Anomaly Detection Of Power Plant Equipment Operating Parameters Based On Deep Learning

With the development of society,different kinds of high technologies and products continue to appear,be civilianized and popularized.People’s lives become more and more convenient and the key to supporting all this is electrical energy.In order to ensure the safe and reliable operation of power plant equipment,it’s important to research the anomaly detection technology of equipment operating parameters.Anomaly detection of power plant equipment operating parameters,that is,using the operating parameters of the unit equipment under normal operating conditions to establish a model,requires that the model can identify abnormalities which occur during productive process.And the model is supposed to detect equipment deterioration trend as early as possible,and provide a reliable basis for failure diagnosis.First,a Deep Autoencoding Gaussian Mixture Model with Reinforced Feature(DAGMM-RF)was proposed.This model has designed a new loss function,which effectively reduces the linear correlation between the dimensions of the compressed feature and helps the compression network extracting more expressive features.Further,a method for calculating the “lowest compressed feature dimension” was proposed.This dimension value could provide a reference for setting the appropriate compressed feature dimension and avoid spending a lot of effort to determine the dimension,which is more conducive to the practical application of the model.The experimental results show that,DAGMM-RF has better performance than Deep Autoencoding Gaussian Mixture Model(DAGMM)and the “lowest compressed feature dimension” is helpful to avoided bad condition.DAGMM-RF doesn’t consider the temporal feature of power plant equipment parameters.Aim at this problem,Isomerism Time-Auto-Encoder Gaussian Mixture Model(ITAEGMM)was proposed.ITAEGMM uses the recurrent neural network toconstruct the encoder of the Auto-encoder,and uses the fully connected layers to construct the decoder.This reduces the training time cost of the model while ensuring the network decoding ability.Then Streaming Peaks-over-Threshold(SPOT)algorithm and time delay method are combined to filter abnormal data.Experiment shows that the introduction of temporal feature brings significant performance improvements and SPOT+delay method helps ITAEGMM get better F1 Score.