Applications of machine learning to machine fault diagnosis: A review and roadmap

The quote only partly states this claim.

The number of publications about intelligent fault diagnosis has rapidly increased since the 2010s.

“According to the results on the topic of machine fault diagnosis by using machine learning, the number of publications has rapidly increased since the 2010s.”

May combine two claims.

Tong et al. employed JDA for bearings in motor and belt conveyor idler and achieved better diagnosis accuracy than TCA-based models.

“With regard to fault diagnosis, Tong et al. [420,421] employed JDA to complete the diagnosis tasks of bearings respectively used in the motor and the belt conveyor idler, and achieve better diagnosis accuracy than TCA-based diagnosis models.”

• cites [420] 420. Z. Tong, W. Li, B. Zhang, F. Jiang, G. Zhou, Bearing fault diagnosis under variable working conditions based on domain adaptation using feature transfer learning, IEEE Access 6 (2018) 76187-76197. sourceNot checked — no full text of the source
• cites [421] 421. Z. Tong, W. Li, B. Zhang, M. Zhang, Bearing fault diagnosis based on domain adaptation using transferable features under different working conditions, Shock Vibrat. (2018) Doi: 10.1155/2018/6714520. sourceNot checked — no full text of the source

May combine two claims.

Shen et al. used cross-domain features to train kNN-based diagnosis models by TrAdaboost, which correctly recognized unlabeled target samples despite few labeled target samples.

“The results showed that the trained diagnosis model could correctly recognize the unlabeled samples in the target domain although there were a small number of labeled samples in the target domain.”

• cites [36] 36. F. Shen, C. Chen, R. Yan, R.X. Gao, Bearing fault diagnosis based on svd feature extraction and transfer learning classification, in: Prognostics and System Health Management Conference, Beijing, China, 2015, pp. 1-6. sourceNot checked — no full text of the source

May combine two claims.

Compared with a diagnosis model just trained with a small number of target-domain samples, the fine-tuned diagnosis model presented a faster convergence rate and higher diagnosis accuracy.

“Compared with the diagnosis model just trained with a small number of target-domain samples, the fine-tuned diagnosis model presented the faster convergence rate and higher diagnosis accuracy.”

• cites [438] 438. R. Zhang, H. Tao, L. Wu, Y. Guan, Transfer learning with neural networks for bearing fault diagnosis in changing working conditions, IEEE Access 5 (2017) 14347-14357. sourceNot checked — no full text of the source
• cites [37] 37. M.J. Hasan, J.M. Kim, Bearing fault diagnosis under variable rotational speeds using stockwell transform- based vibration imaging and transfer learning, Appl. Sci.-Basel 8 (2018) 2357. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Deep learning theories have reformed intelligent fault diagnosis in further releasing artificial assistance since the 2010s, encouraging the construction of an end-to-end diagnosis process.

“Over the recent years, the advent of deep learning theories has reformed IFD in further releasing the artificial assistance since the 2010s, which encourages to construct an end-to-end diagnosis process.”

Probably not a substantial claim: may not pass S2 (checkable).

It needs a new review to summarize the current research progress of intelligent fault diagnosis.

“Therefore, it needs a new review to summary the current research progress of IFD.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Traditional machine learning weakened the contribution of human labor in machine fault diagnosis and pushed it into the era of artificial intelligence.

“With the help of traditional machine learning, the diagnosis models began to establish the relationship between the selected features and the health states of machines, which weakened the contribution of human labor in machine fault diagnosis and pushed it into the era of artificial intelligence.”

Probably not a substantial claim: may not pass S2 (checkable).

Transfer learning theories will promote the research of intelligent fault diagnosis in engineering scenarios in the future.

“In the future, transfer learning theories will promote the research of IFD in engineering scenarios.”

May combine two claims.

Modified SVM and algorithm optimization are the two main branches to improve the diagnosis accuracy of SVM-based models.

“In order to improve the diagnosis accuracy of SVM-based models, researchers mainly focused on two branches, i.e., the modified SVM and the algorithm optimization.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

SVM algorithm is originally used to solve binary classification tasks.

“Third, the SVM algorithm is originally used to solve binary classification tasks.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

PGM-based diagnosis models easily achieve fault diagnosis with multiple health states and can analyze among-class discrepancy conveniently.

“It is easy for PGM-based diagnosis models to achieve fault diagnosis with multiple health states of machines.They can be used to analyze the among-class discrepancy in convenience.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

If probabilistic relationship among variables is not clear, it is difficult to construct PGM-based diagnosis models.

“Furthermore, if the probabilistic relationship among the variables is not clear, it would be difficult to construct the diagnosis models.”

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

Transfer-learning theories are expected to overcome the problems of lacking labeled samples and finally enlarge the applications of intelligent fault diagnosis in engineering scenarios.

“Therefore, transfer-learning theories are expected to overcome the problems of lacking labeled samples and finally enlarge the applications of IFD in engineering scenarios.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

kNN-based diagnosis models are easily implemented.

“The kNN-based diagnosis models are easily-implemented.”

May combine two claims.

Expert systems lack self-learning capability, so their diagnosis knowledge base is difficult to expand and correct.

“Furthermore, the expert system lacks the self-learning capability so that the diagnosis knowledge base is difficult to be expanded and corrected.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

SVM performance is sensitive to kernel parameters; inappropriate kernel parameter cannot induce reliable diagnosis result.

“Second, the performance of SVM-based diagnosis models is sensitive to the kernel parameters.The inappropriate kernel parameter even cannot induce the reliable diagnosis result.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

C4.5 algorithm is widely used to induce decision tree for classification and obtain satisfactory accuracy and easily-understood classification rules.

“Among the proposed methods, the algorithm of C4.5 is widely used to induce a decision tree for classification and obtain satisfactory accuracy and easily-understood classification rules [300].”

• cites [300] 300. J.R. Quinlan, C4.5: Programs for machine learning, Morgan Kaufmann Publishers Inc., San Francisco, 1993.Not checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

kNN takes much computation cost to handle large-volume datasets.

“However, it takes much computation cost to handle the large-volume dataset.”

May combine two claims.

SVM optimization uses convex quadratic optimization, so models can easily obtain global optimal solution and high diagnosis accuracy.

“The optimization objective solution of SVM refers to the convex quadratic optimization so that the diagnosis models could easily obtain the global optimal solution and further get the high diagnosis accuracy.”

May combine two claims.

Expert system-based diagnosis models' performance greatly relies on expert knowledge, which is difficult to acquire and express.

“However, the performance of the diagnosis models greatly relies on the expert knowledge, which is difficult to be acquired and expressed.”

Classified as probably not a substantial claim.

Transfer learning is concerned to overcome such weaknesses by applying the knowledge learned from one or multiple tasks to other related but novel ones.

“Fortunately, transfer learning is concerned to overcome such weaknesses by applying the knowledge learned from one or multiple tasks to other related but novel ones [27].”

• cites [27] 27. S.J. Pan, Q. Yang, A survey on transfer learning, IEEE Trans. Knowl. Data Eng. 22 (2010) 1345-1359. sourceNot checked — no full text of the source

May combine two claims.

Non-naive Bayesian classifier is developed to release the independence assumption and has been used for bearings and gears.

“The non-naive Bayesian classifier [288] is further developed to release the assumption, and it has also been used for fault diagnosis of bearings [289] and gears [290].”

• cites [288] 288. Y. He, R. Wang, S. Kwong, X. Wang, Bayesian classifiers based on probability density estimation and their applications to simultaneous fault diagnosis, Inf. Sci. 259 (2014) 252-268. sourceNot checked — no full text of the source
• cites [289] 289. J. Yu, B. Ding, Y. He, Rolling bearing fault diagnosis based on mean multigranulation decision-theoretic rough set and non-naive Bayesian classifier, J. Mech. Sci. Technol. 32 (2018) 5201-5211. sourceNot checked — no full text of the source
• cites [290] 290. M.Y. Asr, M.M. Ettefagh, R. Hassannejad, S.N. Razavi, Diagnosis of combined faults in rotary machinery by non-naive Bayesian approach, Mech. Syst. Signal Process. 85 (2017) 56-70. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Multi-class classification in IFD requires complicated architectures like OAA and OAO to integrate results from multiple SVM-based models.

“In terms of multi-class classification tasks in IFD, it always needs to use complicated 2 architectures, such as OAA and OAO, to integrate results from multiple SVM-based models.”

Probably not a substantial claim: may not pass S3 (load-bearing).

SVM-based diagnosis models are expected to recognize multiple states but not just the binary states of health and faults.

“For these diagnosis objects, SVM-based diagnosis models are expected to recognize multiple states but not just the binary states of health and faults.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Naive Bayesian classifier applications are subject to an assumption of independence among features.

“It is noted that the applications of naive Bayesian classifier are subject to an assumption of independence among features.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

SVM-based diagnosis models are trained by minimizing structural risk, which improves interpretability due to rigorous theories.

“Different from the ANN, SVM-based diagnosis models are trained by minimizing the structural risk, which is beneficial to improving the interpretability of the models due to the rigorous theories.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Random forest is investigated by integrating decisions from multiple tree-based classifiers to improve generalization performance of decision tree.

“In order to improve the generalization performance of decision tree, the random forest [306] is further investigated by integrating decisions from multiply tree-based classifiers.”

• cites [306] 306. L. Breiman, Random forests, Mach. Learn. 45 (2001) 5-32. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

The one-against-all strategy (OAA) and one-against-one strategy (OAO) are mainly concerned for multi-class SVM.

“Thus, the one-against-all strategy (OAA) and one-against-one strategy (OAO) are mainly concerned [17].”

• cites [17] 17. A. Widodo, B.S. Yang, Support vector machine in machine condition monitoring and fault diagnosis, Mech. Syst. Signal Process. 21 (2007) 2560-2574. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Parameter k is difficult to determine, greatly affecting performance of kNN diagnosis models.

“Furthermore, the parameter 𝑘𝑘 is difficult to be determined, which greatly affects the performance of the diagnosis models.”

Probably not a substantial claim: may not pass S3 (load-bearing).

kNN is a commonly-used supervised learning model for classification.

“The kNN is a commonly-used supervised learning model to complete the classification tasks [13].”

• cites [13] 13. T.M. Cover, P.E. Hart, Nearest neighbor pattern classification, IEEE Trans. Inf. Theory 13 (1967) 21-27. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

kNN performance is subject to indistinguishable neighborhood boundary and difficulty in selecting optimal neighborhood parameter.

“However, the performance of kNN is subject to some problems, such as the indistinguishable neighborhood boundary and the difficulty in selecting the optimal neighborhood parameter.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The diagnosis process of traditional intelligent fault diagnosis includes three steps: data collection, artificial feature extraction, and health state recognition.

“The diagnosis process includes three steps [1], i.e., data collection, artificial feature extraction, and health state recognition, as shown in Fig. 2.”

• cites [1] 1. Y. Lei, Intelligent fault diagnosis and remaining useful life prediction of rotating machinery, Butterworth- Heinemann Elsevier Ltd, Oxford, 2016.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Imbalanced distribution of collected data reduces diagnosis accuracy of kNN-based models.

“In particular, the imbalanced distribution of the collected data would reduce the diagnosis accuracy of this kind of diagnosis models.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Prosperous development of sensor technologies and data transmission, especially IoT and high-speed internet, promotes collecting large number of data with real-time information.

“4) The prosperous development of sensor technologies and data transmission, especially with the advent of IoT and high-speed internet, promotes to collect a large number of data that contain the real-time information.”

• cites [314] 314. G.I. Platforms, The rise of industrial big data, 2012, Available: http://www.geautomation.com/download/rise-industrial-big-data.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [315] 315. S. Yin, X. Li, H. Gao, O. Kaynak, Data-based techniques focused on modern industry: An overview, IEEE Trans. Ind. Electron. 62 (2015) 657-667. sourceNot checked — no full text of the source
• cites [316] 316. J. Lee, Industrial big data: The revolutionary transformation and value creation in industry 4.0 era, Mechanical Industry Press, Beijing, China, 2015.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Multiple AE can be stacked to represent information contained in input data deeply, and further obtain features in deep layers.

“Multiple AE can be stacked to represent the information contained in the input data deeply, and further obtain the features in deep layers, as shown in Fig. 11.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Monitoring big data of machines holds these characteristics and extends to specialized ones: large volume, low value density, multi-source and heterogeneous data structure, monitoring data stream.

“In contrast, monitoring big data of machines holds these characteristics, and further extends to the much-specialized ones.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Stacked AE-based diagnosis models are able to automatically represent health information from input monitoring data, which does not rely on much expert knowledge in feature extraction.

“Stacked AE-based diagnosis models are able to automatically represent the health information from the input monitoring data, which does not rely on much expert knowledge in feature extraction.”

May combine two claims.

Artificial feature extraction in traditional IFD depends on human labor and is unrealistic for large-volume monitoring data due to huge labor cost.

“First, the step of artificial feature extraction depends on the human labor, in which the engineers need to design powerful algorithms to extract sensitive features to health states of machines.However, it is still unrealistic for engineers to extract specialized features from the large-volume monitoring data by expert experience because of the huge labor cost.”

• cites [310] 310. F. Jia, Y. Lei, J. Lin, X. Zhou, N. Lu, Deep neural networks: A promising tool for fault characteristic mining and intelligent diagnosis of rotating machinery with massive data, Mech. Syst. Signal Process. 72-73 (2016) 303-315. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

Big data includes four characteristics: volume, velocity, variety, and veracity.

“Generally, big data includes four characteristics, i.e., volume, velocity, variety, and veracity [312].”

• cites [312] 312. V. Mayer-Schnberger, K. Cukier, Big data: A revolution that will transform how we live, work, and think, Hodder & Stoughton Press, Carmelite House, London, 2013.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

Only minority of collected data is valuable; healthy condition accounts for majority of long-term operation, faults seldom happen, so easier to collect healthy data than faulty.

“First, the healthy condition accounts for the majority of the longterm operation for machines, while faults seldom happen.Consequently, it is more easily to collect healthy data than faulty data.”

• cites [314] 314. G.I. Platforms, The rise of industrial big data, 2012, Available: http://www.geautomation.com/download/rise-industrial-big-data.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [315] 315. S. Yin, X. Li, H. Gao, O. Kaynak, Data-based techniques focused on modern industry: An overview, IEEE Trans. Ind. Electron. 62 (2015) 657-667. sourceNot checked — no full text of the source
• cites [316] 316. J. Lee, Industrial big data: The revolutionary transformation and value creation in industry 4.0 era, Mechanical Industry Press, Beijing, China, 2015.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Diagnosis performance of CNN-based diagnosis models is subject to training with sufficient labeled samples.

“Similar to other deep learning-based models, the diagnosis performance of CNN-based diagnosis models is subject to training with sufficient labeled samples.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Decision tree-based diagnosis models are naturally interpreted, may not rely on expert explanation, and can be easily converted into diagnosis rules.

“The decision tree-based diagnosis models are naturally interpreted, which may not relay on the explanation of experts and can be easily converted into diagnosis rules.”

May combine two claims.

As unsupervised learning method, stacked AE cannot be directly used to recognize health states of machines; a classification layer is usually added, and models need sufficient labeled samples.

“As an unsupervised learning method, the stacked AE cannot be directly used to recognize the health states of machines.Therefore, a classification layer is usually added at the top of the architecture of the model, and the constructed diagnosis models need to be trained with sufficient labeled samples.”

May combine two claims.

ResNet-based diagnosis models inherit advantages of CNN-based models and possibly obtain high diagnosis accuracy, especially for complicated operation conditions such as varying-speed or varying-load.

“Therefore, the ResNet-based diagnosis models inherit the advantages of the CNN-based diagnosis models, and possibly obtain high diagnosis accuracy, especially for the complicated operation conditions such as varying-speed or varying-load conditions.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Decision tree-based diagnosis models could achieve diagnosis tasks with missing data.

“Furthermore, they could achieve the diagnosis tasks with missing data.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Number of training parameters in CNN-based diagnosis models is reduced by sharing weights, which accelerates convergence and restrains overfitting.

“Furthermore, the number of training parameters in diagnosis models is reduced by sharing weights, which could accelerate the convergence and restrain the over fitting.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Output layer is generally ANN-based classifier because of high capability in multi-class classification.

“Furthermore, the output layer is placed after the last extraction layer for health state recognition, generally with an ANN-based classifier because of the high capability in multi-class classification.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

DBN-based diagnosis models could automatically learn features from input data by pre-training a set of stacked RBMs, which solves problem of vanishing gradient in using BP algorithm to fine-tune deep-layer networks.

“Different from the stacked AE, DBN-based diagnosis models could automatically learn features from the input data by pre-training a set of stacked RBMs, which solves the problem of vanishing gradient in using BP algorithm to fine-tune the deep-layer networks.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

It is necessary to use sufficient labeled data to train DBN-based diagnosis models to obtain convinced diagnosis results.

“It is necessary to use sufficient labeled data to train the constructed diagnosis models so as to obtain the convinced diagnosis results.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Deep learning-based diagnosis models automatically learn features from input monitoring data and simultaneously recognize health states according to learned features.

“Deep learning-based diagnosis models automatically learn features from the input monitoring data and simultaneously recognize the health states of machines according to the learned features.”

May combine two claims.

Many monitoring points and multi-source sensors are used, data stored with various structures.

“3) In order to collect sufficient health information of machines, there are many monitoring points on machines.Furthermore, multisource sensors are used to collect different kinds of data.The collected data are stored with various data structures.”

• cites [314] 314. G.I. Platforms, The rise of industrial big data, 2012, Available: http://www.geautomation.com/download/rise-industrial-big-data.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [315] 315. S. Yin, X. Li, H. Gao, O. Kaynak, Data-based techniques focused on modern industry: An overview, IEEE Trans. Ind. Electron. 62 (2015) 657-667. sourceNot checked — no full text of the source
• cites [316] 316. J. Lee, Industrial big data: The revolutionary transformation and value creation in industry 4.0 era, Mechanical Industry Press, Beijing, China, 2015.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Decision tree-based diagnosis models are easily caught by overfitting and low generalization performance, reducing diagnosis performance.

“However, this kind of diagnosis models is easily caught by the over fitting and the low generalization performance, which would reduce the diagnosis performance of the models on the diagnosis tasks.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Advent of edge computing and GPU helps cope with monitoring data stream effectively.

“Furthermore, the advent of state-of-art technologies, such as the edge computing and the applications of GPU, helps cope with monitoring data stream effectively.”

• cites [314] 314. G.I. Platforms, The rise of industrial big data, 2012, Available: http://www.geautomation.com/download/rise-industrial-big-data.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [315] 315. S. Yin, X. Li, H. Gao, O. Kaynak, Data-based techniques focused on modern industry: An overview, IEEE Trans. Ind. Electron. 62 (2015) 657-667. sourceNot checked — no full text of the source
• cites [316] 316. J. Lee, Industrial big data: The revolutionary transformation and value creation in industry 4.0 era, Mechanical Industry Press, Beijing, China, 2015.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Tree-type models are mostly constructed according to expert knowledge.

“In addition, the tree-type models are mostly constructed according to the expert knowledge.”

Classified as probably not a substantial claim.

Deep learning constructs end-to-end diagnosis models to automatically learn features from collected data and recognize machine health states.

“The deep learning theories construct end-to-end diagnosis models to automatically learn features from the collected data, and subsequently recognize the health states of machines.”

May combine two claims.

GAN-based approaches can generate fake labeled data similar to actual target-domain data, which can help train reliable diagnosis models for target domain, and are not affected by availability of labeled target samples, so expected to work for TIM and TDM.

“GAN-based approaches are able to generate fake labeled data that are similar to the actual data in the target domain.These fake data are able to help train reliable diagnosis models for the target domain.Therefore, such approaches are not affected by the condition whether labeled samples in the target domain are available, and they are expected to work for both the TIM scenarios and the TDM”

May combine two claims.

Feature-based and GAN-based approaches are widely concerned because they can accomplish TDM scenarios with serious cross-domain discrepancy; instance-based and parameter-based approaches are easily implemented and mostly focus on TIM scenarios with slight discrepancy; transfer performance may relate to number of labeled target samples.

“Among them, feature-based approaches and GAN-based approaches are widely concerned because they can accomplish TDM scenarios subject to serious cross-domain discrepancy.In contrast, the instance-based approaches and parameter-based approaches are easily-implemented, and mostly focus on TIM scenarios subject to slight cross-domain discrepancy.Furthermore, the transfer performance of these”

May combine two claims.

Instance-based approaches are easily implemented, but their transfer performance relates to number of target-domain samples, and they can only correct discrepancy in TIM scenarios, not serious discrepancy such as TDM, because they may lack powerful data fitting capability.

“Instance-based approaches are easily-implemented for transfer scenarios.However, the transfer performance of them relates to the number of target-domain samples.Furthermore, they just are able to correct discrepancy in TIM scenarios, but incapable to implement transfer scenarios with serious discrepancy, such as TDM scenarios, because these approaches may lack the powerful capability of data”

May combine two claims.

Deep learning-based diagnosis models are black boxes; it is difficult to acquire how they learn diagnosis knowledge from monitoring data, so they are constructed by experimental trials rather than strict theoretical background.

“Although there are state-of-the-art achievements of deep learning-based diagnosis models, an open issue of black box for deep hierarchical networks still confuses the academic researchers.It is difficult to acquire how these models learn the diagnosis knowledge from the monitoring data.As a result, the diagnosis models are constructed by experimental trials once and once again rather than the”

May combine two claims.

Deep learning-based diagnosis models are beneficial to reducing cross-domain discrepancy by extracting deep-layer features, but they are unable to adapt serious cross-domain discrepancy in some scenarios such as TDM.

“Deep learning-based diagnosis models are beneficial to reducing the cross-domain discrepancy by extracting deep-layer features [33].However, they are unable to adapt serious cross-domain discrepancy in some diagnosis scenarios, such as TDM.”

• cites [33] 33. M. Long, J. Wang, Y. Cao, J. Sun, P.S. Yu, Deep learning of transferable representation for scalable domain adaptation, IEEE Trans. Knowl. Data Eng. 28 (2016) 2027-2040. sourceNot checked — no full text of the source

Classified as probably not a substantial claim. May combine two claims.

Selecting optimal transfer approach and guaranteeing transfer success takes much time by experimental trials.

“To select the optimal one and guarantee the transfer success, it must take much time by experimental trials.”

Probably not a substantial claim: may not pass S4 (based).

ANN-based diagnosis models are black-boxed due to the lack of rigorous theoretical supports.

“Second, the ANN-based diagnosis models are black-boxed due to the lack of rigorous 10 theoretical supports.”

May combine two claims.

Imbalanced distribution in machine health states is common in engineering scenarios; healthy data are far more sufficient than fault data; training deep learning models on imbalanced data may shift decision boundary toward minority health states and reduce diagnosis accuracy.

“Imbalanced distribution in health states of machines is a common phenomenon in engineering scenarios.For example, the data collected from the healthy state are far more sufficient than those from the faults.If the imbalanced data are used to train the deep learning-based diagnosis models, the decision boundary of the models might be enforced to shift towards the health states with minority of”

Probably not a substantial claim: may not pass S2 (checkable).

Transfer learning is promising to expand IFD from academic research to engineering scenarios.

“Transfer learning is promising to expand IFD from academic research to engineering scenarios.”

May combine two claims.

Data collected from real-case machines are seriously imbalanced because machines mostly work healthy and faults seldom happen, making healthy data easier to collect than faulty data.

“The fact is that machines mostly work under the healthy state, while the faults seldom happen.Thus, it is easier to collect healthy data than faulty data.As a result, the collected data are seriously imbalanced.”

• cites [26] 26. B. Yang, Y. Lei, F. Jia, S. Xing, An intelligent fault diagnosis approach based on transfer learning from laboratory bearings to locomotive bearings, Mech. Syst. Signal Process. 122 (2019) 692-706. sourceNot checked — no full text of the source

May combine two claims.

The majority of data collected from real-case machines are unlabeled because frequently stopping machines to inspect health states is unrealistic due to huge financial loss.

“The majority of the collected data are unlabeled.It is unrealistic to frequently stop the machines and inspect the health states due to the huge loss of wealth.”

• cites [26] 26. B. Yang, Y. Lei, F. Jia, S. Xing, An intelligent fault diagnosis approach based on transfer learning from laboratory bearings to locomotive bearings, Mech. Syst. Signal Process. 122 (2019) 692-706. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S2 (checkable).

Transfer learning theories are promising to construct diagnosis models in which diagnosis knowledge can be transferred across multiple diagnosis tasks.

“To bridge the gap, transfer learning theories are promising to construct diagnosis models, in which the diagnosis knowledge can be transferred across multiple diagnosis tasks.”

May combine two claims.

Deep learning-based diagnosis models are trained by minimizing empirical risk, which lacks rigorous theory, so their physical meanings are difficult to interpret.

“The deep learning-based diagnosis models are trained by minimizing empirical risk, which really lacks rigorous theory.As a result, the physical meanings of them are difficult to be interpretable.”

May combine two claims.

It is possible to collect data from multiple source domains to match a common target domain, but health information in these source domains is not all related to target domain in engineering scenarios.

“It is possible to collect data from multiple source domains to match a common target domain, but the health information contained in these source domains is not all related to the that in the target domain in engineering scenarios.”

Classified as probably not a substantial claim.

Transfer learning can reuse knowledge from one or more diagnosis tasks to other related but different ones.

“Transfer learning is able to get the above purpose, in which the knowledge from one or more diagnosis tasks can be reused to other related but different ones [27].”

• cites [27] 27. S.J. Pan, Q. Yang, A survey on transfer learning, IEEE Trans. Knowl. Data Eng. 22 (2010) 1345-1359. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based). May combine two claims.

The process of learning features through deep learning is similar to filtering process, so adaptive filter theory is beneficial to analyzing physical meanings of deep learning-based diagnosis models.

“The process of learning features through deep learning is similar to the filtering process.Therefore, the adaptive filter theory is beneficial to analyzing the physical meanings of deep learning-based diagnosis models.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Constructed transfer learning-based diagnosis models are incapable to extract related health information.

“The constructed transfer learning-based diagnosis models are incapable to extract the related health information.”

Probably not a substantial claim: may not pass S4 (based).

Transfer performance of different approaches is different for a given transfer scenario due to diverse performance ceilings.

“According to Section 4.3, several approaches can be used in IFD, but the transfer performance of them is different for a given transfer scenario due to diverse performance ceilings.”

Classified as probably not a substantial claim.

In the past, IFD was implemented by data collection, artificial feature extraction, and health state recognition.

“In the past, IFD is implemented by the steps of data collection, artificial feature extraction, and health state recognition.”

Probably not a substantial claim: may not pass S4 (based).

With big data revolution, imbalanced health states and analytics for data stream bring negative effects on IFD using deep learning.

“With the revolution of big data, two special issues come to bring negative effects on the IFD using deep learning theories, i.e., imbalanced health states and analytics for data stream.”

May combine two claims.

Diagnosis models based on TCA and JDA use simple nonlinear mapping to extract features, which is difficult to fit complicated data distribution, so they may get poor transfer results on target domain due to under-corrected discrepancy.

“Diagnosis models based on TCA and JDA use the simple nonlinear mapping to extract features, which is difficult to fit the complicated distribution of data.As a result, the diagnosis models may get poor transfer results on the target domain because of the under-corrected discrepancy of the source and target domains.”

May combine two claims.

Unreliable data transmission and bandwidth limitations prevent monitoring data stream from arriving in unbroken sequence, and inefficient computation impedes data stream analytics, so IFD using deep learning is just implemented based on off-line historical data.

“Such issue has been concerned by the researchers for years, but the unreliable data transmission and the bandwidth limitations prevent the monitoring data stream from arriving the destination in an unbroken sequence.Furthermore, the inefficient computation capability impedes the success of data stream analytics.As a result, the IFD using deep learning is just implemented based on the off-line”

May combine two claims.

Parameter-based approaches fine-tune pre-trained diagnosis models for target-domain fault diagnosis, consuming less computation resources for massive data, but their transfer performance mostly depends on number of labeled target samples.

“Parameter-based approaches fine-tune the pre-trained diagnosis models for fault diagnosis tasks in the target domain, which consumes less computation resources in handling massive data.However, the transfer performance of these approaches mostly depends on the number of labeled samples in the target domain.”

May combine two claims.

Traditional machine learning diagnosis models can automatically recognize machine health states, but artificial feature extraction still relies on expert knowledge.

“By using traditional machine learning theories, the diagnosis models are able to automatically recognize the health states of machines.However, the literature review presents that the artificial feature extraction still relies on the expert knowledge.”

May combine two claims.

TCA adapts only the marginal probability distribution of input cross-domain data and ignores the conditional probability distribution from feature space to target classes.

“TCA just adapts the marginal probability distribution of the input cross-domain data, but 8 ignores the conditional probability distribution from the feature space to the target classes.Thus,”

• cites [30] 30. M. Long, J. Wang, G. Ding, J. Sun, P.S. Yu, Transfer feature learning with joint distribution adaptation, in: IEEE International Conference on Computer Vision, Sydney, Australia, 2013, pp. 2200-2207. sourceNot checked — no full text of the source

Classified as probably not a substantial claim. May combine two claims.

Statistical learning theories such as SVM and PGM have rigorous theory grounds, promoting construction of diagnosis models with easily-understood parameters, features, and results.

“The statistical learning theories, such as SVM and PGM, have rigorous theory grounds, which promote to constructed diagnosis models with easily-understood model parameters, features, and diagnosis results.”

May combine two claims.

Recollecting sufficient data and labeling them is costly and unpractical for machines in engineering scenarios.

“However, it takes much cost to recollect sufficient data and further label them, which is unpractical for machines in engineering scenarios.”