Dr Genichi Taguchi

Rise to fame
Taguchi's message

Rise to fame

In 1950 he joined the newly founded Electrical Communications Laboratory of the Nippon Telephone and Telegraph Company with the purpose of increasing the productivity of its R and D activities by training engineers in effective techniques. He stayed for more than 12 years, during which period he began to develop his methods.

Whilst working at the Electrical Communications Laboratory, he consulted widely amongst Japanese industry. Accordingly, Japanese companies began applying Taguchi methods extensively from the early 1950s, including Toyota and its subsidiaries. His first book, which introduced orthogonal arrays, was published in 1951.

In 1954-5 Taguchi was visiting Professor at the Indian Statistical Institute. During this visit he met the famous statisticians R A Fisher and Walter A Shewhart.

In 1957-8 he published the first version of his two-volume book on Design of Experiments. His first visit to the United States was in 1962 as Visiting Research Associate at Princeton University, during which time he visited the AT & T Bell Laboratories. At Princeton, Taguchi was hosted by the eminent statistician John Tukey who arranged for him to work with the industrial statisticians at Bell Laboratories. In 1962 he was awarded his PhD by Kyushu University.

In 1964 Taguchi became a Professor at Aoyama Gakuin University in Tokyo, a position he held until 1982. In 1966 Taguchi and several co-authors wrote Management by Total Results which was translated into Chinese by Yuin Wu. At this stage, Taguchi's methods were still essentially unknown in the West, although applications were taking place in Taiwan and India. In this period and throughout the 1970s most applications of his methods were on production processes, the shift to product design being in the last decade.

In the early 1970s Taguchi developed the concept of the Quality Loss Function. He published two other books in the 1970s and the third (current) edition of Design of Experiments. By the late 1970s Taguchi had an impressive record in Japan having won the Deming application prize in 1960 and Deming awards for literature on quality in 1951 and 1953.

In 1980 Taguchi was invited by Yuin Wu, who had emigrated to the United States, to give a presentation at his company. By this time Taguchi was director of the Japanese Academy of Quality. During his visit he arranged to revisit AT & T Bell Laboratories at his own cost where he was hosted by Madhav Phadke. Despite communication problems, successful experiments were run that established Taguchi methods within Bell Laboratories.

Following his 1980 visit to the United States, more and more American manufacturers implemented Taguchi's methodology. Despite an adverse reaction among American statisticians at the methods, and possibly at the way they were being marketed, major US companies became involved in the methods including Xerox, Ford and ITT.

In 1982 Taguchi became an advisor at the Japanese Standards Association. In 1984 he again won the Deming award for literature on quality. In 1986 he was awarded the Willard F Rockwell Medal by the International Technology Institute. With one or two notable exceptions, such as Lucas, his methods had made little impact on Europe until the Institute of Statisticians organised the first conference on the methods in 1987 in London. The UK Taguchi Club (now the Quality Methods Association) was formed later that year.

Taguchi's message

In contrast to Western definitions, Taguchi works in terms of quality loss rather than quality. This is defined as 'loss imparted by the product to society from the time the product is shipped'. This loss includes not only the loss to the company through costs of reworking or scrapping, maintenance costs, downtime due to equipment failure and warranty claims, but also costs to the customer through poor product performance and reliability, leading to further losses to the manufacturer as his market share falls. Taking a target value for the quality characteristic under consideration as the best possible value of this characteristic, Taguchi associates a simple quadratic loss function with deviations from this target.

This loss function shows that a reduction in variability about the target leads to a decrease in loss and a subsequent increase in quality.

With this conception a loss will occur even when the product is within the specification allowed, but is minimal when the product is on target. (If the quality characteristic or response is required to be maximised, eg strength, or minimised, eg shrinkage, then the loss function becomes a half-parabola.) The loss function may be used to evaluate design decisions on a financial basis to decide whether additional costs in production will actually prove to be worthwhile in the market place.

Taguchi methodology can be applied off-line in design or on-line in production.

[Fig] Taguchi breaks down off-line quality control into three stages:

1. System design.

2. Parameter design.

3. Tolerance design.

Parameter design is the crucial step - this is where the Japanese excel at achieving high quality levels without an increase in cost. The nominal design features or process factor levels selected are tested and the combination of product parameter levels or process operating levels least sensitive to changes in environmental conditions and other uncontrollable (noise) factors is determined.

Finally, tolerance design is employed to reduce variation further if required, by tightening the tolerance on those factors shown to have a large impact on variation. This is the stage at which, by utilising the loss function, more money is spent if necessary buying better materials or equipment, emphasising the Japanese philosophy of invest last not invest first.

The potential for these methods within UK and world industry is large. Typically, designs and line calibrations are in reality far from optimal. Much manufacturing folklore is based on the need to 'twiddle' important parameters or settings. Typically we do not understand the correct settings although we do have our prejudices.

Taguchi methodology is fundamentally a prototyping method that enables the engineer or designer to identify the optimal settings to produce a robust product which can survive manufacturing time after time, piece after piece, in order to provide the functionality required by the customer.

There are perhaps two major features of the advantage of Taguchi methodology. Firstly, it was developed by, and is largely used by engineers rather than statisticians. This removes most of the communication gap and the problems of language traditionally associated with many statistical methodologies. In addition, it means that it is tailored directly to the engineering context. The consequence of this is that the importance of noise variables which disrupt production must be considered in addition to the control variables introduced. Optimising a product corresponds not only to getting its quality characteristics on target but also to minimising variability away from that target on a piece-to-piece or time-to-time basis. This is the connection with Statistical Process Control (SPC).

Taguchi may be used to narrow the spread of the quality characteristics distribution and to identify variables to build control on. SPC may then be used to keep quality characteristics on target by making use of the known spread about the target value. Essentially this is the other novel feature of Taguchi methodology: the use of the so-called Signal-To-Noise ratio to choose the control setting that minimises the sensitivity to noise. In addition to this the methods are fundamentally evolutionary.

One major feature, however, is the codifying by Taguchi of the so-called Orthogonal Arrays. These are designs for experiments which were largely previously identified by others but are codified by Taguchi in such a way that an engineer automatically has a route to the minimum number of prototypes necessary for experimentation. The numbers are kept deliberately small by sacrificing all the interaction information that may be present in the design (or almost all of it) since information about interactions can subsequently be found in typical industrial applications by evaluating one more prototype - that corresponding to the predicted optimum setting (the confirmatory trial). This is the difference between industrial application and the agricultural context on which most of the Western statistical methods which foreran Taguchi were based. In agriculture, responses are slow so that leaving out prototype combinations and sacrificing interactions would necessitate an extra year in the agricultural cycle in order to be able to verify that the predicted prototype combination really was best. In the industrial setting responses are usually fast, so that it is possible to go back immediately and try out one additional prototype. Interactions can, however, be incorporated into Taguchi methodology and he presents a simple graphical codification of these (the linear graphs) to enable the analyst to introduce them systematically and easily. However, only limited numbers can be conveniently introduced without leading to great increase in prototype or experimental sizes.

Back to the Business support page

Last Revised: Fri Feb 13 11:58:34 1998