This paper introduces a deep learning approach for estimating fill probabilities of limit orders in limit order books using survival analysis. The model employs a convolutional-Transformer encoder to capture complex temporal dependencies in LOB data, combined with a monotonic neural network decoder to ensure valid survival functions. The authors emphasize the use of proper scoring rules, specifically right-censored log-likelihood, for model evaluation, which avoids pitfalls of improper metrics like time-dependent concordance. The method is tested on Nasdaq data for multiple assets with varying tick sizes and trading activity. Results show that the proposed architecture significantly outperforms traditional survival models (e.g., Cox proportional hazards, DeepSurv) and recent deep learning baselines. A comprehensive statistical analysis reveals that placing limit orders inside the bid-ask spread dramatically increases fill probability and reduces time-to-fill, especially for large-tick stocks. Interpretability analysis using Shapley values indicates that high-frequency microstructural features are most influential for predictions. The work provides practical insights for optimal order placement in electronic markets.