Analysis of heat transfer fouling data using artificial neural networks

在换热设备的设计和操作中，与工艺有关的在换热表面形成沉积物的现象是人们最不了解的现象。这种不令人满意的情况必须部分归咎于传统回归方法不足以将实验数据与分布不均的参数变化联系起来；例如，根据有限的实验观察或仅仅因为它们的可用性来选择自变量，以及缺乏证明所选回归方程是否是最佳回归方程的知识。拟议的研究项目致力于：i)利用来自众多来源的污垢数据分析几个综合数据库，特别是独特的、专有的、迄今无法获得的HTRI(热传递研究公司)。实施人工神经网络，对这些数据库进行训练，以关联/预测结垢行为。目前的研究将集中在普遍存在的问题，从冷却水结垢。如果这项工作成功，可能会在后续项目中调查其他通用的污垢过程，如原油污垢或牛奶污垢。原始的实验污垢数据和相关的操作条件将被转换成由无量纲组组成的工作数据库。选择最合适的无量纲基团组合作为人工神经网络输入以预测污垢阻力是很重要的。除了确定关键参数及其对比例尺形成的影响外，还将使用经过训练的神经网络生成结构良好和分布良好的数据库，然后利用该数据库获得适合在计算机代码中实施的相关性。

The process-related formation of deposits on heat transfer surfaces is the least understood phenomenon in design and operation of heat transfer equipment. Part of this unsatisfactory situation must be blamed on the inadequacy of conventional regression methods to correlate experimental data with an ill-distributed parameter variation; e.g. the selection of independent variables based on limited experimental observations or solely because of their availability, and the lack of knowledge to justify whether the chosen regression equation is the best. The proposed research project endeavours i) to analyse several comprehensive data banks with fouling data from numerous sources, notably the unique, proprietary and hitherto unavailable HTRI (Heat Transfer Research Inc.) database and ii) to implement artificial neural networks trained on these data bases to correlate/predict the fouling behaviour. The present investigation will focus on the ubiquitous problem of scale formation from cooling water. If this work is successful, other generic fouling processes, such as crude oil fouling or milk fouling, may be investigated in subsequent projects. The original experimental fouling data and the related operating conditions will be converted into a working database consisting of dimensionless groups. It is important to select the fittest combination of dimensionless groups to be used as Artificial Neural Network inputs to predict fouling resistances. In addition to the identification of key parameters and their effects on scale formation, the trained neural network will be used to generate a well-structured and well-distributed data base, which can then be used to obtain correlations suitable for implementation into computer codes.