This paper presents a statistical feature approach in fully convolutional time series classification (TSC), which is aimed at improving the accuracy and efficiency of TSC. This method is based on fully convolutional neural networks (FCN), and there are the following two properties: statistical features in data preprocessing and finetuning strategies in network training. The key steps are described as follows: firstly, by the window slicing principle, dividing the original time series into multiple equal-length subsequences; secondly, by extracting statistical features on each subsequence, in order to form a new sequence as the input of the neural network, and training neural network by the fine-tuning idea; thirdly, by evaluating the classification performance about test sets; and finally, by comparing the sample sequence complexity and network classification loss accuracy with the FCN using the original sequence. Our experimental results show that the proposed method improved the classification effects of FCN and the residual network (ResNet), which means that it has a generalization ability to the network structures.