The Air Production Unit (APU) of a metro train is the core unit responsible for compressing and supplying air, which is mainly used to support the pneumatic system of the metro train. Anomaly detection helps to detect faults and errors early before the machine reaches a critical stage. In this paper, a multiscale CNN-Mamba-transformer based prediction model determines future failures in the Air Production Unit (APU) of a Porto Metro train by predicting the oil temperature on the compressor. The model first extracts the surface features of the time series of oil temperature using CNN multiscale with different convolutional kernels, and then performs highly accurate fault prediction by combining the new models mamba and Transformer focusing on the short-term and long-term nature of the time series for efficient train maintenance. Experimental validation shows that the model exceeds the accuracy of the reference models listed in this paper.