Rethinking the definition of set-up time in SMED using artificial intelligence

As a rule, set-up times account for a large proportion of plant productivity losses. The Toyota production system dedicates an important chapter to reducing set-up losses using the SMED method. But how precisely can a set-up time be measured or how do we assess the quality of a set-up process? AI can help in the future.

A practicable definition of set-up time is that the set-up time is between the last product of the previous lot and the first product of the sequential lot. If a plant produces at either 0 or 100% of the output, this definition is valid. However, there are many systems where the quantity curve is different. In operational practice, there are often downtime and start-up losses before or after the actual downtime. A more precise definition of the set-up time would be, for example, the time between the last product of the previous lot at full speed up to the sequential product at full speed to be defined as set-up time.

The typical course of a quantity curve before and after a set-up process is shown in the following figure.

If you add the expectation of full speed to the set-up time definition, the three time components of the rundown, standstill and start-up time are blurred, so that this definition is also not ideal. With its classification capabilities of time series data, artificial intelligence can provide clarification here.

With the help of a trained neural network, deep learning can classify time series data as stable and unstable. This division forms the basis for a precise breakdown of the time components of the set-up process. If the system leaves the stable run (= ridge line), it is in the outlet up to the quantity of “zero”. This is when the standstill begins, during which the conversion takes place. From the start of production and until the stable production level is reached again, the period is defined as a start.

In practice, this data-based classification offers several advantages:

• Even with systems that do not produce an identical number of pieces every minute in full production, the time when the ridge line is left or reached can be clearly defined by the algorithm.
• The run-out time (or outlet angle), and thus the management of the process, can be precisely measured and improved.
• The start-up time (or the starting angle) is a measure of the conversion quality and can be measured and improved precisely.
• Both runoff and start-up losses can be expressed precisely in lost units or lost OEE percentage points.

The following illustration visualizes the points discussed

This procedure is not a solution for advocates of note and pen solutions. However, it creates facts in places that could not be measured without advanced analytics in the past. In this way, you can free yourself from the discussion about the assessment of the facts and get to the actual processing of the process improvement.