Remaining Life Prediction of Water Purification Filter Based on Time Series

Long-term time series forecasting is a widely applicable technique in real-world scenarios. Predicting the remaining lifespan of water filters is crucial for maintaining water quality in both domestic and industrial settings. Accurate predictions facilitate timely filter replacement, thereby preventing water quality degradation and reducing maintenance costs. For such predictions, Informer, an adaptation of the Transformer model, has demonstrated strong performance due to its sparse attention mechanism, which prioritizes the most critical attention distributions. However, this approach faces challenges in capturing complex physical trajectories and trend curves. To address this issue, we enhanced the model by transforming 1D time series data into a 2D tensor set based on multiple periods. Additionally, we incorporated a bi-level routing dynamic sparse attention mechanism, featuring dynamic querying and adaptive sparsity, which allows for more flexible computation and content-aware processing. Experimental results show that this method effectively predicts the remaining lifespan of water filters, exhibiting excellent performance in practical applications.