A learned neural network in weights using at least the four parameters of the engine load, engine speed, intake pressure inside the intake passage downstream of the throttle valve (12), and amount of intake air fed into the engine as input parameters of the neural network and using a target EGR rate as training data is stored. At the time of engine operation, the learned neural network is used to estimate the target EGR rate from the above parameters and abnormality of the exhaust gas recirculation system is detected based on the difference between the estimated value of the target EGR rate and the target EGR rate.