Skip to main content
As for production lines in factories, it is sometimes the case that trouble occurring in upstream processes has an adverse effect on downstream processes and, as a result, production is disrupted. By instantaneously predicting the influence of the upstream trouble and implementing countermeasures in the early stages, it would be possible to create a robust supply chain impervious to such trouble. Accordingly, Hitachi has developed a production-control technology for "visualizing" the status of production in the past, present, and future. Applying this technology makes it possible to forestall the influence of trouble.
NONAKAAs for conventional production control, work has been evaluated by simple accumulation of working hours measured by devices like stopwatches. For example, if the work involved in making a certain component takes ten minutes, since it takes 100 minutes to make ten such components, it is typical that the worker making that component gets ready for handling work for 100 minutes.
The problem with this kind of control is that the fact that "every component cannot be made correctly in ten minutes" is not considered. In the case of skilled workers and apprentices, it is likely that the same work will takes different amounts of time and that even the same worker will have ups and downs at various times of the day and on different days. We refer to this phenomenon as "variability." And without considering this variability, we cannot correctly estimate working hours.
SUGINISHIWhat's more, to perform "production control" correctly, in addition to the variability of individual works the affect of upstream processes on downstream ones must be taken into account. In the case of a simple production line of a factory, determining problematic points is relatively simple. In the case of a factory in which components of electronic devices are made, however, the routes along which manufactured products flow are complicated; for example, the same device is used for multiple manufacturing processes, and various kinds of products are passed along the same lines. In such a complicated case, it is difficult to identify what problems are occurring, and where they are occurring, simply by observations on site. When a problem occurs somewhere, up until now, it has been necessary to come up with a system for determining whether, and in what manner, that problem will affect production.
NONAKAFirst, using theories of probability and statistics, we quantify the working status in consideration of "variability." We then add the parameter, namely, "time," so that it becomes possible to understand the relation between upstream processes and downstream processes.
The working status in a factory is a so-called "nonlinear phenomenon," and there is no easy way to predict it. In the year 2000, it was considered impossible to actually predict a nonlinear phenomenon. At that time, I thought why not use theories of probability and statistics. Although this field that was not researched in Japan, it had been researched at the Massachusetts Institute of Technology (MIT), USA. Consequently, I was sent at the company's expense to study there. The research achievements I made at MIT form the basis of the technology introduced in this article.
NONAKAThis technology has two main features. The first is "visualization," that is, a way of observing production status with your own eyes. For example, a company will make components for cars and supplies them to the car-assembly plant. If some sort of trouble occurs at the company's plant, in the course of time, that trouble will be passed on from upstream processes to downstream ones. And, finally, delivery of components will be disrupted, thereby causing trouble for the customer. A diagram showing the status of the trouble diffusing in this manner is presented on a computer screen.
The second feature is "control." Visualization on its own is not "production control." Given that fact, we developed a simulation technology for investigating "Whether the work can be completed in the manner predicted if certain countermeasures are taken?" on the basis of the visualized production status.
SUGINISHIThe visualized appearance of the statuses of the production processes is shown in Figure 1. The vertical axis represents time, where the upper area is the past, and the lower area is the future. The horizontal axis represents the production processes. The left side shows the upstream processes, and the right side shows the downstream ones. Each cell depicts the production status of "a specific process at a specific time."
Figure 1: Visualization of production status
SUGINISHIAs for the developed production-control technology, the general theory of probability and statistics, such as, means and standard deviations, is applied. First, the statuses of the production processes are quantified by mathematical formula and expressed in either of three colors, namely, blue, yellow, or red. Blue indicates the status that production fluctuation is negligibly small. Yellow indicates the status that a moderate degree of production fluctuation is occurring. Red indicates the status that a large degree of production fluctuation is occurring. If some trouble occurs and the production process starts to stall, the color changes in accordance with the extent of the paralysis. Moreover, mathematical formula for expressing the relationships among these production processes, "upstream/downstream", and "time" are derived and applied.
The figure shows that a failure of some equipment occurring in a past operation, and production fluctuation consequently occurred. With the passage of time, the red and yellow colors diffuse into the downstream productions processes.
NONAKAEven if there are parts of the graph turning yellow or red in the present, as long as the final operations (like delivery of goods) are turning blue, that is no problem. Whether the effect of trouble diffusing to the downstream processes can be erased by performing certain production control now can be verified in consideration of key factors such as number of workers, number of pieces of equipment, and overtime hours. For example, if a delivery date cannot be met with the present operating equipment, we can investigate adding extra equipments. In other words, we can simulate whether adding a certain number of equipments for a certain time will allow us to meet the delivery date. As a result, the number of additional equipments needed (and the length of time for running them) to meet the delivery date can be determined, and appropriate measures to do that can be taken.
SUGINISHIAlthough we have collected data for production control up until now, we could not say that it reflects actual conditions because, for example, error was caused by stopwatch measurement and readjustments were not done after one-time measurement. Correctly evaluating the status of an operation requires precise data on hours worked during each manufacturing operation.
That which we focus on here is log data collected from production lines in factories. In a factory, several million log data—which states when a production operation for making a certain part started and when it finished—are collected per day. We thought that by utilizing this data, we would be able to derive "the genuine time needed for making that part without the effect of variability." Since the log data is collected every day, this derivation can be done on the basis of new data every day.
NONAKAUsing the theory of statistics, we quantify (that is, model) the processing time taken by each operation. We call this method "variability modeling."
In Figure 2, the vertical axis represents information read from the log data, namely, the time taken performing a production operation. The horizontal axis represents the number of parts made during that operation.
Figure 2: Variability modeling (case: mass-produced products)
NONAKAEach dot on the graph is a value representing the actual result of work done by an individual worker or by a particular equipment. When looking at these dots, we can clearly see that there is a certain amount of time below which, regardless of the number of works processed, almost no dots exist. The trend line joining those time periods represents the raw ability of that production process in that factory. And towards the upper part of the graph, the dots become sparsely scattered. This scattering is the variability I mentioned. By incorporating the variability in our calculations, we could elicit the correct number of hours worked.
NONAKASuch a case is a factory making products for social infrastructures such as railway lines and power stations. In one year, since not that many products are made, even if we use the method based on statistical modeling that I just described, reliable values cannot be obtained because sample sizes are too small. With that problem in mind, we decided to use a statistical method called "multivariate analysis."
First, using data measured by the plant in the past, number of hours worked is estimated. For example, if it is supposed that certain operations, namely, "weld 1", "weld 2", "boring", and "bending", are performed in the manufacturing process for making a railway carriage, the operations will be performed in the manner of combinations that differ day by day. In other words, on the first day, the order of work might be "weld 1", "weld 2," "boring;" but on the second day, it might be "weld 1," "boring," "bending." In accordance with this situation, we take the hours worked measured in the plant in the past, and we estimate the hours worked per day by simply adding them up.
Meanwhile, the actual hours worked are reported by the factory every day. Thereupon, there is an error between the hours worked estimated by adding up and the actual hours worked as reported. When we repeat this survey day by day, this error veers from small to large and back on a daily basis. The size of the error is thought to depend on the extent of the difference between the combinations of operations. Given that fact, by comparing and analyzing various patterns of work combination works, we thought of eliciting a predicted value that contains a small error in relation to actual hours worked. That which applies this analysis is "multivariate analysis."
Using this method, we were able to make an estimation that pretty much fitted the true value even in cases in which the difference between the estimated hours worked and the actual hours worked was about two times.
Figure 3: Variability modeling (in the case of products manufactured in small quantity)
Figure 4: Accuracy of variability modeling (in the case of products manufactured in small quantity)
SUGINISHIAt the factory sites, there were people of many different positions (such as workers actually carrying out the operations and their supervisors), and each of them had a different viewpoint. As the side proposing the system, we felt the difficulty of making a proposal from such a broad standpoint. Personally, since I frequently visited factories to better understand the workplace, I spent more time in factories than in my laboratory.
NONAKAThe people actually working at the factory have hunches and knowledge from experience. That said, even if we simply say to workers that "It is now possible to simulate status of operations.", they wouldn't be satisfied without actually inspecting goods to see their condition for themselves. At the moment, we have created a system for updating real statuses on a daily basis, and this system is gaining acceptance from our customers. I think gaining the ability to run a factory while all staff—from the field workers to the plant manager—view the same data will make a great contribution to business.
SUGINISHIWhat's more, in the case of components of electrical devices, the generation of components are changing by improvement. This change is having the effect of significantly changing the contents of logs. And we have worked hard to devise schemes that can properly handle such rapidly changing environments.
SUGINISHISince it is said that the visualization of nonlinear phenomenon was difficult, on making it possible to simulate the circumstances in which a phenomenon propagates from the past into the future, I felt suitably inspired. And as far as I know, the paper concerning visualization of nonlinear phenomenon in production lines that Dr. NONAKA wrote in 2007 was the first of its kind in the world. Under the guidance of Dr. NONAKA, I tried various things, and I am really happy to have been involved in developing this cutting-edge technology.
NONAKAAt Hitachi Group, we are now implementing this technology at operations divisions making electrical devices and at factories making components for power-generation plant. We are also working on applying this technology to manufacturing facilities in other fields. And we are canvassing for sales to corporations other than Hitachi.
SUGINISHITo apply this technology to production lines in other fields, we face the challenge of modeling manufacturing configurations that differ from the types established up until now. I want to unfalteringly focus on that challenge from now onwards. Throughout developing technologies for handling various modeling types, we want to be able to provide an abundance of modeling types when selling packages.
What's more, from the broad viewpoint of "manufacturing," for example, I think we can apply this technology to the development of software. With that in mind, I will do my best to expand the application field of this technology.
(Publication: August 2, 2012)