The invention provides a machine fault diagnosis method and device based on information geometry. The method comprises steps that a Riemann manifold is constructed according to a state probability density function of a machine and a Fisher information matrix; difference of the state probability density function of a present state of the machine and a state probability density function of the machine in a normal state is measured through utilizing a KL divergence of the Riemann manifold, and the difference value is acquired; whether the present state is in a fault state or the normal state can be confirmed through the difference value.