Early fault prediction in chemical processes remains challenging because incipient signatures are weak, evolve gradually, and are often obscured by dominant process dynamics and measurement noise. This paper proposes a temporal–spectral early-warning network (TS-EWNet), which directly maps historical multivariate measurements to fault-related probabilities without intermediate signal forecasting. A coarse-to-fine architecture first captures the global process context and then enhances weak fault-related variations by integrating temporal responses with low-frequency spectral evidence. TS-EWNet is evaluated on the Tennessee Eastman Process (TEP) benchmark, a first-principles continuous stirred-tank reactor (CSTR) simulation, and a laboratory-scale thermo-hydraulic platform. On the TEP dataset, TS-EWNet achieves an accuracy of and a Macro-F1 of at , while retaining an accuracy of at . At , its Early-F1, stable correct warning rate, and penalized normalized stable detection delay are 68.77%, 69.62%, and 0.3286, respectively. After Holm correction applied separately to each metric, 13 of 14 Macro-F1 comparisons and all 14 Early-F1 comparisons remain statistically significant. On the CSTR dataset, the model achieves 96.62% accuracy and 97.89% Macro-F1 at . Under continuous streaming operation on the laboratory platform, it obtains a macro-average accuracy of 77.71% and a Macro-F1 of 70.21%. For a representative F5 trajectory, a stable warning is confirmed 45 s before the reference fault manifestation, while the inference time remains below the 1 s sampling interval. These results demonstrate the effectiveness of TS-EWNet for multi-horizon fault prediction, initial-stage warning, and real-time process monitoring.