The invention discloses a diagnostic method for a pump station unit based on a composite characteristic index and a depth limit learning machine and relates to the technical field of fault diagnosis of hydraulic machinery. The diagnostic method can directly decompose nonstationary time series from original vibrating signals by adopting self-adaptive iterative filtering, so that the nonstationary characteristic of the unit can be extracted effectively. A time domain statistics signal, a frequency domain statistics signal, an energy signal, a sample entropy signal and an arrangement entropy signal are extracted based on each nonstationary signal component. Characteristic learning is performed quickly and effectively by means of the depth limit learning machine, implicit fault information ofeach characteristic is extracted and boundedness of a manual design and characteristic extraction and a complex parameter adjusting process based on an artificial neural network are avoided, so that intelligent diagnosis of faults of a water pump unit is achieved, and therefore the accuracy and the effectiveness of fault diagnosis of the water pump unit are improved.