Hierarchical Deep Neural Network for Fire Detection in Wind Turbine Nacelles

As global demand for renewable energy continues to rise, the number of installed wind turbines is rapidly increasing, leading to a higher incidence of fire accidents in wind turbines. However, conventional fire-detection methods, such as smoke and flame detectors, suffer from low detection accuracy and long response times. To address these limitations, several fire detection methods based on artificial intelligence have been proposed. However, these approaches often rely on object-detection neural networks, that result in high false-alarm rates for pseudo-fire images. In this study, a hierarchical deep neural network for fire-condition monitoring is proposed, which reduces the false alarm rate and accurately identifies the locations of smoke and fire. The proposed neural network initially employs a fire-classification neural network to classify situations into three categories: fire, smoke or normal. By analyzing the overall image information, false alarms are effectively reduced. Based on the classification results, an object-detection neural network specializing in smoke and fire detection is then activated to identify their locations. The inferred information on fire locations can be utilized to operate autonomous targeted fire-suppression systems. Fire image sets are constructed to train and validate the proposed neural network. The performance of the proposed neural network is verified by comparing its classification accuracy, object-detection accuracy, and false-alarm rates with those of other neural networks. The proposed approach can be applicable to the nacelles of wind turbines and also various industrial environments requiring condition monitoring.