In order to effectively diagnose the infeasible linear programming (LP) model of production planning in refinery, the article proposed three stages strategy based on constraints' classification and infeasibility analysis. Generally, infeasibility sources involve structural inconsistencies and data errors, and the data errors are further classified into I, II and III. The three stages strategy are: (1) Check data when they are inputted to detect data error I and repair them; (2) Inspect data whether they are accorded with material balance before solving the LP model to identify data error 11 and repair them; (3) Find irreducible inconsistent system of infeasible LP model and give diagnosis information priority-ranked to recognize data error III and structural inconsistencies. These stages could be automatically executed by computer, and the approach has been applied to diagnose the infeasible model well in our graphic I/O petro-chemical industry modeling system.