基于大数据模型深度玻尔兹曼机的压缩机组状态预测方法研究

发布时间：2018-10-10 13:41

【摘要】：压缩机组不同部分之间互相关联,紧密耦合,使得其故障特征具有不确定性、非线性和并发性。传统模型对机组的故障特征表示能力差,而深度学习本质上是一种采取多层非线性变换的信息提取技术,能够实现对数据间复杂关系的建模。因此以深度学习理论为基础,对压缩机组故障预测进行研究具有理论意义和实用价值。本文就如何将深度玻尔兹曼机(Deep Boltzmann Machine,DBM)模型更好地应用于压缩机组状态预测领域展开了研究:(1)研究了基于高斯深度玻尔兹曼机(Gaussian Deep Boltzmann Machine,G-DBM)的压缩机组振动信号预测的建模方法。通过在DBM可视层增加高斯滤波器实现模型的预处理。并且制定预测前的数据准备规则,构建G-DBM预测模型,依靠经验确定模型结构。(2)依靠经验确定模型工作量大,针对该问题研究了G-DBM预测模型的优化方法。提出带有极值扰动的简化粒子群优化(Ex-tremum Disturbed and Simple Particle Swarm Optimization,tsPSO)算法确定模型的参数,并且采用混合共轭梯度法加速G-DBM模型的训练过程。模型优化后得到的预测序列平均误差值减小,预测计算时间也大大降低。(3)为实现状态的预测,拟对现场机组监测数据的预测序列进行状态识别。由于现场故障状态数据极难获取,因此通过实验模拟获得的轴承监测信号来进行模型搭建,论文搭建非抽样提升小波包和G-DBM结合的轴承状态识别模型,为现场机组的状态预测奠定理论基础。针对状态序列中的噪声信号,采用非抽样提升小波包进行降噪,通过无特征提取的G-DBM模型对轴承信号进行状态识别,经过降噪的G-DBM识别比直接采用G-DBM模型的识别结果准确率提高了三个百分点。
[Abstract]:The different parts of the compressor unit are closely coupled with each other, which makes the fault characteristics uncertain, nonlinear and concurrent. The traditional model has a poor ability to express the fault characteristics of the unit, but depth learning is essentially a multi-layer nonlinear transformation information extraction technology, which can realize the modeling of the complex relationship between the data. Therefore, based on the theory of depth learning, it is of theoretical significance and practical value to study the fault prediction of compressor units. In this paper, how to better apply the depth Boltzmann machine (Deep Boltzmann Machine,DBM) model to the field of compressor unit state prediction is studied. (1) the modeling method of compressor unit vibration signal prediction based on Gao Si depth Boltzmann machine (Gaussian Deep Boltzmann Machine,G-DBM) is studied. Gao Si filter is added to the DBM visual layer to realize the preprocessing of the model. And make the data preparation rules before the prediction, build the G-DBM prediction model, rely on experience to determine the model structure. (2) relying on the experience to determine the model workload, this paper studies the optimization method of the G-DBM prediction model. A simplified particle swarm optimization (Ex-tremum Disturbed and Simple Particle Swarm Optimization,tsPSO) algorithm with extremum perturbation is proposed to determine the parameters of the model, and the hybrid conjugate gradient method is used to accelerate the training process of the G-DBM model. After model optimization, the average error value of the prediction sequence is reduced, and the prediction calculation time is greatly reduced. (3) in order to realize the state prediction, it is proposed to recognize the state of the prediction sequence of the field unit monitoring data. Because the field fault state data is very difficult to obtain, the model is built by the bearing monitoring signal obtained by the experimental simulation. The paper builds the bearing state recognition model which combines the non-sampling lifting wavelet packet and G-DBM. It lays a theoretical foundation for the state prediction of field units. For the noise signal in the state sequence, the non-sampling lifting wavelet packet is used to reduce the noise, and the G-DBM model without feature extraction is used to recognize the state of the bearing signal. The accuracy of G-DBM recognition by denoising is improved by three percentage points compared with the recognition result using G-DBM model directly.
【学位授予单位】：中国石油大学(北京)
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP311.13;TE974

【相似文献】