Geng fraud-detection report: Multi-sensor fusion for high-precision air-gap estimation in electromagnetic maglev systems (Zhang et al., MST 2026)

Academic fraud report · Geng Detector

Summary

Verdict: serious concerns confirmed (实锤), with strongest evidence pointing to internal contradictions and likely AI-assisted writing without author verification. Key issues: (1) Abstract and Conclusion claim fusion EKF air-gap RMSE of 0.12 mm and velocity RMSE of 4.2 mm/s, but Table 2 / Section 8.3 report 0.182 mm and 4.860 mm/s — a direct, irreconcilable data mismatch. (2) Section 5.1 introduces an 'eddy current sensor' that does not exist in the experimental setup (only Hall, inductive, accelerometer are used), a hallmark copy-paste or AI-hallucination artifact. (3) Section 5.2 sets inductive-sensor noise variance σ_L² = 10^-6 (σ_L ≈ 10^-3 H = 1000 μH), while the baseline inductance L0 is only 100 μH, giving SNR ≈ -20 dB and an unrunnable Kalman filter. (4) Equation (16) with parameters a=1.09, b=-0.634, n=0.445 yields a negative base (g−0.634) for g in mm, producing complex values in MATLAB unless the authors silently switched to meters. (5) Accelerometer RMSE of 9.83 m/s² is excused as 'dominated by gravity and bias,' but Eq. (23) already subtracts gravity, indicating an elementary analysis error. Confidence is high on textual and arithmetic issues; image-based reuse was not assessable due to lack of pixel data.

Verdict

🔴 Confirmed serious concerns. The paper contains multiple internal contradictions and mathematical inconsistencies that are individually unlikely and collectively strongly suggestive of unverified AI-assisted generation rather than careful scholarship. A formal institutional investigation is warranted.

Key findings

Internal data contradiction (Finding 1): Abstract/Conclusion air-gap RMSE = 0.12 mm, velocity RMSE = 4.2 mm/s; Table 2 and Section 8.3 report 0.182 mm and 4.860 mm/s. The headline numbers do not match the paper's own table.
'Ghost sensor' (Finding 2): Section 5.1 references an eddy-current sensor that is not part of the experimental setup (only Hall, inductive, and accelerometer sensors are used elsewhere).
Physically impossible noise setting (Finding 3): Inductive-sensor noise σ_L² = 10^-6 ⇒ σ_L ≈ 10^-3 H (1000 μH), while the baseline inductance L0 = 100 μH; noise is ~10× the signal, which would cause EKF divergence.
Complex-domain formula (Finding 4): Substituting a=1.090, b=-0.634, n=0.445 into B(g)=a/(g+b)^n+c with g in mm gives (g−0.634) negative over the simulated range 2–8 mm; MATLAB would return complex values for the 0.445 power, incompatible with standard real-valued EKF updates.
Accelerometer RMSE mishandling (Finding 5): Reported accelerometer RMSE ≈ 9.83 m/s² is attributed to gravity (9.81 m/s²) and bias (0.02 m/s²), but Eq. (23) (a_true = a_measured + g_gravity) already removes gravity, indicating the error was not properly accounted for.
Image reuse (Finding 6): Not assessable from the provided text; no pixel-level analysis was performed.

Evidence highlights

DOI: 10.1088/1361-6501/ae4e03
Paper title: Multi-sensor fusion for high-precision air-gap estimation in electromagnetic maglev systems: an extended Kalman filter approach
Authors: Lindi Zhang, Jiacheng Liu, Shuhao Wang, Yungao Wu, Qirui Ding
Journal: Measurement Science and Technology (MST), year 2026
Quoted Abstract claim: 'fusion EKF achieves an airgap RMSE of 0.12 mm ... velocity RMSE from 5.6 mm/s to 4.2 mm/s'
Quoted Conclusion claim: 'reducing the airgap RMSE from 0.18 mm to 0.12 mm and the velocity RMSE from 5.6 mm s⁻¹ to 4.2 mm s⁻¹'
Quoted Table 2 values: Fusion EKF RMSE(g) = 0.182, RMSE(v) = 4.860
Quoted Section 5.1: 'the measurement noise variance for the eddy current sensor ... our selection of 10^-8 ensures ...'
Quoted Section 5.2: 'σ_L^2 is set to 10^-6 ... standard deviation of approximately 10^-3 H'; L0 = 100 μH per Eq. (18)
Quoted Eq. (16): B(g) = 1.09 / (g − 0.634)^0.445 − 0.151 with simulated g ∈ [2, 8] mm (Section 2.4, 8.2)
Quoted Section 8.1 / Table 1: accelerometer RMSE ≈ 9.83, attributed to 'static gravitational component (9.81 m s⁻²) and the constant bias (0.02 m s⁻²)'
Quoted Eq. (23): a_true = a_measured + g_gravity

Notes

Confidence is high for Findings 1–5 because the quotations are exact and the arithmetic / unit checks are reproducible. Finding 4 could in principle be resolved if the authors actually implemented B(g) using g in meters (0.002–0.008 m), making the base negative but smaller in magnitude and yielding real values; however, the paper as written does not state this unit convention, so the discrepancy stands.
Finding 6 (image reuse) was not assessable; only figure captions were available, and no pixel-level comparison was performed.
The report recommends contacting authors for raw data, posting on PubPeer, and notifying the journal editorial office to review the peer-review process. Institutional investigation is suggested to be reserved pending author response, as the primary indicators are severe sloppiness and unverified AI-generated content rather than provable intentional data fabrication.
Disclaimer: This report is AI-assisted and intended for academic discussion only; final determination of misconduct requires an institutional investigation.