基于图像和改进BP算法的火焰燃烧稳定性判别方法研究

发布时间：2018-05-27 07:44

本文选题：火焰图像 + 改进BP算法　；参考：《合肥工业大学》2017年硕士论文

【摘要】：保证煤炭稳定燃烧是燃煤电站安全运营最根本的要求,炉膛火焰是表征燃烧状态是否稳定的最直接反映,为实现锅炉燃烧状态的自动监测,对燃烧的稳定程度进行量化判定,论文基于火焰图像处理技术,从炉腔火焰图像中提取燃烧参数,建立燃烧参数数据库,为建立的模型提供训练和测试样本,本文提出了两种神经网络模型。针对BP算法抗干扰能力差、学习速率慢且易陷入局部极小点等适用性,本文提出一种基于分层动态调节不同学习速率的BP算法。该算法综合考虑网络训练方式和学习率的不足,改进了隐含层的传递函数,设计新的复合误差函数,同时采用分层动态调整不同学习率的方法以加快传统BP算法的收敛速度,同时对造成陷入局部极小的异常样本进行修复,最后将获取的样本参数分为训练样本和测试样本对所建立的模型进行训练和测试。实验结果表明,所建立模型具有较好的容错能力和映射能力,改进后的算法提高了稳定性的识别率,运算速度较快,可以满足稳定性判定的实时性要求。另外,本文基于多属性区间决策理论得到燃烧决策库,在模糊推理中,为了获取隶属度函数参数和模糊推理规则,提出基于粗糙集简化样本决策库,实现属性简约和属性值简约,增加了网络训练样本参数的可靠性。再结合模糊网络的逻辑推理性和神经网络的学习性、并行计算等优点,提出建立T-S模糊神经网络模型用于燃烧诊断,选择合适的模糊分割数,定义“五四模型”,建立基于“五四模型”的火焰燃烧稳定性判定模型进行仿真。对比训练前后的仿真图参数可得,该模型是可行的,并具有较好的实验效果。对比两种神经网络模型仿真之后的性能参数,由于改进BP神经网络模型的精度性能胜于“五四模型”,因而得出结论,改进BP神经网络模型更适合应用于火焰燃烧稳定性判定的研究。
[Abstract]:To ensure the stable combustion of coal is the most fundamental requirement for the safe operation of coal-fired power stations. The furnace flame is the most direct reflection of whether the combustion state is stable or not. In order to realize the automatic monitoring of the combustion state of the boiler, the degree of combustion stability is determined quantitatively. Based on the flame image processing technology, the combustion parameters are extracted from the flame image of the furnace cavity, and the combustion parameter database is established, which provides training and test samples for the established model. In this paper, two neural network models are proposed. Aiming at the applicability of BP algorithm, such as poor anti-interference ability, slow learning rate and easy to fall into local minimum, this paper proposes a BP algorithm based on layered dynamic adjustment of different learning rates. Considering the deficiency of network training method and learning rate, the algorithm improves the transfer function of hidden layer, designs a new compound error function, and adopts the method of dynamic adjustment of learning rate to speed up the convergence of traditional BP algorithm. At the same time, the abnormal samples which are trapped in local minima are repaired. Finally, the obtained sample parameters are divided into training samples and test samples to train and test the established model. The experimental results show that the proposed model has better fault tolerance and mapping ability, and the improved algorithm improves the recognition rate of stability, and the operation speed is faster, which can meet the real-time requirements of stability determination. In addition, based on the multi-attribute interval decision theory, the combustion decision base is obtained. In fuzzy reasoning, in order to obtain membership function parameters and fuzzy inference rules, a simplified sample decision base based on rough set is proposed. Attribute reduction and attribute value reduction are realized, and the reliability of network training sample parameters is increased. Combining the advantages of logic reasoning of fuzzy network and learning and parallel computing of neural network, this paper proposes to establish T-S fuzzy neural network model for combustion diagnosis, select appropriate fuzzy partition number, and define "May 4th model". A flame combustion stability evaluation model based on the May 4th model was established for simulation. Compared with the parameters of simulation diagram before and after training, the model is feasible and has good experimental effect. Compared with the performance parameters of the two neural network models, the improved BP neural network model has better precision performance than the May 4th model, so it is concluded that the improved BP neural network model is better than the May 4th model. The improved BP neural network model is more suitable for the determination of flame combustion stability.
【学位授予单位】：合肥工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TM621

【相似文献】