For any faults of automotive engines, the diagnosis can be performed based on variety of symptoms. Traditionally, the description of the faulty symptom is just existence or not. However, this description cannot lead to a high accuracy because the symptom sometimes appears in different degrees. Therefore, a knowledge representation method which could precisely reflect the degree of the symptom is necessary. In this paper, the fuzzy logic is firstly applied to quantify the degrees of symptoms. A probabilistic classification system is then constructed by using the fuzzified symptoms and a new technique namely Fuzzy Relevance Vector Machine (FRVM). Moreover, both Fuzzy Probabilistic Neural Network (FPNN) and Fuzzy Probabilistic Support Vector Machine (FPSVM) are used to respectively construct similar classification systems for comparison with FRVM. Experimental results show that FRVM produces higher diagnosis performance than FPNN and FPSVM. © 2011 Springer-Verlag.