The Kohonen self-organizing map (KN) was developed for pattern recognition, and has been extended to fault classification. However, the KN cannot be applied to classify faults from the system output if it contains other factors, such as system state and sensor mounting errors. To overcome this problem, a constrained KN (CKN) is proposed. To eliminate the effect of the system state and the mounting errors, it is proposed that the weight vectors of the CKN are constrained in the parity space. The training algorithm of the CKN is derived, and its convergence discussed. Application of the CKN to fault classification is presented, and its performance is illustrated by an example involving a redundant sensor system with six sensors.