A Robust Health Prognostics Technique for Failure Diagnosis and the Remaining Useful Lifetime Predictions of Bearings in Electric Motors

Round 1

Reviewer 1 Report

Title
- try to avoid using the shortcut in the title of the paper (see RUL).
Abstract
- it is well formulated; the purpose of the research is clear;
- however, maybe the terms (shortcuts?) IMS, FEMTO XJTU-SY need to be clarified (in a single sentence with explanations).
Introduction
- firstly, refers to the causes of the motor failures (bearing, stator, rotor, other);
- then, the author describe the importance of the maintenance actions (reactive, predictive);
- the four steps for RUL prediction are clear;
- also, the purpose of the paper is clear;
- here, I do not know if the "medical" term "health" can be so easy combined with terms as "prognostics", stage etc.,
in technical/engineering domain.
Background
- describes the four steps of the RUL prediction, in the approach of the authors and in connection with different studies from literature.
Proposed solution
- the four steps of the prediction are very suggestive presented in Fig. 1;
- the equations and the normalization procedure are well presented;
- same for health and damaged stages (Fig. 3) and bidirectional LSTM structure (Fig. 4).
Results
- firstly, presents the test phases (motor operation conditions), including the dataset information;
- maybe, from Table 4, the negative values of the prediction error, need additional explanations.
Conclusions and future work
- for me, it is not clear the meaning of the "robust health prognostics" in comparison with "health prognostics".

The References are suggestive for the topic. Random check for: 2. A. J. Bazurto; 6. Y. Li; 8. Y. Lei; 14. Y. Liu;
15. S. Zhang; 22. M. Motahari-Nezhad; 28. R. Cao; 34. J. S. L. Senanayaka; 40. S. Kumar; 51. B.-L. Lu.

Author Response

We would like to thank the reviewer for their comments and constructive suggestions, which really contribute much to the improvement of this manuscript. We have carefully considered them to revise the manuscript. Responses to the reviewer's comments are included in the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

In this paper, a novel procedure for Remaining Useful Life (RUL) of motor bearings is presented. It integrates frequency-domain signal analysis and a stacked autoencoder (SAE) with a bidirectional long short-term memory (BiLSTM)neural network.

The subject of the paper is of high interest, however there are some issues I would like to point to:

1.       The first phase, i.e. feature extraction is based on simple FFT of the collected vibration signal. It is known for a very long time that “pure” FFT is not suitable to obtain frequency spectrum in such cases, because in real situation the useful part of the bearing vibration signal is hidden and covered by noise and disturbances, which usually have much higher amplitude than actual vibration signal needed for fault detection and diagnosis. Therefore, some more advanced techniques are needed for feature extraction. Also, in my opinion, this part does not contain any novelty.

2.       In the text explaining Fig. 2 (p.6, lines 212-224), the coding of features is explained, but there is no explanation for decoding process.

3.       P.7, lines 231-233, how did you adopt this threshold? There is no explanation or reference.

4.       Fig. 8a is not at all in accordance with the text in p. 12, lines 332-335. Additionally, Fig. 8a needs to be checked and improved (the legend covers the important part of the diagram).

Finally, the authors state several times that “not only does the proposed robust health prognostics technique predict the remaining useful lifetime of the bearings of electric motors, but it is also useful for detecting outer-race bearing failures”. However, this technique detects ONLY this type of fault and it is the huge flaw of this procedure, not the advantage. As authors suggest at the end of the paper, this procedure does not detect any other fault and it is not applicable for general use, where bearings may have some other type of fault. Other than that, all presented techniques are well known state-of-the-art and there is no much (or any) novelty in the presented procedure.

Author Response

We would like to thank the reviewer for their comments and constructive suggestions, which really contribute much to the improvement of this manuscript. We have carefully considered them to revise the manuscript. Responses to the reviewer's comments are included in the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

The authors made substantial effort to respond to the review. However, the suggestion 2.1 (as marked in the cover letter) is still an issue. I agree that SAEs are efficient in noise reduction and feature fusion, but not when applied to three numerical values (amplitudes of characteristic frequencies) obtained by simple FFT of the vibration signal. Some more advanced signal analysis technique must be applied to extract these features in order to make this procedure applicable to real fault detection and RUL estimation. Envelope analysis is the least one can introduce to preprocess the raw signal and extract features with higher accuracy and reliability. In the opposite, the input data for SAEs is of insufficient quality for real-world application.

Author Response

We would like to thank the reviewer for their comments and constructive suggestions, which really contribute much to the improvement of this manuscript. We have carefully considered them to revise the manuscript. Responses to the reviewer comments can be seen in the attachment.

Author Response File: Author Response.pdf