A Large Time-Series Model for Cross-Domain Transfer in Multi-Equipment Forecasting and Anomaly Detection

In the operation and maintenance (O&M) stage of industrial and civil equipment, practical challenges—such as data scarcity and contamination, poor coordination among heterogeneous multi-sensor data, and weak cross-equipment generalization of traditional fault detection systems—severely limit real-world deployment. To address these issues, this paper proposes a cross-domain transfer large time-series model for multi-equipment forecasting and anomaly detection. First, we introduce an equipment-oriented pretrain–finetune time-series Transformer architecture to mitigate data scarcity and poor cross-equipment generalization. Second, we construct a forward input module that combines an improved reversible instance normalization (RevIN) with chunking (segmentation) to reduce the impact of data contamination, anomalies, and distribution shift on experimental results. Finally, we design a backward output module that integrates Fast Fourier Transform (FFT)-based temporal processing and time–frequency loss fusion, and a unified generation strategy for both forecasting and anomaly detection. This makes forecasts smoother, improves anomaly detection precision, and enhances scalability and generality. Experimental results show that, after finetuning in few-shot equipment scenarios, our pretrained model achieves better forecasting and anomaly detection performance than prior methods, demonstrating strong generalization and meeting cross-domain transfer requirements across equipment.