Abstract: The diversity of industrial control protocols， unknown specifications， and difficult classification are key challenges in achieving interconnectivity of industrial control systems and ensuring information security. Therefore， a classification method based on multi-scale feature deep learning for unknown industrial control protocols is proposed. Firstly， considering the dense key information in the header field of industrial control protocols， a multi-scale feature extraction method combining both byte and half-byte is proposed to achieve feature extraction without prior knowledge. Furthermore， leveraging the inconsistency of feature bytes in the header field， an automatic feature marking method is proposed to dynamically update the protocol feature set. On this basis， to ensure real-time classification， a deep learning classification method based on a one-dimensional convolutional neural network with stacked gated recurrent units is proposed. Comparative experiments on public datasets demonstrate that the accuracy and the precision achieved by the proposed method are more than 99.5%.