To address the limitations of traditional energy management strategies in hybrid energy storage systems for electric vehicles, including poor adaptability to dynamic conditions and safety risks in online reinforcement learning, this paper proposes an offline reinforcement learning framework integrating a time-aware variational autoencoder and Decision Transformer. Initially, high-quality expert trajectories are generated by a dynamic programming-based energy management strategy. Subsequently, a bidirectional long short-term memory network extracts temporal features from state sequences, while variational autoencoder synthesizes physics-constrained trajectories to mitigate distribution shift. Finally, Decision Transformer employs a self-attention mechanism to conduct multiscale temporal modeling of historical state-action sequences, establishing implicit policy mapping. Experimental results under the Dallas5 driving cycle demonstrate that the energy management strategy trained with a mixed dataset D1 outperforms the strategy trained with a pure expert dataset D2: battery capacity loss is reduced by 5.5%, and the final state of charge of the supercapacitor is stably maintained at 0.7320. This highlights the critical role of data diversity in enhancing generalization, offering a novel pathway for robust EMS design in real-world vehicular applications.

Continuous sign language recognition seeks to identify unsegmented sign language from videos by means of a weakly supervised manner, providing only sentence-level labels. In sign language videos, the gestures are smooth and continuous, and the same word may also correspond to video clips of different scales. Therefore, this poses a challenge in accurately capturing complex temporal dependencies. For hearing-impaired service robots, continuous sign language recognition capability is particularly critical, as the robots need to understand the natural sign language expressions of hearing-impaired users in real time. Previous studies have shown that using methods with a time-invariant receptive field for temporal modeling can partially address this issue, but they are not well-suited to handle video clips of varying scales. In this study, we re-examined the temporal modeling schemes in recent CSLR works and proposed the Two-stage Temporal ConvTransformer (T2CT), which fully leverages the advantages of one-dimensional convolutional neural networks and Transformer encoders, adopting a two-stage structure to capture more comprehensive spatiotemporal features. In particular, each stage of the proposed T2CT consists of two parts: a Local Temporal Modeling Module to capture short-term temporal dependencies, and a Global Temporal Modeling Module for long-term temporal modeling. Experimental results on three challenging CSLR datasets demonstrate that the proposed T2CT achieves competitive performance.

Accurate and effective prediction of battery remaining useful life (RUL) is crucial for the retirement planning of electric buses and the secondary utilization of battery packs. This study utilizes four years of operational data from nine electric buses to achieve precise RUL prediction for power batteries. First, considering the real-world operating characteristics of electric buses, this paper introduces a new RUL definition based on remaining mileage (RML) and remaining days of life (RDL) to characterize the remaining lifespan of battery packs. Subsequently, SOH labeling is conducted using charging data and filtering algorithms, followed by determining the end-of-life point of battery packs from SOH degradation trajectories. Finally, multi-time-scale features—including battery features, historical features, seasonal features, and discharging features are extracted from raw data, and the predictive performance of multiple ensemble learning models is compared. The results indicate that the AdaBoost model achieves the best performance in predicting RML and RDL, with a mean absolute error of 98 days and 16,852 km, respectively.

Under the background of dual carbon, lithium batteries are widely used in the energy field. However, range anxiety limits the popularity of electric vehicles. Optimization strategies for fast charging of lithium-ion batteries have been extensively studied to solve this problem. The huge parameter space of charging protocols and the complex aging mechanism of batteries limit the application of fast charging methods. The ability of reinforcement Learning to learn from the environment and adapt to the stochastic nature of battery behavior is a significant advantage. In this paper, we propose an innovative method for fast charging lithium-ion batteries using the Two-Delay Deep Deterministic Policy Gradient (TD3) algorithm and the Single Particle Model with Electrolyte model. The trained agent dynamically adjusts the charging current according to the state every 5 seconds to optimize the trade-off between fast charging and safety limits and can charge the state of charge (SOC) of the battery from 0.2 to 0.8 in 410 seconds, while protecting against overvoltage and overheating. Meanwhile, it still works well for different initial SOC.