A scaling law calibrated on data pooled from several facilities should be validated on a facility it has never seen, and when that test fails, the failure needs its own diagnosis: either the outcome carries a facility-specific effect that the predictors miss, or the predictor set itself cannot be separated from facility identity. We propose a two-stage protocol. The detection stage consists of algebraic-redundancy screening, leave-one-facility-out (LOFO) cross-validation, a facility-cluster bootstrap and a predictor/facility collinearity check; the diagnosis stage consists of a pre-fit identifiability index and a synthetic two-null discriminator. We apply it to 114 windage-power points from four sources (labelled by first author and year: Guo2024, Vrancik1968, Zheng2024, Liu2024) spanning a 900× Reynolds range. The corpus is heterogeneous: only 64 points are physical measurements, 45 are computational fluid dynamics (CFD) outputs for a facility that has not been built (including 8 rows with an inconsistent power basis and 7 with a frozen radial gap), 5 are reconstructed from a fitted constant, and 20 come from a preprint that has not been peer-reviewed. A four-predictor model selected by the Bayesian Information Criterion reaches an in-sample R2=0.870, but a pooled LOFO R2=−0.885, and no held-out facility is predicted better than its own mean. Repeating the structure selection inside each fold selects the same model in every fold. A pre-fit identifiability index shows that one direction of the predictor block, essentially logΠconfinement, is aliased with facility identity (the largest of the K−1=3 canonical correlations 0.996) and has almost no within-facility variation, also on a verified 94-row subset. With K=4 facilities, the two-null discriminator is fragile: under homoscedastic noise, 3 of 4 observed values fall outside both simulated supports; with per-point reported uncertainty, a facility-offset model reaches the observed values of Guo2024, Vrancik1968 and, in its extreme tail, Zheng2024, whereas a collinearity-only model with the same uncertainty does not reach Zheng2024 and reaches Guo2024 only at its 1st percentile, and Liu2024 is reached by neither; we therefore report it as a documented failure mode on small multi-facility corpora rather than as a standalone diagnostic. Twenty-one further attempts, sixteen of them learning methods, also fail the per-facility rule. Only the sign of the Reynolds effect transfers, and only among the physical facilities: within the CFD rows of the unbuilt facility, it is not detectable once 8 misattributed rows are separated.
Industrial compressed-air systems require continuous operational visibility because pressure regulation, load cycling, thermal conditions, and pneumatic demand affect energy performance. This study develops and field-evaluates a low-cost Internet of Things (IoT) monitoring architecture for a 10 hp lubricated rotary-screw compressor installed in a Colombian microenterprise. The platform combines an ESP32 microcontroller with a PZEM-004T electrical module, an HK1100C pressure sensor, two DS18B20 temperature sensors, a MySQL database, and Power BI visualization. The retained field campaign contains 1014~records across four valid acquisition days between 14 and 18 October 2025. Comparisons with reference instruments under the tested operating conditions yielded mean deviations of 1.4 V, 1.1 A, 1.0 °C, and 4.8 psi for voltage, current, inlet temperature, and pressure, respectively. These comparisons provide an empirical characterization of the monitored channels but do not constitute complete metrological validation of the sensing system. The field data reveal cyclic pressure behavior and pronounced electrical transitions associated with compressor loading. For a representative active-operation subset on 16 October, the mean Estimated Net Receiver-Inventory Rate was 0.0198 m3/s and corresponding exploratory power-to-inventory-rate was 432.2 kW/(m3/s), close in scale to the nominal reference specific power of approximately 416.7 kW/(m3/s). Thermodynamic indicators were calculated using isentropic and isothermal compression relations and pressure-volume-temperature corrections. Because airflow was not independently measured and the electrical channel was not validated with a traceable true-RMS power quality analyzer, these derived indicators are interpreted as exploratory rather than certified or standard-compliant performance measurements. The architecture provides an accessible pathway for persistent energy-relevant and condition monitoring in resource-constrained industrial environments.
This paper presents an energy-efficient operating strategy for a three-phase induction motor drive powered by a proton exchange membrane fuel cell (PEMFC). The proposed approach minimizes hydrogen consumption and improves motor efficiency over a wide operating range by continuously regulating the motor terminal voltage according to the load demand. The optimal reference voltage corresponding to each load torque is determined using the dholes-inspired optimization algorithm (DIOA), a recent swarm-based optimization technique. The DIOA is also employed to determine the optimal parameters of a proportional–integral (PI) controller, enabling accurate and rapid tracking of the reference voltage. To evaluate the effectiveness of the proposed optimizer, its performance is compared with the well-known particle swarm optimization (PSO) algorithm. The reference voltages obtained using the DIOA and PSO differ by less than 1.1%, their hydrogen consumptions differ by less than 0.02%, and a direct voltage sweep confirms the optimum. The dynamic performance is investigated using both an optimally tuned PI controller and a neural network controller (NNC). Simulation results show that the NNC provides faster settling, lower transient deviations, and improved dynamic performance compared with the PI controller. Steady-state results indicate that hydrogen savings are strongly dependent on motor loading. At a load torque of 0.5 N·m (5% of full load), hydrogen consumption is reduced from 1.999121 slpm to 0.706726 slpm, representing a saving of 64.65%. At a load torque of 2 N·m (20% of full load), the hydrogen saving reaches 24.17%. As the load approaches its rated value, the motor operates near its maximum efficiency and rated voltage.
Road-surface ice, snow and standing water can impair winter driving, while irregular boundaries and changing illumination complicate image-based assessment. This study investigates pixel-wise semantic segmentation using HPC-D, a DeepLabv3+ architecture incorporating convolutional block attention (CBAM), parallel axial attention (PAA) and hierarchical feature fusion (HFF), together with self-calibrated illumination (SCI) enhancement for low-light images. The reported dataset contains 5000 image entries from a campus road environment, with a reported training/validation/test allocation of 3000/1000/1000. The original experimental tables report a mean intersection-over-union (mIoU) of 0.96, a reported mean average precision (mAP) of 0.95 and an accuracy of 0.98 for HPC-D. Relative to the other rows in the segmentation comparison, the numerical differences are 2–21 percentage points in mIoU, 0–10 percentage points in mAP and 3–14 percentage points in accuracy. In the illumination comparison, the SCI row exceeds the no-enhancement row by 16 percentage points in mIoU and 17 percentage points in mAP. These are descriptive differences in the reported results. Incomplete documentation of image lineage, acquisition-session separation, metric implementation and comparator configurations limits their interpretation as evidence of independent generalization. The contribution is therefore an architectural integration and a preliminary evaluation for visually annotated road-surface segmentation; cross-camera robustness, shadow rejection and real-time vehicle operation remain unvalidated.
To overcome the thermal loss and interconnect delays of conventional electronic processors under escalating AI computing workloads, analog optical processors have emerged as a compelling low-latency alternative. This work introduces an integrated optical neural accelerator leveraging an electro-optically reconfigurable multimode interference (MMI) framework on a lithium-niobate-on-insulator (LNOI) platform. By harnessing localized spatial mode interference through micro-electrodes strategically placed within the MMI region, the device performs trainable matrix operations with a minimal on-chip insertion loss of 1.42 dB. A hardware-in-the-loop in situ (HIL in situ) calibration framework employing the Nelder–Mead optimization algorithm is developed to adaptively adjust local refractive indices and compensate for physical hardware imperfections. The system’s performance is validated through numerical electromagnetic simulations applied to the famous Iris flower classification benchmark, demonstrating the self-learning capabilities of the integrated optical neural network with an accuracy of 72%. These findings present a compact, low-loss building block for next-generation photonic tensor processors.
