Survishno
Survishno
Survishno
Abstract
Machine fault diagnosis is crucial in industrial systems to enhance reliability, lifetime, and service availability.
Intelligent fault diagnosis (IFD) using artificial intelligence (AI) techniques has emerged as a promising
approach for automating machine health assessment and reducing labor costs. One approach to improve fault
diagnosis accuracy, which has become a highly relevant research topic, is to use multi-data fusion, which
combines information from multiple sources to make a more informed decision. However, there is a lack of
research focusing on the detection of combined machinery faults from multiple sensors. Indeed, when
combined (and emerging) faults happen in different parts of the rotating machines their features are deeply
dependent and the separation of characteristics becomes complex, while multi-sensor information can provide
more comprehensive fault features to deal with the diagnosis and identification of multiple combined faults.
This paper presents a comprehensive methodology for diagnosing combined faults using data fusion and
machine learning techniques. The proposed approach leverages multiple types of sensor data, including
vibration, current, temperature, and acoustic data, sensors to provide a comprehensive picture of the machine's
health. Our proposed methodology incorporates an ensemble learning approach and time-domain features to
improve diagnostic accuracy. The proposed approach is tested on a publicly available dataset of rotating
machinery with multiple faults. The results indicate that the method is viable and achieves good accuracy and
efficiency.
Introduction
Rotating machinery plays a crucial role in modern industries, including manufacturing, transportation, and
power generation. The efficient operation of these machines is vital to ensure the continuity of industrial
processes and the safety of workers. However, rotating machinery is prone to various types of faults that can
result in unexpected downtime, production losses, and safety hazards. Therefore, timely and accurate diagnosis
of machinery faults is essential to reduce maintenance costs and ensure the reliability and availability of
machines. Diagnostic is typically a reactive action performed after a fault or problem has already occurred.
The purpose of fault diagnostic is to identify the cause of the problem or fault so that it can be addressed and
resolved [1].
In recent years, Intelligent Fault Diagnosis (IFD) has emerged as a promising approach to improve the
accuracy and efficiency of machinery fault diagnosis. Several studies have explored the applications of
Artificial Intelligence (AI) techniques such as machine learning, pattern recognition, and signal processing, in
machine fault diagnosis, highlighting its potential for enhancing diagnostic accuracy and enabling automated
fault recognition. Notably, Y. Lei et al. [2] conducted a comprehensive review and roadmap of machine
learning applications in machine fault diagnosis, providing valuable insights into the current state of the field
1
and outlining future research directions. R. Liu et al. [3] also conducted a noteworthy review on the use of
artificial intelligence for fault diagnosis of rotating machinery. Their review focused specifically on the
application of artificial intelligence techniques, including machine learning, in diagnosing faults in rotating
machinery. These reviews collectively underscore the growing interest in applying artificial intelligence
techniques to fault diagnosis in rotating machinery. By leveraging the insights gained from these seminal
reviews, this paper contributes to the existing body of knowledge by presenting a comprehensive methodology
for diagnosing combined faults in rotating machinery using data fusion and signal processing techniques.
While IFD has shown significant success in detecting single faults in rotating machinery, diagnosing
combined faults remains a challenging problem. Combined faults occur when multiple faults occur
simultaneously or sequentially in different parts of the machine, and their features are deeply interdependent
[1], making them difficult to separate and identify. Therefore, detecting combined faults requires integrating
information from multiple sources and fusing it to extract more comprehensive fault features. Multi-source
information fusion is a promising technique for diagnosing combined faults in rotating machinery, it involves
combining data from multiple sensors, such as vibration, current, temperature, and acoustic sensors, to improve
the accuracy and robustness of fault diagnosis. Therefore, combining machine learning and deep learning with
multi-source information fusion techniques can further improve the accuracy and reliability of machinery fault
diagnosis, providing significant benefits to the industry.
This article presents a comprehensive methodology for diagnosing combined faults using data fusion and
machine learning algorithms. The methodology involves data acquisition, data pre-processing including direct
data fusion, and feature extraction. To perform the classification of faults, we use machine learning algorithms
such as Logistic Regression, k-Nearest Neighbor (kNN), and Gaussian Naïve Bayes, which are trained on the
extracted features. Finally, we combine the different decisions obtained from the different algorithms using
stacking ensemble learning technique to increase the overall accuracy of the diagnosis. The proposed approach
is applied to a publicly available dataset to evaluate its effectiveness in diagnosing combined faults. The results
indicate that the method is viable and achieves a good level of accuracy and efficiency. The paper also
discusses the challenges and limitations of data fusion techniques for intelligent fault diagnosis, while
highlighting potential areas for improvement and suggesting future research directions.
Related Works
Fault diagnosis in rotating machinery is crucial for ensuring reliable industrial operation. However,
diagnosing combined faults, where multiple faults occur simultaneously or sequentially in different machine
parts, presents unique challenges. To address these challenges, data fusion techniques have gained
significance. Data fusion combines information from multiple sensors to provide a comprehensive
understanding of machine health. This subsection explores notable works in data fusion techniques for fault
diagnosis in rotating machinery, showcasing advancements in addressing the complexities of diagnosing
combined faults.
2
decision fusion focuses on combining the decisions or outputs from multiple information sources to arrive at
a final decision or inference. Within the domain of fault diagnosis for rotating machinery, several studies have
investigated the application of data fusion techniques at these different levels.
Some studies in fault diagnosis of rotating machinery have explored the fusion of vibration and acoustic
data. For instance, M. Khazaee et al. [7] proposed a feature-level fusion approach using wavelet transform and
artificial neural networks for fault diagnosis of a planetary gearbox. R.S. Gunerkar et al. [8] focused on the
classification of ball bearing faults using vibro-acoustic sensor data fusion. The authors employed time-domain
features. Furthermore, principal component analysis (PCA) was utilized to select the most suitable features
from the feature set. You He et al. [9] developed a deep multi-signal fusion adversarial model based on transfer
learning and residual network for axial piston pump fault diagnosis, where vibration and acoustic data were
fused. The study specifically emphasized the significant benefits of utilizing deep learning techniques,
including the application of a residual neural network. These studies demonstrate the effectiveness of
integrating vibration and acoustic data to improve fault diagnosis accuracy in rotating machinery.
In addition to vibration and acoustic data fusion, researchers have explored the fusion of vibration and
current data for fault diagnosis in rotating machinery. J. Cui et al. [10] proposed M2FN, an end-to-end multi-
task and multi-sensor fusion network for intelligent fault diagnosis. They demonstrated the effectiveness of
fusing vibration and current data in accurately identifying and classifying faults. S. Liu et al. [11] employed a
feature-level fusion approach in their study, the objective of this fusion was to enhance the fault diagnosis
process and improve the accuracy of the results.
Researchers have also explored the fusion of a single type of signal for improved diagnostic performance.
For example, when considering vibration signals, J. Mi et al. [12] proposed a decision-level fusion method
based on evidence theory, integrating multiple sources of vibration data. Similarly, S. Li [13] introduced an
ensemble deep convolutional neural network (CNN) model with improved Dempster–Shafer theory evidence
fusion for bearing fault diagnosis, specifically fusing vibration signals using the D-S evidence theory to
enhance fault classification accuracy and address limitations of single-model-based diagnosis.
Alternatively, in the realm of current signal analysis, specifically using motor current signature analysis
(MCSA), M. Azamfar et al. [14] proposed a data-level multi-sensor fusion approach for gearbox fault
diagnosis. The study utilized a CNN-based fusion approach combined with the MCSA to improve accuracy in
identifying and classifying gearbox faults.
Combined Faults
Diagnosing combined faults in rotating machinery poses significant challenges due to the complexity and
interaction of multiple fault types. Researchers have made notable progress in developing methodologies to
address these challenges. M.Y. Asr et al. [1] used a Non-Naive Bayesian Classifier to successfully identify
gear and bearing failures. U.I. Inyang et al. [15] proposed a comprehensive learning approach using vibration
data to diagnose single and multiple faults across diverse rotating machine components (gearbox, bearing, and
shaft). M. Islam et al. [16] proposed a reliable multiple combined fault diagnosis scheme for bearings based
on acoustic signals, utilizing heterogeneous feature models and an improved one-against-all multiclass support
vector machines classifier. A. Garcia-Perez et al. [17] proposed a condition-monitoring strategy that combines
vibration and current data for accurate identification of single or multiple combined faults in induction motors,
including broken bar (BRB), outer race damaged bearing (BD), and unbalanced pulley (UNB).
In conclusion, while various studies have investigated data fusion techniques for fault diagnosis in rotating
machinery, there is still a lack of research on large-scale sensor fusion, particularly in the context of
determining combined faults in different parts of the machinery such as bearings, gears, shafts, rotors, and
stators. The existing literature has primarily focused on the fusion of vibration and acoustic data, as well as the
fusion of vibration and current data. However, there is a need for more comprehensive and integrated
approaches that encompass a wider range of sensor inputs and address the challenges associated with
diagnosing complex and combined faults in various components of rotating machinery. Future research efforts
should aim to develop novel data fusion methodologies that can effectively handle multiple sensor inputs and
provide accurate and reliable fault diagnosis in diverse operating conditions and fault scenarios.
3
Proposed Methodology
In this section, we present the methodology employed for diagnosing combined faults in rotating
machinery. The methodology, illustrated in Figure 1, provides a comprehensive framework for the fault
diagnosis process. The methodology encompasses data acquisition using multiple sensors, data preprocessing
techniques such as data-level fusion and feature extraction, model training, and decision-making using
machine learning and ensemble learning. By following this systematic approach, we aim to achieve accurate
and reliable fault diagnosis for combined faults in rotating machinery
Data Preprocessing
At this point, the methodology employs data-level fusion to bring together information obtained by different
sensors for a more holistic understanding of the machine's health status. This step plays an instrumental role
in capturing complementary sensing information from every source and aids in the enhancement of overall
equipment insights. Then come to the feature extraction techniques aimed at extracting meaningful attributes
from this fused data set based on time-domain analysis, frequency-domain analysis, and other relevant methods
available. The objective is to identify relevant characteristics that can tell apart different fault conditions by
comparing and analyzing appropriate features extracted from diverse domains.
4
applying a logistic function to a linear combination of the input features, logistic regression estimates the
probability of a data instance belonging to a specific class. The logistic function, represented by Equation (1),
calculates the probability:
1 𝑒 𝑔(𝑋)
𝑃𝑟𝑜𝑏(𝑒𝑣𝑒𝑛𝑡) = 𝑃(𝑋) = = (1)
1 + 𝑒 −𝑔(𝑋) 1 + 𝑒 𝑔(𝑋)
Where P(X) refers to the probability of a particular event or outcome occurring given a set of input features
𝑋 = (𝑥1 , 𝑥2 , … , 𝑥𝑘 ), and g(X) represents the logit model. The logit model, as shown in Equation (2), transforms
the linear combination of input features and their coefficients into the log-odds or logit value:
𝑃(𝑋)
𝑔(𝑋) = log ( ) = 𝛼 + 𝛽1 𝑥1 + 𝛽2 𝑥2 + ⋯ + 𝛽𝑘 𝑥𝑘 (2)
1 − 𝑃(𝑋)
Here α, 𝛽1 , 𝛽2 … 𝛽𝑘 , are the regression coefficients estimated using maximum likelihood estimation.
5
accuracy, precision, recall, and F1-score to assess the effectiveness of the methodology in accurately
identifying combined faults.
(c) (d)
Figure 3: Experimental setup of the Figure 2: Row vibration signals (a) healthy condition, (b) Inner
bearing motor of the CWRU Dataset Race (IR) fault 0.014 inch, (c) IR 0.021 inch (d) IR 0.007 inch
6
Algorithm Accuracy Precision Recall F1-score
Logistic Regression 0.942 0.94 0.94 0.94
KNN 0.932 0.93 0.93 0.93
Gaussian Naïve Bayes 0.924 0.93 0.92 0.92
Stacking Classifier 0.972 0.97 0.97 0.97
The study employed a methodology for diagnosing combined faults in rotating machinery, incorporating
data acquisition, preprocessing techniques, machine learning, and ensemble learning. Despite lacking multi-
source data fusion, the CWRU dataset still allowed for the evaluation of the proposed methodology by
considering multiple fault classes within the vibration data alone, resulting in a perplexing yet innovative
approach.
The results showed promising performance across different algorithms. Figure 5 presented a bar plot
showcasing the accuracy of each algorithm used in the study. The mean accuracy values, represented by blue
dots, served as a visual reference for comparing the performance of different classifiers. Additionally, the error
bars accompanying the mean accuracy values offered an indication of the standard deviation, capturing the
variability around each mean. This information provided a comprehensive perspective on the accuracy
measurements, helping to assess the robustness of the classifiers.
Among the base classifiers, the Logistic Regression achieves the highest accuracy of 94.2% followed
closely by the KNN algorithm with 93.2% accuracy. Gaussian Naïve Bayes achieved an accuracy of 92.4%.
However, the most impressive results were obtained from the Stacking Classifier, which achieved an accuracy
of 97.2% and high precision, recall, and F1-score. The stacking approach, combining predictions from multiple
base classifiers, proved highly effective in accurately diagnosing the 10 classes of faults based on the testing
dataset. To gain deeper insights into the Stacking Classifier's performance, a confusion matrix is presented in
Figure 4. The confusion matrix revealed a high number of correct predictions along the diagonal, indicating
that the classifier accurately identified the 10 fault classes in the testing dataset. Furthermore, the convergence
of the learning curves given in Figure 6 indicates that the model gradually improved its generalization
capability, thereby reducing the performance gap between the training and test sets. This convergence is a
7
positive indication that the model effectively learned from the training data and was able to generalize well to
unseen test data. These findings highlight the effectiveness of the methodology and the potential of ensemble
learning for accurately diagnosing multiple combined faults in rotating machinery, contributing to the
maintenance and operational improvements in various industrial applications.
Conclusion
In conclusion, this study presented a comprehensive methodology for intelligent fault diagnosis of
combined faults in rotating machinery using data fusion and machine learning techniques. Although the
application of the methodology was carried out on the CWRU dataset, which did not include multi-source data
fusion as it only utilized vibration signals, it successfully addressed the challenge of diagnosing combined
faults by considering 10 fault classes in the bearing. Otherwise, by integrating multiple sensor data through
the direct fusion model and leveraging machine learning algorithms and ensemble learning techniques, the
proposed methodology can achieve a high level of accuracy and efficiency in diagnosing the multiple fault
classes. The evaluation results on the vibration dataset demonstrated promising performance, with the
ensemble learning model, particularly the Stacking Classifier, achieving an impressive accuracy of 97.2% and
exhibiting high precision, recall, and F1-score.
While the dataset did not encompass the case of multi-source data fusion, the study effectively demonstrated
the potential of the proposed methodology in diagnosing combined faults using vibration data alone. However,
it is important to acknowledge that challenges still exist in accurately diagnosing combined faults, highlighting
the need for further research to enhance accuracy and reliability. Overall, this study significantly contributes
to the field of intelligent fault diagnosis by providing a comprehensive methodology and valuable insights for
future research endeavors in this area.
8
References
[1] M. Y. Asr, M. M. Ettefagh, R. Hassannejad, et S. N. Razavi, « Diagnosis of combined faults in Rotary
Machinery by Non-Naive Bayesian approach », Mech. Syst. Signal Process., vol. 85, p. 56‑70, févr. 2017,
doi: 10.1016/j.ymssp.2016.08.005.
[2] Y. Lei, B. Yang, X. Jiang, F. Jia, N. Li, et A. K. Nandi, « Applications of machine learning to machine
fault diagnosis: A review and roadmap », Mech. Syst. Signal Process., vol. 138, p. 106587, avr. 2020, doi:
10.1016/j.ymssp.2019.106587.
[3] R. Liu, B. Yang, E. Zio, et X. Chen, « Artificial intelligence for fault diagnosis of rotating machinery: A
review », Mech. Syst. Signal Process., vol. 108, p. 33‑47, août 2018, doi: 10.1016/j.ymssp.2018.02.016.
[4] M. Huang, Z. Liu, et Y. Tao, « Mechanical fault diagnosis and prediction in IoT based on multi-source
sensing data fusion », Simul. Model. Pract. Theory, vol. 102, p. 101981, juill. 2020, doi:
10.1016/j.simpat.2019.101981.
[5] L. Wald, « Some terms of reference in data fusion », IEEE Trans. Geosci. Remote Sens., vol. 37, no 3, p.
1190‑1193, mai 1999, doi: 10.1109/36.763269.
[6] Z. Duan, T. Wu, S. Guo, T. Shao, R. Malekian, et Z. Li, « Development and trend of condition monitoring
and fault diagnosis of multi-sensors information fusion for rolling bearings: a review », Int. J. Adv. Manuf.
Technol., vol. 96, no 1‑4, p. 803‑819, avr. 2018, doi: 10.1007/s00170-017-1474-8.
[7] M. Khazaee, H. Ahmadi, M. Omid, A. Banakar, et A. Moosavian, « Feature-level fusion based on wavelet
transform and artificial neural network for fault diagnosis of planetary gearbox using acoustic and
vibration signals », Insight - Non-Destr. Test. Cond. Monit., vol. 55, no 6, p. 323‑330, juin 2013, doi:
10.1784/insi.2012.55.6.323.
[8] R. S. Gunerkar et A. K. Jalan, « Classification of Ball Bearing Faults Using Vibro-Acoustic Sensor Data
Fusion », Exp. Tech., vol. 43, no 5, p. 635‑643, oct. 2019, doi: 10.1007/s40799-019-00324-0.
[9] Y. He, H. Tang, Y. Ren, et A. Kumar, « A deep multi-signal fusion adversarial model based transfer
learning and residual network for axial piston pump fault diagnosis », Measurement, vol. 192, p. 110889,
mars 2022, doi: 10.1016/j.measurement.2022.110889.
[10] J. Cui, P. Xie, X. Wang, J. Wang, Q. He, et G. Jiang, « M2FN: An end-to-end multi-task and multi-sensor
fusion network for intelligent fault diagnosis », Measurement, vol. 204, p. 112085, nov. 2022, doi:
10.1016/j.measurement.2022.112085.
[11] S. Liu et al., « Multi-feature fusion for fault diagnosis of rotating machinery based on convolutional neural
network », Comput. Commun., vol. 173, p. 160‑169, mai 2021, doi: 10.1016/j.comcom.2021.04.016.
[12] J. Mi, X. Wang, Y. Cheng, et S. Zhang, « Multi-Source Uncertain Information Fusion Method for Fault
Diagnosis Based on Evidence Theory », in 2019 Prognostics and System Health Management Conference
(PHM-Qingdao), Qingdao, China: IEEE, oct. 2019, p. 1‑6. doi: 10.1109/PHM-
Qingdao46334.2019.8942946.
[13] S. Li, G. Liu, X. Tang, J. Lu, et J. Hu, « An Ensemble Deep Convolutional Neural Network Model with
Improved D-S Evidence Fusion for Bearing Fault Diagnosis », Sensors, vol. 17, no 8, p. 1729, juill. 2017,
doi: 10.3390/s17081729.
[14] M. Azamfar, J. Singh, I. Bravo-Imaz, et J. Lee, « Multisensor data fusion for gearbox fault diagnosis using
2-D convolutional neural network and motor current signature analysis », Mech. Syst. Signal Process.,
vol. 144, p. 106861, oct. 2020, doi: 10.1016/j.ymssp.2020.106861.
[15] U. I. Inyang, I. Petrunin, et I. Jennions, « Diagnosis of Multiple Faults in Rotating Machinery Using
Ensemble Learning », Sensors, vol. 23, no 2, p. 1005, janv. 2023, doi: 10.3390/s23021005.
[16] M. M. Manjurul Islam et J.-M. Kim, « Reliable multiple combined fault diagnosis of bearings using
heterogeneous feature models and multiclass support vector Machines », Reliab. Eng. Syst. Saf., vol. 184,
p. 55‑66, avr. 2019, doi: 10.1016/j.ress.2018.02.012.
[17] A. Garcia-Perez, R. de J. Romero-Troncoso, E. Cabal-Yepez, et R. A. Osornio-Rios, « The Application
of High-Resolution Spectral Analysis for Identifying Multiple Combined Faults in Induction Motors »,
IEEE Trans. Ind. Electron., vol. 58, no 5, p. 2002‑2010, mai 2011, doi: 10.1109/TIE.2010.2051398.
[18] W. Caesarendra, A. Widodo, et B.-S. Yang, « Application of relevance vector machine and logistic
regression for machine degradation assessment », Mech. Syst. Signal Process., vol. 24, no 4, p. 1161‑1171,
mai 2010, doi: 10.1016/j.ymssp.2009.10.011.
[19] I. Ouachtouk, S. El Hani, S. Guedira, et K. Dahi, « Detection and classification of broken rotor bars faults
in induction machine using K-means classifier », in 2016 International Conference on Electrical and
Information Technologies (ICEIT), Tangiers, Morocco: IEEE, mai 2016, p. 180‑185. doi:
10.1109/EITech.2016.7519586.
9
[20] T. Cover et P. Hart, « Nearest neighbor pattern classification », IEEE Trans. Inf. Theory, vol. 13, no 1, p.
21‑27, janv. 1967, doi: 10.1109/TIT.1967.1053964.
[21] B. Mahesh, « Machine Learning Algorithms - A Review », vol. 9, no 1, 2018.
[22] S. E. Pandarakone, S. Gunasekaran, Y. Mizuno, et H. Nakamura, « Application of Naive Bayes Classifier
Theorem in Detecting Induction Motor Bearing Failure », in 2018 XIII International Conference on
Electrical Machines (ICEM), Alexandroupoli: IEEE, sept. 2018, p. 1761‑1767. doi:
10.1109/ICELMACH.2018.8506836.
[23] Z. Cao, Z. Li, J. Zhang, et H. Fu, « A Homogeneous Stacking Ensemble Learning Model for Fault
Diagnosis of Rotating Machinery With Small Samples », IEEE Sens. J., vol. 22, no 9, p. 8944‑8959, mai
2022, doi: 10.1109/JSEN.2022.3163760.
10