This paper introduces a dual-input high step-up converter with soft-switching operation. The proposed converter is a combination of two boost structures, voltage multiplier and passive clamp cell to improve the voltage gain and reduce the switch voltage stress with clamping the switch voltage. Soft-switching operation under zero-voltage switching at turn-on and turn-off is achieved using a resonant operation, which eliminates switching and capacitive turn-on losses and results in high converter efficiency. It is noted that all diodes in the converter are reverse-biased under zero-current switching conditions. This structure benefits from continuous input currents that would be an essential feature for renewable energy systems. The 24,12V/100V 100kHz-200 W converter is analyzed, and tested which shows 95.2% efficiency and confirms high step-up gain, fully soft-switching operation for all switches and eliminated reverse recovery problem of diodes due to providing ZCS at turn-off instant for diodes. In addition, the advantages of the converter are confirmed through a comparative study.
Apple Silicon’s unified memory can hold large tensors, but PyTorch’s Metal Performance Shaders (MPS) backend silently returns wrong results for batched matrix multiplication (torch.bmm) with more than $2^{32}$ elements. In every release from 2.4.1 to 2.14.0, some layouts give relative errors above 1 without any warning, and the failures reach torch.matmul and eager attention. Sweeping bmm over dtypes, layouts and shapes around $2^{31}$ and $2^{32}$ elements against a float64 CPU reference, we find three rules that explain every outcome on 2.14.0. If the output exceeds $2^{32}$ elements, a transposed operand makes the entire output wrong, matching a computation that ignores its strides, and anything else is correct. If it does not, a view with at least $2^{31}$ elements raises an exception, and a contiguous input above $2^{32}$ elements corrupts exactly the batches that start beyond element $2^{32}$ , matching a computation whose batch offset wraps at $2^{32}$ . A slightly larger problem can thus turn an explicit error into a silent failure. The rules extend to the backward pass, where a correct forward pass can return silently wrong gradients. A second machine with another chip reproduces all 6156 results, wrong values included, under two macOS versions, while an NVIDIA A100 is correct in all 2530 runs. In a public sentiment classifier, one oversized batch corrupts a third of the outputs, which collapse onto one class. All findings are black-box observations. The harness, raw results and a guard rejecting MPS tensors with $2^{32}$ or more elements are available at https://github.com/jniimi/mps-silent-failures
This study investigates local stabilization of a period-1 unstable periodic orbit (UPO) in a recently proposed three-dimensional continuous-time chaotic system using Ott-Grebogi-Yorke (OGY) control and input-constrained linear Poincaré model predictive control (MPC). The target UPO, return period, local Poincaré map, and OGY gain are recomputed, and numerical reliability is checked with respect to integration step size and finite-difference increment. To remove a control-authority asymmetry, the source-logic OGY implementation, which discards commands exceeding the admissible bound, is complemented by a Saturated OGY implementation using the same gain and the same bound, $|\Delta p|\leq 0.5$ . For 60 independent local initial conditions, Saturated OGY achieves a strict-capture rate of 75%, whereas MPC with a three-step horizon achieves 85%. The paired difference is +10 percentage points (95% confidence interval: 3.33–18.33; exact McNemar $p=0.03125$ ). Among 45 common-success cases, capture return shows a significant paired test but zero median difference; capture time is not significantly different, and MPC has lower time-weighted control effort. In 100 independent ±5% structural-model-mismatch scenarios, all scenario-specific UPOs are successfully continued; the median UPO displacement is 0.0294. MPC provides a small reduction in median scenario-specific terminal error and control effort, although strict capture is generally lost and upper-tail errors do not support universal superiority. Under Poincaré-section coordinate measurement noise, MPC generally yields lower true-state terminal error and lower control effort. The results therefore isolate the effect of bounded-command handling from the additional effect of finite-horizon optimization implemented in receding-horizon form under matched local conditions.
Accurate segmentation of pulmonary nodules in computed tomography (CT) images is an important step in computer-aided diagnosis of lung cancer, which directly affects evaluation of disease, planning of treatment, and clinical decision making. However, there is a large variation in nodule size, shape, texture and boundary characteristics, which makes automatic segmentation a challenging task. In this paper, we propose a hybrid three-dimensional CNN-Transformer network with multi-scale feature calibration, namely HCTC-Net, for automatic lung nodule segmentation. The proposed framework focuses on segmentation of pulmonary nodules within localized three-dimensional regions of interest (ROIs), while nodule detection and localization are outside the scope of this study. The proposed architecture is a hybrid of Residual 3D CNN encoder and Swin UNETR encoder to jointly learn the complementary local anatomical features and global contextual representations. To improve the interaction of features at multiple semantic levels, we propose an adaptive residual feature fusion strategy, a bottleneck cross-attention mechanism and a multi-scale feature calibration module. In addition, we propose an attention-guided decoder with boundary refinement and deep supervision to improve the boundary localization and feature learning for training. We evaluate the proposed framework on the public LIDC-IDRI dataset with a patient-wise split. Nodule-centered ROIs are defined using the available consensus annotations to isolate the segmentation task from the preceding localization problem. The experimental results demonstrate that HCTC-Net achieves competitive segmentation performance and outperforms representative deep learning methods, effectively delineating pulmonary nodules of various sizes and morphology. The proposed framework provides a robust and computationally efficient solution for automated pulmonary nodule segmentation, which can potentially lead to reliable computer-aided diagnosis of lung cancer in clinical practice.
In this article, the problem of detecting undesirable transitions in deterministic state transition systems in the absence of complete and reliable anomaly labels is considered. In contrast to classical methodologies of anomaly detection in time series, which model the distribution of trajectories as a stochastic process, the analyzed system is deterministic - each state is uniquely determined by the previous state and the operation performed. The proposed framework, designated as USTAD (Uncertainty-aware State Transition Anomaly Detection), integrates deterministic state-trajectory reconstruction, deep sequential modeling, and a threshold-based decomposition of the model score into confidence components. This approach facilitates the identification of redundancies and structural returns in the trajectory, without manual annotation or domain-specific rules. The reconstructed trajectories are used only to generate training labels. The trained model works on a window of past states and moves, so it can indicate a redundant path before the return to an earlier state occurs. The score is decomposed into the certainty of the normal class, the certainty of the anomalous class and a residual uncertainty, which indicates borderline decisions. This is of particular importance in conditions of strong class imbalance and limited supervision. The method has been experimentally verified in the context of speedcubing, a complex yet deterministic system characterised by state transitions. The findings suggest that the proposed solution exhibits superior performance on metrics sensitive to data imbalance, and that the confidence decomposition supports selective decisions on borderline windows.
