基于改进神经网络的抽油机故障智能诊断研究

Study on Intelligent Fault Diagnosis of Pumping Unit Based on Improved Neural Network

针对目前抽油机示功图故障诊断人工分析方法效率低,而智能方法计算量大且识别类型少的问题,提出了抽油机故障智能诊断新方法,通过改进的自组织特征映射网络(SOM神经网络)对抽油机示功图进行故障诊断。对比了蚁群算法(ACO)、遗传算法(GA)、粒子群算法(PSO)对SOM神经网络的改进效果,PSO算法的最优个体适应度值最小,且所需的迭代次数少,收敛速度快。同时对比了神经网络BP、LVQ、常规SOM和PSO-SOM诊断模型效果,PSO-SOM神经网络在识别时间、诊断准确率和诊断方差方面展现出了明显的优势。采用傅里叶描绘子和不变矩复合的形状识别方法提取抽油机样本示功图特征参数,利用PSO-SOM神经网络技术进行学习训练和仿真分析,实现了不同工况模式的分类以及待测示功图的工况诊断。研究表明:对于XX油田现场抽油机700井次的9种不同工况示功图,PSO-SOM方法有效识别正确率高达90.3%,是实现智能化高效诊断示功图的一种有效途径。

At present,the artificial analysis method for fault diagnosis of pumping unit indicator diagram is inefficient,while the intelligent method has a large amount of calculation and few types of faults can be recognized.In order to solve this problem,a new intelligent pumping unit fault diagnosis method is proposed,that is,pumping unit indicator diagram fault diagnosis is carried out by using improved self-organizing feature mapping network(SOM neural network).The improved effects of ant colony algorithm(ACO),genetic algorithm(GA)and particle swarm optimization(PSO)on SOM neural network are compared.It is shown that PSO algorithm has the smallest optimal individual fitness value,small iteration number and fast convergence speed.At the same time,the fault diagnosis effects of BP,LVQ,conventional SOM and PSO-SOM are compared.The results show that PSO-SOM neural network has obvious advantages in recognition time,diagnostic accuracy and diagnostic variance.The classification of different working modes and the working condition diagnosis of indicator diagrams are finished by extracting the characteristic parameters of indicator diagram samples of pumping units using the shape recognition method of Fourier descriptors and moment invariants and learning,training and simulating using PSO-SOM neural network technology.The research shows that the PSO-SOM method has an effective recognition accuracy of 90.3%for nine types of indicator diagrams of 700 well times of pumping units in XX Oilfield,which is an effective way to achieve intelligent and efficient diagnosis of indicator diagrams.