Multi-Source Data Fusion with Deep Neural Networks for Safety Behavior Recognition in Power Plant Environments

Abstract

A neural network-based method for identifying safety behaviors of power plant personnel is presented, integrating optical-flow, RGB, and environmental sensor data through a CNN–LSTM architecture. The network includes three convolutional blocks with 64 and 128 filters (kernel size 3×3) using ReLU and batch normalization, followed by a two-layer LSTM with 128 and 64 hidden units and a softmax classifier for final prediction. To enhance generalization, dropout of 0.5 is applied, and training is optimized using Adam with a learning rate of 0.001 and batch size of 64.The dataset consists of 5,000 labeled samples collected from multiple areas of a power plant, divided into 70%training,15%validation,and 15%testing.Experimental evaluation demonstrates consistent superiority over traditional approaches, with accuracy of 0.923,recall of 0.913,F1 score of 0.918,and precision of 0.928,while achieving an average accuracy of 0.928 across different scenarios such as high temperature, humidity, and explosive environments. These results confirm that the proposed method can provide accurate and reliable recognition of safety behaviors, offering practical support for improving personnel management and risk prevention in modern power plants.