This study presents a novel method based on the convolutional neural network to evaluate knock probability. In this way, lots of data sets are extracted from the real driving conditions of a turbocharged gasoline vehicle. The trained network shows 92 % accuracy in knock detection and 87 % accuracy in normal combustion recognition. This trained network is then used to forecast the knocking probability of the vehicle in different operating conditions. The results revealed that the exhaust gas temperature sensor considerably affects the knock recognition precision by 3 %. Most of the network errors in knock probability determination would be associated with the full load-high speed region and the network errors in normal combustion detection would be related to the low-end torque zone. It is also found that the knock probability decreases with the vehicle's speed increase and its probability at the constant vehicle speed would be higher at low engine rotational speed. This new approach can predict engine knock zones for different vehicle parameters and can be implemented by the engine control unit to suppress knock before happening in an active manner rather than the current passive method.