基于PSO-Nadam-BP神经网络的多传感器火灾检测算法  

Multi-sensor Fire Detection Algorithm Based on PSO-Nadam-BP Neural Network

作　　者：孙苗 张雪凡[1] 林阳 何理旭 Sun Miao

机构地区：[1]上海大学通信与信息工程学院特种光纤与光接入网重点实验室,上海200072

出　　处：《工业控制计算机》2022年第6期1-3,7,共4页Industrial Control Computer

基　　金：国家自然科学基金(61871261)项目资助。

摘　　要：由于采用传统单传感器检测火灾时存在精度低、误报率高等不足,提出了一种基于PSO-Nadam算法优化BP神经网络的多传感器火灾检测算法。算法模型以火灾环境参数作为BP神经网络的输入,以不同火灾状态的概率作为输出。但是传统BP神经网络算法中存在两个问题,一是因初始权值和阈值随机选定而导致的可能陷入局部最优解,二是因反向误差传播过程中学习率α固定所导致的收敛速度慢。因此分别采用粒子群算法(PSO)和Nadam算法来优化BP神经网络模型。最后在MATLAB软件上进行算法仿真模拟,相较于BP神经网络模型,该优化模型输出结果与样本实际值吻合度更高,样本预测均方根误差更小,并且收敛速度更快。该算法在一定程度上提高了检测准确率和检测速度,为后续更完善的火灾检测算法打下基础。Due to the shortcomings of low precision and high false alarm rate when using traditional single sensor to detect fire,this paper proposes a multi-sensor fire detection algorithm based on PSO-Nadam algorithm to optimize BP neural network.The model takes the fire environment parameters as the input of BP neural network and the probability of different fire states as the output.However,there are two problems in the traditional BP neural network algorithm,one is that the initial weights and thresholds are randomly selected,and the other is that the convergence speed is slow due to the fixed learning rateαin the process of backward error propagation.Therefore,this paper adopts particle swarm optimization(PSO)and Nadam algorithm to optimize the BP neural network model.Finally,the algorithm simulation is carried out in MATLAB software.Compared with BP neural network model,the output result of the optimized model in this paper is in higher agreement with the actual sample value,the root mean square error of sample prediction is smaller,and the convergence speed is faster.

关 键 词：多传感器 火灾检测 BP神经网络 粒子群算法 Nadam算法 

分 类 号：X932[环境科学与工程—安全科学] TP183[自动化与计算机技术—控制理论与控制工程] TP212[自动化与计算机技术—控制科学与工程]