Fault Prediction Of Thermal Generator Set Based On Sequential Machine Learning Approach

In the future,thermal power generation will still be the main form of electricity production in China.The failure of large thermal power units will cause serious adverse effects.Due to the complex system structure and multiple heterogeneous parameters of large thermal power units,it is difficult for traditional fault prediction methods to accurately identify abnormal conditions,diagnose abnormal causes and predict abnormal evolution trend.At present,large thermal power generating units have preliminary acquired the ability of full coverage of the condition measurement data,which contains a lot of information related to the health condition of large thermal power generating units.In this paper,the machine learning approaches based on sequence features are employed,the typical unplanned shutdown process and important parts of the thermal generator set bearing are taken as the main research objects,and fault prediction approaches suitable for large thermal power units are proposed.The specific contents and main contributions of this paper are as follows:· A fault prediction approach based on similarity discrimination of uni-variate degradation trend is proposed to solve the critical uni-variate monitoring problem in large thermal power units.Firstly,according to the existing historical degradation cases,the temporal degradation trends of different degradation stages are extracted by using time window technology.In the online test phase,the dynamic time warping algorithm is employed to measure the similarity between the change trend of parameters in the current time period and the historical degradation trend,and the degree of degradation in the current time period is measured according to the most matched degradation trend.The effectiveness of the proposed approach in fault prediction is proved by case analysis.· A fault prediction approach based on Gate Recurrent Unit(GRU)is proposed to solve the problem of multi-variable monitoring of large thermal generating units.Firstly,GRU is employed to extract high-dimensional nonlinear sequence features between multiple variables.In order to enhance the ability of sequence information extraction and avoid the omission of sequence information,a non-parametric attention mechanism based on dynamic time warping is proposed.Secondly,in order to quantify the uncertainty in the process of equipment degradation,the weight of traditional network is converted from a specific value to a variable subject to a specific distribution by using variational inference algorithm.Through case study,the proposed approach can enhance the extraction ability of sequence information from traditional GRU network,and measure the uncertainty of fault prediction.· In order to improve fault prediction performance,transfer machine learning training is carried out by using the degradation trend of the current monitoring equipment itself,but the degradation trend of the current monitoring equipment can not be completely obtained.To solve this problem,a fault prediction approach based on transfer learning framework is proposed.Firstly,the original network is obtained through training based on the existing historical full life cycle degradation cases.Then,according to the current monitoring data,the particle filter algorithm is employed to generate degradation data based on the current degradation trend.Finally,the generated data is used for transfer learning.Through case analysis,the proposed approach can effectively improve the fault prediction performance;· The preventive maintenance strategy of large thermal power generating units leads to fewer degradation cases,which has a great impact on the training process of machine learning approach.Therefore,a fault prediction approach based on incremental learning framework is proposed.Firstly,the original network model is obtained by using owned degenerate cases,and a new sequence input vector is constructed.Later,when a new degradation case is obtained,the original model is updated and the weight update direction is constrained to avoid catastrophic forgetting problems.Through case study,this method can effectively carry out incremental learning function.