散状物料连续累计称重系统精度补偿研究

发布时间：2018-04-19 17:24

本文选题：散状物料 + 电子皮带秤　；参考：《南京理工大学》2016年博士论文

【摘要】：随着中国经济的高速发展,各种散状物料尤其是大宗工业原、燃料的贸易运输量急剧上升,使得散料贸易中动态称量的要求越来越高。目前,如何长期保持≤0.1%的计量精度已成为国内外诸多散料贸易中动态计量专家和科技工作者亟需解决的难题。机器学习的发展给传统行业带来了不一样的色彩,随着"互联网+"战略计划的提出,传统衡器行业将面临新的转型升级。本文以目前应用最广泛的散状物料连续累计称重设备——电子皮带秤为研究对象,结合各种机器学习方法,对其累计计量精度所涉及到的"皮带效应"、局部性故障、输送带跑偏、温度变化等问题展开研究。在电子皮带秤结构组成和称重原理基础上,结合以往的实验经验,以累计计量的测量信号流程为研究脉络,对其累计计量精度的误差源和耐久性问题进行了深入讨论研究,对影响精度最主要误差源以及耐久性误差源进行了总结。通过研究总结发现,需对运行中电子皮带秤的多个误差因素以及各个故障状态进行实时在线监测,并针对这些误差因素变化以及不同故障状态的不同程度建立一个具有较好泛化性能和鲁棒性的精度补偿模型,以真正提高累计计量的耐久性。针对"皮带效应",从梁理论出发对称重力误差进行了机理研究,对阵列式皮带秤"内力理论"的理论公式进行了推导。然后,以"内力理论"对QPS皮带秤全性能试验中心的4#皮带秤进行精度补偿试验。通过试验分析得出:无故障时,精度可达OIMLR50 2014(E)中的0.2级精度等级,即累计称重误差±≤0.1%;但当存在输送带跑偏、称重架卡料等一些故障时,精度很不理想,需对故障状态进行实时在线监测,并加以补偿。针对皮带秤称重区域内的故障对称重精度的影响,对故障的在线监测方法进行了研究。首先针对皮带秤不同流量称重数据密度的不均匀,分别提出改进型DENCLUE和改进型DBSCAN,并都应用于的称重区域故障在线检测,其中改进型DENCLUE采用动态阈值法替代爬山法大大降低了算法复杂度,相比较于改进型DBSCAN具有更好的聚类精度和更快的聚类速度;然后采用BRNN和改进型BTSVM对检测出来的故障进行在线识别,最后将识别出来故障码、故障位置(即哪个称重单元)、当前托辊传感器数据以及同一时刻正常数据的平均值作为故障特征。阵列式皮带秤故障试验表明:"基于改进型DENCLUE的在线检测+基于改进型BTSVM在线识别"模型具有更好的称重区域故障在线监测性能。针对皮带秤输送带跑偏故障对称重精度的影响,对输送带跑偏的在线监测方法进行了研究。研究引入流形学习和深层神经网络,分别建立了基于LTSA+GRNN+SVM和基于CDBN+SVM的在线跑偏监测模型,模型能产生显性非线性映射将原始称重数据压缩成3维的跑偏特征。二者的训练结果以及试验测试结果表明,皆具有很好的跑偏识别精度,可取代传统硬件检测设备,但适合于不同工作场合,LTSA+GRNN+SVM很适用于皮带秤称重标定较为频繁、跑偏检测实时性要求不是很高的情况,而CDBN+SVM非常适用于标定不是很频繁、但实时性和识别精度要求很高的情况。最后依据散状物料连续称重累计流量累加法计量原理,引入过程神经网络建立累计计量精度综合补偿模型,并以"内力理论"、称重区域内故障特征以及输送带跑偏特征为准定义了补偿模型的输入。研究基于过程神经网络精度补偿模型的训练算法,通过融合正则化极限学习机和误差最小化极限学习机算法,提出一种基于EM-RPELM的精度补偿模型。张力变化、有故障状态、温度变化的试验结果表明,基于EM-RPELM的精度补偿模型具有良好的鲁棒性、泛化性能以及一定的不平衡数据处理能,补偿后的精度总体达到0.2级。
[Abstract]:With the rapid development of China's economy, all kinds of bulk materials, especially large industrial ones, have increased the amount of fuel trade and transportation, which makes the demand for dynamic weighing in the bulk trade higher and higher. At present, how to maintain the measurement accuracy of less than 0.1% for a long time has become the urgent need for many dynamic measurement experts and scientists and technicians in the bulk of bulk materials trade. To solve the problem. The development of machine learning brings a different color to the traditional industry, with the advance of "Internet plus" strategic plan, traditional weighing industry will face a new transformation and upgrading. In this paper the bulk material is currently the most widely used continuous cumulative weighing equipment - electronic belt scale as the research object, combined with a variety of machine learning square On the basis of the structure composition and weighing principle of the electronic belt scale, the accumulative measurement signal flow is taken as the research vein, and the error source of the accumulative measurement accuracy is discussed. The problem of durability is discussed in depth, and the main error sources of the influence precision and the source of the durability error are summarized. Through the study, it is found that the multiple error factors of the electronic belt scale in the operation and the real time on-line monitoring of each fault state are needed, and the changes of these error factors and the different fault states are taken. A precision compensation model with better generalization performance and robustness is set up to improve the durability of accumulative measurement. Based on the "belt effect", the mechanism of symmetrical gravity error is studied from the beam theory, and the theoretical formula of "internal force theory" of the array belt scale is derived. Then, the "internal force theory" is used for the QP. The precision compensation test of the 4# belt scale of the S belt scale full performance test center is carried out. Through the test analysis, it is concluded that the precision can reach the 0.2 grade precision grade of OIMLR50 2014 (E) without fault, that is, the accumulative weighing error is less than 0.1%, but when there are some faults such as the running belt deviation and the weighing frame card, the precision is very unsatisfactory and the fault state should be realized. On line monitoring and compensation, the on-line monitoring method of the fault is studied in view of the influence of the fault symmetry heavy precision in the weighing area of the belt weigher. Firstly, the improved DENCLUE and the improved DBSCAN are put forward respectively for the uneven density of the weighing data of the belt weigher. Test, the improved DENCLUE uses the dynamic threshold method instead of mountain climbing method to greatly reduce the complexity of the algorithm. Compared with the improved DBSCAN, it has better clustering accuracy and faster clustering speed. Then, BRNN and improved BTSVM are used to identify the detected faults online, and the fault location (i.e., where the fault location) will be identified. A weighing unit), the current roller sensor data and the average value of the normal data at the same time as the fault characteristics. The array type belt scale fault test shows that the "online detection based on improved DENCLUE + based on the improved BTSVM online recognition" model has better performance on the on-line monitoring of weighing area fault. On the basis of the influence of the symmetrical heavy precision of the fault, the on-line monitoring method of the belt running deviation is studied. In this paper, we introduce the manifold learning and the deep neural network, and establish the on-line running deviation monitoring model based on LTSA+GRNN+SVM and CDBN+SVM respectively. The model can produce the running deviation characteristic of the original weighing data into 3 dimension by the explicit nonlinear projection. The training results of the two and the test results show that all of them have good accuracy of deviation recognition and can replace the traditional hardware detection equipment, but it is suitable for different working situations. LTSA+GRNN+SVM is very suitable for the weighing scale of the belt weigher, and the real-time requirement of deviation detection is not very high, and the CDBN+SVM is very suitable for calibration. It is very frequent, but the real time and recognition precision are very high. Finally, according to the accumulative flow cumulation principle of the continuous weighing of the bulk material, the synthetic compensation model of accumulative measurement precision is established by introducing the process neural network, and the compensation model is defined by the "internal force theory", the characteristic of the fault in the weighing area and the characteristic of the belt running deviation. A training algorithm based on the precision compensation model of process neural network is studied. A precision compensation model based on EM-RPELM is proposed by integrating the regularization limit learning machine and the error minimization limit learning machine algorithm. The test results of tension change, failure state and temperature change show that the precision compensation model based on EM-RPELM is shown. It has good robustness, generalization performance and a certain degree of imbalance data processing energy, and the accuracy of compensation reaches 0.2 levels.

【学位授予单位】：南京理工大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TH715.1

【相似文献】