AI Quality Systems

When gauging or human visual inspection can detect defects, prevention can generally be accomplished by a “Cause and Effect” LSS analysis (Chapter 8, pp 141-197 of the “Lean Six Sigma Pocket Toolbook” by George et al). These powerful tools are generally effective in prioritizing, identifying and solving  the vast majority of  manufacturing quality problems.

Scrap after Setup:  However, consider the case a process that is under-control after setup is completed, but which suddenly generates scrap for no explainable reason. How did it happen? How can it be prevented?  Everyone in the factory will have an opinion, but the human mind can only consider a few possibilities, and are captive to previous experience of individuals.

Most processes can be well instrumented and provide data on scores or even hundreds of process parameters that may contribute to scrap. Buried in this “Big Data” output of the instrumentation is an often counter-intuitive answer which could be dependent on scores of parameters which result in a “classification” of failure together with root cause. The Neural Network will search for and “learn” from common patterns in the data of failure events, and classify the events versus parameters of  the process. The key input is the early detection and timing of the defect which will allow a strong correlation to be made of possible causes. The detection of potential problems is the key to evolving a solution and may be beyond human visual and logical capability.

One example of the use of AI in quality inspection is in the production of Very Large Scale Integrated (VLSI) circuits. The complexity of Integrated Circuits has grown from 20 transistors per chip in 1963 to 8 Million at present. While it was possible to test the performance of each transistor in 1963, it is clearly impossible at present. How then can a company know it is producing good chips? The CEO of Intel has written:

“Almost every company you can think of, every application, it’s going to be affected by Artificial Intelligence. You are going to be using Artificial Intelligence, or you’re going to be outpaced by people who are. .…We produce about a million chips per day, and take about 1.6 million photos of each chip as it progresses through the production line. Why didn’t this chip work? We get that answer as AI compares the 1.6 million photos for patterns that do not agree with a good chip in a matter of hours. You’d be surprised at how many companies have access to the data but don’t put the investment in place (to use it).

“Companies must use AI-or else” Wall Street Journal, October 24 2017

At each step of the manufacturing process, abnormalities are noted in the hundreds of chips on each wafer. A map of the wafer is then constructed showing the position of each abnormality. A library of such wafer maps is created with root causes.  Other wafers with similar maps are an indication of the same root cause. Non-random patterns of defects carry more information as to the process problem than do purely random patterns, consistent with Shannon’s definition of Information.

Prior to the development of Deep Neural Networks in 2012, people attempted to classify each map defect pattern based on probabilistic models. Neural Networks, with several internal layers (Deep learning), can automatically generate the classifications1, and does not require any specific engineering knowledge of the process. The Neural Network can use the classifications for automatic image retrieval. The “training sets” for the Neural Network can use real data patterns to create a library of many thousands of simulations, and uses a sigmoid activation function2. It should be pointed out that the number of layers to be used in a Neural Network is more an art than a science at present. As a general rule, the more the layers, the more the possible classifications of defects. However, the more the layers, the slower the database search. Thus, in one example, we used 10,000 training images, and found that the defect classification among 1000 actual defect images had an accuracy of 99.9%. Juran has found that human inspection accuracy is less than 90%3. AIT experts will develop the AI Neural Network for your quality application.

References: “Lean Six Sigma in the Age of Artificial Intelligence” 148, Table 7.1

2.Page 148, Figure 7.5 150