This study introduces a digital platform for the oil and gas industry, leveraging industrial big data and machine learning to improve assessment and forecasting precision. We introduce an improved model from the Time-Masked Autoencoder (TimeMAE) for prediction of oil and gas production anomaly time series. This architecture inserts time series information into a low-dimension latent space to embed major features and utilises a self-enforcing attention module to aggregate deep multi-modal data for superior extraction of temporal patterns and outlier signals. Moreover, we advocate a cloud-edge collaborative paradigm where deep cloud models and shallow edge models collaborate under their own expertise to best execute multi-modal data processing, economize on training costs, and optimize task assignment and inference pathways. This solution offers a cost-effective and adaptable approach to implementing intelligent systems in upstream oil and gas operations, with a focus on real-time surveillance and forecasting capabilities that can scale with increasingly demanding and intricate production environments.