The reasoning illusion of large language model is the core bottleneck restricting its safe landing. The existing research mostly focuses on post-event detection and lacks pre-prediction mechanism. In order to improve the accuracy of illusion prediction and generalization performance, a BiLSTM-Attention classification algorithm based on RBMO optimization is proposed in this paper. The model uses a two-way long-short-term memory network to capture the pre-and post-sequence dependencies, focuses on key features with the help of attention mechanism, and introduces the red-billed blue magpie optimization algorithm to adaptively optimize the network weights and hyperparameters, so as to alleviate the local optimal problem of gradient descent. A standardized data set containing 1729 samples was constructed in the experiment, and a variety of machine learning algorithms were selected for comparison and verification. The results show that the proposed algorithm is optimal in accuracy, recall, accuracy, F1 value and AUC value, and all indexes are 0.975 or above, which is overall superior to traditional machine learning and single deep learning models, and can provide effective technical support for large model hallucination pre-prediction and safe deployment.

Show more

Rapid urbanization has promoted the continuous expansion of urban built-up space. Green spaces, water bodies, woodlands, and open spaces have been increasingly divided by roads, residential areas, commercial land, and high-intensity functional zones. This process has caused ecological patch fragmentation, blocked ecological flows, and reduced the accuracy of corridor identification. To support the optimization of urban landscape ecological networks, this study takes the central urban area of Guangzhou as the case area. It integrates Sentinel-2 remote sensing imagery, 30 m land-use data, POI data, and road data from 2020 to 2024. A multi-source spatial data-driven method is developed to evaluate urban landscape ecological connectivity. NDVI and MNDWI are used to extract vegetation and water information. MSPA is combined to identify ecological sources. Land-use resistance, road proximity resistance, built-up intensity, POI kernel density, and the inverse NDVI indicator are integrated into a comprehensive ecological resistance surface. The MCR model, circuit theory, and graph-theoretic indicators are then used to extract potential ecological corridors and identify pinch points, barrier points, and key restoration nodes. The results show that ecological spaces in central Guangzhou present a pattern of relative continuity in peripheral areas and clear fragmentation in the urban core. The Pearl River green belts, Haizhu Wetland, the southern edge of Baiyun Mountain, and large parks form the main ecological connection framework. Road intersections, high-density commercial areas, and hardened riverbanks still create significant barriers. The study provides a computational basis for green-space restoration, ecological corridor planning, and landscape spatial optimization.

Show more

Aiming at the problems of insufficient prediction accuracy, weak interpretability and inconsistent data caliber of traditional models, this paper constructs a Bayesian Optimization-LightGBM model, and combines SHAP values to carry out grain yield prediction and influencing factor analysis. The research results show that the coefficient of determination is 1.0000, the root mean square error is 10.7267, and the mean absolute error is 6.4542, indicating excellent model fitting accuracy. The national grain yield in 2026 will still be highly concentrated in major producing areas, with Heilongjiang, Henan and Shandong ranking the top three steadily. The prediction results are consistent with the actual production pattern. SHAP value analysis shows that sown area indicators are the core driving factors. This study provides effective methods and reference basis for accurate grain yield prediction and policy formulation.

Show more

Cracks in mechanical structures cause numerous hazards, such as stress concentration, reduced structural strength and fatigue performance; continuous crack propagation easily leads to component fracture and equipment failure; it results in downtime maintenance and increased costs, and even safety accidents in severe cases. Therefore, accurate crack detection is crucial. Traditional detection relies on manual work, which has pain points such as low efficiency, strong subjectivity, obvious missed detection of micro-cracks, and difficulty in identifying internal hidden cracks, failing to meet the needs of real-time large-scale detection. This study takes YOLOv5 as the core and combines deep learning technology to carry out research on mechanical crack detection. A mechanical crack dataset is constructed, and sample annotation and data augmentation are completed to improve sample diversity; parameters are modified based on the YOLOv5 model, and after multiple rounds of training iteration and verification, the model finally achieves a mAP@0.5 of 31.37%, effectively reducing the missed detection rate and false detection rate; after model deployment, it can realize fast, automatic and high-precision identification and location of mechanical surface cracks, effectively alleviating the core pain points of traditional detection, and providing an efficient and feasible technical solution for mechanical structure crack detection.

Show more