Accurate detection of mental state is vital for sleep assessment and anesthesia monitoring. However, challenges exist, including 1) limited data sources caused by few modalities, few channels, biased sample distribution, or large individual differences; 2) multiscale dynamic feature of mental state signals, containing complex and dynamic fluctuations. In this paper, we proposed a multiscale temporal convolution and attention improved neural network (MTA-NN) to handle these two challenges, including: a preprocessing method using a clustering-first and label-generation approach for data augmentation involves filtering the production samples to mitigate the impact of noise on model performance; multiscale temporal-convolution module for detecting different scales of temporal and frequency features; an attention module for decoding key features in different spatial and temporal channels, and for setting the importance of historical dependencies; a full-connection layer is designed to recognize different sleep stage. The proposed MTA-NN is then verified in three typical sleep-state recognition tasks, including SleepEDF-20, SleepEDF-78, and ISRUC3. Compared to some state-of-the-art algorithms, the proposed MTA-NN achieves higher performance. Further ablation experiments show the necessary and importance of these designed sub modules in MTA-NN. We consider the same algorithm can also be extended to further similar mental state recognition tasks such as anesthesia monitoring.