Study On Multiple Information Fault Diagnosis And State Trend Forecasting For Hydroelectric Generator Unit

With the steady development of renewable green and clean energy such as wind,solar and tidal energy in China,the capacity of conventional hydropower units and pumped storage unit has also increased rapidly.In the same time,hydropower will take on more tasks about peak regulation and frequency modulation in sate grid to weaken the impact when renewable energy connect to the grid.In order to guarantee the safe and stable operation of power grid,the reliable operation and health management of hydroelectric generating set are facing higher requirements.As the key equipment of hydropower station,hydroelectric generating set consists coupled components,which show a trend of complexity and integration.Traditional fault diagnosis methods of hydroelectric generating set has shown their weakness in practical engineering.Therefore,to ensure the stable and safe operation of hydroelectric units,this paper aims at the scientific issue in fault diagnosis and trend prediction of hydroelectric generators,analyzes the limitations of existing theories and research methods and proposes a method for correlative analysis of operation parameters of hydropower units which takes multi-source information fusion of hydropower units as the breakthrough point.Based on the correlative analysis method,a fault diagnosis system for hydropower units with high accuracy and robustness under supervised and unsupervised conditions is constructed combined with advanced technological means such as machine learning and adversarial learning.In addition,this paper proposes a multi-step non-linear trend forecasting method through signal processing and deep learning,which can accurately predict the sate change trend of hydropower units.The main research contents and innovation achievement are as follows:(1)In order to effectively utilize the massive operation monitoring data of hydropower units and mining association relationships of them,this paper proposes the operating parameters relevant analysis method of hydroelectric units through data discretization and data mining.In consideration of that there are some outlier when fault occurs,K-Mediods is utilized to discrete operating parameters data.Through comparing silhouette coefficient,the number of clustering is optimized.The fact physical meaning of each dispersion interval are also given in this paper.Through collecting the fault operating data of hydropower station,a discrete transaction set is constructed.FP-growth is utilized to mining the frequent itemsets and association relationships.Finally we provide some practical guidance suggestions for operation and maintenance personnel of hydropower station.(2)The results of units parameters association relationships indicate the potential fault characteristics among different parameters.Therefore,this paper proposed a fault diagnosis method based on GRU-NP-DAE through recurrent neural networks.Traditional fault diagnosis methods ignore the time-series relationship of units vibration and are unable to extract the correlation features of different sensors.Through recurrent neural network,GRU-NP-DAE can storage the information of time-series data and take multiple signal as input to constitute different vibration modes.The diagnosis results are determined by comparing the reconstruction error of different modes.In addition,the denoising auto-encoder and variable input length are introduced to improve the robustness of method.Classic rotating machinery datasets and practical hydropower operating data are employed to testify the effectiveness of proposed diagnosis method.The experiment results indicate that the proposed method achieves high classification accuracy with strong robustness.(3)In the application of fault diagnosis method,the lack of fault labels leads to the failure of supervised mode training.In order to break through the dependence of labeled data,this paper proposed an unsupervised fault diagnosis method for hydropower units based on generative adversarial networks.The operating hydropower units data in high dimension space are reduced into low dimension feature space through auto-encoder.Then adversarial training are employed to imposed priori distribution into the feature space.The loss functions of classifier in unsupervised learning are discussed to constitute Cat AAE model.The latent coding space of rotate bearing fault data show the effectiveness of proposed data to cluster data in different categories.Finally,the experiments of hydropower operating data indicates that the proposed Cat AAE can classify them with 100% accuracy which verifies the engineering value.(4)Simple fault diagnosis method for hydropower units can not satisfy the pre-maintenance strategy of units.The trend of units status in fault evolution process should be analyzed in detail.Although existing trend prediction methods which belong to single-step forecasting can predict the next step of units status,it is hard for them to accurately predict the trend of unit state change over a long period of time.In order to realize the condition-based maintenance of the unit and discover the early fault symptoms of the unit,this paper proposed a multiple-step non-linear forecasting method for hydropower units vibration data based on VMD and CNN.This approach decomposes the strong non-linear and non-stationary vibration signals of hydroelectric units into a series of intrinsic mode functions,which are considered as each channel of input in CNN.The relationships of different intrinsic mode functions and local features are extracted through convolution kernel.The multiple forecasting result is directly output through architecture of CNN.The vibration data from coping of hydropower unit are employed and the forecasting results show the low fitting error and high correlation coefficient of proposed method.(5)Based on the above theoretical achievements,this paper designs and developed a service-oriented multi-source information mining and fault diagnosis system for hydropower units.By fusing multi-source heterogeneous data of hydropower units,a unified knowledge management platform for large data hydropower units is constructed.The prior knowledge can be updated or supplemented in time according to the operation data and inspection reports of power stations.Functional modules such as unit association analysis,fault diagnosis,fault warning,trend prediction and condition assessment are realized.The system has been successfully applied to Hu Bei Bailianhe pumped storage power station to provide maintenance guidance and decision-making suggestions for power station operators.