A Computationally Efficient Method for the Diagnosis of Defects in Rolling Bearings Based on Linear Predictive Coding
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsMonitoring and diagnosis of rolling bearing defects have been extensively studied. This paper presented an approach based on Linear Predictive Coding and a simple feedforward artificial neural network (NN). Both methods are not new thus making the contribution to the research very weak. A main issue is the lack of any insight into the fault detection mechanisms and corresponding signatures. A sound justification of the proposed approaches to this application is missing. The presentation of the verification results is not solid or convincing enough. Details of the algorithm implementation are not provided.
Some further comments:
The most effective and widely used bearing diagnosis approach is envelope analysis based. Authors should review and benchmark it with the proposed method.
In Eq. (1), why is the predicted value assigned with a "-" sign? When it comes to Eq. (2), the prediction error becomes actually a summation. Please double check it.
There are some errors in the reference list. Some DOI codes point to different articles than the intended.
More details for the signals shown in Fig. 2 should be provided, such as, What are the rotating speeds? fundamental parameters of the bearings?
The statement of "It can also be observed that the spectrum of these signals resembles a speech-like spectrum" must be justified. Do you mean in the time domain or frequency domain, or time-frequency domain? The validity is very doubtful actually.
Line 199: "At the first stage, features are extracted from the signals using the LPC algorithm 199 described in Section 2.1" - unfortunately I could not find what the mentioned features are in Section 2.1?
"the signal is 200 divided into short overlapping time frames" - what's the specific requirement for the time frame length determination? time? data points? relationships with bearing operational conditions? sampling rate?
Line 245: "The motor speed varies according to the applied load: 1797, 245 1772, 1750, and 1730 RPM respectively" - please clarify if the speed is constant for a recorded data sample, or the motor was just relating with a various speed?
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 2 Report
Comments and Suggestions for AuthorsManuscript ID: algorithms-3378486
Type of manuscript: Article
Title: A computationally efficient method for the diagnosis of defects 2 in rolling bearings based on Linear Predictive Coding
Comments:
The authors propose a classification method to fill the gap of available labeled defect data. In this paper, the authors propose a solution that uses a simple feedforward neural network (NN) for classification and adopts the linear predictive coding (LPC) algorithm for feature extraction. According to the authors, the LPC algorithm finds several coefficients for a given signal segment containing information about the signal spectrum, which is sufficient for further classification.
Authors should talk about AR model because it is an AR model. Also, authors should do a thorough study on the selection criteria of the order of AR model. Indeed, the number of coefficients depends on the number of spectral lines. If the number of coefficients is underestimated, there is a lack of significant lines in the spectrum. Conversely, if the number of coefficients is overestimated, parasitic spectral lines appear in the spectrum. It is therefore necessary to optimally estimate the number of coefficients of the model. The authors should provide more explanation on the choice of the LPC algorithm in the calculation of the coefficients. The authors should compare this algorithm to other algorithms such as Levinson, Burg, etc. There are several criteria for estimating the order of the model, including the AIC, FPE and MDL criteria.
In Figure 3, we see a modulation phenomenon around the frequency 1000 Hz that is no longer seen in the LPC spectrum. The authors should explain the phenomena that appear in the LPC spectrum to justify the number of coefficients of the model which is equal to 50. The authors should also show LPC spectra with defects to show the differences between the different spectra (healthy and defective).
The dimension reduction method should be justified and compared with other methods.
The authors used the CWRU and SUSU datasets. The authors should explain why the number of coefficients in both datasets remains at 50. The authors should also show the effect of the number of coefficients used at the input of the NN.
Certainly, the results obtained seem very satisfactory both from the point of view of the computation time and the precision obtained. Nevertheless, the authors should provide more explanation on the analysis of the number of coefficients of the LPC model at the input of the NN used.
This paper cannot be published as is. The authors should provide major explanations in the use of the LPC algorithm.
In general, I would suggest an improvement of the English of the document (but I myself am not a native English speaker).
I invite authors to consult the article below:
- Autoregressive Model-Based Structural Damage Identification and Localization Using Convolutional Neural Networks, KSCE Journal of Civil Engineering Volume 24, pages 2173–2185, (2020).
Comments for author File: Comments.pdf
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThe authors have addressed most of my concerns in the revision.
Reviewer 2 Report
Comments and Suggestions for AuthorsIn this second version, the authors have made changes corresponding to the remarks I made.
In particular :
The authors confirmed that the LPC model is an autoregressive model.
The authors provided explanations regarding other algorithms such as Levinson, Burg and other criteria to estimate the order of the model including AIC, FPE and MDL criteria.
The authors added LPC spectra to show the differences between the different spectra (healthy and defective).
The dimension reduction method used was justified and compared with other methods.
The authors showed the effect of the number of coefficients used at the input of the NN.
The work done by the authors corresponds well to the current problem of bearing defect detection by neural networks. In this paper, the authors propose a solution that uses a simple forward propagation artificial neural network (NN) for classification and adopts the linear predictive coding (LPC) algorithm for feature extraction. This method seems promising. Also, for this reason this article deserves to be published.
Comments for author File: Comments.pdf