As a means of further improving combustion efficiency of gasoline engine, an increase in compression ratio, which enhances the risk of knocking, is thinkable. To optimize engine combustion parameters, a technology that can precisely detect knocking is desirable. Presently skillful experts have been evaluating knocking subjectively by listening to radiation noise so far.The authors developed a device that can precisely detect knocking by means of processing sound signals, which are captured by a high-performance microphone that is sensitive in the wide frequency range. Shock waves induced by knocking cause in-cylinder gas vibrations that emits metallic hit noises from the outer engine wall. We studied how to identify the feature values of frequency characteristics when knocking occurs, under the assumption that the engine radiation noise includes more than 2nd-order harmonic components with respect to the basic frequency of the in-cylinder gas vibration mode.In extracting sound data near the TDC, a feature value is calculated with the bispectrum analysis, which can easily grasp harmonic components and is tough against Gaussian noise. In applying a statistical technique and machine learning methods, threshold values for detecting knocking are determined. It is now possible to detect knocking even under a circumference where several disturbance noises exist, such as combustion noises from multi-cylinder engines, auxiliary equipment noises, valve train noises, gear noises, etc. The measurement results coincide well with the subjective ratings made by experts. Furthermore knocking is detectable even at a high-speed region more than 5,000 rpm. The method can identify the cylinder, in which knocking occurs, using only one microphone.From the above results, we conclude that the method is effective for the objective rating of knocking and automatic optimization of engine combustion parameters.