Objective. Current autism clinical detection relies on doctor observation and filling of clinical scales, which is subjective and prone to misdetection. Existing autism research of functional magnetic resonance imaging (fMRI) over-compresses the time-scale information and has poor generalization ability. This study extracts multiple time scale brain features of fMRI, providing objective detection. Approach. We first use least absolute shrinkage and selection operator to build a sparse network and extract features with a time scale of 1. Then, we use hidden markov model to extract features that describe the dynamic changes of the brain, with a time scale of 2. Additionally, to analyze the features of the potential network activity of autism from a higher time scale, we use long short-term memory to construct an auto-encoder to re-encode the original data and extract the features at a higher time scale, with a time scale of T, and T is the time length of fMRI. We use recursive feature elimination for feature selection for three different time scale features, merge them into multiple time scale features, and finally use one-dimensional convolution neural network for classification. Main results. Compared with well-established models, our method has achieved better results. The accuracy of our method is 76.0%, and the area under the roc curve is 0.83, tested on completely independent data, so our method has better generalization ability. Significance. This research analyzes fMRI sequences from multiple time scale to detect autism, and it also provides a new framework and research ideas for subsequent fMRI analysis.