People's daily lives involve numerous periodic behaviors, such as eating and traveling. Local-life platforms cater to these recurring needs by providing essential services tied to daily routines. Therefore, users' periodic intentions are reflected in their interactions with the platforms. There are two main challenges in modeling users' periodic behaviors in the local-life service recommendation systems: 1) the diverse demands of users exhibit varying periodicities, which are difficult to distinguish as they are mixed in the behavior sequences; 2) the periodic behaviors of users are subject to dynamic changes due to factors such as holidays and promotional events. Existing methods struggle to distinguish the periodicities of diverse demands and overlook the importance of dynamically capturing changes in users' periodic behaviors. To this end, we employ a Frequency-Aware Multi-View Interest Modeling framework (FIM). Specifically, we propose a multi-view search strategy that decomposes users' demands from different perspectives to separate their various periodic intentions. This allows the model to comprehensively extract their periodic features than category-searched-only methods. Moreover, we propose a frequency-domain perception and evolution module. This module uses the Fourier Transform to convert users' temporal behaviors into the frequency domain, enabling the model to dynamically perceive their periodic features. Extensive offline experiments demonstrate that FIM achieves significant improvements on public and industrial datasets, showing its capability to effectively model users' periodic intentions. Furthermore, the model has been deployed on the Kuaishou local-life service platform. Through online A/B experiments, the transaction volume has been significantly improved.

The rapid and accurate detection of faults for lithium-ion batteries plays a critical role in ensuring the safe operation of electric vehicle systems. This paper proposes a fault detection method for electric vehicle batteries by exploiting the temporal smoothness and spatial similarity of battery pack data. Firstly, a temporal convolutional network (TCN) is utilized to learn the latent spatial and temporal features of the data. Then, a self-attention mechanism is employed to capture the correlations and importance between different features. Furthermore, an autoencoder is employed to reconstruct input data based on the extracted spatial-temporal features. This encoder-structured approach is trained only using normal data and the anomalies are detected as conspicuous differences between the input data and the reconstructed data. The detection performance of the proposed method is validated by utilizing real-world electric vehicle operational datasets.

Temperature heavily affects the lifespan and performance of batteries. High temperatures accelerate capacity degradation and can cause thermal runaway, highlighting the importance of battery cooling strategies for electric vehicles. This paper proposes the battery cooling strategy for electric vehicles with LiFePO4batteries using the twin delayed deep deterministic policy gradient (TD3) algorithm. Firstly, the electric-thermal-aging and active battery thermal management system models are introduced, and the thermal management problem is formulated as a continuous Markov decision process. Then, a reward function is designed based on prolonging battery life, reducing refrigeration cost, and maintaining battery temperature. Finally, a TD3 algorithm based on the double-delayed update mechanism is designed to obtain the optimal battery cooling strategy in the continuous state-action space. Simulation results demonstrate that the proposed strategy outperforms the traditional methods and closes to the optimal benchmark, offline dynamic programming, with battery capacity loss of less than 3.32% and an SoC consumption of less than 0.55%, significantly reducing the operational cost and mitigating the battery aging.

Extracting motion pattern implied in the history trajectory is important for the pedestrian trajectory prediction task. The motion pattern determines how a pedestrian moves, including but not limited to reaction of interaction, tendency of speed and direction change. Although the motion pattern is a comprehensive concept and can’t be described concretely, it is clear that it contains both long-term and short-term factors. Inspired by this, we introduce SpectrumNet which enables more effective encoding of historical motion patterns for trajectory prediction. Different from existing methods, which consider the history trajectory as a time sequence of position, SpectrumNet represents it in the frequency space by applying Fourier Transform (FT) to decompose the historical information on different time scales. SpectrumNet consists of two sub-networks, the Multi-Frequency Combination (MFC) encoder, which models the historical information by combining multiple frequency feature in the spectrum; and the Frequency Interaction (FI) encoder, which captures the interaction between pedestrians in the frequency domain. To validate the effect of SpectrumNet, we build a CVAE-based prediction system to predict stochastic future trajectory. Experiments conducted on ETH-UCY dataset show that our prediction system with SpectrumNet out-performs the previous state-of-the-art model and achieves a new record on ADE metric.

Modeling scene-agent interaction is required in reliable and feasible for pedestrian trajectory prediction. Most existing methods utilize segmentation result of scene images as semantic information and concatenate it with trajectory features to model scene-agent interaction by concatenating semantic features with trajectory features. Since the image and trajectory are heterogeneous data, directly concatenating features proves inadequate for modeling scene-agent interaction. We notice that agent trajectory is the result of the interaction between the agent movement decision and the scene. They could be regarded as examples that demonstrate the influence between the scene and pedestrians’ motion in different locations. Compared with semantic image, historical trajectories of neighboring agents do not cover the whole scene, but they are more direct and efficient information for scene-agent interaction modeling. We propose a novel scene-agent interaction modeling method which uses Fourier Transform to convert the temporal historical trajectories into the frequency domain, utilizes transformer to encode the scene-agent interaction feature(SAIF), and designs a spectral position embedding to represents the scene influence on the agents. We combine the spatio-temporal features of trajectories and inter-agent interaction features with SAIF for generating predicted trajectories. Experiments demonstrate that our method achieves the state-of-the-art on both ETH-UCY and SDD.

The distribution of pedestrian trajectories is highly complex and influenced by the scene, nearby pedestrians, and subjective intentions. This complexity presents challenges for modeling and generalizing trajectory prediction. Previous methods modeled the feature space of future trajectories based on the high-dimensional feature space of historical trajectories, but this approach is suboptimal because it overlooks the similarity between historical and future trajectories. Our proposed method, TrajCLIP, utilizes contrastive learning and idempotent generative networks to address this issue. By pairing historical and future trajectories and applying contrastive learning on the encoded feature space, we enforce same-space consistency constraints. To manage complex distributions, we use idempotent loss and tightness loss to control over-expansion in the latent space. Additionally, we have developed a trajectory interpolation algorithm and synthetic trajectory data to enhance model capacity and improve generalization. Experimental results on public datasets demonstrate that TrajCLIP achieves state-of-the-art performance and excels in scene-to-scene transfer, few-shot transfer, and online learning tasks.

The SOTA face swap models still suffer the problem of either target identity (i.e., shape) being leaked or the target non-identity attributes (i.e., background, hair) failing to be fully preserved in the final results. We show that this insufficient disentanglement is caused by two flawed designs that were commonly adopted in prior models: (1) counting on only one compressed encoder to represent both the semantic-level non-identity facial attributes(i.e., pose) and the pixel-level non-facial region details, which is contradictory to satisfy at the same time; (2) highly relying on long skip-connections [50] between the encoder and the final generator, leaking a certain amount of target face identity into the result. To fix them, we introduce a new face swap framework called "WSC-swap" that gets rid of skip connections and uses two target encoders to respectively capture the pixel-level non-facial region attributes and the semantic non-identity attributes in the face region. To further reinforce the disentanglement learning for the target encoder, we employ both identity removal loss via adversarial training (i.e., GAN [18]) and the non-identity preservation loss via prior 3DMM models like [11]. Extensive experiments on both FaceForensics++ and CelebA-HQ show that our results significantly outperform previous works on a rich set of metrics, including one novel metric for measuring identity consistency that was completely neglected before.

A power electronics transformer (PET) based on dual active bridge (DAB) with input series and output parallel (ISOP) cascaded topology is studied in this paper. Fe-based nanocrystalline alloys transformer cores and Silicon Carbide (SiC) power MOSFETs are adopted to improve the power density and efficiency. Mathematical derivation of the input voltage distribution is deduced with the conclusion that the voltage equalization cannot be achieved autonomously. A feedback control method is proposed to balance the input voltage. The measurement of the dynamic B-H loops of the nanocrystalline core under the excitation of square wave and sinusoidal wave is conducted. The high frequency isolation transformer (HFIT) is designed based on the loss performance calculated using the dynamic B-H loop and Ohm's law. The optimized operating flux density is around ±0.68 T at 20 kHz, which is much higher than the empirical value around ±0.3 T to take full advantage of the nanocrystalline core. Finally, a 13.2 kVA prototype is built to verify the above analysis and design.

The multiplex 3-phase winding motors suffer from the problem of the magnetic coupling between the normal winding and the fault winding. To investigate the behaviors of the dual-winding PMSMs on the short-circuited fault conditions, a fault analysis model is developed that includes a dual-winding PMSM circuit model with a short-circuited fault coil within some phase coil and an inverter model with a vector control. The 1/10 and 1/4 turn-to-turn short-circuited faults are simulated with a customized dual-winding PMSM of 4.7 kW. The results show that even with a short-circuited current up to three times of the rated value, the damping torque produced by the fault current can be compensated if the motor windings and their driving inverters have a back-up capacity of about 30% the rated value. An experimental platform with the 4.7kW dual-winding PMSM is built for verifying the analysis model. The simulated waveforms are very close to the measured ones not only for the turn-to-turn short-circuited but also for phase-to-phase short-circuited faults. With the short-circuited current being limited to the rated value, the damping torque produced by the phase-to-phase short-circuited current is too large to be compensated.