Regression Model for the Prediction of Total Motor Power Used by an Industrial Robot Manipulator during Operation

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

In this paper, ML methods for modeling the motor power for robots were investigated. This paper achieved its goal by mainly focusing on empirical evidence, although it lacks theoretical innovation. The experimental data may enlighten readers who are interested in the area. The methods were correctly designed and the result looks convincing. However, the writing must be improved before publication. Also, Fig 2 can be made larger with higher resolution. 

Comments on the Quality of English Language

The writing should follow the logic. The key point of the logic is to state the goal clearly. In the abstract, the authors wrote: "This paper focuses on using a machine learning (ML)-based model, using 2 the multilayer perceptron (MLP) method on the dataset collected using the ABB IRB 120 industrial 3 robotic manipulator." You can see that the authors only claimed that they are going to focus on ML, but only declare what to do with the ML. Combining the sentence before it, people might be able to believe that the authors implied that the ML is used for modeling the motor power. But this is not the convention for English writing. I hope the author can consider it. 

Author Response

We would like to thank the reviewer for their review of our manuscript. We have provided detailed point-by-point answers to the points posed by the reviewer below. Within the manuscript we have marked the changes made to the text thanks to this reviewer’s comments with green font color.

 

• This paper achieved its goal by mainly focusing on empirical evidence, although it lacks theoretical innovation. 

 

The main theoretical innovation the authors wished to focus on is the testing of ability to use regression based techniques, such as MLP, instead of time-series-based techniques such as LSTM. This allows for the modeling of the instantaneous motor power based on the values of connected, but independent variables (such as robot position, speed, acceleration…). We have added the following line to the introduction in order to clarify our intentions:

“A regression model based on variables that do not include a set of previous values of the targeted variable would allow for shorter-term prediction and instantaneous modeling of energy power.”

 

• The experimental data may enlighten readers who are interested in the area. 

 

The authors fully agree and were planning on publishing the data anyways after the publication of the paper. So, we have published the dataset and made it publicly available using Kaggle.

 

We have added the following text to the end of the section describing the dataset collection process::

 

“The dataset collected in this part of the research is made publicly available [19]”

 

Where citation 19 is:

 

Baressi Šegota, S. Robot Motor Power Dataset, 2024. https://doi.org/10.34740/KAGGLE/DSV/7874548. 

 

The DOI being the link to the dataset. We have also updated the “Data availability statement” given at the end of the paper, as follows:

The original data presented in the study are made openly available by the authors in “Robot Motor Power Dataset” at https://doi.org/10.34740/KAGGLE/DSV/7874548.

 

• However, the writing must be improved before publication. 

 

As the authors are not native English speakers, we have used a writing quality service to note which changes need to be performed in the manuscript, along with asking a native English speaker to check the manuscript. We have marked the changes in the manuscript specifically related to improving the quality of English with a red text color.

 

• Also, Fig 2 can be made larger with higher resolution. 

 

Figure 2 was made significantly larger in order to allow for better visibility.

 

In addition to the above, extensive english editing was performed in the paper, with all the changes marked with red font color. .

 

We sincerely hope that the reviewer will consider the provided answers appropriate and that their concerns with the manuscript have been addressed fully.

 

Kind regards, 

the authors

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Comments on: “Regression Model for the Prediction of Total Motor Power Used by an Industrial Robot Manipulator During Operation” by Sandi Baressi Segota, Nikola Andelic, Jelena Stifanic and Zlatan Car


The manuscript is solidly written and there are no special technical issues.

However, the authors should further clarify from a methodological point of view the differences and contributions of this work with their previously published work [1] (“Dynamics Modeling of Industrial Robotic Manipulators: A Machine Learning Approach Based on Synthetic Data”, Mathematics 2022, 10, 1174).

Furthermore, the presentation of the results itself lacks scientific soundness and contribution, as well as a comparison with similar cited methods. Since this is actually an integration of known methods into one procedure for modelling a robotic manipulator based on a measured data set, the authors should focus their efforts on a more rigorous evaluation of the robustness of the proposed procedure. This should include what the authors state as future work, which is the generalization of the procedure to more different robotic manipulator structures.

Comments on the Quality of English Language

Moderate editing of English language required.

Author Response

We would like to thank the reviewer for assisting us in addressing the issues with our manuscript. We have provided the answers to the questions posed below. Authors have marked the changes done in the text using blue font color.

 

1. However, the authors should further clarify from a methodological point of view the differences and contributions of this work with their previously published work [1] (“Dynamics Modeling of Industrial Robotic Manipulators: A Machine Learning Approach Based on Synthetic Data”, Mathematics 2022, 10, 1174).

 

We have addressed the differences compared to the cited work in the Introduction, with the following text:

 

“In the previous work by the authors [1], an attempt was made to model the dynamics of an industrial robotic manipulator using a similar methodology to the one applied in this paper (an MLP regression ANN). In that paper, the authors applied the MLP to regress the moments of torsion that appear on the motors during the operation. Compared to that work, the work that the authors present in this paper addresses the following gaps in knowledge:

 

• Utilization of real, experimentally collected data, compared to the mathematical models, should create a more robust model, as the data may include noise and other minor measurement errors not present in the data created by a mathematical model.

• The testing of the importance of individual features was not performed in the previous work, mostly because the model was based on the mathematical model developed with a method requiring predetermined variables (speeds, accelerations, and positions of the robot joints). With the models newly developed in the presented work, features can be simply eliminated, allowing authors to test the possible benefits of that type of preprocessing.

• The model developed in the aforementioned paper did not make use of certain variables that were not available to the model but may have a certain influence on the output — such as kinematic variables pertaining to the limits and singularities.

 

To summarize, the presented work aims to improve the existing research by applying a similar methodology to a real, laboratory environment. The dataset pruning methods are applied to simplify the dataset collection process, which is now performed on an actual robot and the created models may potentially be applied to the real data measured directly from the industrial robot.”

2. Furthermore, the presentation of the results itself lacks scientific soundness and contribution, as well as a comparison with similar cited methods. 

 

The comparison to similar cited methods has been added to the discussion, referencing a newly added Table 1 in the introduction, as follows:

 

“Comparing the results to the ones achieved by the previous researchers, as shown in table 1., the applied methodology achieves significantly better results, when the average scores are observed, with an improvement of almost 3% when MAPE is observed and 0.02 increase in the R2 score. As the improvement is present even when the focus is given to the minimal scores, it can be concluded that the given method of using regression techniques instead of focussing on time-series modeling has a definite merit. The same holds true when the additional validation is performed on different simulated robots as shown in the following section.”

 

3. Since this is actually an integration of known methods into one procedure for modelling a robotic manipulator based on a measured data set, the authors should focus their efforts on a more rigorous evaluation of the robustness of the proposed procedure. This should include what the authors state as future work, which is the generalization of the procedure to more different robotic manipulator structures.

 

While the authors have attempted to perform the more rigorous evaluation with the use of more cross-validation procedure, we agree with the reviewer that additional validation should be performed on different robots.

 

Sadly, as the authors don’t have, at this time, access to different manipulators within their laboratory, the authors have performed this by using the simulation data within the RobotStudio package. Only the validation was performed, because the additional training with different robots using comparable dataset sizes would not be possible within the time provided for this review, and we believe that the validation illustrates the applicability of the developed models nicely. The newly added section “Validation of results on different industrial robotic manipulators” has been added, with the text provided below:

 

“To further validate the obtained results, the authors have evaluated different industrial robotic manipulators in a simulated environment. The simulated environment was chosen, due to lack of access to similar robots to the authors. It has to be noted that this process has some limitations — for example, any simulated measurements will not include any noise that may appear during the process of measurement on the real industrial robotic manipulator. Additionally, a real industrial manipulator may have certain influences which affect its motor power use — such as the environment temperature, or the component condition, which are not included within a simulated environment. Despite this, the measurement should be similar enough that a prediction can be established. The process of performing the measurement is done in much the same way as the original data collection. The main change lays in the fact that there is no connection to the robot controller. Instead, an internal virtual controller, included in the ABB RobotStudio \cite{RobotStudio} is used to control the robot and simulate the measurement. The code used to generate the simulation points is the same, except the limits of the robots are adjusted to reflect the limits provided by the manufacturer in the technical documentation and reference of each robot used for validation — as there are no potential collision points with the environment (e.g. fences, cables, pneumatic lines…) that exist in the real laboratory environment. The robots selected were: IRB 1010, IRB 1100, IRB 1200, IRB 1410, IRB 2600, IRB 4600, IRB 5710. The comparison between the robots is given in figure 5. The robots were selected as ones having the same virtual controller and overall configuration of joints and degrees-of-freedom as ABB IRB 120 — in other words, robots capable of providing the same measurements, with multiple different sizes of robots selected.

 

Due to this being validation, only ten random points were selected for the database, resulting in datasets that ranged between 300 and 400 individual data points. The measurement was performed only on the variables remaining after the pruning, according to the previously presented methodology. Then, the prediction was performed using the best-performing MLP model given in the previous subsection. The average $R˘2$ and $MAPE$ were calculated between the simulated Total Motor Power and the predicted one, and given in Table 6, below. In the table, the results for IRB 120 are shown both on the separate validation set collected in the laboratory environment and within the simulation.

 

The results show that there is a significant drop in performance between the real and simulated IRB 120 robot data, with the MAPE increasing from 0.33 to 0.70 — more than doubling. Other robots used in the validation procedure also show drops in performance. With the error increasing the more different the robot configuration is (this is mostly addressing the size of the robot). Interestingly, the IRB 1100, which is the ABB replacement for IRB 120 has the lowest error, after the simulated IRB 120. All the other robots, even though the error is increased, still fall within a satisfactory range for $MAPE$ and $R^2$. For simplicity the data is also visually represented in figure 6.”

 

We sincerely hope that the reviewer will consider the provided answers appropriate and that their concerns with the manuscript have been addressed fully.

 

Kind regards,

the authors

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

1/ For  prediction engery consumption, there are alot of article using machine learning to ppredict the power before.What is your supperivor of your method comapred to previous studies

 

2/The the multilayer perceptron is now a new twchnique nowadays. Uisng 1DCNN, LSTM, Bi-LSTM seems more efetive, Why do you use MLP?

 

3/You mention that "The scores demonstrate that while both MLPs achieve good scores, the model trained on the pruned dataset has a slightly higher  performance". The big quesion is that with "slight hiher performance", it is worth to use "the random forest (RF) model" and " feature permutation score difference (FP), along the Pearson’s correlation coefficient" to eliminate the variables with a low influence from the dataset.

I think we can use Good model and the performance can improve without using additional steps to find "Feature importance"

4/In machine learning, removing variables (also known as features or attributes) that have little influence on the target variable or outcome can assist simplify models, prevent overfitting, and increase computing efficiency.

There are other approaches, such as:

Principal component analysis (PCA)

Tree-Based Methods

Techniques like Ridge (L2 regularization) and Lasso (L1 regularization) 

What are the advantages of your methods compared to the methods mentioned above?

Comments on the Quality of English Language

The English is just fine.

Author Response

We would like to thank the reviewer for assisting us in addressing the issues with our manuscript. We have provided the answers to the questions posed below. Authors have marked the changes done in the text using blue font color.

 

• For  prediction engery consumption, there are alot of article using machine learning to ppredict the power before.What is your supperivor of your method comapred to previous studies

 

The following text was used to address this in the introduction:

 

“We can see that, while there is interest by researchers in modeling the power use of industrial robotic manipulators, most research focuses on using LSTM networks for predictions. These networks are used in the prediction of one-dimensional time-series [15], and while this approach has its uses, there may be shortcomings to it based on the desired application. There may be instances where the power use of the robot may need to be tested in the given instant, based on other variables except past values of the energy use. This is especially interesting in cases where the movement of the robot requires rapid changes in speed and direction. Due to this, the research performed in this paper will focus on the creation of a dataset that is meant to be used in a regression task, where the instantaneous total motor power of an industrial robot manipulator can be predicted based on connected variables.” 

 

With the addition of the following for further clarification and simplification of the explanation: 

 

“A regression model based on variables that do not include a set of previous values of the targeted variable would allow for shorter-term prediction and instantaneous modeling of energy power.”

 

In addition, the results analyzed in the introduction were separated in table 1 with the following text:

 

“For later comparison, the state-of-the-art results are included in the table 1. In studies with ranges and multiple results presented, the best results were selected.”

 

The table was later referenced when discussing the results, showing the improvement of the results given in table 1.:

 

“Comparing the results to the ones achieved by the previous researchers, as shown in table 1., the applied methodology achieves significantly better results, when the average scores are observed, with an improvement of almost 3% when MAPE is observed and 0.02 increase in the R2 score. As the improvement is present even when the focus is given to the minimal scores, it can be concluded that the given method of using regression techniques instead of focussing on time-series modeling has a definite merit.”

 

• The the multilayer perceptron is now a new twchnique nowadays. Uisng 1DCNN, LSTM, Bi-LSTM seems more efetive, Why do you use MLP?

 

The use of MLP was explained by the following text added to the section “Regression Methodology”:

 

“As stated in the introduction, the goal of using a regression technique instead of an LSTM or similar approaches is the ability to perform the modeling of the instantaneous power. In other words, the authors want to establish a model which will be able to predict the motor power of the industrial robot at any moment — with this prediction being based on the other measured variables such as speed and position, and not the previous values of motor power, as this may enable different optimization schemes. While many regression techniques could have been applied to the dataset of this type authors have selected MLP based on two features, and those are the high performance in similar modeling tasks in the previous research [1] and a high computational speed of the trained models.”

 

With the text that is referred in the introduction being:

 

“We can see that, while there is interest by researchers in modeling the power use of industrial robotic manipulators, most research focuses on using LSTM networks for predictions. These networks are used in the prediction of one-dimensional time-series [15], and while this approach has its uses, there may be shortcomings to it based on the desired application. There may be instances where the power use of the robot may need to be tested in the given instant, based on other variables except past values of the energy use. This is especially interesting in cases where the movement of the robot requires rapid changes in speed and direction. Due to this, the research performed in this paper will focus on the creation of a dataset that is meant to be used in a regression task, where the instantaneous total motor power of an industrial robot manipulator can be predicted based on connected variables. A regression model based on variables that do not include a set of previous values of the targeted variable would allow for shorter-term prediction and instantaneous modeling of energy power.”

 

• You mention that "The scores demonstrate that while both MLPs achieve good scores, the model trained on the pruned dataset has a slightly higher  performance". The big quesion is that with "slight hiher performance", it is worth to use "the random forest (RF) model" and " feature permutation score difference (FP), along the Pearson’s correlation coefficient" to eliminate the variables with a low influence from the dataset.

 

The authors believe that the best way of demonstrating the worth of using the dataset pruning as described in the dataset is by demonstrating the amount of time necessary for these methods to be applied. The following text was added as an explanation:

 

“One of the key issues with the application of any pre-processing to the ML-based modeling is the issue of increased time and computational complexity added to the modeling process. A benefit of the used pre-processing approach to the dataset pruning is that it is completed in advance. Compared to the application of methods such as PCA the variable selection is performed in advance. All the future modeling can be based on the subset of variables collected, which can even simplify the dataset collection and processing due to the smaller dataset size (notably, the not-pruned dataset is 33.38 MB in size, while the pruned dataset is 11.00 MB in size — a 67\% decrease in size. When it comes to the preprocessing time that is additionally taken up by the dataset pruning, the average of time needed for the application of all three feature importance methods (RF with FP and MDI, and Pearson’s correlation) was 4 minutes and 11 seconds, with the standard deviation of 9 seconds. It should be noted that the analyzed dataset is very large, and the modeling was performed on a desktop computer with an Intel(R) Core(™) i5-6400 CPU (six cores and six logical processors, base clocked at 2.9 GHz), and 32 GB of RAM. Scikit-learn Python library [25] was used to determine the feature importances based on RF, while Pandas Python library [36] was used to calculate the correlation. Considering that the MLP models took approximately two days of training on a workstation consisting of an AMD Epyc 7532 processor (24 cores and 48 threads, base clock of 2.3GHz) and 128 GB of RAM, the time added by the processing of the dataset using feature importance analysis is practically negligible. Considering the possible benefits (lower model complexity, lower training time, a smaller dataset, and most importantly improved scores with the same method), it can be concluded that there is a clear advantage to the application of the suggested method.”

 

• I think we can use Good model and the performance can improve without using additional steps to find "Feature importance"

 

While the authors agree with the statement, the focus of this paper was to test two things - and that is whether the score improvement can be achieved with feature importance-based pruning and whether a regression environment can be used for instantaneous prediction. The improvement to the scores could be achieved by applying more novel techniques than MLP, which was noted in the future work:

 

“Future work should also focus on the application of more advanced regression techniques to test if the results could be improved, even without resorting to feature importance-based pruning.”

 

• In machine learning, removing variables (also known as features or attributes) that have little influence on the target variable or outcome can assist simplify models, prevent overfitting, and increase computing efficiency. There are other approaches, such as: Principal component analysis (PCA), Tree-Based Methods, Techniques like Ridge (L2 regularization) and Lasso (L1 regularization)  What are the advantages of your methods compared to the methods mentioned above?

 

We have discussed the reasons why we selected the techniques we used in comparison to other techniques in the following text added to the manuscript:

 

“In addition to the described techniques for dataset feature preprocessing, several other techniques can be used to remove features with a low influence on the output. One of the main benefits of the applied RF method, which is a tree-based method is that the method provides a simple list of variables that have a low influence and may be removed, allowing for an easy application in conjunction with correlation-based analysis. Decomposition methods are also commonly applied by researchers, such as the principle component analysis (PCA) methods which serve to transform the dataset into a set of principal components [23,24]. Compared to the used approach, the PCA approach does not provide a clear list of variables to be removed (or simply not collected), but the transformation needs to be performed every time new data is collected, before using the developed regression model, possibly adding time to the very fast prediction time of the developed MLP [25]. A more direct approach would be the application of Lasso or Ridge regression to determine the coefficients of the variables and lower them to zero to eliminate the low-influenced ones from the dataset [26]. Still, RF does have some benefits in comparison to these methods such as robustness to non-linearity and multicollinearity, lower sensitivity to outliers and unscaled features, and automatic variable interaction capturing [27,28].”

 

We sincerely hope that the reviewer will consider the provided answers appropriate and that their concerns with the manuscript have been addressed fully.

 

Kind regards,

the authors

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

This version of the manuscript “Regression Model for the Prediction of Total Motor Power Used by an Industrial Robot Manipulator During Operation” has been improved and additionally explained based on the reviewers' comments.

I have no further specific comments or suggestions.

Comments on the Quality of English Language

Minor editing of English language required.

Author Response

We would like to thank the reviewer for reviewing the resubmitted revision of our manuscript.

Regarding the minor editing of the language, it was performed with the changes marked in the manuscript.

Kindest regards,
the authors.

Reviewer 3 Report

Comments and Suggestions for Authors

The authors provided appropriate answers. It is advised that it be published on the journal.

Author Response

We would like to thank the reviewer for re-reviewing the revision of our manuscript.

Kindest regards,
the authors