The design of sequential control programs for manufacturing systems is a difficult task which is traditionally carried out by engineers. Artificial intelligence planning techniques have proved to be very useful in the building process of such programs obtaining error-free programs and saving engineering time, which makes it an area of increasing interest in the AI community. However, there are some features of manufacturing systems that are not considered or are not considered in enough detail. Initially, our work focused on these features building a planning system called MACHINE, able to deal with them. The reason for doing this is to show that the reasoning process about the actions that take place in a manufacturing system is slightly different from the process followed in most known artificial intelligence planners and that the results obtained could be more realistic if these features were considered.
MACHINE has been able to solve different and interesting problems for real size manufacturing system. For example, from a knowledge base describing the manufacturing system of Figure 1 (a description based on the devices and interconnections among them, their initial state and raw products) and the description of the manufactured product (as a phase level recipe), MACHINE obtains the skeleton for the algorithm of the sequential control program (Figure 2). This basic structure can be directly translate into a GRAFCET diagram (Figure 3) or a live and safe Petri net (Figure 4), in such a way, that a control engineer can directly implement the sequential control program.
However, MACHINE has several limitations for a direct industrial use, and the extension of MACHINE in order to improve the system is our current goal. At present, we are collaborating with two industrial companies in the development of these goals, one corresponding to a control consumer company, a spanish national dairy company, and a control producer company, a multinational company of industrial systems.
A more detailed description of MACHINE can be found in the following Postcript file.
We have worked on several improvements about the architecture of MACHINE, these improvements basically correspond to the efficiency in the search mechanism and an enrichment of the description language. In any case, there are a lot of possible improvements that can be made in MACHINE in order to increase the usefulness of the algorithm. Many of these improvements are related with the inclusion of uncertainty management in the planning algorithm or the inclusion of knowledge acquisition techniques.
The first goal is the extension of the previous model of action and plans, in order to include the properties of reaction to sensing information and adaptability or failure tolerance. This extension can consider characteristics as distribute processing, hierarchical design, temporal restriction or uncertainty management.
The second goal is the extension of the current architecture of MACHINE to include the new action model and plan representation. In this goal, we are investigating in conditional reasoning in order to obtain conditional plans that can work with different sensing information. Another important aspects are