PURPOSE: To obtain the title apparatus easily preparing a fault discriminating dictionary by allowing a neural network to learn so as to output a category showing the characteristics of a pattern shape to a representative code.
CONSTITUTION: A neural network 60 is equipped with an inlet layer (8 units) 601, an intermediate layer (16 units) 602 and an output layer (9 units) 603 and allowed to learn by changing the weight of the bonding between the units of the different layers on the basis of the error between the output signal to an input signal and a teacher signal according to a back propagation system. The input layer 601 receives the supply of four-bit radial codes A-D from an encoding circuit 26A and the length bits L1, L2 of the codes are supplied to one unit as length codes with respect to each set and a pair bit P and a direction bit Dr are supplied to the other unit as balance codes. The category outputted from the network 60 is supplied to a fault discrimination circuit 62 and discriminated to output a result. When a quality judging mistake is generated, the network 60 is allowed to learn so as to output the correct category.
HIZUKA TETSUO
ANDO MORITOSHI