Thermal-Imaging-Based PCA Method for Monitoring Process Temperature

Round 1

Reviewer 1 Report

The 2000 sample number is considered enough?

More real models and tests are required, including real industrial tanks.

Which are the limitations of your SIPCA system, in terms of resolution?

What other applications are suitable for SIPCA?

Which is the minimal temperature gradient recorded?

Author Response

Reviewer 1:

1. The 2000 sample number is considered enough?

Reply: The amount of training data depends on the complexity of the monitored object. The experimental tank system is not very complex and hence 2000 samples are enough for training model.

 

2. More real models and tests are required, including real industrial tanks.

Reply: That’s a good suggestion. Indeed, the experimental tank system is very close to the real chemical heating process, and this paper verified the feasibility of SIPCA. Of course, our future work will apply SIPCA in real industrial processes.

 

3. Which are the limitations of your SIPCA system, in terms of resolution?

Reply: That’s a good question. Indeed, the SIPCA method proposed in this paper is based on 2-D image, and hence it cannot detect the faults which occur on the back of the monitored object. We have added the following words in section conclusion:

“One drawback of SIPCA is that it’s based on 2-D image, and hence it cannot detect the faults which occur on the back of the monitored object. For handling this issue, our future work will focus on the combination SIPCA with the multi-view 3D reconstruction.”

 

4. What other applications are suitable for SIPCA?

Reply: Though this paper adopted SIPCA in thermal monitoring, it can also be used in detection of faults which occur on the process surface.

 

5. Which is the minimal temperature gradient recorded?

Reply: Due to the fluctuation of water flow, the change of water temperature is also fluctuating. In the experiment test, the normal temperature is around 45°C, and it will change to 40°C in 20 seconds after code water inlet changes. As such, the minimal temperature gradient recorded is -0.25°C/s.

Reviewer 2 Report

The authors researched a new method for monitoring and thermal fault detection in processes. The method is based on a thermal camera as a sensor, and the newly proposed method called a spatial information-based principal component analysis (SIPCA). To validate the proposed method, they compared the results between the regular PCA method (data obtained from six temperature sensors) and SIPCA (used a thermal camera for data processing).

 

This is an interesting research report. In my opinion, the subject is worth pursuing. The method is interesting and has great potential, but the manuscript is rushed and needs changes before publication.

 

The major concern:

1. Performing the experiment, you compared SIPCA with a "regular" PCA method, but for the SIPCA experiment, you used a thermal camera, and for the other method, you used temperature sensors (six temperature sensors). Also, from your experiment, it is unclear whether the thermal camera is moving. If not, then what is the point of determining the fault location? Further, there are no metrics that confirm the supremacy of the proposed method.

Some other concerns:

1. L82-L85; What is the point of accurately (pixel-level) detecting the source of temperature anomaly in processes? From my experience, such precision is unneeded. Please, state what processes you refer to for such precision.
2. Can you explain equation 2 in more detail? I checked literature reference 22, and there is no such term. Also, it could be confusing to denote the score matrix (T) the same as the matrix operation of transpose. Please, for the sake of clarity, consider renaming it. 
3. L112; What is the purpose of determining the coefficient "k" if you already know the dimensions of matrices T and P?
4. L116; Please, define metrics T^2 and SPE for equations 4 and 5.

 

Chapter 4 needs more polishing.

1. Please, write down what equipment you used in the experiment. (thermal camera; acquisition equipment; software for processing, etc.)
2. What are the regular parameters for water? Further, what is threshold temperature?
3. L187; What are the temperature lower and upper values?
4. L188; Please, define "certain temperature range"?
5. L195; How did you secure that fault (for all three experiments) always occurred in the 501st sample?
6. Figure 4 is the same as Figure 5. Please, correct it.
7. Please, explain figures 4 to 5 in more detail. For instance, the y-axis values on the left and right figures are different, and it can be observed that threshold values are different. The authors need to explain.
8. L219; Please, define what does mean the following "...and hence it detects the fault in time."? Define the term "in time"? Do other algorithms fail to detect a fault in time?
9. L236-L238; First, Figure 8 is blurred. 
10. What is the accuracy of located area (referring to L82)?
11. Did you determine the confusion matrix? Please, provide it.
12. Please, include the potential limits of the SIPCA in conclusion.
13. Please, mention the future work also.

 

Author Response

Reviewer 2:

 

The authors researched a new method for monitoring and thermal fault detection in processes. The method is based on a thermal camera as a sensor, and the newly proposed method called a spatial information-based principal component analysis (SIPCA). To validate the proposed method, they compared the results between the regular PCA method (data obtained from six temperature sensors) and SIPCA (used a thermal camera for data processing).

This is an interesting research report. In my opinion, the subject is worth pursuing. The method is interesting and has great potential, but the manuscript is rushed and needs changes before publication.

Reply: Thanks to the reviewer for his/her rigorous work. We revised the article according to the suggestions of the reviewer.

 

The major concern:

1. Performing the experiment, you compared SIPCA with a "regular" PCA method, but for the SIPCA experiment, you used a thermal camera, and for the other method, you used temperature sensors (six temperature sensors). Also, from your experiment, it is unclear whether the thermal camera is moving. If not, then what is the point of determining the fault location? Further, there are no metrics that confirm the supremacy of the proposed method.

Reply: We thank the reviewers for their professional and responsible questions. Firstly, the "regular" PCA method cannot be used for handling the thermal image because “the spatial information is missing” (explained in section 3.2), so we just use the temperature sensors (six temperature sensors) for PCA. Secondly, the thermal camera is fixed, and hence the pixel position and the device position are in one-to-one correspondence. Thirdly, for traditional monitoring methods, “(a) first, most traditional temperature sensors are contact-type, which cannot be installed in equipment/process with extreme high temperature, so the temperature values in some positions is unmeasurable; (b) second, for monitor equipment/process comprehensively, one needs to install temperature sensors over all positions of equipment/process in high density, which is expensive and difficult to implement; (c) third, heat leakage on the surface of the equipment/process cannot be detected by traditional PCA because usually sensors are only installed in the internal of devices and processes.” As such, the fault location ability of thermal image is meaningful, and this is also the supremacy of the proposed method.

 

Some other concerns:

1. L82-L85; What is the point of accurately (pixel-level) detecting the source of temperature anomaly in processes? From my experience, such precision is unneeded. Please, state what processes you refer to for such precision.

Reply: That’s a good question. For large scale industry processes, which contains hundreds of units, the pixel-level precision is meaningful, because each device only occupies a few pixels and hence we may fail to locate the fault when the precision is very low.

 

2. Can you explain equation 2 in more detail? I checked literature reference 22, and there is no such term. Also, it could be confusing to denote the score matrix (T) the same as the matrix operation of transpose. Please, for the sake of clarity, consider renaming it.

Reply: Reference 22 is for singular value decomposition (SVD) rather than for equation 2 (we have added reference 23 at the end of equation 2). T and P can be obtained by using singular value decomposition (SVD) to S and extract the eigenvectors corresponding to large eigenvalues. For PCA related papers, usually the score matrix is named as T in bold, and transpose is denoted as superscript T, and they will not be confused normally.

 

3. L112; What is the purpose of determining the coefficient "k" if you already know the dimensions of matrices T and P?

Reply: The main idea of PCA is to extract the important information in X, i.e., the score matrix T. As such, the dimension of T is much lower than that of X, i.e., k<s. The rest information is noise and they will be denoted as E. Because of the dimension reduction operation, we need to determine coefficient "k".

 

4. L116; Please, define metrics T^2 and SPE for equations 4 and 5.

Reply: We have added the following words after equations 4 and 5

“Statistic T^2 represents the distance between the location of the new data projected onto the subspace and the origin of subspace; statistic SPE is a measure of the ap-proximation error of the new data within the PCA subspace.”

 

Chapter 4 needs more polishing.

1. Please, write down what equipment you used in the experiment. (thermal camera; acquisition equipment; software for processing, etc.)

Reply: Good suggestion, we have added the following words:

“In this experiment, the following equipment are used: thermal camera for obtaining the thermal image; the heating module; and the temperature detection module, which includes 6 temperature sensors”

 

2. What are the regular parameters for water? Further, what is threshold temperature?

Reply: We have added the temperature description in the paper:

“the upper value (46°C)” and “the lower limit (44°C)”, which are also the threshold temperatures.

 

3. L187; What are the temperature lower and upper values?

Reply: We have added the temperature description in the paper:

“upper value (46°C)” and “the lower limit (44°C)”.

 

4. L188; Please, define "certain temperature range"?

Reply: for the position around the heating module, the temperature range is 44°C to 46°C. As the temperature decreases from the right part to the left part, the temperature range varies with position.

 

5. L195; How did you secure that fault (for all three experiments) always occurred in the 501st sample?

Reply: The fault is artificially controlled, and we can control the time when the fault occurs.

 

6. Figure 4 is the same as Figure 5. Please, correct it.

Reply: We have deleted the same figure.

 

7. Please, explain figures 4 to 5 in more detail. For instance, the y-axis values on the left and right figures are different, and it can be observed that threshold values are different. The authors need to explain.

Reply: Usually for data-driven monitoring algorithms (such as PCA), the y-axis value represents not a physical quantity, but a statistic, so the y-axis values and the threshold values vary with different algorithms. The definition has been given in equations (4) and (5).

 

8. L219; Please, define what does mean the following "...and hence it detects the fault in time."? Define the term "in time"? Do other algorithms fail to detect a fault in time?

Reply: For avoiding confusing, we replaced “in time” with “earlier”.

 

9. L236-L238; First, Figure 8 is blurred.

Reply: We adjusted the contrast of the image to make it clearer.

 

10. What is the accuracy of located area (referring to L82)?

Reply: the “accurately locate” means that the infrared thermal imaging can accurately grasp the high brightness pixels, and tell us that there is a heat leak at the corresponding device location.

 

11. Did you determine the confusion matrix? Please, provide it.

Reply: No “confusion matrix” in this paper.

 

12. Please, include the potential limits of the SIPCA in conclusion.

Reply: That’s a good suggestion. We have added the following words in section conclusion:

“One drawback of SIPCA is that it’s based on 2-D image, and hence it cannot detect the faults which occur on the back of the monitored object.”

 

13. Please, mention the future work also.

Reply: That’s a good suggestion. We have added the following words in section conclusion:

“For handling this issue, our future work will focus on the combination SIPCA with the multi-view 3D reconstruction.”

Round 2

Reviewer 2 Report

Dear authors,
Thank you for addressing all concerns and comments and improving the manuscript considerably.

Before publication, I still have one minor task for you to do.

Please, add the name and characteristics of the used equipment (thermal camera, temperature sensors, data acquisition equipment), which are important experimental setup information, especially for the scholars who will read (and try to reproduce the experiment) and compare your results with their work.

Author Response

Thanks. we have added the following words:

"In this experiment, the following equipment are used: thermal camera for obtaining the thermal image (Guide sensmartMobIR air, 25Hz, and 0.06°C accuracy); the heating module (CHB401, <0.5% full scale); and the temperature detection module, which includes 6 temperature sensors (SXPT100, 0.15°C accuracy)."