DeepSIP: A Deep Learning Approach for Sensor Data Imputation and Time Series Forecasting

Large-scale datasets collected from sensor networks in domains such as industrial IoT, healthcare, transportation, and environmental monitoring often contain significant temporal and spatial gaps caused by sensor failures, communication losses, or maintenance outages. Such missing data can introduce bias, reduce reliability, and limit the performance of predictive models. To address this challenge, we propose DeepSIP (Deep Sensor Imputation and Prediction), a novel deep learning framework that unifies imputation and forecasting for multivariate time series data. DeepSIP employs a cluster-based training approach on fully observed sensor data to identify contextual and temporal patterns prior to imputation. Its autoencoder-based module learns latent representations from correlated sensor variables and contextual information (e.g., time and date) to reconstruct missing values. Subsequently, a deep predictive network with multiple fully connected layers is trained on the imputed datasets to model complex temporal and cross-sensor dependencies for accurate forecasting. Our experiments across various missingness scenarios demonstrate that DeepSIP consistently achieves the lowest reconstruction errors (MSE, MAE, RMSE) and the highest forecasting accuracy and $$R^2$$ scores compared to K-Nearest Neighbors (KNN), Multiple Imputation by Chained Equations (MICE), and Linear Interpolation. These results validate DeepSIP’s robustness and adaptability for sensor-driven applications, highlighting the importance of high-quality imputation in improving downstream predictive performance.