Electronics

Inverter current control can significantly weaken power-frequency fault signatures in distribution networks with high photovoltaic (PV) penetration. This paper proposes an active fault detection method using positive-to-negative-sequence perturbation responses. A small-signal positive-sequence voltage is injected on the low-voltage side of the PV step-up transformer, and the target-frequency ratio of negative-sequence voltage to positive-sequence current (K_np = V_n/I_p) measured at the medium-voltage point of common coupling (PCC) quantifies the resulting cross-sequence coupling response. Sequence-domain analysis combined with nodal admittance reduction explains the off-diagonal coupling terms introduced by single-phase ground and series faults and extends the observation principle to intermittent arcing faults. Sliding-window phasor estimation, a calibrated and locked normal-operation baseline, and consecutive confirmation form the online protection criterion. Co-simulation demonstrates a 13.133–22.638 dB index change across 650–850 Hz and a 37.43–43.35 dB separation between healthy and faulted states across the tested operating and grid strength variations. A 32-point fault resistance sweep characterizes the baseline-crossing mechanism responsible for the intermediate magnitude–sensitivity gap near 3.68 kΩ and verifies detection recovery up to 24 kΩ under the nominal 3 dB threshold. RT-Box real-time tests confirm online trip generation within 20 ms for all three fault types. Furthermore, robustness evaluations verify reliable detection of the representative 10 Ω fault down to a 20 dB target-frequency SNR under a fixed 14.5 dB anti-noise threshold, while explicitly quantifying the selectivity and sensitivity trade-offs under non-fault switching, channel mismatch, and the reduced kilohm-range sensitivity caused by the elevated anti-noise threshold.

ByteTrack remains a strong tracking-by-detection framework for online multi-object tracking. Its BYTE association splits detections at a high-score threshold $T_{\mathrm{high}}$ that decides which boxes enter the first matching pass and which unmatched boxes may start a track. ByteTrack nevertheless freezes $T_{\mathrm{high}}$ for an entire sequence. Occlusion, crowding, camera motion, and detector confidence keep changing, so the $T_{\mathrm{high}}$ that best preserves identities can change as well. When labeled frames from earlier in the same sequences are available, that mismatch can be absorbed by grid search; the later half cannot be searched in the same way. Therefore, we argue that a preset sequence-wide $T_{\mathrm{high}}$ is not a reasonable operating assumption. To bridge the gap, we propose a lightweight $T_{\mathrm{high}}$ Predictor that updates $T_{\mathrm{high}}$ online from the recent detection state. We report two findings. (1) The identity-preserving $T_{\mathrm{high}}$ varies over time within a sequence. A window-level greedy oracle on val-half yields a non-constant $T^{*}$ on every MOT17 and MOT20 sequence. (2) The $T_{\mathrm{high}}$ Predictor improves ByteTrack’s robustness to the initial $T_{\mathrm{high}}$. Across initializations on val-half, MOT17 IDF1 moves from $78.03\pm 0.82$ to $79.06\pm 0.15$, HOTA from $64.68\pm 0.42$ to $65.25\pm 0.04$, and MOTA from $76.47\pm 0.96$ to $77.44\pm 0.17$ (mean ± standard deviation across initializations). On MOT20, IDF1 moves from $89.54\pm 0.41$ to $89.64\pm 0.01$, HOTA from $76.58\pm 0.31$ to $76.68\pm 0.01$, and MOTA from $92.91\pm 0.31$ to $93.16\pm 0.02$. The cross-initialization standard deviation of IDF1, HOTA, and MOTA falls by $82\%$, $90\%$, and $82\%$ on MOT17, and by $98\%$, $97\%$, and $94\%$ on MOT20.

Mechanical-load emulation enables motor-drive testing under different load conditions without physically replacing the load. However, forward-dynamics-based mechanical-load emulation is challenging when a commercial load-motor drive has fixed proprietary firmware. This paper presents a real-time architecture that relocates the desired-load model, position-tracking controller, and compensation algorithm to a programmable test-motor drive, while the commercial load-motor drive operates in its standard torque-control mode. A TMS320F28388D multicore MCU distributes test-motor control, load-emulation computation, and EtherCAT communication among two DSP cores and a Connectivity Manager core. The load-emulation algorithm executes at 1 kHz, while EtherCAT process data are exchanged every 250 µs. The discrete-time realization and its nominal stability are analytically and numerically verified. Experiments with two inertia–viscous-friction load conditions show close agreement between measured and simulated motion, with position RMSEs of 0.043 and 0.069 rad under the 250 µs EtherCAT communication period and 0.1 ms torque-reference input-filter setting. The maximum measured on-board processing delay through EtherCAT process-data write completion was 19.597 µs. Additional tests demonstrate that shortening the EtherCAT communication period improves or maintains emulation accuracy, while variations in the torque-reference input-filter time constant have a limited effect. These results demonstrate the feasibility of implementing forward-dynamics-based mechanical-load emulation using a commercial servo drive without modifying its internal control firmware.