Online prognostic failure AIoT system for industrial generators maintenance service based two-stage deep learning algorithm

Abstract

Intelligent anomaly diagnosis for industrial generators is essential in providing appropriate maintenance service, which makes it challenging to identify machine failures due to a complicated operational environment. For these reasons, an AIoT framework for anomaly diagnosis of industrial 125kW/250 kW generators is developed to provide indicators in maintenance services based on a two-stage deep learning convolution neural network and gate recurrent unit (CNN-GRU). In the proposed AIoT system, the IoT module collects different working features of 125kW/250 kW diesel generators in the experimental setup, including three-phase current, frequency, vibration, three-phase voltage, engine temperature, starting battery DC voltage, and power factor to generate labeled anomaly conditioning representative data. The convolution neural network is firstly deployed to reduce the dimensionality of 2D historical data, and then all the extracted valuable features are transferred to the gate recurrent unit to process sequential information. The developed algorithm was evaluated with different deep learning techniques, including the recurrent neural network (RNN), GRU, CNN, and long short-term memory (LSTM) by various benchmarks and data sequential horizons. Experiments prove that the developed CNN-GRU contains superior diagnosis capability and improved accuracy compared to other state-of-the-art deep learning models in a 10-second sample frequency dataset.