Prediction and intelligent decision-making for smart lamp poles: a DSADN framework based on multimodal enhanced machine learning with physical constraints

Abstract

This study addresses the challenge of fault prediction in smart lamp post systems. It proposes a fault-prediction and intelligent decision-making framework (DSADN) based on an improved machine-learning algorithm. By integrating a support vector machine (SVM) optimized with a kernel function, a long short-term memory (LSTM) enhanced with attention, a graph neural network, and a Transformer architecture, a multimodal feature extraction channel is constructed. By integrating the physical-constraint mechanism for domain knowledge embedding with the three-stage adaptive training strategy, the model’s robustness in complex environments is significantly enhanced. Experiments show that the system achieves an LED module attenuation prediction accuracy of 96.3% on 326,784 multi-source data samples, with an F1 score of 0.932. The performance degradation rate in extreme environments is only 7.6%, which is more than 35% lower than that of traditional models. Edge deployment achieves a 15 ms real-time response and reduces memory usage by 62%. The research results address three major problems of the existing smart light pole system: low detection efficiency, high false alarm rates during monitoring, and prediction lag. It also provides an efficient solution for the predictive maintenance of smart city infrastructure.