【作者】 曹翌佳；

【导师】 阳永荣； 王靖岱；

【作者基本信息】 浙江大学 ， 化学工程， 2010， 博士

【摘要】 过程工业是我国国民经济的重要支柱产业之一,而多相流反应器在过程工业的连续生产中更是起到了举足轻重的作用,对其进行准确的在线故障检测与诊断是保障大型工业装置长期连续稳定运行的重要手段。随着现代工业生产的发展和科学技术的进步,多相流反应器的设计正朝着多功能、高集成、自动化的方向飞速发展。随之产生的问题是反应器内高度复杂的流体动力学行为,进而形成具有非线性、多尺度、动态性等特征的复杂系统,给多相流反应器的实时故障检测与诊断带来不小的困难。另一方面,传统检测手段(温度、压力等)仅针对过程的工艺参数进行采集,难以达到工业故障检测灵敏、准确、及时的高标准,不能满足现代工业生产的需求。因此,针对上述问题,系统研究过程工业中多相流反应器故障的产生机理、特征、影响因素以及应对策略,开发一种简单快捷、灵敏准确的故障早期检测技术不仅极富挑战性,同时也具有重大的理论意义和实用价值。本论文针对化工过程中三种应用广泛的多相流反应器(聚丙烯卧式搅拌床反应器、聚乙烯流化床反应器和催化裂化循环流化床反应器),以先进的无损声发射技术为在线测量手段,综合应用非线性动力学分析和多元统计理论,系统地研究了颗粒性质(大小、组成)与各类故障(结块、结焦)之间的定性与定量关系。论文不仅建立了以声发射信号为基础的新型检测方法和诊断模型,而且提出了各类故障的应对策略,研究结果对于工业多相流反应器的连续安全生产、优化工艺操作和提高产品质量都具有重要的指导意义。论文主要开展了以下四方面的研究工作：1.深入研究了颗粒性质改变前后,多相流系统流体动力学信息发生的特定变化,建立了针对颗粒结块故障的声发射早期结块识别系统(Acoustic Emission-Early Agglomeration Recognition System, AE-EARS)、以及针对颗粒结焦的催化剂结焦量在线检测模型。1) AE-EARS以非线性动力学分析和现代统计理论为依托,主要由基于吸引子比较法的统计值S与基于复杂性分析的故障系数C两大模块构成。其设计思路是从状态比较的角度出发,通过比较当前状态下声发射信号与正常工况下的差异,最终判断多相流反应器内的流体动力学行为是否发生了状态改变。2)吸引子比较法通过构建声发射信号的混沌吸引子以描述反应器内的流体动力学行为,在多维空间中计算当前状态与正常状态下吸引子之间的“距离”,并设计统计量S以实现结块的预警。当S值接近于0时,在零假设的前提下,说明当前的评估状态与正常生产状态相似；当S值大于3时,说明当前评估状态下反应器内的流体动力学行为已经发生了改变,即产生了结块。3)故障系数法通过引入复杂性分析中关联维与Kolmogorov熵理论,定义故障系数C为当前评估状态与已知正常状态下特征参数(关联维、K-熵)的相对状态比值,并通过设定阈值函数对当前的评估状态进行故障判别。4)为建立催化剂结焦量在线检测模型,在深入分析颗粒撞击壁面产生声发射信号的功率谱特征与催化剂结焦量之间的复杂关系基础上,结合多元数据处理中偏最小二乘法(Partial Least Squares, PLS)实现了声发射信号中有用信息的提取,成功地建立了声发射信号与催化剂结焦量之间的定量预测模型。2.采用AE-EARS系统,对聚丙烯卧式搅拌床反应器内的结块进行在线预警,并结合生产工艺,针对性地提出了结块故障的应对策略。1)针对吸引子比较法在实际应用过程中各主要参数(窗口时间Tw、嵌入维数m、卷积带宽(,、片段长度L)的选择问题,提出了一种以优化理论为基础的参数设置方法,它以计算得到最大的S值、重复实验S值的标准方差最小为优化目标,以S值随大颗粒质量分数的增加而增加为约束条件,采用网格搜索法为优化手段,具有简洁、易于操作等优点。2)通过各种针对性的方法降低了外界操作参数对S值产生的干扰,极大地提升了AE-EARS对外界操作变量波动的鲁棒性,使之能适用于复杂工业装置的结块预警。基于离散小波除噪法对由搅拌桨转速的波动而产生的干扰进行了消除,并使得在小范围内(±10%)S值基本不受搅拌桨转速的影响；通过对原始声发射信号进行标准归一化处理来消除大部分由床层重量波动而产生的影响；通过合理选择声发射传感器的位置来大大降低S值对底部循环气波动的敏感性,声发射传感器可接受的范围为距离釜底1／8D至1／2D高度,最佳位置为1／4D,其中D为卧式搅拌釜的内径。3)基于S值的AE-EARS系统能敏感地检测卧式搅拌床反应器内床层颗粒粒径的微小变化,在结块发生的早期提供精确的预警。在此基础上,AE-EARS不仅能够指示在牌号切换过程中多相体系是否达到了稳定状态,而且借助多传感器技术能够准确定位“异常事件”的发生位置,并在大型年产25 wt聚丙烯工业热模装置上进行了验证。3.考察了AE-EARS在气相聚乙烯流化床反应器内对于结块的检测能力,并根据流化状态下颗粒运动的特点,通过调整参数、降低误报率等方法优化了AE-EARS的模型性能。1)根据细颗粒在加热状态下发生团聚的机理,在实验室冷模装置内模拟流化床内颗粒结块的动态过程,并获得了各故障系数C的阈值(aD2=0.3,aK2=1.2)。在此基础上,对于工业聚乙烯流化床内结块的故障进行了检测,实验结果证明,基于复杂性分析的故障系数C能用于结块的早期检测,且上述阈值具有一定的通用性。2)综合比较了状态统计值S与故障系数C在流化床反应器中对于结块早期识别的能力。经实验室冷模和工业热模装置的全方位测试,证明统计值S无论是在结块识别的精度(S:95%,C:90%),还是预警的及时性(工业测试中S比C提前20-30 min)方面都显示出更加优异的性能。3)通过“过滤”方法对S值进行处理,能够有效提高检测精度,由原先95%提高至99%。4.以催化裂化循环流化床中颗粒撞击壁面的声发射信号为基础,催化剂结焦量在线检测模型为分析手段,分别在实验室冷模与工业热模装置上针对FCC颗粒的结焦量进行了在线检测,为长期的故障诊断积累了宝贵的工业数据。1)对不同结焦量催化剂颗粒进行了各项颗粒性质的表征,结果发现：结焦前后催化剂颗粒的粒径分布没有发生较大改变,而颗粒的比表面积、孔容以及孔径在结焦后均有不同程度的下降。结合扫描电镜结果推测,焦物大部分沉积于固体颗粒表面的孔道内,且呈现出非连续性的分布状态。2)借助多次累加平均法对传统的功率谱分析进行了改进,获得了高分辨率、低噪音的功率谱图。在实验室冷模装置内,对不同结焦量的FCC颗粒进行了流化实验,采集并计算得到了不同表观气速下催化剂颗粒的声发射信号功率谱图。实验结果表明：功率谱特征峰的频率位置随催化剂颗粒结焦量的增大而向低频方向线性移动。由此,将催化剂颗粒的结焦量与特征峰的频率位移相关联,获得了较好的预测效果,不同气速下的平均相对误差不超过8.12%。3)在频率位移模型的基础上,借助PLS回归对实验室内30组不同结焦量颗粒声发射信号的功率谱进行了建模,并进行了独立验证。实验结果表明：不同气速下模型的相关系数为0.906,交叉验证均方根(RMSECV)为0.0535,满足工业检测的要求。4)在工业催化裂化的循环流化床中对催化剂颗粒结焦量的在线检测模型进行了验证。工业测试结果表明：基于PLS回归模型的声发射功率谱信号能便捷、准确地实现再生器进出口两端催化剂结焦量的在线检测,与传统采样分析法相比,具有明显优势。更多还原

【Abstract】 Process industry is one of the most important mainstay industries in our national economy, and multiphase flow reactors even play an important role in the continuous production within process industry.The accurate on-line faults detection and diagnosis of the reactors is an indispensable technique which is used to ensure the stable operation of the industrial plants. With the development of chemical engineering and modern scientific technology, the design of the multiphase flow reactors is evolving to the direction of multifunction, highly integrated and automation. The followed problem is that the hydrodynamic behaviors in the reactors turn to be much more complex, which causes big troubles on the on-line detection and diagnosis for the multiphase flow reactors. On the other hand, the traditional detection means, such as temperature and pressure, were only designed to record the process parameters, and could not meet the high standards (for examples, sensitive, accurate, and early warning) for the faults detection in modern industry. Therefore, in order to address these issues, it is not only challenging, but with practical and theoretical significance to develop a novel measurement tool and systematically investigate the mechanisms, characteristics, factors and corresponding strategies for the faults detection and diagnosis for the multiphase flow reactors.Based on the advanced non-invasive acoustic emission technique, combining the experimental study and theoretical analysis, this thesis had an in-depth study of the faults (agglomeration, coking) which were generated because of the qualities changing of particles (size, composition) in the three main multiphase flow reactors in chemical production, such as horizontal stirred bed reactors for polypropylene, fluidized beds for polyethylene and circulating fluidized beds for fluid catalytic carking. From this study, we not only established the new methods and diagnosis models based on the acoustic signals, but also proposed several corresponding means for the individual faults, which had an important guiding significance on the safely continuous production, optimization of the process and development of the new products. The results in this thesis can be summarized as following.1. After the in-depth research of the hydrodynamic information changes between the state that the particles quality had changed and the state in normal operation, we built up an Acoustic Emission-Early Agglomeration Recognition System, AE-EARS, for the chunks and large agglomerates detection and an on-line quantitative detection model for the amount of coke deposit on catalyst particles.1) AE-EARS, which was based on the nonlinear dynamic analysis and modern statistical theory, was composed with the statictical value S based on the attractor comparison method and the malfunction coefficients C. The inspiration of the system was from the idea of states comparison. It was based on the finding or calculating the differences between the evaluated state and the normal state, then the system could decide the current process conditions whether or not had changed.2) The attractor comparison method was based on a general distance concept between multidimensional distributions. By comparing the attractor of a reference AE time series at the normal operating conditions with the attractor of evaluated AE time series acquired during operation of HSBR, we were able to get the "distance" (which was called S value in a statistical way) between the two distributions. When the S value was close to zero, under the null hypothesis, it means that evaluated situation is similar to the normal operating condition; when the S value was larger than 3 the hydrodynamic behaviors in the evaluated situation had changed and agglomeration might have formed. 3) Based on the reconstruction of the attractors in multi-dimension space, correlation dimension and Kolmogorov entropy were introduced to monitor the dynamic process of malfunction in multiphase flow reactors.4) The on-line model for coke deposit measurement was established based on the improved power spectrum density, whose data transformed from the acoustics signals by FFT. The partial least square, PLS, method was used to quantitatively correlate the PSD data and coke deposit amount.2. By using the AE-EARS, we were able to give early warnings of agglomeration in both lad-scale and industrial horizontal stirred bed reactors (HSBR). Furthermore, by combining the detection results and the background of the production process, several corresponding means were proposed to resolve the particle problems.1) The parameters of this method for AE measurement were set by an optimization method. Based on the analysis of the general range of these four parameters, we had taken the biggest S values and the smallest stand deviations of S as optimization objective, and the S values increase with increasing fraction of PP4 as constraint function.2) External factors affecting the performance of the method had been reduced. A denoising algorithm based on a discrete wavelet transform was used to remove a part of noise from the agitator which made the method insensitive to small changes in the agitator speed. Normalization the AE signals before calculating the S values was able to remove most of the influence of the bed mass. By carefully choosing the positions for AE sensors could decrease the sensitivity of the S values to the air flow rate which means that a novel detection technique was absolutely necessary for selecting an optimal position for sensors during a lab-scale research. The allowable range for AE sensors was 1/8 D to 1/2 D above the bottom; and the optimal position for sensors was 1/4 D, where D was the diameter of the HSBR. 3) The AE technique based on the attractor comparison method was sensitive to small changes in the particle size distribution, which meant it could offer "early and accurate warning" of agglomeration in HSBR. Furthermore, the monitoring method might not only be used to indicate if a stationary situation had reached during a grade transition, but also had the probability to locate the agglomeration in HSBR with multiple-sensors. However, further research was needed before applying the method in an industrial HSBR for agglomeration detection.3. The performance of AE-EARS in the fluidized beds agglomeration detection had been investigated. According to the unique characteristics in fluidized state, we had optimized the output of AE-EARS by adjusting the parameters and reducing the false alarm rate.1) According to the theory that under heating conditions fine particles of polyethylene will adhere to each other to form larger chunks, lab-scale experiments were carried out to simulate the situation in real reactors. Meanwhile, the threshold functions for the parameters were obtained (aD2=0.3, aK2=1.2).2) The performance of the statistical S and malfunction coefficient C were compared in both lad-scale and industrial fluidized bed reactors. The results showed that whether on the sensitivity to agglomeration, or the timeliness to send warnings, the S method showed higher quality than C.3) Based on the "filtering" process for the original S, the accuracy of the proposed method was highly improved, from 95% to 99%.4. Based on the acoustic signals emitted by the collisions between the FCC particles and the internal wall of the circulating fluidized beds, supplemented with on-line quantitative detection model, the experiments both in lab-scale and industrial plants to measure the amount of coke deposit on catalyst particles were studied.1) Particle properties of the catalyst which had different amount of coke were characterized by several measurements. The particle size distribution showed no big changes between the fresh catalyst and the coked one. The BET tests told us that the surface area, pore volume and size had all dropped after the reaction. Therefore, it could be inferred that most of the coke deposited in the pores on the surface of the catalyst particles, and the distribution showed non-continuous status.2) A model to predict the amount of coke deposit on catalyst was established to describe the correlation between the characteristic frequency of AE signals from the particles-wall collisions and different amount of coke deposit. Experiments in a cold mode fluidized bed showed that this model was useful for online measurement of the amount of coke deposit on catalyst. The average absolute relative deviation (AARD) was below 8.12% when the predicted amount of coke deposit was compared with the true one.3) The correlation between acoustic power spectrum density and the amount of coke deposit were investigated in lab-scale experiments by PLS regressions method. The best PLSR results, obtained under the different superficial gas velocities showed for a correlation coefficient of 0.906, and a root mean square error of cross validation (RMSECV) of 5.35%, which fully met the requirements of industrial applications.4) The PLS model had been verified in the industrial plants, and the final results could be used to illustrated that the proposed method could meet the requirements of on-line measurement of the amount of coke deposit on catalyst in circulating fluidized bed reactors.更多还原