A Road Condition Classification Algorithm for a Tire Acceleration Sensor using an Artificial Neural Network

Round 1

Reviewer 1 Report

Remarks:

It seems that the method of citing other authors is not appropriate. For example, in line 49 the name of author (Kanawar) appears at the beginning of the sentence, whereas reference to its work appears at the end of the sentence. Line 122 – “Multilayer perceptron” – rather multilayered perceptron Line 122 – MLP is a type of neural network, it is not neither optimization algorithm nor learning algorithm. MLP is trained with an external learning algorithm, the most famous is Back-propagation algorithm which is gradient decent optimization algorithm. Inertial unit used by the authors is three-axis sensor what is shown in Figure 5. The paper says that classification of road conditions is performed based on K samples, after FFT. Which samples, from which axis, from ”Leading edge” to ”Trailing edge” points, or in other way? It should be better explained how vector of K samples is constructed.   Line 171 – “with a learning model composed of three layers”: Again, MLP is not learning algorithm. Information about applied learning algorithm is missing. Line 174 – Input layers? There were more MLPs applied? If only one MLP was tested, it should be “layer”. Line 176 – The learning rate equal to 0.05 seems to be very small. In principle, the algorithm practically stands still, or moves very slowly…..at first glance, it looks strange. How many iterations were performed to achieve the final result? Line 194 – The problem seems to be very simple in classification, so maybe it would be useful to try other MLP architectures – more/few hidden neurons, one more hidden layer. Line 220 – If the input space is nine-fold, and simple MLP coped almost perfectly with the current problem, then deep learning seems to be completely unnecessary. If the network would have to cope with rough FFT, then deep networks could be considered. 

Author Response

Thank you very much for your detailed and thoughtful comments.

Point 1: It seems that the method of citing other authors is not appropriate. For example, in line 49 the name of author (Kanawar) appears at the beginning of the sentence, whereas reference to its work appears at the end of the sentence.

Response 1: According to reference 1 and 2, we have revised as “Kanwar et al.” (Line 49), “Niskanen et al.” (Line 52), “Hanstsuka et al.” (Line 57).

 

Point 2: Line 122 – “Multilayer perceptron” – rather multilayered perceptron Line 122 – MLP is a type of neural network, it is not neither optimization algorithm nor learning algorithm.

Response 2: A The author has made a mistake. As you commented, we have revised as “Back-propagation algorithm [21], an optimization technique, was used as the ANN learning algorithm.” (Line 120-121).

 

Point 3: Inertial unit used by the authors is three-axis sensor what is shown in Figure 5. The paper says that classification of road conditions is performed based on K samples, after FFT. Which samples, from which axis, from ”Leading edge” to ”Trailing edge” points, or in other way?

Response 3: Since the data were experimented in the straight section, only z-axis data was used. And, regardless of the “leading edge” and the “trailing edge”, the entire data of the experimented section was used as shown in Table l.

 

Point 4: Line 171 – “with a learning model composed of three layers”: Again, MLP is not learning algorithm. Information about applied learning algorithm is missing. Line 174 – Input layers? There were more MLPs applied? If only one MLP was tested, it should be “layer”.

Response 4: As you commented, we have revised as “In this paper, the relationship between the acceleration data measured through iTire and the road surface condition was modeled using MLP, with a learning model composed of three layers.”, “Nine values obtained through pre-processing (at 0 Hz, and then at each 50 Hz interval from 100–500 Hz, where these data were summed) and a bias value were selected as input variables for the input layer.”(Line 166-170).

 

Point 5: The learning rate equal to 0.05 seems to be very small. In principle, the algorithm practically stands still, or moves very slowly at first glance, it looks strange. How many iterations were performed to achieve the final result?

Response 5: We have iterated about 47,170 times when set the learning rate at 0. In addition, 40% of the total data in Table 1 was used for training.

Reviewer 2 Report

The paper is well-written and concise in its descriptions. This is very suitable for the straightforward approach. Since a novel technology is used for the data collection, many parts of the paper are product-oriented. I suggest to set a broader view on the topic in section 1 and especially section 2. Fig. 1 should be less abstract and, since it has plenty of space, would benefit from depictions of vehicle parts and an example road surface. If a more intuitive schematic is possible here, it would highly benefit the general understanding of the sensor. The description of the algorithm in fig. 2 is understandable, but would also benefit from adjusting the flowchart; the font size is too small and widening the chart would allow to increase it. Same goes to fig. 3, since widening and especially increasing the vertical axis would achieve more overview and less overlapping of the symbols. Section 4 has a typo in the title starting with "discussion", I would delete it. In fig. 5 it is hard to compare the plots of the 3 axes; therefore, I suggest to stack them one above each other together with widening the respective horizontal axis. Conclusions section is fine.

Author Response

Thank you very much for your detailed and thoughtful comments.

 

Point 1: I suggest to set a broader view on the topic in section 1 and especially section 2. Fig. 1 should be less abstract and, since it has plenty of space, would benefit from depictions of vehicle parts and an example road surface. If a more intuitive schematic is possible here, it would highly benefit the general understanding of the sensor.

Response 1: As you commented, we have revised Fig,1. (Line 94-95)

 

Point 2: The description of the algorithm in fig. 2 is understandable, but would also benefit from adjusting the flowchart; the font size is too small and widening the chart would allow to increase it. Same goes to fig. 3, since widening and especially increasing the vertical axis would achieve more overview and less overlapping of the symbols.

Response 2: As you commented, we have revised Fig, 2 (Line 104-105) and 3(Line 115-116).

 

Point 3: In fig. 5 it is hard to compare the plots of the 3 axes; therefore, I suggest to stack them one above each other together with widening the respective horizontal axis. 

Response 3: As you commented, we have revised Fig,5. (Line 175-180)