This project focuses on the three core stages of water quality monitoring: data acquisition, intelligent analysis, and proactive warning. An intelligent water quality detection system integrating an AI-based prediction model was designed and implemented. The system adopts a dual-chip architecture using STM32F103 and STM32F407, which are respectively responsible for multi-parameter acquisition and edge intelligent computing. It achieves high-precision, millisecond-level synchronized acquisition of key indicators including water temperature, pH, TDS, and turbidity. To address resource constraints on embedded platforms, the LSTM model was pruned, quantized, and structurally optimized. Combined with the CMSIS-NN acceleration library, the model was successfully compressed to 48 KB, maintaining high accuracy (R² ? 0.92, RMSE ? 0.35) while achieving real-time prediction within 800 ms. To enhance robustness and interpretability, a hybrid “Physics + AI” prediction model was proposed by integrating physical-mechanism features, improving pH anomaly prediction accuracy by approximately 38% compared with conventional methods. Furthermore, the system integrates local audio-visual alarms and remote Wi-Fi synchronization, forming a prediction-driven closed-loop warning mechanism that can issue early warnings up to 2 hours in advance. Experimental results show that the system achieves excellent stability, accuracy, and real-time performance, with strong potential for application in aquaculture, wastewater management, and environmental monitoring.