A New Model for Remaining Useful Life Prediction Based on NICE and TCN-BiLSTM under Missing Data

Round 1

Reviewer 1 Report

General comments

The work presented is very interesting. However, there are some comments related to this work. 

Section dedicated comments

Abstract

·         Full words that define abbreviations should be capitalized. For example: Remaining Useful Life (RUL)

 

Introduction

·      Add another section ‘’Literature Review’’ in order to refer others’ work and methodologies in order to identify important gaps that the present approach deals with. It should be direct to the reader, at the end, what are the gaps and the key contributions of the present work in comparison to the identified gaps of other methods. Here, for example “Variant networks based on the structure of an RNN [27] include LSTM [28], BiRNN [29], and Bi-LSTM [13,30], and variant networks based on the CNN structure are mainly DCNN [31] and TCN [32,33]. Wang et al. [13] proposed a new data-driven method using the BiLSTM network for RUL estimation, which can make full use of bidirectional sensor data sequences.” is not clear to the reader what these methods do or why your method is better. Maybe you should add more papers regarding RUL prediction methods and investigate more the steps that these methods propose.

 

Experimental Research

·        You could add another section ‘’Implementation’’ in order to describe your software system you implemented. For example, what technologies did you choose?

 

Conclusion

·         Next steps and future work should be explained in detail.

 

References

It is suggested to add a more extended literature review. This literature can strengthen the significance of the paper in academia and industrial worlds. Focus should be given in providing an adequate review in the following topics:

o   https://doi.org/10.3390/s21030972

o   https://doi.org/10.1016/j.ress.2021.107560

o   https://doi.org/10.1016/j.cie.2019.106024

o   https://doi.org/10.1016/j.compind.2020.103383

Author Response

Response to Reviewer 1 Comments

 

General comments

The work presented is very interesting. However, there are some comments related to this work.

Section dedicated comments

 

Point 1 (Abstract): Full words that define abbreviations should be capitalized. For example: Remaining Useful Life (RUL).

 

Response 1: Thanks to the reviewers for their valuable comments, we have carefully checked and all abbreviations have been revised.

 

Point 2 (Introduction): Add another section ‘’Literature Review’’ in order to refer others’ work and methodologies in order to identify important gaps that the present approach deals with. It should be direct to the reader, at the end, what are the gaps and the key contributions of the present work in comparison to the identified gaps of other methods. Here, for example “Variant networks based on the structure of an RNN [27] include LSTM [28], BiRNN [29], and Bi-LSTM [13,30], and variant networks based on the CNN structure are mainly DCNN [31] and TCN [32,33]. Wang et al. [13] proposed a new data-driven method using the BiLSTM network for RUL estimation, which can make full use of bidirectional sensor data sequences.” is not clear to the reader what these methods do or why your method is better. Maybe you should add more papers regarding RUL prediction methods and investigate more the steps that these methods propose.

 

Response 2: Thanks to the valuable advice of the reviewers, the ' 2 Literature Review ' section was added through content adjustments.( Because of the length is too long, here is not repeated)

 

Point 3 (Experimental Research): You could add another section ‘’Implementation’’ in order to describe your software system you implemented. For example, what technologies did you choose?

 

Response 3: Thanks to the valuable comments of the reviewers, the ' 5.1 Implementation ' section has been added through content adjustments, as follows :

“In order to test and validate the potential contribution of the proposed approach for future real-world applications, this method has been implemented into a prototype software system using Python 3.6.13. In particular, the NICE and TCN-BiLSTM model are implemented using Keras 2.3.1, a Python library for developing and evaluating deep learning models. The resources used in order to integrate the aforementioned system were a computer with an Intel i7 processor (Intel(R) Core(TM) i7-10770K CPU @3.80 GHz 3.79 Ghz), regarding the processing power, and an 128 gigabyte RAM memory. The operating system that the proposed system was hosted and tested on was Microsoft Windows 10.”

 

 

Point 4 (Conclusion): Next steps and future work should be explained in detail.

 

Response 4: Thanks to the valuable comments of the reviewers, the Next steps and future work section has been added through content adjustments, as follows :

“In the future research, we will further consider the influence of the complex relationship of the actual environment on the generation of unbalanced and incomplete non-ideal data. Although this method has obtained good experimental results, further architecture optimization is still necessary, because the current training time is longer than most shallow networks in the literature.”

 

Point 5 (References): It is suggested to add a more extended literature review. This literature can strengthen the significance of the paper in academia and industrial worlds. Focus should be given in providing an adequate review in the following topics:

o   https://doi.org/10.3390/s21030972

o   https://doi.org/10.1016/j.ress.2021.107560

o   https://doi.org/10.1016/j.cie.2019.106024

o   https://doi.org/10.1016/j.compind.2020.103383

 

Response 5: Thanks to the valuable comments of the reviewers, Papers submitted by experts are reflected in references.

 

 

Author Response File: Author Response.pdf

Reviewer 2 Report

This paper proposes a multivariate degradation device based on nonlinear independent components estimation (NICE) and the temporal convolutional network–bidirectional long short-term memory (TCN-BiLSTM) network for the RUL prediction requirements in the case of missing data. RUL prediction of aeroengines is an example to perform multivariate degradation data-filling and prediction tasks.

-          Problem is clear formulated in sec. 2.

-          Proposed method is detail presented in sec. 3-4.

-          In tables 1 and 3, how to separate the data into subdata.

-          Caption size of figures should be similar to the text.

-          Resolution and quality of figures should be enhanced.

-          Advantages and disadvantages should be provided in discussion section.

-          References should be reformatted in the list.

Author Response

Response to Reviewer 2 Comments

 

This paper proposes a multivariate degradation device based on nonlinear independent components estimation (NICE) and the temporal convolutional network–bidirectional long short-term memory (TCN-BiLSTM) network for the RUL prediction requirements in the case of missing data. RUL prediction of aeroengines is an example to perform multivariate degradation data-filling and prediction tasks.

 

-          Problem is clear formulated in sec. 2.

 

-          Proposed method is detail presented in sec. 3-4.

 

-          In tables 1 and 3, how to separate the data into subdata.

 

-          Caption size of figures should be similar to the text.

 

-          Resolution and quality of figures should be enhanced.

 

-          Advantages and disadvantages should be provided in discussion section.

 

-          References should be reformatted in the list.

 

 

Thanks for the seven valuable opinions of the reviewers, the first two points are the affirmation of my work. At the same time, the following five revised responses are as follows :

 

 

Point 1 : In tables 1 and 3, how to separate the data into subdata.

 

Response 1: Both Table 1 and Table 3 use the published aero-engine guideline experimental dataset C-MAPSS, where Table 1 includes four sets of data sets for different operating conditions and failure modes. Table 3 takes the data set numbered FD001 as an example, and refers to [1] .Specifically, 100 engines are processed by multi-dimensional sliding time window with window size of 30 through programming. (see Figure 9)The training set is the life cycle data from operation to failure, which is 26631 cycles before processing and 17731 cycles after processing. The test set is the cycle data running to a certain location, with 13096 cycles before processing and 100 cycles after processing.

 

• Li X, Ding Q, Sun J Q. Remaining useful life estimation in prognostics using deep convolution neural networks[J]. Reliability Engineering & System Safety, 2018, 172: 1-11.

 

 

Point 2: Caption size of figures should be similar to the text.

 

Response 2: Thanks to the valuable comments of the reviewers, specifically, the Caption size of figures1, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12 have been resized to look similar to the text.

 

 

Point 3: Resolution and quality of figures should be enhanced.

 

Response 3: Thanks to the valuable comments of the reviewers, specifically, the resolution and quality of figures1, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12 have been resized to look similar to the text.

 

 

Point 4: Advantages and disadvantages should be provided in discussion section.

 

Response 4: We added an analysis of the advantages and disadvantages of the proposed method in the experimental analysis of the revised draft. For details, see the revised draft:

“The experimental results in Table 5 show that the proposed TCN-BiLSTM structure is very suitable for multidimensional degradation data prediction. The TCN network is the first structure to use multidimensional feature extraction, which can quickly extract local features in parallel. The BiLSTM network is the second structure using circular information flow, and the integrated BiLSTM layer helps to improve the learning ability of the network. Because the proposed model is a combined model, the training time is longer than most shallow networks in the literature, so there is a problem of slow training speed. However, in terms of prediction accuracy and reliability, the proposed combined model further confirms the superiority of using parallel structure to extract original degradation features and cyclic structure to mine sequence hidden features.”

 

 

 

Point 5: References should be reformatted in the list.

 

Response 5: We re-added some literature and checked the format.

 

 

Author Response File: Author Response.pdf

Reviewer 3 Report

Authors insist that the paper proposes a data generation way entrenched on the NICE model, which may attain good generation results. That is,

-      NICE technology may be the true distribution laws behind lost data, map training data to a standard normal distribution, produce realistic data by sampling, and then fill up in lost values. So, multivariate degradation data may be produced in the full-time sense

-      Method can catch both long-term and short-term dependencies, successfully making sure that the taken-out attributes fully indicate the health status of the device

 

After reading this paper, I recommend to modify it as following:

1)         First of all, abstract and introduction is not concise and hard to figure it out. It is required to make it concise or structured. That is, it gives more paragraph after separating it for study issues in introduction.

2)         Methodology, that is, section 2, section 3, section 4, section 5, should be combined with one topic. Results such as section 5.4 are mixed and combined in section 5. So, it is hard to figure it out. I recommend that results and discussions part should be added.

3)         Conclusions part make concise and add more paragraph.

4)         Additionally, you should have to check spelling. I found a lot of errors.

Author Response

Response to Reviewer 3 Comments

 

Authors insist that the paper proposes a data generation way entrenched on the NICE model, which may attain good generation results. That is,

-      NICE technology may be the true distribution laws behind lost data, map training data to a standard normal distribution, produce realistic data by sampling, and then fill up in lost values. So, multivariate degradation data may be produced in the full-time sense

 

-      Method can catch both long-term and short-term dependencies, successfully making sure that the taken-out attributes fully indicate the health status of the device

 

After reading this paper, I recommend to modify it as following:

 

1)         First of all, abstract and introduction is not concise and hard to figure it out. It is required to make it concise or structured. That is, it gives more paragraph after separating it for study issues in introduction.

2)         Methodology, that is, section 2, section 3, section 4, section 5, should be combined with one topic. Results such as section 5.4 are mixed and combined in section 5. So, it is hard to figure it out. I recommend that results and discussions part should be added.

3)         Conclusions part make concise and add more paragraph.

4)         Additionally, you should have to check spelling. I found a lot of errors.

 

 

 

Point 1 : First of all, abstract and introduction is not concise and hard to figure it out. It is required to make it concise or structured. That is, it gives more paragraph after separating it for study issues in introduction.

 

Response 1: Thanks to the valuable opinions of reviewers, the ' 2 literature review ' part was added through content adjustment. ( Due to the length is too long, here will not be repeated )

 

Point 2: Methodology, that is, section 2, section 3, section 4, section 5, should be combined with one topic. Results such as section 5.4 are mixed and combined in section 5. So, it is hard to figure it out. I recommend that results and discussions part should be added.

 

Response 2: Thanks to the valuable advice of the reviewers, the ' 2 Literature Review ' section was added ( Because of the length is too long, here is not repeated) , and adjust the methodology to reclassify the full text structure :“Section 1 and 2 are Introduction and Literature Review, respectively. Section 3 describes the Problem Formulation. Section 4 introduces the multivariate degradation data filling model based on NICE model and the multivariate degradation data forecasting model based on TCN-NICE model. Section 5 is the example validation part, choosing multivariate degra-dation data set C-MAPSS as validation data set. Section 5.1 gives the implementation of this experiment. Section 5.2 carries out the data set and processing work. Section 5.3 ap-plies the NICE model to the multivariate degradation data filling task. Section 5.4 applies the TCN-BiLSTM model to the multivariate degradation data filling task. Section 6 is the conclusion part, which summarizes the innovation points of this paper.”

 

And results and discussion sections were added as follows :(see section 5.5.4)

“The experimental results in Table 5 show that the proposed TCN-BiLSTM structure is very suitable for multidimensional degradation data prediction. The TCN network is the first structure to use multidimensional feature extraction, which can quickly extract local features in parallel. The BiLSTM network is the second structure using circular infor-mation flow, and the integrated BiLSTM layer helps to improve the learning ability of the network. Because the proposed model is a combined model, the training time is longer than most shallow networks in the literature, so there is a problem of slow training speed. However, in terms of prediction accuracy and reliability, the proposed combined model further confirms the superiority of using parallel structure to extract original degradation features and cyclic structure to mine sequence hidden features.”

 

 

Point 3: Conclusions part make concise and add more paragraph.

 

Response 3: For the conclusion, we add the following :

“In the future research, we will further consider the influence of the complex relationship of the actual environment on the generation of unbalanced and incomplete non-ideal data. Although this method has obtained good experimental results, further architecture optimization is still necessary, because the current training time is longer than most shallow networks in the literature.”

 

 

Point 4: Additionally, you should have to check spelling. I found a lot of errors.

 

Response 4: Thanks to the valuable comments of the reviewers, we have carefully checked, modified and corrected some spelling errors.

 

 

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

All issues are raised. I recommend to publish it as current form.