Advanced Analytics in Manufacturing

Data driven decision making has a long history in manufacturing. Indispensable to any introductory level statistics course, the “t-test” is the result of data driven quality assurance procedures developed at Guinness brewery in early 1900’s. In the following years various manufacturing-oriented analytic techniques have been developed and used. Statistical process control, for example, is still one of the key analytical tools in manufacturing. However, compared to other industries manufacturing seems to be lagging behind in adopting advanced analytical tools which have been rapidly growing and evolving since early 2000’s. We will outline the difficulties in developing and deploying advanced analytical solutions in manufacturing.

Data Content

It is common knowledge that we do need data for analytic solutions but which data, at which precision and resolution? There are no one size fits all type of answers to this question. It depends very much on the type of problem you want to solve and the manufacturing process at hand. We can, however, generally say that binary variables are the least useful for modern analytic techniques. Useful in certain data driven approaches, binary variables — such as a variable indicating whether or not a certain production parameter is within specification limits — are usually not quite as valuable as continuous variables in modern analytic models. Data collection cannot be considered as a separate issue from the questions we are trying to answer and the relevant domain expertise. Temperature data sampled once a minute, reliable up to 1 degree can be useful in some manufacturing problems in some areas but we can’t expect it to be sufficient for any analytic solution we want to develop.

Data Infrastructure

The high-level architecture of analytic solutions in manufacturing is a 3-layer infrastructure which can be visualized as below.

The data is produced at the physical data layer through IoT devices, PLCs etc. and transferred to various big data infrastructures or cloud platforms. The analytical layer consumes the data from the digital data layer and produces the results — usually these results are fed back into some visualization/automation tool, which is omitted here for simplicity. When we go into the details of this architecture, there are many different tools to choose from each of which are optimized for different tasks. These choices should be made according to the needs of the analytical solution and how it will be used in the field. Each component has to work without compromising security, reliability and timeliness. Most manufacturing plants work on tight schedules and as such all the components in the preferred architecture should be closely monitored to provide uninterrupted service. We should also note that in the recent years, various applications which require high-speed implementations are moving towards edge-analytics, that is the analytic layer is actually located right next to the physical data layer.

Analytic Methods

Even if we have the right data content and an appropriate analytic architecture, the results rely on the analytic methods employed in solving the problem at hand. A significant source of complexity in manufacturing analytics is that the data generated by manufacturing processes often have a time-series structure. Time-series data points are not independent of each other, which renders many techniques inapplicable. Solutions which disregard this fact fail to deliver high impact solutions. Another common issue with manufacturing data is the problem of imbalance. For example, when monitoring an asset for predictive maintenance which breaks down rarely the data we obtain contains very few points indicating unusual behaviour. This makes using supervised learning methods difficult to apply. We can use data augmentation methods to resolve this issue or switch to unsupervised learning methods such as anomaly detection.

Sustainability

Analytic solutions need to be up and running just as much as the production line itself. To assure this, the analytic infrastructure should be monitored continuously for reliability and maintenance strategies should be set in place. Moreover, production processes evolve both on their own and according to the outcomes of analytic models. This may result in data drift which in turn may render analytic model outcomes unreliable. Both model and process outcomes should be monitored closely to avoid such cases.