Temperature Estimation in Induction Motors using Machine Learning

Dinan Li; Panagiotis Kakosimos

REVIEW 3 major objections 6 minor 20 references

Temperature Estimation in Induction Motors using Machine Learning

T0 review · 3 major / 6 minor · reviewed 2026-08-16 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A multilayer perceptron trained on drive signals can estimate induction motor winding and bearing temperatures with mean squared errors below one degree under the tested transients.

desk verdict The paper's ML comparison is competent, but the headline precision claim rests on a single operating profile; the missing aggregate leave-one-out results are a fixable but real reporting gap. read the letter →

arxiv 2504.18105 v1 pith:LFNOWQGN submitted 2025-04-25 cs.LG

Dinan Li , Panagiotis Kakosimos This is my paper · ORCID

classification cs.LG

keywords inductionmotormachinelearningtemperatureestimationmonitoringmultilayerperceptronconvolutionalneuralnetworktransientoperationstatorwinding

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that data-driven models, trained on signals already available in a modern electric drive, can estimate internal motor temperatures accurately enough for commercial monitoring. The authors compare linear regression, a multilayer perceptron, and a temporal convolutional network on experimental data from a test bench running eighteen operating profiles of slow, medium, and fast dynamics. On the representative profile they report, the linear baseline has MSE 1.94, the MLP reaches 0.31, and the CNN 0.54. The claim matters because winding and bearing overheating is a major failure cause, and accurate temperature estimates without additional sensors would support condition monitoring and control optimization.

What carries the argument

The central mechanism is a supervised regression pipeline built on drive-available inputs: motor speed, stator current, and shell temperature, each normalized and expanded with exponentially weighted moving averages over eight span values so that non-sequential models such as linear regression and the MLP still carry memory of past behavior. The MLP is configured with two hidden layers of 90 and 20 neurons, and the temporal CNN uses 1D convolution with small filters and dilation, followed by global pooling and a dense regression head. The load-bearing work is done by the feature-expansion step together with a leave-one-profile-out training scheme that removes each of the eighteen operating profiles from training once, so every profile serves as test data at least once.

What would settle it

Run the same leave-one-profile-out evaluation and report the mean and standard deviation of MSE across all eighteen folds; if the average is close to the single-profile values (linear 1.94, MLP 0.31, CNN 0.54), the central claim holds, whereas a large gap or an obviously favorable choice of reported profile would show the claimed precision does not generalize.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that a two-layer multilayer perceptron, fed with normalized speed, current, and shell temperature expanded by exponentially weighted moving averages, estimates stator winding and bearing temperatures under transient operation with high precision. It beats both a linear baseline (MSE 1.94) and a three-layer temporal CNN (MSE 0.54) on the reported operating profile, achieving MSE 0.31. The authors state that neural networks performed satisfactorily even under fast transient conditions and that errors remained mostly within a few degrees Celsius, which they regard as acceptable for commercial applications. The paper frames this as opening the way to deploying similar models in commercial drives using only signals the drive already collects.

Load-bearing premise

The load-bearing premise is that the single operating profile used for the reported results is representative of all eighteen test profiles, since the quoted MSE values would overstate typical performance if that profile was selected for its favorable errors.

Editorial extensions

If this is right

If the reported results hold, a two-layer MLP can reduce temperature-estimation error by roughly a factor of six compared with a linear baseline (MSE 0.31 versus 1.94) on a profile mixing slow and fast transients.
Because the same input set, speed, current, and shell temperature, is already logged by modern drives, the approach could be deployed without installing internal temperature sensors.
Because the CNN performs slightly worse than the MLP despite seeing sequences, the explicit memory features appear to carry most of the temporal information needed for this task.
Accurate internal temperature estimates would let a drive monitor winding and bearing health continuously and act before thermal limits are exceeded, supporting condition monitoring and control optimization.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper reports detailed errors for only one representative profile; the leave-one-profile-out design implies that the other seventeen folds also produced results, but their mean and spread are not shown, so the quoted MSE values should be read as one draw rather than a verified average.
The feature expansion uses eight exponentially weighted moving-average spans, which suggests the model's memory length is a key tuning factor; one could test whether the MLP's advantage over the CNN shrinks or grows with the number of spans.
If extended beyond the reported test bench, the approach could be checked for robustness to motor-to-motor variation, cooling condition changes, and aging, none of which the current experimental design varies.
A controlled comparison matching model size and training budget would clarify whether the MLP's edge over the CNN comes from the architecture or from the hyperparameter search, a distinction the paper does not separate.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

The paper is a straightforward engineering comparison of linear, MLP, and CNN models for estimating induction-motor winding and bearing temperatures from drive-available signals. Good points first: the leave-one-profile-out protocol is the right way to test generalization, the feature expansion with exponentially weighted moving averages is a sensible and small practical contribution, and the reported numbers on the shown profile are in line with what you'd expect — MLP beats linear, CNN is in between. The experimental data are real lab data, and the authors only use signals a drive would have, which is the right constraint for deployment.

Where it gets soft: Section V shows results for one 'representative' operating profile and nothing else. The paper says all eighteen profiles were used as test data at least once, but there are no aggregate statistics — no mean or median MSE, no variance, no worst case. So the 'high precision' claim in the introduction and conclusion is only actually demonstrated for one profile. If that profile was picked because it looks good, the quoted MSE values overstate typical performance. The stress-test note is correct. This is a reporting gap rather than a modeling flaw, and it's fixable, but it is the difference between a supported claim and a plausible one. The hyperparameter description also has holes (Table III's contents are largely missing from the text I have), and there is no code or data release, which is a real limitation for an empirical engineering paper.

The citation pattern is fair: the prior ML temperature-estimation work is properly acknowledged, and this paper does not oversell its novelty. No circularity that I can see: hyperparameters were chosen separately and the evaluation is on held-out profiles.

Bottom line: this is a competent, honest subfield paper. It deserves referee time, not a desk reject. But the review should require the authors to report leave-one-out aggregate results across all eighteen profiles and explain how the representative profile was selected. That is a reasonable revision, and with that addition the paper would make a useful data point for anyone building drive-integrated thermal monitoring.

Referee Report

3 major / 6 minor

Summary. The paper develops and compares three machine-learning models (elastic-net-regularized linear regression, a two-layer MLP, and a temporal CNN) for estimating stator-winding and bearing temperatures of an induction motor from drive-available signals (speed, current, shell temperature). Data were collected on a lab test bench under eighteen operating profiles with slow, medium, and fast dynamics; a leave-one-profile-out scheme is used for evaluation. A hyperparameter search is performed for each model. In Section V, quantitative results are reported for one representative profile, giving MSE values of 1.94 (linear), 0.31 (MLP), and 0.54 (CNN). The paper concludes that neural networks estimate the motor's internal temperatures with high precision, including under transient conditions.

Significance. If the reported performance is representative, the results are practically significant: a data-driven model using only signals already available in a drive could replace dedicated internal temperature sensors, with errors mostly within a few degrees Celsius. The leave-one-profile-out protocol is a sound approach for assessing generalization to unseen operating conditions, and the comparison of a linear baseline against nonlinear networks is appropriate. The lower MSE of the MLP relative to the linear model is consistent with the expected nonlinearity of motor thermal dynamics. However, the evidence presented in the manuscript is incomplete in two load-bearing respects: only one test profile is analyzed in detail, with no aggregate statistics across the eighteen folds, and key tables describing the experimental setup and final model configurations are referenced but not present in the manuscript text.

major comments (3)

[Section V, Section IV-C] The central claim of 'high precision' (Introduction and Conclusion) is supported only by results on 'one representative operating profile' in Section V. Section IV-C states that all eighteen generated profiles were used as test data at least once in a leave-one-out scheme, but Section V provides no aggregate statistics for the other seventeen folds: no mean or median MSE, no variance, no worst-case error, and no per-fold table. If the selected profile is not typical, the quoted values (linear 1.94, MLP 0.31, CNN 0.54) overstate typical performance. Please report the full distribution of per-profile metrics, or explicitly justify the representativeness of the selected profile with quantitative evidence.
[Table III and Section IV-B] The manuscript text refers to Table I (motor specifications), Table II (model inputs and outputs), and Table III ('the finally selected model configurations'), but none of these tables appears in the manuscript. Without the final hyperparameters (regularization penalty and mixing ratio for the linear model, layer sizes and dropout for the MLP, layer count, filter sizes, dilation rate, and input sequence length for the CNN), the exponentially weighted moving average span values, and the exact input/output variable list, the experiments cannot be reproduced or independently checked. Please include the full content of all three tables.
[Section IV-C] The hyperparameter search is described only in broad terms: 'a lot of parameters are fine-tuned' and 'another small set of profiles has been taken away as the validation set.' The number of validation profiles, the split strategy, and the search protocol (grid or random search, number of trials, early-stopping criterion) are not specified. Because the final models are selected by this search, the absence of these details weakens the claim that the reported configurations are optimal and makes the comparison between models less interpretable.

minor comments (6)

[Section II-A] There is a typo immediately before Eq. (1): 'I ts mathematical formulation' should read 'Its mathematical formulation.'
[Throughout] The term 'multilayer perception' should be 'multilayer perceptron,' and 'Convolutional neuron network' should be 'Convolutional neural network.'
[Section IV-B] The sentence 'eight span values ranging' is incomplete; the actual span values are not listed.
[Section V] Figures 7–9 show error and input signals, but axis labels and units are not visible in the manuscript text; please ensure all axes are labeled.
[Section V-B] The statement that the MLP 'shows an obvious difference between slow and fast dynamics' is not quantified; a per-segment error comparison would make the claim about transient performance precise.
[References] References [1] and [13] are the same work; please consolidate to avoid duplication.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the temperature estimates are genuine out-of-sample predictions from models fitted to measured drive signals and temperatures.

full rationale

The paper's derivation chain is self-contained as an empirical ML study. Input features (speed, current, shell temperature) and targets (winding and bearing temperatures) are independent measurements from an experimental test bench. Models are trained by minimizing MSE on training profiles, hyperparameters are chosen by a separate search, and the reported test MSE values (linear 1.94, MLP 0.31, CNN 0.54) are computed on a held-out operating profile after training. No equation defines a model output in terms of the target it is supposed to predict, and no fitted parameter is renamed as a prediction. The only author self-citation, reference [10] on machine-learning-based adjustments of thermal networks, appears in the introduction as background and is not load-bearing for any step of the present derivation; no uniqueness theorem or ansatz is imported from that work. The main legitimate concern is reporting evidence from a single representative profile rather than aggregating all eighteen leave-one-out folds, but that is a potential selection-bias or reporting-completeness issue, not circularity. The reported errors are genuinely predictive in the sense that the evaluation data were not used to fit the models.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim rests on standard ML background, three domain assumptions about sensor availability and data representativeness, and a set of hyperparameters selected by search. No new physical entities are introduced. The main burden is that the evaluation depends on the representativeness of the lab profiles and the single reported test profile.

free parameters (7)

Linear regularization penalty (alpha) = 0.43
Selected by hyperparameter search (Section V-A); controls the strength of L1/L2 penalty in the SGD regressor.
Linear L1 mixing ratio (l1_ratio) = 0.99
Selected by hyperparameter search (Section V-A); the model is almost pure Lasso.
MLP hidden layer sizes = [90, 20]
Selected by hyperparameter search (Section V-B); number of layers fixed to two.
MLP dropout ratio = not reported in text
Listed as a hyperparameter in Section III-B, final value is not given in the available text.
CNN architecture (layers, filter sizes, dilation rate) = 3 layers, small filters, large dilation on first layer (exact values not reported)
Selected by hyperparameter search (Section V-C); details are summarized in Table III, which is not reproduced in the text.
CNN input sequence length = not reported
Hyperparameter for the temporal CNN (Section III-C); final value not given.
EWMA span values = eight spans, values not reported
Exponentially weighted moving average spans for feature expansion (Section IV-B) are hyperparameters tuned during the search; values are not reported.

assumptions (5)

standard math Universal approximation and trainability of MLPs with sufficient hidden units
Section II-C asserts that an MLP with a single hidden layer can estimate any target, and stacking layers improves generalizability; this is background ML theory.
standard math Gradient descent with backpropagation converges to a useful minimum for the chosen architectures and hyperparameters
Section II-A and II-C rely on gradient descent minimization without convergence guarantees.
domain assumption The selected inputs (speed, current, shell temperature) and EWMA-expanded features are sufficient to determine winding and bearing temperatures
Section IV-B selects features based on a correlation matrix; no physical thermal model or causality argument is provided.
domain assumption The 18 lab-generated operating profiles are representative of real-world operating conditions
Section IV-A describes profiles generated under predetermined slow/medium/fast dynamics; no field data or external validation is presented.
domain assumption A shell/reference temperature sensor is available as a model input at deployment
Section IV-B states shell temperature provides boundary conditions and environmental context; the paper does not discuss deployments where this signal is unavailable.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Temperature Estimation in Induction Motors using Machine Learning." pith.science (2026). https://pith.science/paper/LFNOWQGN

@misc{pith2026250418105,
  author       = {Pith},
  title        = {Pith review of: Temperature Estimation in Induction Motors using Machine Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LFNOWQGN}},
  note         = {Machine review of arXiv:2504.18105}
}

read the original abstract

The number of electrified powertrains is ever increasing today towards a more sustainable future; thus, it is essential that unwanted failures are prevented, and a reliable operation is secured. Monitoring the internal temperatures of motors and keeping them under their thresholds is an important first step. Conventional modeling methods require expert knowledge and complicated mathematical approaches. With all the data a modern electric drive collects nowadays during the system operation, it is feasible to apply data-driven approaches for estimating thermal behaviors. In this paper, multiple machine-learning methods are investigated on their capability to approximate the temperatures of the stator winding and bearing in induction motors. The explored algorithms vary from linear to neural networks. For this reason, experimental lab data have been captured from a powertrain under predetermined operating conditions. For each approach, a hyperparameter search is then performed to find the optimal configuration. All the models are evaluated by various metrics, and it has been found that neural networks perform satisfactorily even under transient conditions.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 13 canonical work pages

[2]

A Review of Thermal Monitoring Techniques for Radial Permanent Magnet Machines,

G. Carbone, B. Vogel -Heuser, D. Zhang, T. Meng, and P. Zhang, “A Review of Thermal Monitoring Techniques for Radial Permanent Magnet Machines,” Mach. 2022, Vol. 10, Page 18 , vol. 10, no. 1, p. 18, 2021, doi: 10.3390/MACHINES10010018

work page doi:10.3390/machines10010018 2022
[3]

Critical issues of applying machine learning to condition monitoring for failure diagnosis,

F. Q. Yuan, “Critical issues of applying machine learning to condition monitoring for failure diagnosis, ” in IEEE International Conference on Industrial Engineering and Engineering Management , Dec. 2016, pp. 1903–1907. doi: 10.1109/IEEM.2016.7798209

work page arXiv 2016
[4]

Supervised Machine Learning Algorithm Selection for Condition Monitoring of Induction Motors,

N. Rajapaksha, S. Jayasinghe, H. Enshaei, and N. Jayarathne, “Supervised Machine Learning Algorithm Selection for Condition Monitoring of Induction Motors, ” in IEEE Southern Power Electronics Conference , Dec. 2021 , pp. 1 –10. doi: 10.1109/SPEC52827.2021.9709436

work page arXiv 2021
[5]

Machine Learning for Condition Monitoring and Innovation,

N. H. Pontoppidan, T. Lehn -Schioler, and K. B. Petersen, “Machine Learning for Condition Monitoring and Innovation, ” in IEEE International Conference on Acoustics, Speech and Signal Processing, May 2019, pp. 8067 –8071. doi: 10.1109/ICASSP.2019.8683545

arXiv 2019
[6]

Estimating Electric Motor Temperatures with Deep Residual Machine Learning,

W. Kirchgassner, O. Wallscheid, and J. Bocker, “Estimating Electric Motor Temperatures with Deep Residual Machine Learning, ” IEEE Trans. Power Electron., vol. 36, no. 7, pp. 7480 –7488, Jul. 2021, doi: 10.1109/TPEL.2020.3045596

arXiv 2021
[7]

Deep Residual Convolutional and Recurrent Neural Networks for Temperature Estimation in Permanent Magnet Synchronous Motors,

W. Kirchgassner, O. Wallscheid, and J. Bocker, “Deep Residual Convolutional and Recurrent Neural Networks for Temperature Estimation in Permanent Magnet Synchronous Motors, ” in IEEE International Electric Machines & Drives Conference, May 2019, pp. 1439–1446. doi: 10.1109/IEMDC.2019.8785109

work page arXiv 2019
[8]

Towards NN -based online estimation of the full -chip temperature and the rate of temperature change,

M. Rapp, O. Elfatairy, M. Wolf, J. Henkel, and H. Amrouch, “Towards NN -based online estimation of the full -chip temperature and the rate of temperature change, ” ACM/IEEE Work. Mach. Lea rn. CAD, pp. 95–100, Nov. 2020, doi: 10.1145/3380446.3430648

arXiv 2020
[9]

IGBT Junction Temperature Estimation Based on Machine Learning Method,

J. G. Miao, Q. Yin, H. Wang, Y. Liu, H. Li, and S. Duan, “IGBT Junction Temperature Estimation Based on Machine Learning Method,” 9th Int. Power Electron. Motion Control Conf. , vol. 2020 - January, Nov. 2020, doi: 10.1109/IPEMC - ECCEASIA48364.2020.9395640

work page arXiv 2020

Show all 20 references

[10]

Machine learning-based adjustments of thermal networks,

K. R önnberg, P. Kakosimos, Z. Kolondjovski, and E. Nordlund, “Machine learning-based adjustments of thermal networks, ” in 11th International Conference on Power Electronics, Machines and Drives, Jun. 2022, pp. 424–428. doi: 10.1049/icp.2022.1087

2022
[11]

Batteries State of Health Estimation via Efficient Neural Networks with Multiple Channel Charging Profiles,

N. Khan, F. U. M. Ullah, Afnan, A. Ullah, M. Y. Lee, and S. W. Baik, “Batteries State of Health Estimation via Efficient Neural Networks with Multiple Channel Charging Profiles, ” IEEE Access, vol. 9, pp. 7797–7813, 2021, doi: 10.1109/ACCESS.2020.3047732

2021
[12]

Temperature Estimation of PMSM Using a Difference-Estimating Feedforward Neural Network,

J. Lee and J. I. Ha, “Temperature Estimation of PMSM Using a Difference-Estimating Feedforward Neural Network, ” IEEE Access , vol. 8, pp. 130855 –130865, 2020, doi: 10.1109/ACCESS.2020.3009503

2020
[13]

Machine Learning for Sensorless Temperature Estimation of a BLDC Motor,

D. Czerwinski, J. G ęca, and K. Kolano, “Machine Learning for Sensorless Temperature Estimation of a BLDC Motor,” Sensors 2021, Vol. 21, Page 4655 , vol. 21, no. 14, p. 4655, Jul. 2021, doi: 10.3390/S21144655

2021 doi
[14]

C. M. Bishop and N. M. Nasrabadi, Pattern recognition and machine learning, vol. 4, no. 4. Springer, 2006

2006
[15]

Raschka and V

S. Raschka and V. Mirjalili, Python machine learning: Machine learning and deep learning with Python, scikit-learn, and TensorFlow
[16]

Packt Publishing Ltd, 2019

2019
[17]

An overview of gradient descent optimization algorithms,

S. Ruder, “An overview of gradient descent optimization algorithms,” arXiv Prepr. arXiv1609.04747, 2016

2016 arXiv
[18]

Yan and X

X. Yan and X. Su, Linear regression analysis: theory and computing. world scientific, 2009

2009
[19]

Model selection and multi -model inference,

D. Anderson and K. Burnham, “Model selection and multi -model inference,” Second. NY Springer -Verlag, vol. 63, no. 2020, p. 10, 2004

2020
[20]

Deep sparse rectifier neural networks,

X. Glorot, A. Bordes, and Y. Bengio, “Deep sparse rectifier neural networks,” in Proceedings of the fourteenth international conference on artificial intelligence and statistics, 2011, pp. 315–323

2011
[21]

Temporal convolutional networks for anomaly detection in time series,

Y. He and J. Zhao, “Temporal convolutional networks for anomaly detection in time series, ” in Journal of Physics: Conference Series , 2019, vol. 1213, no. 4, p. 42050

2019

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.

[1] [2]

A Review of Thermal Monitoring Techniques for Radial Permanent Magnet Machines,

G. Carbone, B. Vogel -Heuser, D. Zhang, T. Meng, and P. Zhang, “A Review of Thermal Monitoring Techniques for Radial Permanent Magnet Machines,” Mach. 2022, Vol. 10, Page 18 , vol. 10, no. 1, p. 18, 2021, doi: 10.3390/MACHINES10010018

work page doi:10.3390/machines10010018 2022

[2] [3]

Critical issues of applying machine learning to condition monitoring for failure diagnosis,

F. Q. Yuan, “Critical issues of applying machine learning to condition monitoring for failure diagnosis, ” in IEEE International Conference on Industrial Engineering and Engineering Management , Dec. 2016, pp. 1903–1907. doi: 10.1109/IEEM.2016.7798209

work page arXiv 2016

[3] [4]

Supervised Machine Learning Algorithm Selection for Condition Monitoring of Induction Motors,

N. Rajapaksha, S. Jayasinghe, H. Enshaei, and N. Jayarathne, “Supervised Machine Learning Algorithm Selection for Condition Monitoring of Induction Motors, ” in IEEE Southern Power Electronics Conference , Dec. 2021 , pp. 1 –10. doi: 10.1109/SPEC52827.2021.9709436

work page arXiv 2021

[4] [5]

Machine Learning for Condition Monitoring and Innovation,

N. H. Pontoppidan, T. Lehn -Schioler, and K. B. Petersen, “Machine Learning for Condition Monitoring and Innovation, ” in IEEE International Conference on Acoustics, Speech and Signal Processing, May 2019, pp. 8067 –8071. doi: 10.1109/ICASSP.2019.8683545

arXiv 2019

[5] [6]

Estimating Electric Motor Temperatures with Deep Residual Machine Learning,

W. Kirchgassner, O. Wallscheid, and J. Bocker, “Estimating Electric Motor Temperatures with Deep Residual Machine Learning, ” IEEE Trans. Power Electron., vol. 36, no. 7, pp. 7480 –7488, Jul. 2021, doi: 10.1109/TPEL.2020.3045596

arXiv 2021

[6] [7]

Deep Residual Convolutional and Recurrent Neural Networks for Temperature Estimation in Permanent Magnet Synchronous Motors,

W. Kirchgassner, O. Wallscheid, and J. Bocker, “Deep Residual Convolutional and Recurrent Neural Networks for Temperature Estimation in Permanent Magnet Synchronous Motors, ” in IEEE International Electric Machines & Drives Conference, May 2019, pp. 1439–1446. doi: 10.1109/IEMDC.2019.8785109

work page arXiv 2019

[7] [8]

Towards NN -based online estimation of the full -chip temperature and the rate of temperature change,

M. Rapp, O. Elfatairy, M. Wolf, J. Henkel, and H. Amrouch, “Towards NN -based online estimation of the full -chip temperature and the rate of temperature change, ” ACM/IEEE Work. Mach. Lea rn. CAD, pp. 95–100, Nov. 2020, doi: 10.1145/3380446.3430648

arXiv 2020

[8] [9]

IGBT Junction Temperature Estimation Based on Machine Learning Method,

J. G. Miao, Q. Yin, H. Wang, Y. Liu, H. Li, and S. Duan, “IGBT Junction Temperature Estimation Based on Machine Learning Method,” 9th Int. Power Electron. Motion Control Conf. , vol. 2020 - January, Nov. 2020, doi: 10.1109/IPEMC - ECCEASIA48364.2020.9395640

work page arXiv 2020