The paper describes a new technique for on-line process fault diagnosis using fuzzy neural networks. The fuzzy neural network considered in this paper is obtained by adding a fuzzification layer to a conventional feed-forward neural network. The fuzzification layer converts the increment in each on-line measurement and controller output into three fuzzy sets; ‘increase’, ‘steady’ and ‘decrease’ with corresponding membership functions. The feed-forward neural network then classifies abnormalities, represented by fuzzy increments in on-line measurements and controller outputs, into various categories. The fuzzification layer can compress training data, and thereby ease training effort. Robustness of the diagnosis system is enhanced by adopting a fuzzy approach in representing abnormalities in the process. Applications of the proposed technique to the fault diagnosis of a continuous stirred tank reactor system demonstrate that the technique is robust to measurement noise, capable of diagnosing incipient faults, and requires fewer training data examples than a conventional network approach.