【作者】 王琼；

【导师】 任伟建；

【作者基本信息】 东北石油大学 ， 油气信息与控制工程， 2011， 博士

【摘要】 抽油机故障诊断的关键是实现从故障征兆空间到故障空间的映射,从而实现对故障的识别和诊断,它是一个复杂的非线性问题。神经网络的自学习能力、非线性映射能力、对任意函数的逼近能力、并行计算能力和容错能力等为故障诊断提供了有力手段。但是由于实际生产设备工况复杂,故障类型种类繁多,致使在诊断时出现网络规模庞大、学习训练时间超长、易于陷入局部最小点等问题,降低了神经网络的实用性。神经网络与遗传算法、进化机制等结合,形成计算智能,将成为神经网络用于故障诊断的趋势。本文就是将神经网络与其它优化算法结合,改善神经网络性能,从而用于抽油机的故障诊断。主要研究内容如下：1、设计一个适合故障诊断的双权连接可拓神经网络,该网络以抽油机的状态数据为输入层,以故障类型为输出层,输入输出采用双权连接,权值分别为故障的上限数据和下限数据。提出一种能够自适应改变交叉率和变异率的自适应遗传算法,以可拓距离为评价函数,利用遗传算法的全局搜索能力,对建立的可拓神经网络的权值进行优化,克服BP算法的训练神经网络收敛性差,容易陷入局部极值的缺点。2、设计一种免疫遗传RBF神经网络。对免疫遗传算法进行改进,给出一种基于抗体矢量距离的亲和度计算方法,在抗体的促进和抑制环节增加基于密度的调节因子,保留优秀抗体,保证抗体的多样性,避免未成熟收敛现象。用改进的免疫遗传算法优化RBF神经网络的隐层中心,提高其逼近精度,克服传统算法需要预先指定隐含层节点数或者通过大量实验获得节点数、学习效率差的缺点。3、设计一种基于粒子群优化的神经网路。根据传统粒子群算法在训练后期容易陷入局部极值的缺点,对基本粒子群算法的速度方程进行更新,在现有的速度更新机制上加入非常小的扰动项,并动态调整加速系数,使算法能够分别调整进化初期和后期的性能。采用粒子群优化算法对神经网络的权值和阈值进行优化。4、提出一种新的故障诊断融合方式,并将以上神经网络组成功能相容并具有选择机制的软件包,利用抽油机无线巡检数据实现智能在线故障诊断,并对诊断结果进行分析比较。更多还原

【Abstract】 The key of pump-jack fault diagnosis is realizing the mapping from fault symptom space to fault space.It is a complex nonlinear problem and it can realize the identification and diagnosis of the fault.Neural network offer the way for fault diagnosis,because it have the ability of self-learning,nonlinear mapping,arbitrary function approximation,parallel computing and fault-tolerant. But because of complex working condition and various fault type,so when diagnose, existing problems such as network vast,long learning time and falling into local extreme value and so on.Neural network combining with genetic algorithm,evolutionism and other algorithm will become the trend of the fault diagnosis.In this paper, neural network and optimization algorithm are combined to improve the properties of neural network,so then used for pump-jack fault diagnosis.The details are follows:1、Design a double weights extension neural network for fault diagnosis.The inputs of the network are status data of the pump-jack and the outputs of the network are fault type. Inputs and outputs are connected by double weights,the weights are the upper and lower limit of the status data. Put forward adaptive ,it can change the crossover and mutation adaptively and it’s evaluation function is extenics distance.Making use of the genetic algorithm’s global search ability to optimize the neural network weights. Overcame the BP’s disadvantage of premature convergence and falling into local extremum.2、Design an immune genetic neural network. Improve the immune genetic algorithm and give a calculational methods of affinity degree based on antibody vector distance. An adjusting factor based on density is increased in the process of promoting and suppressing of antibody. Thus, the best individual can be preserved, the diversity can be ensured, and the phenomenon of premature convergence can be avoided. Optimize the hide centers of RBF network by using the improved genetic immune algorithm. The characteristic of low learning efficiency of RBF neural network can be overcome.The approximation accuracy can be also improved and the number of constructing the center of the hide layer of network is dispensable.3、Design a particle-swarm-optimization-based neural network. The traditional PSO algorithm is easy to fall into local optimum situation in the later stage.In this paper the rate equation of the PSO is updated by introducing a minuscule disturbing term and a dynamic changing acceleration factor.The new PSO can adjust the PSO’s performance in the earlier and later stage respectively.The new PSO algorithm is used to train the weights and thresholds of the BP network.4、Give a new fusion diagnosis method. Make a software package,it synthesize the neural netwok in the paper. Using wireless patrol data realize the fault diagnosis of pump-jack,analyze and contrast the result.更多还原