For the above reasons, some might argue that for certain tasks, the scientific or artificial intelligence of machines is more likely to provide reliable results than does human intelligence. One of these tasks can be the model-based predictions of the longer range outcomes of present trends. One might argue that a process control model, such as the ANN of the global warming process shown in Figure 2, can be trained on historical data to predict future trends. (Later in this article I will show, how ANN can be used to determine the gains, time constants, dead times, inertias and interactions of various processes.)
By using ANN and other intelligent tools, we can obtain scientifically reliable predictions of the future consequences of present actions. Some might argue that these scientific predictions are based on incomplete sets of data or on biased facts. This can be a valid argument, as no expert system knows what it does not know! Others believe that the past provides unbiased and reliable data and therefore, by “training” ANN models on historical data one can accurately predict the timing and size of future events. This debate is likely to continue for some time, and process control professionals can make major contributions, because they can point to the many AI applications already in successful use in a variety of industrial, transportation and other processes.
It is too early to say where the limits of AI applications are (*7). It seems reasonable that AI can predict the outputs of either industrial or nonindustrial processes on the basis of their inputs, as long as it is good enough to “train” the ANN relationship on past data.
Intelligence is seldom exact or mathematical. A tennis champion seldom understands Newton’s laws or the principles of aerodynamics that determine the path of a tennis ball. Tennis is learned by a trial-and-error process (similarly to training an ANN model). This is done by observing the input variables involved in the process of hitting the ball and memorizing the consequences of those inputs. There might be 25 input variables, from spin to force and from direction to height, which all contribute to a perfect serve or a perfect return.
The “AI model” developed by years of training in the brain of the tennis player, is similar to the ANN model that is developed by training an AI controller on the past performance data of say a refinery. Once the ANN is trained, you “just use it,” because, as Yogi Berra put it: “You can’t think and hit at the same time!”
White, Gray and Black AI Boxes
AI is an attempt to reproduce human reasoning and learning. It is a relatively new field: The first meeting of the American Association of Artificial Intelligence was held only in 1980. Yet, in some areas of application, such as process control, substantial progress has been made. The various AI strategies can be grouped into white, grey and black box categories.
The models of well-understood processes are referred to as white-box or first-principle models, because they are constructed from a priori knowledge and physical insight. Here, the dynamic models of the process are derived mathematically from the mass, energy and momentum balances of the process. Good examples of the early white-box models are the various feed-forward control schemes.
|FIGURE 3: FEED-FORWARD OPTIMIZATION
|In feed-forward optimization of steam heaters, major load variations (T1 and W) are corrected by the feed-forward portion of the loop, leaving only the minor load variables for feedback correction.
In case of a water heater (See Figure 3), we know how much steam will be needed to heat the water to match a new demand, so that the temperature of the hot water will remain constant. Therefore, by “feed-forwarding” that quantity as the set point to the steam flow controller (FIC), we can prevent an upset.
For those processes that are not understood the so-called black-box or empirical models are used. Here, because no physical insight is available, the process models are developed on the basis of measurement data. This means that input-output data is collected for some extended period of operation for both the steady state and the dynamic responses of nonlinear processes. The tools of black-box modeling include mathematical approximation via estimation theory, nonparametric regression, wavelet algorithms, fuzzy models and artificial neural networks (ANNs).
Fuzzy logic (*8) and gray-box modeling are non-mathematical. They are linguistically interpretable formations of rule-based models developed on the basis of the available expert knowledge and the measured data for the process. ANN can be embedded within databases or expert-system applications, can act as preprocessors or postprocessors to other systems, or be linked in a distributed architecture. From the process-modeling point of view, two main integration approaches can be distinguished: bias modeling and semi-mechanistic modeling.
The bias or parallel modeling approach assumes that a first-principle (white-box) model of the process can be obtained, but it is not possible to identify all the sources and statistical characteristics of the disturbances. In this case, a neural network is trained to predict the difference (residual) between the process and its first-principles model.
The semi-mechanistic model is based on a first-principle model and the unknown parts of the white-box model, such as the otherwise difficult to calculate parameters are represented by black box elements.