Bearing fault diagnosis in modern industrial applications frequently encounters severe challenges due to data privacy fragmentation, domain shifts induced by varying working conditions, and the insufficiency of single-sensor information. To address these critical bottlenecks, this paper proposes a multi-modal domain generalization federated learning framework. Specifically, a dual-branch 1D-CNN feature extraction network is constructed to independently process horizontal and vertical vibration signals, which are then integrated using multi-scale convolutional blocks and a Squeeze-and-Excitation (SE) channel attention mechanism for robust feature fusion. Furthermore, to mitigate the client drift issue caused by non-IID data distributions across different rotational speeds and loads, a FedProx-based federated optimization strategy with a dynamic proximal term regularization is implemented. Extensive experiments conducted on the XJTU-SY bearing dataset demonstrate that the proposed framework achieves an overall classification accuracy of 67.20% and an F1-score of 67.32% under strict differential privacy constraints (DP-SGD noise multiplier = 0.05). Compared with the baseline federated model, the optimized version provides a significant performance gain of +12.37% in diagnostic accuracy. The proposed approach successfully strikes an optimal balance between data privacy protection and cross-condition domain generalization, offering a reliable paradigm for distributed industrial intelligent maintenance.

The report gives a very clear, systematic exposition of the major shortcomings in current Large Language Model (LLM)-based automated reward design methods for reinforcement learning, and after a careful analysis of EUREKA and LEARN-Opt, it naturally and elegantly identifies two fundamental issues: dependence on environment source code and absence of mathematical guarantees for policy optimality. In view of the existing limitations, this paper proposes a well-designed framework that combines Vision-Language Models (VLMs) for visual observation, Graph-of-Thoughts for hierarchical task decomposition, and Potential - based Reward Shaping for theoretical safety guarantees, hence naturally solving the problem of code dependency by using pure visual feedback and reducing generation variance by structured decomposition. The paper presents a method that achieves policy invariance by means of mathematically constrained reward generation, and accordingly gives a clear, logically organized discussion of the problem analysis, theoretical background, proposed approach, implementation architecture, current results, and experimental validation plan. It makes a very natural and important contribution to building robust, generalizable, and theoretically sound automated reward design systems.

It remains unclear whether health benefits differ meaningfully across PA volume levels among individuals who already exercise. To examine this, PA volume data from 787 gym members were analyzed using K-Means clustering. The optimal cluster count was identified using within-cluster sum of squares (WCSS), Silhouette Coefficient, and Davies-Bouldin Index, which indicated four PA volume groups: Low, Moderate, High, and Very High Activity. After the clustering step, one-way ANCOVA was applied to examine differences in body fat percentage and resting heart rate across the groups, with age serving as a covariate and gender taken as a fixed factor. Body fat percentage varied significantly by cluster, and the pattern was non-linear. Resting heart rate, by contrast, did not show visible group differences. These findings, taken together, underscore the utility of coupling clustering with adjusted group comparisons to identify PA volume subgroups, and may in turn guide more individualized exercise assessment and activity recommendations.

In the age of rapid digital transformation, electric vehicles have become more and more common, and have gradually displaced the status of diesel. It is essential for people to pay more attention to the safety of lithium batteries. To illustrate, a lithium-ion battery is prone to error under long-term charging or discharging and high temperature. If the fault diagnosis is not at the right time, it might lead to the declining performance of the lithium-ion battery and shorten the servicing life. This study aims to discuss the technologies for detecting the error in lithium batteries, which are categorized into three forms: Model-based method, Signal processing-based method and Machine learning method. Meanwhile, these technologies have been widely applied in the area of electric vehicle and battery management systems, detecting the state of batteries in real-time. Overall, this study considers that there are no any fault diagnosis methods that are suitable for all the application scenarios, and researchers should choose the approach according to their needs. The future development trend is the combination of various fault diagnosis techniques, improving the accuracy, reliability and performance.

Currently, diabetes affects around 589 million adults worldwide and is one of the more pressing chronic disease burdens in modern public health. Machine learning is widely accepted in diabetes risk modeling, although most published work tends to evaluate models in isolation rather than using them side by side in matching conditions. In particular, how LightGBM is compared to simple deep neural network structures in small, structured clinical datasets is not well described. This study directly addressed this question using a publicly available Kaggle dataset of 1,879 patient records and 46 clinical variables. The paper evaluated two models from three dimensions: differentiation, calibration, and interpretability. LightGBM is better overall than the DNN --AUC 0.965 vs. 0.913, accuracy 0.941. 0.870 – resulting in a significant reduction in missed diabetes cases. To explore what the model actually relies on, it applied SHAP analysis; fasting blood glucose and HbA1c were key drivers, consistent with clinicians' expectations. Overall, the results suggest that LightGBM is a more dependable option in situations where data are limited and interpretability is important, at least for this type of structured tabular input.

Intrusion detection systems built on hand-written rules are losing ground as attackers find new ways to bypass fixed signatures. Working with the NSL-KDD benchmark, the paper put three classical machine learning methods—Decision Tree (DT), Random Forest (RF), and K-Nearest Neighbors (KNN)—through the same set of binary and five-class detection tests. Random forest handled the diverse five-class label set best, hitting 76.3% accuracy with an F1 of 0.720. In the simpler binary task, however, the decision tree pulled ahead (81.2%, F1 0.813). The reason was straightforward: the ensemble model grew too cautious when samples were scarce, suppressing false alarms at the cost of missing real attacks. DoS floods were caught almost without fail (F1 0.884). R2L and U2R, on the other hand, evaded detection almost entirely—their training presence was vanishingly small, at 0.83% and 0.04% of the training pool. Feature importance scores from the random forest pointed overwhelmingly to traffic volume and rate-based statistics as the strongest signals, while class skew stood out as the single largest obstacle. Rather than chasing marginal accuracy gains, the results argue for treating rare-class detection as the main priority in future work.

Accurate hourly bike-sharing demand prediction is important for bicycle rebalancing and resource allocation. This study evaluates the individual and combined contributions of environmental and temporal features to hourly bike-sharing demand prediction using 17,414 hourly records from the London Bike Sharing Dataset. A time-ordered 80:20 split is adopted to mimic forecasting future demand from historical observations. A distance-weighted k-nearest neighbors (KNN) regressor is adopted as the baseline, and a tuned random forest (RF) regressor serves as the main model. RF achieves MAE=173.08 rentals/h, RMSE=304.76 rentals/h, and R²=0.927, reducing MAE and RMSE by 65.7% and 60.2% compared with KNN. Feature-combination experiments show that environmental features alone have limited explanatory power (R²=0.132), whereas temporal features capture the main demand structure (R²=0.868). After feature fusion, MAE and RMSE decrease by 29.0% and 25.6% relative to the temporal-only setting. Visual analyses and permutation importance indicate that temporal variables establish the main demand baseline, while environmental variables provide complementary corrections. The results offer interpretable evidence for feature selection and short-term bike-sharing operations.

Short-term taxi order forecasting is a key reference for dynamic allocation of urban transportation capacity. New York taxi passenger flow exhibits significant intraday cyclical fluctuations, making it difficult for manual dispatch to predict changes in passenger flow. This paper takes hourly taxi order time-series data in New York City as the research object, selecting the Seasonal Autoregressive Integrated Moving Average (SARIMA) model and the Support Vector Regression (SVR) model to conduct a short-term order prediction comparison. Based on AIC and BIC criteria, the optimal SARIMA parameters are traversed and screened, and SVR data are subjected to Min-Max normalization. Both types of models use multi-step rolling forecasting to generate a complete 7-day prediction sequence, relying on MAE, RMSE, MAPE, and R2 indicators to quantify prediction accuracy. Experimental results show that the optimal SARIMA (1,0,1)(1,1,24) model can accurately fit peak and valley fluctuations in passenger flow, with an R2 of 0.815. SVR models can only capture overall trends and lack the ability to characterize extreme passenger flows. The study demonstrates that for taxi time series data with strong intraday cycles, the SARIMA model offers better prediction stability and accuracy, providing a reference for taxi capacity scheduling.

Accurate one-hour-ahead electricity load forecasting sustains dispatch, reserve planning, and dependable power-system operation, but additional inputs do not always improve predictions. This study assesses how different parameter settings contribute to power demand forecasting. Four long short-term models with a unified architecture were trained on 48048 hourly observations and a 24-hour window. Input configurations include historical demand alone, demand with twelve weather variables, demand with two calendar indicators, or all inputs. Each configuration was trained five times and evaluated on a chronological test set. The weather-based model achieved the lowest mean errors: a mean absolute percentage error of 1.599% and a root mean square error of 26.333. The history-only model remained competitive, while calendar indicators and the all-input model offered no improvement. Recent demand therefore provides most of the useful information at this horizon, with lagged weather adding a modest signal. Careful feature selection can reduce complexity and support interpretable, dependable operational forecasts.

Fake-news detection needs both semantic and social evidence: text-only systems miss coordinated diffusion, while graph-only systems may confuse virality with falsehood. We propose a framework that combines BERT text embeddings with Graph Attention Network (GAT) representations of propagation graphs. Projected features are aligned by a cross-modal consistency loss and combined through an adaptive attention gate; when graphs are unavailable, the structural branch is masked. FakeNewsNet and Twitter15/16 evaluate full fusion, while LIAR tests text-only operation. On FakeNewsNet, the model reaches 93.4% accuracy, 92.8% Macro-F1, and 0.97 AUC. Ablations support both fusion components, and token- and node-level attribution assists human review.