Pointer-type instruments are widely deployed in petrochemical environments, and the accurate and rapid automatic reading of such instruments is critical to on-site safety monitoring. To address the limitations of existing recognition algorithms, such as heavy dependence on prior information and poor adaptability to complex harsh environments, this paper proposes a deep learning-driven intelligent reading method for pointer-type instruments. The overall framework consists of three core modules: First, a dual-task deep learning architecture is employed to simultaneously implement instrument detection and scale value recognition. Second, an enhanced geometric analysis strategy is proposed to identify instruments and their pointer positions, featuring binary image processing with scale line extension, voting-based center determination, and semi-fitting pointer localization through Progressive Probabilistic Hough Transform with least squares optimization. Third, a confidence-aware vector angle method is developed by incorporating zero-scale determination and angular measurement rules to achieve accurate instrument readings. Comprehensive experiments demonstrate that the proposed method exhibits strong robustness under complex operating conditions, including image degradation (noise and blur), weather interference (fog, snow, and rain), visual obstructions (occlusions, contamination, and reflections), and geometric variations (rotations and tilts). Evaluated across 300 instruments with 50 representative cases, the method achieves an overall mean reading error of 1.26% with maximum errors remaining below 2.5% and a single-frame processing time of 146 milliseconds. This performance confirms its utility for the “stop-and-read” robotic inspection mode in complex and hazardous environments.
Self-supervised transfer learning can reduce the number of labeled examples needed to calibrate a multigas electronic nose, although its benefits vary across gases and experimental conditions. We investigate this approach by first training a shared temporal encoder to reconstruct masked sensor responses without concentration labels, then fine-tuning it on a small set of labeled concentration pairs. Across ten random seeds, transfer learning reduces mean absolute error relative to training from scratch by 13.9% for NH3, 21.5% for H2, 27.5% for CO, and 25.7% for ethylene. These results come from six held-out identifier partitions for NH3/H2 using a predefined early-response window and three rolling chronological folds for CO/ethylene. To express label efficiency more directly, we also compare models trained with different calibration budgets. Using 10 labeled pairs, including validation labels, fine-tuning achieves approximately the NH3 and ethylene errors of scratch training with 20 pairs and a lower mean CO error, but does not match its H2 performance. Sensor identity and cross-sensor attention contribute unevenly, with NH3 showing the weakest evidence of benefit. Transfer also improves concentration regression in two of three future drift partitions when gas identity is known. Overall, the findings support selective use of self-supervised transfer for label-efficient calibration within the same sensor system. However, the proposed model does not outperform the strongest full-budget baseline on any target, and the NH3/H2 evaluation does not establish generalization across acquisition days.
To address transmission efficiency degradation caused by coil misalignment in electric vehicle wireless power transfer (EV-WPT) systems, this paper proposes a one-dimensional misalignment detection and positioning method for the receiver coil based on the phase difference between two detection-coil currents. Two small detection coils are symmetrically arranged on the transmitter-coil plane. Receiver-coil misalignment causes asymmetric variations in the mutual inductances between the receiver coil and the two detection coils, thereby changing the phase difference between their currents. An analytical relationship between the phase difference and receiver-coil displacement is derived, demonstrating an approximately linear relationship under small-displacement conditions. Finite-element simulations and experiments are conducted to validate the proposed method. Simulation results show a strong linear relationship over -20 to 20 mm, with R2 = 0.99902 and a sensitivity of 0.245 deg/mm. Experimental results demonstrate a strong linear relationship over −20 to 20 mm, with R2 = 0.99355 and a sensitivity of 1.68 deg/mm. The proposed method requires no modification to the receiver-side structure and introduces negligible interference with the main power-transfer channel, providing position information for transmitter-side active alignment in EV-WPT systems.
This paper proposes DLBAcalib (Dual LiDAR Bundle Adjustment calibration), a targetless framework for extrinsic calibration when two LiDARs have no instantaneous field-of-view (FoV) overlap. The method first uses Overlap-Propagated Local Plane Adjustment (OP-LPA) to correct the target-LiDAR trajectory and construct a reference map, then freezes that map and estimates the source-LiDAR extrinsic from temporally distributed observations. Thus, the temporal overlap of consecutive windows from one target LiDAR is not an assumption of instantaneous overlap between the two sensors. In CARLA, 50 trials across five non-overlapping configurations yielded mean translation errors of 0.00509–0.00580 m and rotation errors of 0.00336–0.00418 rad; under initial perturbations of ±0.4 m and ±30◦, the five configurations yielded DLBAcalib success rates of 96–100%. Backpack experiments yielded a mean reference-relative error of 0.0062 m and 0.0051 rad, and X30 experiments evaluated portability across another carrier and LiDAR configuration. The results support the method under the tested configurations, while its accuracy remains dependent on shared static geometry, motion compensation, and reference quality. The source code is available at https://github.com/Silentbarber/DLBAcalib.
Tapered optical fibers or microfibers (MF) decorated with nanostructures offer enhanced evanescent fields and direct light-matter interaction through a vast effective surface area. Various high-performance optical sensors have been developed based on this configuration. In this study, a functionalized single-mode guiding MF with a radially ordered array of ZnO nanowires is demonstrated, exhibiting high sensitivity along with stable and smooth sensing characteristics toward humidity and temperature. The utilized MF is an adiabatically tapered single-mode fiber, with ZnO nanowires synthesized on its surface via the hydrothermal method. Experimental results indicate that the output intensity of the decorated microfiber with a 20 μm waist diameter exhibits a highly linear response (R2~0.99) across a relative humidity (RH) range of 18% to 80%, achieving an average sensitivity of −0.016 a.u./%RH. Moreover, the MF sensor shows a smooth quadratic thermal characteristic (R2~0.99) over the 20 – 150 °C range, with a sensitivity of 0.0025 a.u/°C. Finally, to gain insight into the complex sensing behavior of the decorated MF, full-wave simulations are utilized, which support and explain the experimental observations. Compared to previously reported multimode configurations, the proposed single-mode design offers significant advantages, making it a promising platform for humidity/temperature sensing applications.
