The deep learning model we developed estimates knocking sound pressure and knocking component superimposed on in-cylinder pressure from engine radiation noise measured by a microphone. This model obtains the time frequency mask and frequency response from many pair data of engine radiation noise and in-cylinder pressure. The time frequency mask extracts the knocking sound from engine radiation noise. The frequency response converts the extracted knocking sound into the knocking component superimposed on in-cylinder pressure. We propose an improved model to separate the knocking sound from the engine radiated sound in order to evaluate the magnitude of the knocking sound.