Data-Driven Feedforward Force Control of a Single-Acting Pneumatic Cylinder with a Nonlinear Hysteresis Characteristic
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsPlease consider incorporating the following aspects into the paper:
Hysteresis and Error Analysis: Include a discussion on hysteresis effects and provide a thorough error analysis to enhance the technical depth of the study.
Performance Evaluation Under Varied Conditions: Assess the trained neural network’s performance across a broader range of operating conditions, such as different force profiles and speeds, to evaluate its generalization capability.
Comparison with Conventional Control Methods: Compare the proposed neural network-based control approach with other established methods, such as PID control or model-based control. This comparison will help demonstrate the advantages and potential limitations of the proposed method.
Real-Time Performance Analysis: Provide more details regarding the real-time computational performance of the neural network. Specifically, report the computation time required for a single force control step and assess whether the system meets the real-time requirements of the application.
Discussion on Alternative Pressure-Driven Actuators: Expand the discussion to include other types of pressure-driven actuators, such as soft actuators and hydraulic actuators. Justify the selection of the actuator used in this study and compare its design and performance against alternative designs.
Additionally, consider referencing and comparing findings with the following relevant papers:
- Le XT, Gunawardane PH, Mallikarachchi S, Chiao M, Godage IS. *Hybrid Control of 3D-Printed Multimodal Soft Pneumatic Actuators.* IEEE International Conference on Mechatronics and Automation (ICMA), 2024.
- Gunawardane PD, Cheung P, Zhou H, Alici G, de Silva CW, Chiao M. *A Versatile 3D-Printable Soft Pneumatic Actuator Design for Multi-Functional Applications in Soft Robotics.* Soft Robotics, 2024.
- Yao Z, Xu F, Jiang GP, Yao J. *Data-Driven Control of Hydraulic Manipulators by Reinforcement Learning.* IEEE/ASME Transactions on Mechatronics, 2023.
- Hamon P, Michel L, Plestan F, Chablat D. *Model-Free Based Control of a Gripper Actuated by Pneumatic Muscles.* Mechatronics, 2023.
General Improvements:
Addressing Drawbacks in Real-Time Applications:** While data-driven neural network modeling presents a promising approach for controlling nonlinear pneumatic systems, it is crucial to acknowledge potential drawbacks. Expand the discussion to cover challenges such as data dependency, model validation, and robust control design. Provide insights into balancing accuracy and real-time feasibility in practical applications.
Proofreading and Grammar: Carefully review the paper for any grammatical errors or typographical mistakes to enhance readability and clarity.
Figure Quality: Ensure that all figures are clear, well-labeled, and visually comprehensible.
Citations: Verify that all citations are accurate, correctly formatted, and complete.
Author Response
Addressing concerns of reviewer #1:
- Reviewer’s comments: Include a discussion on hysteresis effects and provide a thorough error analysis to enhance the technical depth of the study.
Author’s reply: Thank you for your valuable suggestion. In response to the reviewer’s comments, we have incorporated a detailed discussion on the hysteresis effects and conducted a thorough error analysis to enhance the technical depth of our study. This discussion is now included in Section 4.1 of the revised manuscript, where we elaborate on the origins and impacts of hysteresis, as well as present comprehensive error metrics to quantify the control performance.
- Reviewer’s comments: Performance Evaluation Under Varied Conditions: Assess the trained neural network’s performance across a broader range of operating conditions, such as different force profiles and speeds, to evaluate its generalization capability.
Author’s reply: Thank you for your insightful comments. We fully agree that evaluating the generalization capability of the trained neural network across a broader range of operating conditions is crucial for understanding its robustness and adaptability. While our current study focuses on specific scenarios, we acknowledge the importance of expanding the evaluation to include diverse force profiles and speeds. This will provide a more comprehensive assessment of the neural network’s performance and its ability to handle varied operational demands.
We plan to address this in our future work by conducting extensive experiments and simulations under different conditions. This will help us further optimize the neural network architecture and training process to enhance its generalization capability.
Thank you again for your valuable suggestions. We believe that this direction of research will significantly contribute to the development of more robust and adaptive control models for pneumatic systems.
- Reviewer’s comments: Comparison with Conventional Control Methods: Compare the proposed neural network-based control approach with other established methods, such as PID control or model-based control. This comparison will help demonstrate the advantages and potential limitations of the proposed method.
Author’s reply: at present Thank you for your valuable suggestion regarding the comparison of our proposed neural network-based control approach with conventional methods such as PID control or model-based control. We agree that such a comparison is essential to highlight the advantages and potential limitations of our method.
At present, we have not yet implemented PID control or model-based control algorithms in our experimental setup, which limits our ability to provide a direct comparison in this study. However, we recognize the importance of this comparison and plan to address it in our future work.
- Reviewer’s comments: Real-Time Performance Analysis: Provide more details regarding the real-time computational performance of the neural network. Specifically, report the computation time required for a single force control step and assess whether the system meets the real-time requirements of the application.
Author’s reply: Thank you for your insightful comment regarding the real-time performance analysis of our neural network model. We appreciate your interest in understanding the computational efficiency of our approach, especially in the context of real-time applications. To address your query, we have conducted a detailed evaluation of the real-time computational performance of our neural network models. Specifically, we have compiled the neural network models with varying numbers of hidden neurons into C code and executed them on an Arduino UNO microcontroller. This setup allows us to assess the average computing cost over 1000 runs for each configuration. Our results, as shown in Figures 10a and 10b, provide valuable insights into the trade-off between modeling accuracy and computational efficiency. Specifically, the analysis demonstrates that the optimal balance is achieved when the number of hidden neurons is set to 5 and 7 for the respective neural network models. These configurations can be executed on a microcontroller within 2 ms, thereby fulfilling the real-time requirement for actuating force control.
- Reviewer’s comments: Discussion on Alternative Pressure-Driven Actuators: Expand the discussion to include other types of pressure-driven actuators, such as soft actuators and hydraulic actuators. Justify the selection of the actuator used in this study and compare its design and performance against alternative designs.
Author’s reply: Thank you for your valuable suggestion regarding the discussion on alternative pressure-driven actuators. In response to your comment, we have expanded the discussion in Section 4.2 of the revised manuscript to include a detailed comparison of various pressure-driven actuators, such as soft actuators and hydraulic actuators.
- Reviewer’s comments: Additionally, consider referencing and comparing findings with the following relevant papers: ////Le XT, Gunawardane PH, Mallikarachchi S, Chiao M, Godage IS. *Hybrid Control of 3D-Printed Multimodal Soft Pneumatic Actuators.* IEEE International Conference on Mechatronics and Automation (ICMA), 2024. ////Gunawardane PD, Cheung P, Zhou H, Alici G, de Silva CW, Chiao M. *A Versatile 3D-Printable Soft Pneumatic Actuator Design for Multi-Functional Applications in Soft Robotics.* Soft Robotics, 2024. ////Yao Z, Xu F, Jiang GP, Yao J. *Data-Driven Control of Hydraulic Manipulators by Reinforcement Learning.* IEEE/ASME Transactions on Mechatronics, 2023. ////Hamon P, Michel L, Plestan F, Chablat D. *Model-Free Based Control of a Gripper Actuated by Pneumatic Muscles.* Mechatronics, 2023.
Author’s reply: Thank you for your suggestion. We have carefully reviewed the suggested papers and have incorporated them into our manuscript. We have also compared our findings with the relevant works to provide a more comprehensive context for our contributions.
- Reviewer’s comments: Addressing Drawbacks in Real-Time Applications: While data-driven neural network modeling presents a promising approach for controlling nonlinear pneumatic systems, it is crucial to acknowledge potential drawbacks. Expand the discussion to cover challenges such as data dependency, model validation, and robust control design. Provide insights into balancing accuracy and real-time feasibility in practical applications.
Author’s reply: Thank you for your insightful comment regarding the real-time performance analysis of our neural network model. We appreciate your interest in understanding the computational efficiency of our approach, especially in the context of real-time applications. To address your query, we have conducted a detailed evaluation of the real-time computational performance of our neural network models. Specifically, we have compiled the neural network models with varying numbers of hidden neurons into C code and executed them on an Arduino UNO microcontroller. This setup allows us to assess the average computing cost over 1000 runs for each configuration. Our results, as shown in Figures 10a and 10b, provide valuable insights into the trade-off between modeling accuracy and computational efficiency. Specifically, the analysis demonstrates that the optimal balance is achieved when the number of hidden neurons is set to 5 and 7 for the respective neural network models. These configurations can be executed on a microcontroller within 2 ms, thereby fulfilling the real-time requirement for actuating force control.
- Reviewer’s comments: Proofreading and Grammar: Carefully review the paper for any grammatical errors or typographical mistakes to enhance readability and clarity.
Author’s reply: We sincerely thank the reviewer for pointing out the importance of proofreading and ensuring grammatical accuracy. In response to this comment, we have conducted a thorough review of the entire manuscript to identify and correct any grammatical errors or typographical mistakes.
- Reviewer’s comments: Figure Quality: Ensure that all figures are clear, well-labeled, and visually comprehensible.
Author’s reply: We thank the reviewer for highlighting the importance of high-quality figures. To address this concern, we have thoroughly reviewed and revised all figures in the manuscript to ensure they are clear, well-labeled, and visually comprehensible. We have enhanced the resolution and clarity of each figure, standardized labeling and annotations for consistency, and simplified complex layouts to improve visual comprehensibility. Additionally, we have ensured that all figures are properly referenced in the text and that captions accurately describe the content. We believe these improvements significantly enhance the overall quality and readability of the manuscript.
- Reviewer’s comments: Citations: Verify that all citations are accurate, correctly formatted, and complete.
Author’s reply: We appreciate the reviewer's attention to the accuracy and completeness of our citations. In response, we have meticulously reviewed all references in the manuscript to ensure they are accurate, correctly formatted, and complete. We have cross-checked each citation with the original sources to verify their accuracy and have updated any formatting inconsistencies to adhere to the journal's citation style.
We extend our sincere gratitude to the reviewer for your professional and insightful suggestions regarding our manuscript. These valuable inputs have greatly enhanced the clarity and precision of our work.
Reviewer 2 Report
Comments and Suggestions for AuthorsThis paper presents a data-driven neural network modeling and feedforward force control method for a single-acting pneumatic cylinder with nonlinear hysteresis characteristics. The authors developed a neural network model to accurately predict the actuating force and voltage instruction for the pneumatic system. The model's input layer was optimized to include essential variables, and the number of hidden neurons was balanced to ensure real-time execution speed. Experimental results demonstrated that the proposed method achieved ideal force tracking with a control error of less than 0.5N during loading/unloading processes and reduced steady-state error in step force control through instruction smoothing. This paper is well-structured and well-written. Before publication, there are some questions to be solved.
- The author mentioned “Pneumatic servo system has broad application background in automation field, such as industrial polishing, robotic grasping, and humanoid robot, due to its advantages of small size, cleanness, safe, and cost-effective.”, more state-of-the-art can be cited: DOI: 10.34133/cbsystems.0025; DOI: 10.34133/cbsystems.0008.
- The scale bar of the figures 1,2,5,6 are missing. This would help audiences identify the sizes.
- More details on the architecture of the neural network used for modeling the actuating force are required, the layers and functions.
- How about the real-time performance?
- The details on the experimental setup are required to help audiences understand the calibration process.
- When the environment changes, e.g., humidity or temperature, how to ensure the dynamic performance of the control strategy.
Author Response
Addressing concerns of reviewer #2:
- Reviewer’s comments: The author mentioned “Pneumatic servo system has broad application background in automation field, such as industrial polishing, robotic grasping, and humanoid robot, due to its advantages of small size, cleanness, safe, and cost-effective.”, more state-of-the-art can be cited: DOI: 10.34133/cbsystems.0025; DOI: 10.34133/cbsystems.0008.
Author’s reply: Thank you for your suggestion. We have carefully reviewed the suggested papers and have incorporated them into our manuscript.
- Reviewer’s comments: The scale bar of the figures 1,2,5,6 are missing. This would help audiences identify the sizes.
Author’s reply: Thank you for pointing out this important detail. We apologize for the oversight regarding the scale bars in Figures 1, 2, 5, and 6. We fully agree that including scale bars is essential for providing context and aiding the interpretation of the figures. In the revised version of the manuscript, we have added clear and appropriately sized scale bars to each of these figures. Additionally, we have ensured that the scale bars are clearly labeled to help readers accurately identify the sizes and dimensions depicted.
- Reviewer’s comments: More details on the architecture of the neural network used for modeling the actuating force are required, the layers and functions.
Author’s reply: We appreciate the reviewer's suggestion to provide more details on the architecture of the neural network used for modeling the actuating force. In our manuscript, we have now included a more comprehensive description of the neural network architecture, including the specific layers and activation functions used.
The neural network model is designed with an input layer, one or more hidden layers, and an output layer. The input layer consists of variables such as actuating pressure, piston displacement, piston velocity, and motion state, which are identified through our analysis as critical factors influencing the actuating force. The hidden layers employ neurons with Rectified Linear Unit (ReLU) activation functions to capture the nonlinear relationships within the system, while the output layer utilizes a linear activation function.
- Reviewer’s comments: How about the real-time performance?
Author’s reply: Thank you for your insightful comment regarding the real-time performance analysis of our neural network model. We appreciate your interest in understanding the computational efficiency of our approach, especially in the context of real-time applications. To address your query, we have conducted a detailed evaluation of the real-time computational performance of our neural network models. Specifically, we have compiled the neural network models with varying numbers of hidden neurons into C code and executed them on an Arduino UNO microcontroller. This setup allows us to assess the average computing cost over 1000 runs for each configuration. Our results, as shown in Figures 10a and 10b, provide valuable insights into the trade-off between modeling accuracy and computational efficiency. Specifically, the analysis demonstrates that the optimal balance is achieved when the number of hidden neurons is set to 5 and 7 for the respective neural network models. These configurations can be executed on a microcontroller within 2 ms, thereby fulfilling the real-time requirement for actuating force control.
- Reviewer’s comments: The details on the experimental setup are required to help audiences understand the calibration process.
Author’s reply: Thank you for your valuable feedback regarding the experimental setup and calibration process. We have revised and expanded the description of the experimental setup in the manuscript to provide a clearer and more detailed explanation of the calibration process. Specifically, we have enhanced Figure 4 to illustrate the overall modeling procedure, including data acquisition, neural network training, and validation steps. Additionally, we have added detailed text descriptions to explain the calibration methods used, that is the Bezier Calibration Method (BCM), and how they contribute to improving the accuracy of the measurements. We believe these revisions will help readers better understand the calibration process and its significance in our study.
- Reviewer’s comments: When the environment changes, e.g., humidity or temperature, how to ensure the dynamic performance of the control strategy.
Author’s reply: Thank you for raising this important point regarding the dynamic performance of the control strategy under changing environmental conditions such as humidity and temperature. While our current study focuses on the development and validation of the neural network-based control approach under controlled laboratory conditions, we recognize the need to address environmental variability in practical applications.
In our future work, we plan to investigate the impact of environmental factors on the control performance by conducting experiments under varying humidity and temperature conditions. We will explore adaptive control strategies that can dynamically adjust the control parameters in real-time based on sensor feedback, ensuring robustness against environmental changes. Additionally, we aim to incorporate environmental sensors into the system to monitor and compensate for these variations proactively. We believe these enhancements will further improve the reliability and adaptability of our control strategy in real-world applications.
We extend our sincere gratitude to the reviewer for your professional and insightful suggestions regarding our manuscript. These valuable inputs have greatly enhanced the clarity and precision of our work.
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThe authors have made the necessary revisions to the paper accordingly.
Comments on the Quality of English LanguageThe authors have made the necessary revisions to the paper accordingly.
Reviewer 2 Report
Comments and Suggestions for AuthorsThis paper can be published