Wind Power Transmission Chain Prognostics And Health Management Based On Multi-deep Learning Model

The development of wind power is of great significance to the coordination of the ecological environment system and is also an important way to achieve the goal of carbon neutrality.At present,our country already has the world’s largest cumulative wind power installed capacity,and the scale of single-unit capacity is constantly increasing.At the same time,wind power equipment is facing problems such as high failure rate,high maintenance costs,and poor operational safety.The downtime due to failure accounts for more than 25% of the rated power generation time.The wind power transmission chain is the key object of condition monitoring and fault diagnosis.Fault prediction and health management of the transmission chain will be an important guarantee for the efficient operation of wind power equipment.With the construction of the wind power big data platform and the development of deep learning technology,learning from the data can reduce the reliance on professional knowledge in the health management process,and at the same time accurately and comprehensively grasp the health status of the key components of wind power,before the critical component fails.Provide early warning and treatment to ensure the safe operation of wind turbines.Based on this,this paper studies the health management of wind power transmission chain based on the multi-source depth model.The main contents are as follows:(1)The components of the wind turbine transmission simulation system and the corresponding signal acquisition system are introduced,and the failure modes of rolling bearing components and the fault sensitivity of vibration signals and acoustic emission signals(AE)are analyzed.(2)Based on the gradient boosting tree,a fault diagnosis and identification method for the transmission chain is proposed.Aiming at the problem of the traditional method’s low dimension of feature set and high dependence on manual experience,this paper semiautomatically extracts the feature of bearing vibration signal,and combines feature selection technology to reduce redundancy,and faults ten working states of bearing simulation Diagnosis and identification;At the same time,acoustic emission signals and model fusion technology are used to effectively classify shaft rubbing,bearing cracks and normal conditions.Experiments have proved that the gradient boosting tree can accurately identify failure modes based on vibration and acoustic emission signals.(3)Based on the generative countermeasure network,a method for abnormal detection and prediction of transmission chain faults is proposed.Aiming at the problem of the imbalance between normal samples and abnormal samples of the transmission chain and the complicated failure modes,the deep convolution auto-encoder and normal samples are used to train the generation of confrontation network,and the latent vector mapping of the auto-encoder is used to evaluate the type of samples and the failure development status.First,organize the vibration signal into a numerical matrix and convert it into a grayscale image;then initialize the generator and discriminator weights that generate the confrontation network,and use the confrontation loss,image loss and coding loss to punish the generator during the iterative training process until the game Balanced,the generator can generate normal samples that tend to be true;finally,the weight of the encoder in the generator is extracted,and the normal samples and abnormal samples are input at the same time for detection,and the fault development status of the abnormal samples is determined.The results show that for the various failure modes of the transmission chain,the method can detect the abnormal state and restore the failure development stage.(4)Aiming at the problem of data missing in remote data monitoring system due to sensor failure and network congestion,a method for repairing missing data based on time series feature mining and time series modeling is proposed.In the data preprocessing stage,the collection timeline of each unit in the wind farm is aggregated and aligned first,the characteristics of other units are constructed for the missing parameters,and the Word2 Vec feature of the wind condition of the wind farm is introduced.In the model construction stage,the time series data is modeled through the combination of GRU and LSTM models and Encoder-Decoder to fully mine data auto-correlation and cross-correlation information.The results show that the original data can be better restored based on the time series forecasting model and comprehensive utilization of the wind field global information.