The invention discloses a multi-source information-based grinding machine vibration and vibration sound characteristic extraction method which comprises the following steps: A, multi-source signals are collected, signals representing grinding machine load states are obtained via a data collection system, the collected signals are multi-source time domain signals and are denoted as S, and N samples of the multi-source time domain signals S are denoted as a formula shown in the description; B, shallow-layer characteristic information selection operation is conducted, and grinding machine vibration and vibration sound shallow characteristic information selection operation is performed based on non-stationary and non-linear signal self-adaptive decomposition and analysis technologies; C, depth characteristic knowledge is extracted, and grinding machine vibration/vibration sound multi-scale depth characteristic knowledge is extracted based on an optimal coupled depth neural network characteristic extraction and mutual information characteristic measurement algorithm. Via the method disclosed in the invention, support can be provided for building a grinding machine load state identification and grinding machine load parameter soft measurement model.