The advent of Large Language Models (LLMs) holds promise for revolutionizing various fields traditionally dominated by human expertise. Urban planning, a professional discipline that fundamentally shapes our daily surroundings, is one such field heavily relying on multifaceted domain knowledge and experience of human experts. The extent to which LLMs can assist human practitioners in urban planning remains largely unexplored. In this paper, we introduce a comprehensive benchmark, UrbanPlanBench, tailored to evaluate the efficacy of LLMs in urban planning, which encompasses fundamental principles, professional knowledge, and management and regulations, aligning closely with the qualifications expected of human planners. Through extensive evaluation, we reveal a significant imbalance in the acquisition of planning knowledge among LLMs, with even the most proficient models falling short of meeting professional standards. For instance, we observe that 70% of LLMs achieve subpar performance in understanding planning regulations compared to other aspects. Besides the benchmark, we present the largest-ever supervised fine-tuning (SFT) dataset, UrbanPlanText, comprising over 30,000 instruction pairs sourced from urban planning exams and textbooks. Our findings demonstrate that fine-tuned models exhibit enhanced performance in memorization tests and comprehension of urban planning knowledge, while there exists significant room for improvement, particularly in tasks requiring domain-specific terminology and reasoning. By making our benchmark, dataset, and associated evaluation and fine-tuning toolsets publicly available at https://github.com/tsinghua-fib-lab/PlanBench, we aim to catalyze the integration of LLMs into practical urban planning, fostering a symbiotic collaboration between human expertise and machine intelligence.

In response to the challenges posed by imbalanced failure diagnosis samples, limited labeled data, and significant computational costs in actual industrial production settings, this paper introduces a high-precision, low-resource, end-to-end fault diagnosis framework. On one hand, we propose a data augmentation method based on GCGAN, which combines CNN and GRU to construct core network structures for the generator and discriminator. We integrate a novel Smoothed Hinge-Cross-Entropy loss function to facilitate the training process, effectively mitigating mode collapse and vanishing gradient issues. On the other hand, we design a lightweight fault diagnosis model based on MDSCNN-ICA-BiGRU. By substituting standard convolutions with depthwise separable convolutions on deeper channels, the model complexity is significantly reduced, facilitating effective extraction of multiscale spatial features. The improved Coordinate Attention (CA) mechanism filters out noise and enhances the extraction of high-frequency characteristics. Combined with BiGRU, the model captures global temporal associations, achieving a fusion of spatiotemporal features. Experimental results demonstrate that the proposed approach performs well on both publicly available simulation datasets and private laboratory datasets. Compared to other benchmark methods, the GCGAN module significantly enhances data augmentation, improving classification accuracy on CNNs by 10%. When compared with classic convolutional networks such as DRSN and WDCNN, our MDSCNN-ICA-BiGRU shows faster and more stable convergence rates, with near-100% accuracy on test sets and an average computation cost reduction of approximately 70%. Even in noisy environments, our method maintains high accuracy with a slow rate of precision decay, indicating robustness and generalization capabilities.