This paper introduces in detail the design and implementation process of a real-time monitoring and fault diagnosis system for coking waste gas treatment. By constructing a comprehensive data acquisition system, combined with distributed sensor layout, the continuous monitoring of key emission parameters in coking process was realized. For data processing, feature selection, data cleaning and other preprocessing measures are adopted, and the innovative A-LSTM model is introduced. The model enhances the ability of LSTM network to capture key information in time series data by introducing attention mechanism, and significantly improves prediction accuracy and response speed. In terms of fault diagnosis, CNN-RNN fusion framework is developed, which effectively integrates the advantages of two deep learning models and strengthens the recognition ability of complex fault modes. In addition, model fusion and optimization strategies, such as weighted average and hyper-parameter tuning, are used to further improve the overall system performance.