testing set,
demonstrating good generalization ability.
Overall, this paper provides a valuable contribution to the
field of fault detection in gearboxes, and the developed
framework has the potential to improve the performance and
reliability of various systems that rely on gearboxes. The future
research in this area can focus on further enhancing the
proposed framework's accuracy and efficiency by integrating
other advanced techniques, such as machine learning and Mechatronics, Aug. 2011, pp. 102–106. doi:
artificial intelligence. 10.1109/FPM.2011.6045738.
[11] Y. Dai, X. Huang, and Z. Chen, “Application of
ACKNOWLEDGMENT Wavelet Denoising and Time- Frequency Domain Feature
The authors would like to acknowledge the support of AI tools Extraction on Data Processing of Modulated Signals,” in 2021
(ChatGPT, Writefull) for language support in the preparation 2nd International Seminar on Artificial Intelligence,
of this paper. The use of these tools helped improve the clarity Networking and Information Technology (AINIT), Oct. 2021,
and accuracy of our written communication. pp. 611–615. doi: 10.1109/AINIT54228.2021.00123.
[12] C. S. Hlaing and S. M. M. Zaw, “Plant diseases
REFERENCES recognition for smart farming using model-based statistical
[1] V. Sharma and A. Parey, “A Review of Gear Fault features,” in 2017 IEEE 6th Global Conference on Consumer
Diagnosis Using Various Condition Indicators,” Procedia Electronics (GCCE), Oct. 2017, pp. 1–4. doi:
Eng., vol. 144, pp. 253–263, Jan. 2016, doi: 10.1109/GCCE.2017.8229343.
10.1016/j.proeng.2016.05.131. [13] B. Schölkopf and A. J. Smola, “Support Vector
[2] M. Nie and L. Wang, “Review of Condition Machines,” in Learning with Kernels: Support Vector
Monitoring and Fault Diagnosis Technologies for Wind Machines, Regularization, Optimization, and Beyond, MIT
Turbine Gearbox,” Procedia CIRP, vol. 11, pp. 287–290, Jan. Press, 2001, pp. 187–188. Accessed: Mar. 09, 2023. [Online].
2013, doi: 10.1016/j.procir.2013.07.018. Available: https://ieeexplore.ieee.org/document/6282654
[3] A. S. Sait and Y. I. Sharaf-Eldeen, “A Review of [14] S. Parvathy, T. Safni Usman, and K. Harikrishnan,
Gearbox Condition Monitoring Based on vibration Analysis “Energy Management via Anomaly Detection for
Techniques Diagnostics and Prognostics,” in Rotating Manufacturing Enterprises,” in 2020 International Conference
Machinery, Structural Health Monitoring, Shock and for Emerging Technology (INCET), Jun. 2020, pp. 1–7. doi:
Vibration, Volume 5, T. Proulx, Ed., in Conference 10.1109/INCET49848.2020.9153989.
Proceedings of the Society for Experimental Mechanics [15] P. Chandran, F. Thierry, J. Odelius, S. M. Famurewa,
Series. New York, NY: Springer, 2011, pp. 307–324. doi: H. Lind, and M. Rantatalo, “Supervised Machine Learning
10.1007/978-1-4419-9428-8_25. Approach for Detecting Missing Clamps in Rail Fastening
[4] T. Hao, T. Liwei, and Y. Tongqiang, “Application of System from Differential Eddy Current Measurements,” Appl.
Acoustic Testing for Gear Wearing Fault Diagnosis,” in 2007 Sci., vol. 11, no. 9, Art. no. 9, Jan. 2021, doi:
8th International Conference on Electronic Measurement and 10.3390/app11094018.
Instruments, Aug. 2007, pp. 3-635-3–638. doi: [16] T. Nagano and S. Miyajima, “A neural network
10.1109/ICEMI.2007.4350997. model for the developmental process of hypercomplex cells,”
[5] Y. Shao and C. K. Mechefske, “Gearbox vibration IEEE Trans. Syst. Man Cybern., vol. SMC-13, no. 5, pp. 847–
monitoring using extended Kalman filters and hypothesis 851, Sep. 1983, doi: 10.1109/TSMC.1983.6313078.
tests,” J. Sound Vib., vol. 325, no. 3, pp. 629–648, Aug. 2009, [17] G. B. Coleman and H. C. Andrews, “Image
doi: 10.1016/j.jsv.2009.03.029. segmentation by clustering,” Proc. IEEE, vol. 67, no. 5, pp.
[6] J. Yao, C. Liu, K. Song, C. Feng, and D. Jiang, “Fault 773–785, May 1979, doi: 10.1109/PROC.1979.11327.
diagnosis of planetary gearbox based on acoustic signals,” [18] S. Parvathy, N. R. Patne, and T. Safni Usman,
Appl. Acoust., vol. 181, p. 108151, Oct. 2021, doi: “Optimal Battery Charging Forecasting Algorithms for
10.1016/j.apacoust.2021.108151. Domestic Applications and Electric Vehicles by
[7] P. Chandran, M. Lokesha, M. C. Majumder, K. Fathi, Comprehending Sustainable Energy,” in Control Applications
and A. Raheem, “Application of Laplace wavelet kurtosis and in Modern Power System, A. K. Singh and M. Tripathy, Eds.,
wavelet statistical parameters for gear fault diagnosis,” Int. J. in Lecture Notes in Electrical Engineering. Singapore:
Multidiscip. Sci. Eng., vol. 3, no. 9, pp. 1–8, 2012. Springer, 2021, pp. 29–43. doi: 10.1007/978-981-15-8815-
[8] C. Kar and A. R. Mohanty, “Vibration and current 0_3.
transient monitoring for gearbox fault detection using [19] P. Chandran, “Train Based Automated Inspection for
multiresolution Fourier transform,” J. Sound Vib., vol. 311, Railway Fastening System,” 2022, Accessed: Mar. 31, 2023.
no. 1, pp. 109–132, Mar. 2008, doi: 10.1016/j.jsv.2007.08.023. [Online]. Available: http://urn.kb.se/resolve?
[9] P. Geethanjali, Y. K. Mohan, and J. Sen, “Time urn=urn:nbn:se:ltu:diva-89090
domain Feature extraction and classification of EEG data for [20] A. Van Rotterdam, “Limitations and Difficulties in
Brain Computer Interface,” in 2012 9th International Signal Processing by Means of the Principal-Components
Conference on Fuzzy Systems and Knowledge Discovery, May Analysis,” IEEE Trans. Biomed. Eng., vol. BME-17, no. 3, pp.
2012, pp. 1136–1139. doi: 10.1109/FSKD.2012.6234336. 268–269, Jul. 1970, doi: 10.1109/TBME.1970.4502744.
[10] J. Wanlu, Z. Sheng, and W. Shengqiang, [21] R. Kamimura and S. Nakanishi, “Information
“Experimental research on sensitivity of frequency domain maximization for feature detection and pattern classification
feature parameters to hydraulic pump faults,” in Proceedings by autoencoders,” in Proceedings of ICNN’95 - International
of 2011 International Conference on Fluid Power and Conference on Neural Networks, Nov. 1995, pp. 985–989
vol.2. doi: 10.1109/ICNN.1995.487554.
[22] Z. Wang, M. Chen, Y. Guo, Z. Li, and Q. Yu, [25] W. Teng, N. Wang, H. Shi, Y. Liu, and J. Wang,
“Bridging the Domain Gap in Satellite Pose Estimation: A “Classifier-Constrained Deep Adversarial Domain Adaptation
Self-Training Approach Based on Geometrical Constraints,” for Cross-Domain Semisupervised Classification in Remote
IEEE Trans. Aerosp. Electron. Syst., pp. 1–14, 2023, doi: Sensing Images,” IEEE Geosci. Remote Sens. Lett., vol. 17,
10.1109/TAES.2023.3250385. no. 5, pp. 789–793, May 2020, doi:
[23] S. Z. K. Khine, T. L. Nwe, and H. Li, “Singing voice 10.1109/LGRS.2019.2931305.
detection in pop songs using co-training algorithm,” in 2008 [26] Enshaei, H. Chen, F. Naderkhani, J. Lin, S. Shahsafi,
IEEE International Conference on Acoustics, Speech and S. Giliyana, M. Mirzaei, S. Li, C. Hansen, J. Rupe.
Signal Processing, Mar. 2008, pp. 1629–1632. doi: “ICPHM’23 Benchmark Vibration Dataset Applicable in
10.1109/ICASSP.2008.4517938. Machine Learning for Systems’ Health Monitoring”,In
[24] T. Li, H. Luo, and C. Wu, “A PSO-based fine-tuning proceeding of IEEE Conference on Prognostics and Health
algorithm for CNN,” in 2021 5th Asian Conference on Management (ICPHM ),2023.
Artificial Intelligence Technology (ACAIT), Oct. 2021, pp.
704–709. doi: 10.1109/ACAIT53529.2021.9731225.