Research On Low Frequency Vibration Fault Diagnosis Of 1000MW Thermal Power Unit Based On Blind Source Separation And Deep Neural Network

China is in a critical period of energy structure transformation.Large-scale new energy access to the power grid brings large power and frequency fluctuations,resulting in 1000 MW thermal power units need to be absorbed by peak-load regulation.Deep peak-load regulation will cause large fluctuation of load,which is easy to cause low-frequency vibration fault of 1000 MW thermal power unit.Vibration coupling is easy to occur between low-frequency vibrations,which makes the vibration characteristics pollute each other.In the big data environment,the vibration data of 1000 MW thermal power unit reaches 100000 levels in one day.There is a problem of sample imbalance,that is,the normal sample is much larger than the fault sample.The traditional vibration fault diagnosis method is greatly affected by feature pollution,and the feature mining ability of large data samples is insufficient.There is a short board in the low-frequency vibration fault diagnosis of 1000 MW thermal power unit.From the perspective of low-frequency vibration separation and thermal power big data vibration samples,this paper studies the low-frequency vibration fault diagnosis of 1000 MW thermal power unit based on blind source separation and deep neural network.The blind source separation method is used to realize the separation and location of low-frequency vibration,and the improved deep neural network is used to improve the diagnosis accuracy of low-frequency vibration fault of 1000 MW thermal power unit in the environment of big data.Aiming at the problem of large data sample imbalance,the data augmentation method based on generative adversarial network is used to reduce the impact of sample imbalance on accuracy and stability and improves the accuracy of specific low-frequency vibration faults.The fault diagnosis method proposed in this paper has good engineering practical significance in improving the safety and economy of 1000 MW thermal power unit.Firstly,combined with engineering practice and vibration theory,the low-frequency vibration is located in the actual fault sources such as oil film oscillation,steam excitation,bearing seat looseness,dynamic and static rub impact and so on.It is found that the vibration has the characteristics of blind source,and there is vibration aliasing between low-frequency vibration faults;Secondly,the improved blind source separation method based on maximum approximate negative entropy is used to study the coupled low-frequency vibration.The blind source separation method is used for vibration unmixing and separation.This method reduces the feature pollution,and the similarity of the separated signals is more than 98%;Then,the deep convolution neural network is used to study the fault diagnosis of low-frequency vibration.The deep network is not clear enough to diagnose the coupled low-frequency vibration.The traditional deep convolution neural network is not clear enough to diagnose low-frequency vibration,and there is model degradation.After using the residual network and blind source separation,the degradation of the model is improved,the classification results of vibration sources are clear,and the accuracy is more than 95%;Finally,the specific low-frequency vibration fault is studied by data amplification.The traditional data augmentation method is easy to over fit or not converge.Therefore,an improved generative adversarial network model based on conditional convolution and Laplace pyramid model is proposed.The convergence time is reduced by 50%,and the diagnosis accuracy is improved from 75% to more than 95%.This method has passed the test of many evaluation indexes and performed well in engineering application,which shows that this method has strong anti over fitting ability and engineering application potential.The method proposed in this paper effectively completes the coupling and separation of lowfrequency vibration faults of 1000 MW thermal power unit.The diagnosis accuracy of lowfrequency vibration faults is high,and the location of vibration sources and fault causes is clear.In the environment of big data and unbalanced samples,this method improves the diagnosis accuracy and convergence speed of big data vibration fault samples and strengthens the diagnosis ability of specific low-frequency vibration faults.The vibration fault diagnosis method proposed in this paper provides a new idea of taking the separation of vibration sources as the object of fault diagnosis and puts forward a new method of data expansion using generative adversarial network.This method has reference value and engineering significance for enhancing the safety and stability of 1000 MW thermal power unit.