The invention relates to an industrial internet patent identification method based on multi-instance learning, which utilizes a natural language processing technology to segment summary information in a patent into sentences, and utilizes a text subject sentence extraction algorithm based on a sentence relation graph to extract subject sentences in a summary, so that the calculation overhead can be effectively reduced. Meanwhile, by combining the topic sentences extracted from the title and the abstract, the patents are converted into sentence packages, each patent is regarded as a package, and each sentence in the package is regarded as an example. And finally, the category of the new sample is predicted by adopting a K Neighbor Networks (KNN) algorithm (K Neighbor Networks, KNN). By means of the method, the industrial internet patent recognition effect can be effectively improved, the manual review cost is greatly reduced, and the method has very important significance on patent retrieval.