As a typical reciprocating engine power machinery, complex structure determines its failure brings about the complexity and diversity, it shows the uncertainties of operating environment, system noise and sensor accuracy, and engine fault diagnosis accuracy rate is reduced, taking into account the limitations of traditional BP neural networks, improved BP algorithms include statistical algorithms, additional momentum method, variable learning rate method and conjugate gradient method are studied. Finally, the engine is as an example, engine fault diagnosis experimental system is set, the vibration signals are measured in the normal state, left one and right six cylinders off the oil and air filter blockage in the load of 2565Nm, and the speed of 1500r/min, 1800r/min and 2200r/min. The test and analysis by comparing above mentioned methods indicate it is verified the superiority improved BP neural network with the conjugate gradient method.