Intelligent Self-tuning Active EMI Filtering for Electrified Automotive Power Systems Using Reinforcement Learning

Kelin Jia; Mahuizi Lu; Rajib Goswami; Yukun Hu

arxiv: 2604.28084 · v1 · submitted 2026-04-30 · 📡 eess.SY · cs.SY

Intelligent Self-tuning Active EMI Filtering for Electrified Automotive Power Systems Using Reinforcement Learning

Mahuizi Lu , Kelin Jia , Rajib Goswami , Yukun Hu This is my paper

Pith reviewed 2026-05-07 05:26 UTC · model grok-4.3

classification 📡 eess.SY cs.SY

keywords electromagnetic interferenceactive EMI filterreinforcement learningautomotive power systemsself-tuning filterelectric vehiclesvariational autoencoder

0 comments

The pith

Reinforcement learning lets active EMI filters self-tune in real time to suppress interference 25-30 dB better than fixed or passive methods in automotive power systems.

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

The paper shows how to frame active EMI filtering as a Markov decision process so a reinforcement learning agent can adjust filter settings on the fly as interference changes with vehicle operation. A variational autoencoder compresses the high-dimensional EMI state into a compact representation that supports generalization, while noise-based exploration prevents the agent from getting stuck in poor policies during training. The method is tested in co-simulation that replays experimentally measured spectra from an automotive electric drive unit. It produces consistent 25-30 dB attenuation gains over a wide frequency range relative to conventional fixed controllers and passive filters. This matters because oversized passive filters add weight, volume, and losses; an adaptive active solution can shrink those penalties while still protecting safety-critical electronics and communications under dynamic conditions.

Core claim

The authors formulate EMI mitigation as a Markov decision process and train a reinforcement learning agent to adapt the parameters of an active EMI filter in response to time-varying interference. They augment the agent with a variational autoencoder for compact state representation and a noise-based exploration mechanism to improve robustness under non-stationary conditions. When the trained policy is evaluated inside a MATLAB/Simulink co-simulation driven by experimentally measured EMI spectra from an automotive electric drive unit, it delivers 25-30 dB greater attenuation across a broad frequency band than conventional fixed-control strategies or passive filtering solutions.

What carries the argument

A reinforcement learning agent that models EMI mitigation as a Markov decision process and continuously updates active filter parameters, using a variational autoencoder to compress the observed interference spectrum into a compact state representation.

If this is right

Active EMI filters can be made smaller and lighter because the RL policy compensates for changing interference instead of relying on worst-case passive components.
Electromagnetic compatibility performance remains high across the full range of vehicle operating points without manual retuning of filter parameters.
Energy losses associated with oversized passive filters decrease, supporting higher overall efficiency in electrified powertrains.
Safety-critical in-vehicle communications and control functions receive more reliable protection from switching noise under transient conditions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same MDP-plus-RL formulation could be applied to other switched-mode converters in industrial drives or renewable inverters where load and interference profiles also vary continuously.
If the policy proves robust on hardware, it opens the possibility of embedding the agent directly in the vehicle control unit for predictive adaptation based on driving mode or speed forecasts.
Longer-term, successful field validation might encourage standards bodies to consider performance requirements for adaptive rather than purely passive EMI mitigation in future vehicle homologation.

Load-bearing premise

The reinforcement learning agent trained in co-simulation with measured spectra will generalize and remain stable under the full range of real-vehicle non-stationary operating conditions, sensor noise, and hardware imperfections not captured in the MATLAB/Simulink environment.

What would settle it

A hardware-in-the-loop test or on-vehicle measurement campaign that records EMI spectra at the electric drive unit under varying speeds, torques, and auxiliary loads and checks whether the 25-30 dB improvement over baseline filters persists outside the simulation.

Figures

Figures reproduced from arXiv: 2604.28084 by Kelin Jia, Mahuizi Lu, Rajib Goswami, Yukun Hu.

**Figure 1.** Figure 1: The schematic depicting the application of our proposed method workflow 3 view at source ↗

**Figure 2.** Figure 2: The circuit with damping and compensation components view at source ↗

**Figure 3.** Figure 3: The structure of VAE: the left-side encoder maps input to the view at source ↗

**Figure 4.** Figure 4: The framework of the MLP constitution Following the feature extraction and data encoding process, the output layer of the MLP is directly linked to the estimation of action values. The agent uses the observed states to predict the expected values for each action and selects actions based on these predictions. Simultaneously, it updates the MLP’s weights using rewards and new state information obtained from… view at source ↗

**Figure 5.** Figure 5: The flowchart of the proposed EQRL algorithm view at source ↗

**Figure 6.** Figure 6: Spearman correlation heatmap and significance analysis view at source ↗

**Figure 7.** Figure 7: Noise measurement from EDU LV harness In this figure, EMI noise is detected from the LV harness using a high-bandwidth current transformer (CT). The LV harness comprises cables originating from the Inverter, Motor, and other active sub-components of the EDU. The Line Impedance Stabilisation Network (LISN) lies between the LV DC power supply and the component side of the harness. In the present work, the … view at source ↗

**Figure 8.** Figure 8: Extraction cable bundle dataset insertion loss and EMI signal filtering view at source ↗

**Figure 10.** Figure 10: Cable 44 dataset insertion loss and EMI signal filtering performance view at source ↗

**Figure 9.** Figure 9: Cable 44 dataset insertion loss and EMI signal filtering performance view at source ↗

**Figure 11.** Figure 11: Cable 47 dataset with disturbance signal insertion loss and EMI view at source ↗

read the original abstract

The rapid electrification and intelligence of modern transportation systems place stringent demands on the electromagnetic compatibility, reliability, and adaptability of automotive power electronics. In electric and autonomous vehicles, electromagnetic interference (EMI) generated by high-frequency switching power converters can compromise safety-critical functions, in-vehicle communications, and system efficiency under dynamic operating conditions. Conventional passive EMI filters, while robust, are often oversized and lack adaptability, leading to increased weight, volume, and energy losses. This paper proposes an intelligent self-tuning active EMI filtering approach for electrified automotive power systems based on reinforcement learning (RL). The EMI mitigation problem is formulated as a Markov decision process, enabling an RL agent to continuously adapt filter parameters in response to time-varying interference characteristics. To improve robustness and generalisation under complex and non-stationary conditions, a variational autoencoder is employed for compact state representation, while a noise-based exploration mechanism enhances learning efficiency and prevents suboptimal convergence. The proposed method is evaluated using experimentally measured EMI spectra from an automotive electric drive unit within a MATLAB/Simulink co-simulation framework. Results demonstrate consistent EMI attenuation improvements of 25-30 dB across a wide frequency range compared with conventional control strategies and passive filtering solutions. By reducing reliance on oversized passive components and enabling adaptive EMI suppression, the proposed framework supports lightweight, energy-efficient, and reliable power-electronic systems for intelligent and green transportation applications.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper applies RL with a VAE state encoder to adaptively tune active EMI filters and reports 25-30 dB gains in co-simulation with measured spectra, but the results never leave MATLAB/Simulink.

read the letter

The main thing to know is that the authors cast active EMI filter tuning as an MDP, use a variational autoencoder to compress the state, and add noise-based exploration, then test the policy in co-simulation driven by real spectra from an automotive drive unit. They claim consistent 25-30 dB better attenuation than conventional methods and passive filters. That combination of RL elements is not a routine extension of existing work in this domain, and grounding the simulation in measured data is a clear step up from synthetic spectra alone. The formulation itself is sensible for handling non-stationary interference in vehicle power systems. They also avoid obvious circularity by treating attenuation as an outcome rather than a fitted quantity. Those are the concrete strengths. The central limitation is that everything stays in co-simulation. The policy is never run closed-loop on hardware, so the reported gains do not account for sensor noise, PWM timing jitter, component temperature drift, or the filter's effect on DC-link voltage and motor currents. These mismatches are known to change both the disturbance and the achievable cancellation in real automotive converters. The abstract gives no training curves, error bars, statistical tests, or detailed baseline comparisons, which makes it hard to judge how reliable the 25-30 dB number actually is. This work is aimed at power-electronics researchers who are willing to consider RL for adaptive control and at RL people looking for a grounded application. A reader focused on practical EMI solutions in EVs would find the MDP-plus-VAE setup useful to examine. The paper shows honest engagement with the problem and a non-trivial technical approach, so it deserves a serious referee. I would send it out for review but would ask the referees to press on hardware validation or a quantified sim-to-real gap analysis.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes a reinforcement learning (RL) approach for self-tuning active EMI filters in electrified automotive power systems. The EMI mitigation task is formulated as a Markov decision process, with a variational autoencoder used for compact state representation and a noise-based exploration strategy to improve learning. The method is evaluated exclusively in a MATLAB/Simulink co-simulation environment driven by experimentally measured EMI spectra from an automotive electric drive unit, with results claiming consistent 25-30 dB attenuation improvements across a wide frequency range relative to conventional control strategies and passive filters.

Significance. If the simulation results prove robust, the work could contribute to more compact, adaptive EMI mitigation in electric vehicles, potentially reducing the size and weight of passive components while addressing dynamic operating conditions. The combination of RL with a VAE for non-stationary interference handling represents a reasonable application of machine learning to power-electronics control. The grounding in measured spectra is a positive aspect that distinguishes the evaluation from purely synthetic simulations.

major comments (2)

[Results section] Results section: The headline performance claim of 25-30 dB attenuation is obtained solely from co-simulation. No hardware-in-the-loop testing, physical prototype validation, or analysis of real-vehicle effects (sensor quantization, PWM synchronization jitter, temperature drift of passives, or DC-link voltage back-action) is presented, despite the paper targeting automotive power systems where these factors are known to alter both the disturbance spectrum and achievable cancellation.
[Abstract and Results section] Abstract and Results section: The reported attenuation improvements lack accompanying quantitative details on the exact frequency bands over which the 25-30 dB figure holds, the specific conventional control strategies and passive filter designs used as baselines, error bars, number of independent trials, or any statistical significance tests. Without these, the claim of 'consistent' improvement across 'a wide frequency range' cannot be rigorously evaluated.

minor comments (2)

[Method section] The MDP formulation (state space, action space, and reward function) should be presented with explicit equations rather than high-level description to allow reproducibility of the RL setup.
[Results section] Include training curves for the RL agent (reward vs. episodes) and overlaid EMI spectra plots (with/without the proposed filter) to support the performance claims and demonstrate convergence.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the constructive and detailed comments on our manuscript. We address each major comment point by point below, indicating the revisions we will incorporate.

read point-by-point responses

Referee: [Results section] Results section: The headline performance claim of 25-30 dB attenuation is obtained solely from co-simulation. No hardware-in-the-loop testing, physical prototype validation, or analysis of real-vehicle effects (sensor quantization, PWM synchronization jitter, temperature drift of passives, or DC-link voltage back-action) is presented, despite the paper targeting automotive power systems where these factors are known to alter both the disturbance spectrum and achievable cancellation.

Authors: We acknowledge that the evaluation is performed exclusively in co-simulation using experimentally measured EMI spectra from an automotive electric drive unit. This approach was selected to enable controlled, repeatable assessment across diverse operating conditions while grounding the disturbance in real data. In the revised manuscript we will add a new subsection in the Discussion that incorporates simplified models of key real-vehicle effects (sensor quantization, PWM jitter, temperature drift of passives, and DC-link voltage variation) into the existing Simulink framework and quantifies their influence on achievable attenuation. We will also expand the Limitations and Future Work section to explicitly discuss the gap between co-simulation and hardware validation and to outline a concrete plan for hardware-in-the-loop experiments. Full physical prototype testing, however, requires specialized high-power automotive test facilities and safety approvals that cannot be completed within the revision cycle. revision: partial
Referee: [Abstract and Results section] Abstract and Results section: The reported attenuation improvements lack accompanying quantitative details on the exact frequency bands over which the 25-30 dB figure holds, the specific conventional control strategies and passive filter designs used as baselines, error bars, number of independent trials, or any statistical significance tests. Without these, the claim of 'consistent' improvement across 'a wide frequency range' cannot be rigorously evaluated.

Authors: We agree that additional quantitative rigor is required. In the revised manuscript we will: (i) specify the exact frequency bands (150 kHz–30 MHz, aligned with CISPR 25) over which the 25–30 dB improvement is observed; (ii) provide explicit descriptions and component values for the passive filter baselines (e.g., two-stage LC topologies) and the conventional active-filter controllers (fixed-gain PI and adaptive notch); (iii) report mean attenuation curves with error bars derived from 10 independent training runs using different random seeds; and (iv) include statistical significance testing (paired t-tests and Wilcoxon rank-sum tests) with p-values. These details will appear in the Abstract, Results section, and in new tables/figures. revision: yes

standing simulated objections not resolved

The request for hardware-in-the-loop testing and physical prototype validation, which cannot be fulfilled without new experimental hardware and safety approvals outside the scope of the current revision.

Circularity Check

0 steps flagged

No significant circularity in derivation or performance claims

full rationale

The paper formulates EMI mitigation as an MDP, applies standard RL with VAE state encoding and noise exploration, then reports 25-30 dB attenuation from direct comparison of simulation outputs against conventional and passive baselines in MATLAB/Simulink driven by measured spectra. No equations define a quantity in terms of itself, no fitted parameters are relabeled as independent predictions, and no load-bearing claims rest on self-citations or imported uniqueness results. The reported improvements are simulation outcomes, not tautological redefinitions of the inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The approach rests on the untested assumption that the simulated EMI environment sufficiently represents real-vehicle dynamics and that standard RL convergence guarantees apply to this non-stationary control task; no explicit free parameters or invented physical entities are named in the abstract.

axioms (2)

domain assumption The EMI mitigation task can be faithfully represented as a Markov decision process whose state is observable from frequency-domain spectra.
Directly invoked when the problem is formulated as an MDP for the RL agent.
domain assumption A variational autoencoder yields a compact yet sufficient state representation that preserves the information needed for effective policy learning.
Stated as the mechanism used to improve robustness and generalization.

pith-pipeline@v0.9.0 · 5553 in / 1544 out tokens · 47932 ms · 2026-05-07T05:26:41.066301+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

6 extracted references · 6 canonical work pages

[1]

Darisi, M., Caldognetto, T., Biadene, D., Stellini, M., 2024

URL: http://ieeexplore.ieee.org/document/ 7332787/, doi:10.1109/TPEL.2015.2502218. Darisi, M., Caldognetto, T., Biadene, D., Stellini, M., 2024. Digi- tal active emi filter for smart electronic power converters. Elec- tronics 13. URL: https://www.mdpi.com/2079-9292/ 13/19/3889, doi:10.3390/electronics13193889. Du, L., Yan, D., Ma, D.B., 2023. A condition-...

work page doi:10.1109/tpel.2015.2502218 2015
[2]

Goswami, R., Wang, S., 2018

URL: https://ieeexplore.ieee.org/document/ 10067594/, doi:10.1109/ISSCC42615.2023.10067594. Goswami, R., Wang, S., 2018. Modeling and stability analysis of active differential-mode emi filters for ac/dc power converters. IEEE Transactions on Power Electronics 33, 10277–10291. doi:10.1109/TPEL.2018.2806361. Hamza, D., Qiu, M., 2013. Digital active emi cont...

work page doi:10.1109/isscc42615.2023.10067594 2023
[3]

Narayanasamy, B., Luo, F., 2019

URL: http://ieeexplore.ieee.org/document/ 5444974/, doi:10.1109/TPEL.2010.2047734. Narayanasamy, B., Luo, F., 2019. A survey of active emi filters for conducted emi noise reduction in power electronic converters. IEEE Transactions on Electro- magnetic Compatibility 61, 2040–2049. URL: https: //ieeexplore.ieee.org/document/8915991/, doi: 10. 1109/TEMC.2019...

work page doi:10.1109/tpel.2010.2047734 2010
[4]

Sakthivel, A., Vijayakumar, P., Senthilkumar, A., Lakshmi- narasimman, L., Paramasivam, S., 2015

URL: http://ieeexplore.ieee.org/document/ 1174072/, doi:10.1109/TIE.2002.807658. Sakthivel, A., Vijayakumar, P., Senthilkumar, A., Lakshmi- narasimman, L., Paramasivam, S., 2015. Experimental investigations on ant colony optimized pi control algo- rithm for shunt active power filter to improve power quality. Control Engineering Practice 42, 153–169. URL: ...

work page doi:10.1109/tie.2002.807658 2002
[5]

IEEE Transactions on Power Electronics 20, 523–

Integrating active, passive and emi-filter functions in power electronics systems: A case study of some tech- nologies. IEEE Transactions on Power Electronics 20, 523–

work page
[6]

Vedde, A., Neuburger, M., Reuss, H.C., 2021

URL: http://ieeexplore.ieee.org/document/ 1427810/, doi:10.1109/TPEL.2005.846553. Vedde, A., Neuburger, M., Reuss, H.C., 2021. An op- timized high-frequency emi filter design for an automo- tive dc/dc-converter, in: 2021 National Power Electron- ics Conference (NPEC), IEEE. pp. 01–06. URL: https: //ieeexplore.ieee.org/document/9672512/, doi: 10. 1109/NPEC...

work page doi:10.1109/tpel.2005.846553 2005

[1] [1]

Darisi, M., Caldognetto, T., Biadene, D., Stellini, M., 2024

URL: http://ieeexplore.ieee.org/document/ 7332787/, doi:10.1109/TPEL.2015.2502218. Darisi, M., Caldognetto, T., Biadene, D., Stellini, M., 2024. Digi- tal active emi filter for smart electronic power converters. Elec- tronics 13. URL: https://www.mdpi.com/2079-9292/ 13/19/3889, doi:10.3390/electronics13193889. Du, L., Yan, D., Ma, D.B., 2023. A condition-...

work page doi:10.1109/tpel.2015.2502218 2015

[2] [2]

Goswami, R., Wang, S., 2018

URL: https://ieeexplore.ieee.org/document/ 10067594/, doi:10.1109/ISSCC42615.2023.10067594. Goswami, R., Wang, S., 2018. Modeling and stability analysis of active differential-mode emi filters for ac/dc power converters. IEEE Transactions on Power Electronics 33, 10277–10291. doi:10.1109/TPEL.2018.2806361. Hamza, D., Qiu, M., 2013. Digital active emi cont...

work page doi:10.1109/isscc42615.2023.10067594 2023

[3] [3]

Narayanasamy, B., Luo, F., 2019

URL: http://ieeexplore.ieee.org/document/ 5444974/, doi:10.1109/TPEL.2010.2047734. Narayanasamy, B., Luo, F., 2019. A survey of active emi filters for conducted emi noise reduction in power electronic converters. IEEE Transactions on Electro- magnetic Compatibility 61, 2040–2049. URL: https: //ieeexplore.ieee.org/document/8915991/, doi: 10. 1109/TEMC.2019...

work page doi:10.1109/tpel.2010.2047734 2010

[4] [4]

Sakthivel, A., Vijayakumar, P., Senthilkumar, A., Lakshmi- narasimman, L., Paramasivam, S., 2015

URL: http://ieeexplore.ieee.org/document/ 1174072/, doi:10.1109/TIE.2002.807658. Sakthivel, A., Vijayakumar, P., Senthilkumar, A., Lakshmi- narasimman, L., Paramasivam, S., 2015. Experimental investigations on ant colony optimized pi control algo- rithm for shunt active power filter to improve power quality. Control Engineering Practice 42, 153–169. URL: ...

work page doi:10.1109/tie.2002.807658 2002

[5] [5]

IEEE Transactions on Power Electronics 20, 523–

Integrating active, passive and emi-filter functions in power electronics systems: A case study of some tech- nologies. IEEE Transactions on Power Electronics 20, 523–

work page

[6] [6]

Vedde, A., Neuburger, M., Reuss, H.C., 2021

URL: http://ieeexplore.ieee.org/document/ 1427810/, doi:10.1109/TPEL.2005.846553. Vedde, A., Neuburger, M., Reuss, H.C., 2021. An op- timized high-frequency emi filter design for an automo- tive dc/dc-converter, in: 2021 National Power Electron- ics Conference (NPEC), IEEE. pp. 01–06. URL: https: //ieeexplore.ieee.org/document/9672512/, doi: 10. 1109/NPEC...

work page doi:10.1109/tpel.2005.846553 2005