Abstract:Abstract Prognostics and Health Management (PHM) Systems have been widely applied in online monitoring of modern mechanical devices, and models in PHM systems are of vital importance. Traditional mechanism models are able to simulate parameters and evaluate the overall machine state accurately, but their defect is the high requirements in understanding the objects. Therefore, data driven models could be a suitable substitute. Data driven models take the advantage of machine learning technique and are able to be established only on the basis of past data. Models applied in this paper are based on Artificial Neural Network (ANN), but the complexity of network structures will occupy huge amount of computing resources and time in training. For online monitoring, spending too much time on training will postpone real time prediction and decrease the reaction speed of PHM systems in abnormal conditions. In this paper distributed training based on computer cluster is proposed to decrease training time. By dispatching computing task into several workers, the severs in the cluster will only undertake a small fraction of the total computing load and therefore accelerate the overall training process. On the other hand, there are some risks of losing prediction accuracy and stability because of the nonlinear gradient deviation in data parallelism. Aggregation period (AP) is an important factor in balancing the requirements of both ends. This paper analyzes the influence of AP on training speed, accuracy and stability, then proposes a novel distributed training algorism according to the regulations achieved. Then a distributed training and online monitoring process for a typical two-shaft gas turbine is taken as an example in the result and discussion section. It turns out that the experimental results fit the theoretical regulations well, and the revised distributed learning algorism is able to meet all the requirements of training speed, prediction accuracy and stability.

ON-LINE TRAINING METHOD OF ANN IN DMS

Non-linear Model Predictive Control Based on Neural Network Model with Modified Differential Evolution Adapting Weights

Direct Prediction of Steam Cracking Products from Naphtha Bulk Properties: Application of the Two Sub-Networks ANN

On-line soft sensing for end point of naphtha based on neural network

A New Method For On-Line Demand Calibration Of Wds Hydraulic Model

Development and Application of Online Raman Analyzer for Methanol Gasoline

Neural network intelligent system for the on-line optimization in chemical plants

On-line Data Rectification for Chemical Process Measurements Based on Artificial Neural Networks

Dynamic monitoring system for FCCU product qualities

CFD and Neural Network-Based Predicting for Oil Delivery Performance of Oil Pump in Engines

An On-Line Amending Optimal Control Method Based on Ann

An Online Fault Diagnosis Strategy for Full Operating Cycles of Chemical Processes

APPLICATION OF RPROP ANN BASED FAULT DIAGNOSIS MODEL FOR TRANSFORMER DISSOLVED GAS-IN-OIL ANALYSIS

An Online Wear State Identification Method for Axial Piston Pump Key Friction Pair Based on FSANN

Distributed Training for Data Driven Models in Power Machinery Online Monitoring

On-Line Estimation of Vapour Pressure of Stabilized Gasoline Via Ann'S

Online DGA Sensor Calibration Using GANN and Data Augmentation

An Improved Artificial Neural Network Method for Monitoring and Diagnosis of Engineering Processes With Applications

Research on an ANN System for Monitoring Hydrostatic Turntable Performance Based on ODNE Training

Dynamic Liquid Level Prediction in Oil Wells During Oil Extraction Based on WOA-AM-LSTM-ANN Model Using Dynamic and Static Information

A DSS for Oil Spills Based on ANN