In the hot strip rolling production process, mechanical properties are detected offline. The process data is used for clustering analysis to acquire the state in advance. During the mechanical properties detection, the outliers in the same steel grade are regarded as the focal points. Spectral clustering is one of the advanced methods, where the Euclidean distance as the common similarity measure, can only extract the features of the spherically distribution data and can not express the complex distribution data. In this paper geodesic distance is introduced to spectral clustering, which is used to do the production state clustering. The Tennessee Eastman and hot strip rolling process data are used for model validation, as a result the proposed method has better performance on clustering, compared with the Euclidean distance.