Персона: Трофимов, Александр Геннадьевич
Email Address
Birth Date
Результаты поиска
Синтез нейросетевых структур для моделирования управляемых объектов с распределенными параметрами
A method of choosing a pre-trained convolutional neural network for transfer learning in image classification problems
© Springer Nature Switzerland AG 2020.A method of choosing a pre-trained convolutional neural network (CNN) for transfer learning on the new image classification problem is proposed. The method can be used for quick estimation of which of the CNNs trained on the ImageNet dataset images (AlexNet, VGG16, VGG19, GoogLeNet, etc.) will be the most accurate after its fine tuning on the new sample of images. It is shown that there is high correlation (ρ ≈ 0.74, p < 0.01) between the characteristics of the features obtained at the output of the pre-trained CNN’s convolutional part and its accuracy on the test sample after fine tuning. The proposed method can be used to make recommendations for researchers who want to apply the pre-trained CNN and transfer learning approach to solve their own classification problems and don’t have sufficient computational resources and time for multiple fine tunings of available free CNNs with consequent choosing the best one.
Abnormal operation detection in heat power plant using ensemble of binary classifiers
© Springer Nature Switzerland AG 2019. The problem of abnormal operation detection is considered for prediction of malfunctions appearance and their progress in the equipment of power plant. Abnormal operation detection method based on multivariate state estimation technique (MSET) along with machine learning algorithms is proposed. The ensemble of linear regression models is used for feature construction. The ensembles of binary classifiers (logistic regressions) together with the multilayer neural network are used for the abnormal operation index calculation based on the constructed features. The method was applied to abnormal operation detection in turbo feed pump (TFP 1100-350-17-4) at Kashirskaya heat power plant (Moscow region, Kashira). It is shown that the abnormal operation index of the pump starts to increase a few days before accidents appear and stays close to zero during the normal operation periods. The obtained results demonstrate that the developed model can be used to detect and predict operation anomalies in the power plant equipment.
Predictive Model for Calculating Abnormal Functioning Power Equipment
© 2020, Springer Nature Switzerland AG.A method of early detection of defects in technological equipment of energy facilities is proposed. A brief analysis of the Russian market of cyber-physical industrial equipment monitoring systems was carried out. Special attention is paid to the problems of preparing initial data for training a model, in particular, the problem of obtaining adequate data on accidents that have occurred. A mathematical problem is formulated for modeling the anomaly index, which takes values from 0 (normal operation) to 1 (high probability of an accident). The model is based on well-known statistical methods. A method for dividing the periods of operation of technological equipment into “normal” and “anomalous” is proposed. The method of binary classification AUC ROC allows you to limit the number of signs involved in the formation of the anomaly indicator, signs that have a good “separation” ability. Using the Spearman’s rank correlation criterion, signs are selected that are most sensitive to the development of process equipment malfunctions. As an anomalous indicator, it is proposed to consider the ratio of the densities of distribution of the final signs, estimated in the anomalous and normal areas of operation of the process equipment. A method is proposed for generating an alarm for detecting the anomalous operation of the technological equipment of power units. It is shown that the proposed model made it possible to identify the beginning of the development of an emergency, while individual measurements did not detect any features of the operation of equipment of energy facilities in the pre-emergency time interval.
Model for Early Detection of Emergency Conditions in Power Plant Equipment Based on Machine Learning Methods
© 2019, Pleiades Publishing, Inc. Abstract—: The article discusses a method for early detection and prediction of abnormality in operation of power-unit process equipment taking as an example the PTN 1100-350-17-4 turbine driven feedwater pump of a 300 MW power unit. The importance of the problem of predicting possible process equipment malfunctions at an early state of their occurrence is determined, and the specific features of solving it in the power industry are explained. The range of process equipment defects that can be efficiently detected using the predictive analytics methods is outlined. The fundamental assertion stating that the scope of analog and discrete measurements available in the process control system’s set of computerized automation tools is sufficient for applying the predictive analytics methods is emphasized. Modern predictive analytics methods are briefly reviewed, and the specific features of model training algorithms are mentioned. Separate attention is paid to the problems of preparing initial data for training the model. The mathematical problem of modeling an abnormality indicator taking the values from 0 (normal operation) to 1 (abnormal operation) is formulated. In turn, this problem is formulated as the binary classification problem of attribute vectors characterizing the equipment state at the given moment of time. An original approach is suggested, which combines the multivariate state estimation technique (MSET), in which the degree of abnormality in a technical state is determined from the extent to which the Hotelling criterion exceeds a threshold level (which is automatically calculated in the algorithm), and machine learning methods, the use of which makes it possible to overcome a number of difficulties inherent in the MSET. For solving the problem of determining the composition of the most informative attributes from the values of which early development of an emergency can be detected, it is proposed to use an ensemble of regression models. A method for selecting the modeled variable and the set of regressors is substantiated. An abnormality indicator calculation method based on composing an ensemble of linear regression models is proposed, and the advantage of using an ensemble over a single classifier is shown. A method for producing an alarm in response to detected abnormality in the operation of power unit process equipment is suggested. It is shown that it became possible by using the proposed model to detect the onset of the emergency development process, whereas individual indicators failed to reveal pump operation singularities in the preemergency interval of time.