An automatic system of ultrasonic inspection with a multi-layered neural network software to adhesive surface defects between CBN segment chips and a disk periphery has been developed to serve the guarantee of the quality of the grinding wheel. The network was used to contrive the accuracy improvement of the inspection. The waveforms reflected from the adhesive location of either prescribed artificially exfoliated defect or non-defect were investigated in detail to distinct the characteristics of the waveform. The network learned preferentially with both defect and non-defect waveforms, and also the improvement of the network learning compensated the amplitude of the wave near the edges was implemented. The inspection was performed to the grinding wheel with being unknown defects by using the learned network. Inspection results supported that the inspection system contributes the decision of the adhesive integrity of the CBN grinding wheel.