Onboard estimation of lithium-ion battery state-of-health (SOH) is essential for safe electric-vehicle operation. Nonlinear capacity fade, scarce training data and pronounced local regeneration phenomena, however, complicate reliable SOH forecasting. This paper presents LMM-PatchTSSA, a lightweight framework that couples a Local Mixing Module (LMM) with a Patch Time Series Transformer (PatchTST) backbone. The LMM extracts short-term, degradation-sensitive signatures from charge–discharge curves, exposing early-stage ageing indicators that standard transformers overlook. Within PatchTST, a Token-Statistics Self-Attention (TSSA) mechanism replaces conventional quadratic self-attention; by formulating attention through second-order moment statistics, TSSA reduces computational complexity from [[EQUATION]]O(n2) to [[EQUATION]]O(n), while the patch-based structure captures long-range dependencies at linear cost. A frozen-backbone transfer strategy further mitigates chemistry-induced domain shifts. Trained and evaluated on the public CALCE dataset, the model achieves an average RMSE of 0.0118. Under zero-shot cross-chemistry transfer from CALCE (NMC) to NASA (LCO) cells without retraining, the framework records [[EQUATION]]R2>0.95; after fine-tuning on a minimal subset of target data, [[EQUATION]]R2 reaches 0.9913 on NASA B0007. These results demonstrate accurate SOH prediction, robust cross-chemistry generalisation and low inference overhead, making the framework suitable for real-time battery management systems.

Wind turbines operating in harsh environments are prone to surface defects that compromise efficiency and safety. Traditional convolutional neural networks lack sufficient multi-scale feature representation, while Transformer-based methods suffer from excessive computational complexity. This study proposes HAMD-DETR, an end-to-end detection framework for wind turbine defect identification. The framework consists of three key components: an Adaptive Dynamic Multi-scale Perception Network (ADMPNet), a Hierarchical Dynamic Feature Pyramid Network (HDFPN), and a Dynamic Frequency-Domain Feature Encoder (DFDEncoder). Firstly, ADMPNet integrates multi-scale dynamic integration fusion and adaptive inception depthwise convolution for feature extraction. Then the HDFPN balances deep semantic and shallow detail features through pyramid adaptive context extraction and gradient refinement modules. At last, DFDEncoder enhances feature discrimination through frequency-domain transformation. Experiments on wind turbine datasets demonstrate that HAMD-DETR achieves 58.6% mAP50 and 31.7% mAP50-95, representing improvements of 3.1% and 2.1% over the baseline RT-DETR. The proposed method reduces computational complexity by 27.2% and parameters by 30% while achieving a 151.9 FPS inference speed. These results validate HAMD-DETR’s effectiveness for wind turbine defect detection and demonstrate its potential for intelligent operation and maintenance applications.

Accurate detection of early minor faults in electric vehicle traction batteries is important for preventing thermal runaway under complex operating conditions. Aging-related capacity degradation and measurement noise can mask the weak voltage distortions caused by early faults, leading to false alarms in data-driven diagnostic models. To improve robustness to these disturbances, this paper develops an early multi-fault decoupling and diagnosis framework based on PatchTSSA, a lightweight Transformer architecture that adapts time-series patching and Token Statistics Self-Attention (TSSA) to battery diagnostic sequences. The framework combines static–dynamic feature fusion with time-series patching to capture both global voltage drift and local morphological gradients. Within this adapted framework, TSSA replaces quadratic dot-product attention with second-order moment pooling, giving linear complexity with respect to the number of tokens and supporting future investigation of embedded Battery Management System (BMS) implementation. A physics-informed fault-injection strategy is used to construct a five-class dataset comprising the healthy state (E00), minor internal short circuit (E01), severe internal short circuit (E02), penetration fault (E03), and sensor drift (E04) from public Center for Advanced Life Cycle Engineering (CALCE) and National Aeronautics and Space Administration (NASA) battery-aging data. Across five raw-cycle-grouped splits and training seeds under 5 mV Gaussian white noise, the complete PatchTSSA configuration achieves $91.5\pm 1.9\%$ overall accuracy, $94.1\pm 1.2\%$ macro recall, and $85.4\pm 5.3\%$ E01 recall for the simulated fault patterns. The direct E01–E04 confusion rate is $0.09\pm 0.20\%$, whereas the E00-to-E01 false-alarm rate is $15.1\pm 6.9\%$. CALCE–NASA protocol differences are used only to describe cross-dataset domain shift; no transfer-performance claim is made without a matched capacity-free evaluation. The results indicate the potential of the framework for online fault-pattern discrimination, while validation using real fault data and embedded hardware remains necessary.