Abstract: In order to realize real-time fault diagnosis, a method for engine fault diagnosis based on exhaust density analysis and support vector machines (SVM）was put forward. Under typical fault working conditions of the engine, firstly, the data of exhaust densities of HC, CO, CO2, O2, NOX were gotten by using NHA-500 exhaust density analysis instrument. Then the data were normalized, and feature vectors were extracted from the data as learning samples and then used in designing and training multielement classifier based on support vector machines for fault pattern recognition. Experimental results showed that error correction coding classification method based on support vector machines was better in classification ability and had stronger anti-jamming capability than neural networks. In the case of small samples, accuracy rate of this fault diagnostic method could reach 98.5%. The result means that the method can effectively describe the complex relationship between exhaust compents changes and fault states.