Review Reports
- Desislava Stoitseva-Delicheva * and
- Snejana Yordanova
Reviewer 1: Anonymous Reviewer 2: Anonymous Reviewer 3: Zonglai Mo
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe following issues need to be discussed:
1. The paper claims to have proposed a novel design method for gain-scheduled PID control systems. However, in what aspects is this novel method reflected? A comparison with baseline methods needs to be supplemented.
2. In the validation experiments, the number of neurons (3) seems rather small. Extensive studies have shown that with only 3 nodes, the performance of neural networks shows little difference from traditional adaptive methods and similar approaches.
3. PLC, PID and PWM do not belong to the same category of concepts. Is it appropriate to group them together?
4. The frequency-domain analysis presented in the paper exhibits no obvious improvements or distinctions; Equations (2) and (3) are both derived from conventional methods.
5. It is recommended to use standard symbolic notation instead of plain text such as tau to represent variables. In addition, several grammatical errors exist in the paper.
Author Response
- The paper claims to have proposed a novel design method for gain-scheduled PID control systems. However, in what aspects is this novel method reflected? A comparison with baseline methods needs to be supplemented.
The novelty of the suggested approach is explained in lines 124÷144. It is based on data from experiments with a laboratory dryer where the real world signals are limited and noisy due to ambient influences. Firstly, we step on derived and validated from experiments in previous research MTSK plant and PID controller models to move investigations from real world to the field of simulations and to trust the results from simulations as near to reality. Secondly, we compare the step responses of the real time PLC PID control against the simulated designed BANN-PID and the benchmarked PID (the validated model of the PLC-PID) and SPID systems to assess improvements in the BANN-PID system.
A novelty is how the BANN training dataset is provided from simulation of the MTSK step responses and their ZN approximation – MTSK is a validated nonlinear plant model, the simulations allow to freely collect in short time reliable, smooth, various and rich in magnitude and frequency dataset for training of the BANN. Simulations enable safe for the plant and fast experiments which can be difficult, forbidden or impossible for the real world system and also take a long time, have restrictions and are noisy.
All methods in the literature review use idealistic mathematical models, simulations and comparison with linear PID control system for a simple nonlinear plant, pls see lines 112÷115. No relation with experiments on real world industrial plant, inconsistent and far in concept comparison with linear PID control and linear plant model. ANN, FL and adaptation enhance the PID facilities in handling plant nonlinearity! There a many closer in concept approaches to use for comparison.
- In the validation experiments, the number of neurons (3) seems rather small. Extensive studies have shown that with only 3 nodes, the performance of neural networks shows little difference from traditional adaptive methods and similar approaches.
We use a two-layer BANN with input the operation point, represented by the control action, and the three parameters of the plant ZN model as 3 outputs. We selected 5 hidden neurons, i.e. the most typical BANN architecture, and in our case it ensures the desired accuracy in assessing of the ZN model parameters as seen in Fig.6.
- PLC, PID and PWM do not belong to the same category of concepts. Is it appropriate to group them together?
The abbreviation PLC-PID-PWM is an abbreviation for the type of controller or system investigated. Line 110 “real time PID temperature control via programmable logic controller (PLC).” explains the relation between PLC and PID, i.e. this is PID implemented in a PLC. Line 154 “The present research is based on the PLC-PID with pulse-width modulation (PWM)” explains the relation of PID with PWM.
- The frequency-domain analysis presented in the paper exhibits no obvious improvements or distinctions; Equations (2) and (3) are both derived from conventional methods.
The robustness analysis is necessary to show if the designed system with changing plant’s and controller’s parameters can operate in industrial environment, where reality differs from models, preserving stability and performance. The derivation of (2) and (3) steps on the principles of Morari [22] where only plant uncertainty is assumed. Our approach differs from {22] in considering also the uncertainty of a gain-scheduled controller and in the way the uncertainty is assessed both for the controller and the MTSK plant.
- It is recommended to use standard symbolic notation instead of plain text such as tau to represent variables. In addition, several grammatical errors exist in the paper.
Thank you - corrected
Reviewer 2 Report
Comments and Suggestions for AuthorsThis paper focuses on adaptive PID control for nonlinear industrial processes under wide operating conditions, and proposes a gain scheduling scheme based on offline training BANN. However, the following issues need to be improved:
- The object modeling and condition coverage are limited, only covering a temperature range of 25~95℃ and a control quantity of 2~20%, without considering disturbances such as load, humidity, and air volume. It is recommended to add experiments with variable load and variable environmental temperature to expand the training set and the range of conditions.
- It is suggested that in the introduction section, research on control methods for nonlinear disturbances such as those in the vehicle chassis be included. such as “Adaptive fuzzy PID control strategy for vehicle active suspension based on road evaluation. Electronics, 2022.”,“A sky-hook positive real network based ISD seat suspension for commercial vehicle. Journal of Vibration and Control. 2025. DOI:10775463251407519.”
- It is suggested to compare the performance differences of different structures (RBF, recursive NN, deep network), and increase the sensitivity analysis of network structure to explain why the current configuration was chosen.
- Only the use of GA algorithm to optimize A/B/C coefficients and PID parameters is described, without specifying the optimization objective and constraints.
- It is recommended to supplement the GA parameter settings and optimization convergence curve to enhance reproducibility.
- The images in the text are not clear, and the axis titles are missing.
Author Response
- The object modelling and condition coverage are limited, only covering a temperature range of 25~95℃ and a control quantity of 2~20%, without considering disturbances such as load, humidity, and air volume. It is recommended to add experiments with variable load and variable environmental temperature to expand the training set and the range of conditions.
The drying of fruits and the hardware used put limits on the signals. This is typical for operation in industrial environment and with industrial controllers.
The disturbances are accounted in several ways:
- They are present in the measured signals used in the MTSK modelling in [18, 19], explained in Section 2, and in the real time experimental PLC-PID-PWM control (Fig.10).
- The accuracy achieved (Fig.6) with the training dataset from Fig.5 ensures system robustness (Fig.7) which means preserving system stability and performance in case of disturbances and plant changes some provoked via extra experiments and training dataset.
- Simulation with a varied plant in Fig.8 (dotted lines) which is a result of changes in the plant and its MTSK model (Fig.5) respectively due to ambient and industrial factors, including disturbances, proves system’s robustness.
Real time experiments with the designed system are planned for future work, pls see the Conclusion section.
- It is suggested that in the introduction section, research on control methods for nonlinear disturbances such as those in the vehicle chassis be included. such as “Adaptive fuzzy PID control strategy for vehicle active suspension based on road evaluation. Electronics, 2022.”,“A sky-hook positive real network based ISD seat suspension for commercial vehicle. Journal of Vibration and Control. 2025. DOI:10775463251407519.”
Thank you, we included as [18] Han, S. Y., Dong, J. F., Zhou, J., Chen, Y. H., Adaptive fuzzy PID control strategy for vehicle active suspension based on road evaluation. Electronics. 11:1-21, 921. 2022, https://doi.org/10.3390/electronics11060921
- It is suggested to compare the performance differences of different structures (RBF, recursive NN, deep network), and increase the sensitivity analysis of network structure to explain why the current configuration was chosen.
Thank you, we will consider for future work.
- Only the use of GA algorithm to optimize A/B/C coefficients and PID parameters is described, without specifying the optimization objective and constraints.
Added missing fitness functions in red in line 193 “The basic values of the PID gains are GA optimised qSPID = [Kps=2.66, Kis=0.009, Kds=32] to minimise the mean squared error of the system.”
- It is recommended to supplement the GA parameter settings and optimization convergence curve to enhance reproducibility.
The GA optimisation details (in cited previous research), the ANN convergence curve and reproducibility are not a part of the present investigation. The focus is on the accuracy of the final BANN model which training is based on Matlab algorithm that includes also check for generality and validation.
- The images in the text are not clear, and the axis titles are missing
Thank you, corrected
Reviewer 3 Report
Comments and Suggestions for AuthorsThis paper addresses a meaningful industrial problem: gain-scheduled PID control for nonlinear plants using a backpropagation neural network. The work is well‑organized, technically solid, and the proposed BANN–PID method shows clear advantages in dynamic performance, robustness, and energy efficiency. The simulation comparisons with PID and FL–SPID are comprehensive. However, experimental validation is relatively limited, and several parts can be strengthened to improve rigor and readability.
1.Add quantitative performance improvements (overshoot, settling time, robustness) versus PID and FL–SPID to make contributions more concrete.
2.Add 2–3 recent references (2023–2026) on neural gain-scheduling and robust PID for thermal/industrial processes.
3.Justify the choice of 5 hidden neurons with a brief comparison (3/5/7/9 nodes) in terms of error, convergence, and generalization.
4.Add plain‑language interpretation of Equations (2) and (3) to help engineering readers understand the physical meaning.
5.Add a short sensitivity study for parameters A, B, C to show their influence on performance and robustness.
6.Add simulation results under external disturbances (load, ambient temperature, noise) to fully validate robustness.
7.The manuscript relies mainly on simulation. Please add real-time experimental results from the actual fruit dryer to verify the proposed BANN–PID in practice. Provide a direct, quantitative comparison between simulated responses and measured experimental responses under the same reference commands. Include experimental validation under real operating variations (e.g., different fruit loads, ambient conditions) to prove industrial feasibility.
8. Grammar check:
- Missing articles: based an offline trained → based on an offline-trained
- Plural consistency: parameter → parameters
- Hyphenation: gain scheduled → gain-scheduled; offline trained → offline-trained
- Split long sentences for better readability
- Use italic for variables, Roman for units and functions
- Final proofread for minor typos and spacing
Author Response
- Add quantitative performance improvements (overshoot, settling time, robustness) versus PID and FL–SPID to make contributions more concrete.
Table 1 is revised and explained in the text above it “The ratios of mean values (in blue) in Table 1 indicate how much higher the PID and SPID systems’ indicators are compared to the BANN-PID system. Overall, the BANN-PID system performs on average 4.8 times better than the PID system and 2.2 times better than the SPID system.”
- Add 2–3 recent references (2023–2026) on neural gain-scheduling and robust PID for thermal/industrial processes.
Three papers are added with corresponding text in the Introduction:
[16] Almachi, J.C.; Vicente, R., Bone, E.; Montenegro, J.; Cando, E, Reina, S. Implementation of a neural network for adaptive PID tuning in a high-temperature thermal system. Energies, 18: 1-25, 3113, 2025. https://doi.org/10.3390/en18123113
p.2 end:” In [16] an adaptive PID control system based on complex three-layered BANN is developed for the temperature in an industrial furnace. The BANN predicts the proportional and derivative controller’s gains from the current measured system error, temperature and time. The training dataset is obtained from experiments and simulations and standardized. The BANN is validated and tested for generalization, overfitting, variance, bias, sensitivity to inputs’ perturbations, execution time and ability to distinguish between heating and cooling. The system is validated in Python-LabView integrated real time control for following a desired temperature ramp profile of heating, stabilization and cooling sections. It outperforms the PID and fuzzy logic (FL) PID systems designed in previous research”
[18old] Han, S. Y., Dong, J. F., Zhou, J., Chen, Y. H., Adaptive fuzzy PID control strategy for vehicle active suspension based on road evaluation. Electronics. 11:1-21, 921. 2022. https://doi.org/10.3390/electronics11060921 - recommended by a reviewer
p.3 “In [18] a PID controller for an active suspension system is developed and validated via simulations to auto-tune its parameters based of FL estimation of the road conditions”
[24] Karlova-Sergieva, V. Robust Performance Assessment of Control Systems with Root Contours Analysis. Cybernetics and Information Technologies, Vol. 25, 2025, No 2, pp. 83-99.
Now the references after [16] have different numbers
- Justify the choice of 5 hidden neurons with a brief comparison (3/5/7/9 nodes) in terms of error, convergence, and generalization.
Added text in red in lines 210-211 “It has five hidden neurons, the commonly selected, and three output neurons for the three assessed PID controller gains, as depicted in Fig.4”
Comparison with other architectures is not a part of the present investigation. The accuracy achieved with 5 hidden neurons is accurate enough and simple for PLC implementation for the purpose of gain schedule control.
- Add plain‑language interpretation of Equations (2) and (3) to help engineering readers understand the physical meaning.
The explanation of (2) and (3) is related with Fig.7 – the paragraph before Fig.7 and the text after it. It is well known for control engineers and one of the other reviewers who find this very simple..
- Add a short sensitivity study for parameters A, B, C to show their influence on performance and robustness.
The long-term engineering practice shows that the impact of A, B, C on the performance is equivalent. A, B, C in the given ranges are used for years for fast engineering PID tuning that ensures system stability and acceptable performance.
Here we GA optimise A, B and C within their ranges to minimise the MSE in the closed loop system.
- Add simulation results under external disturbances (load, ambient temperature, noise) to fully validate robustness.
Simulated disturbance step responses are added to Fig.8b) and explanations in the text.
- The manuscript relies mainly on simulation. Please add real-time experimental results from the actual fruit dryer to verify the proposed BANN–PID in practice. Provide a direct, quantitative comparison between simulated responses and measured experimental responses under the same reference commands. Include experimental validation under real operating variations (e.g., different fruit loads, ambient conditions) to prove industrial feasibility.
The future work will focus on implementation via PLC of the developed gain-scheduled PID system for real time control – pls. see lines 446-448
Now we trust simulations which yield close to real world processes, because of the use of validated MTSK and Simulink PID models against experimental data in previous research–pls see lines 177-178 and also Fig. 10 where the step responses of the simulated PID and the real-time PLC systems are nearly identical, thus validating the models.
Because the BANN gain-scheduled PID outperforms the traditional PID and SPID systems in simulations, a similar improvement of the temperature control can be expected when the BANN gain-scheduled PID is implemented on an industrial PLC and used in real time control.
- Grammar check:
- Missing articles: based an offline trained → based on an offline-trained
- Plural consistency: parameter → parameters
- Hyphenation: gain scheduled → gain-scheduled; offline trained → offline-trained
- Split long sentences for better readability
- Use italic for variables, Roman for units and functions
- Final proofread for minor typos and spacing
Thank you - corrected
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsNo further comments.
Reviewer 2 Report
Comments and Suggestions for AuthorsThe paper has been revised and I have no question