PLM, Lean, Manufacturing Systems.
Manufacturing environments are dynamic and random in nature. In order to maximize productivity, people and processes must respond to changing conditions as they occur. Real-time scheduling techniques can be used to identify and respond to the current behavior of a manufacturing system. However, accurate predictions concerning the future behavior of the manufacturing system would also be useful information for engineers and managers as they try to optimize the performance of a system.
A key performance measure for most manufacturing systems is the throughput (# parts/unit time). Our (Professor Sengupta and myself) current research considers the construction of a throughput prediction tool. This tool will collect real-time performance data from the manufacturing system and use it to make a prediction concerning future behavior.
For example, the predictive tool could collect data during the first 4 hours of a production shift and use it to make a prediction of the total system throughput at the end of the 8 hour shift. Such a prediction is useful to manufacturing management personal in making both short and long term production decisions.
There are several powerful data-driven modeling techniques that are being studied for constructing such a predictive tool, these include neural networks, statistical regression and case-based reasoning. Note that to successfully use these techniques, a large set of input-output data reflecting the system's performance is required.
This set of input-output performance data will be referred to as historical performance data. For example, each element in this set of historical performance data could contain the throughput and Work-In-Progress (WIP) collected at multiple locations in the manufacturing system for various times during the first 4 hours of a given production shift (the input) as well as the total throughput at the end of the shift (the output). Unfortunately for many modern manufacturing systems, a sufficiently large set of such historical performance data required to use these data-driven modeling techniques is usually not available.
There are two main reasons for this shortage of historical performance data. First, the manufacturing system may be too new to have accumulated a large enough set of historical performance data. However, a more important reason is due to the continuous improvement process applied to most manufacturing systems. Most manufacturing systems are modified on a weekly, if not daily, basis. Hence, the current behavior of a manufacturing system is usually not reflected in the input-output performance data collected during the previous month due to system modifications.
Our solution is to use a verified simulation model of the manufacturing system to produce a set of simulated historical performance data. This set of simulated historical performance data is used to train the throughput prediction tool via one of the data-driven modeling techniques. Once the throughput prediction tool is trained and verified, it will use real-time performance data collected from the manufacturing system in order to make its prediction.