To best leverage manufacturing data, it first pays to collect the right data. One inclination might be to measure “everything” and figure out what it means later. Most organizations can’t afford this approach, either in terms of sensor capital cost, or the IT infrastructure and human resources to manage it all. Instead, a targeted approach of focusing on the right data to collect at the right time for the right reasons will ultimately pay out instead of a “shotgun” approach. Here are some ways to bring focus to the effort involved with collecting and analyzing data and making business decisions based on the insights gained.
Explore the Data You Have
Existing data can sometime be mined to understand the current state of a manufacturing process. Some thought was already put into measuring something about the process, after all. It might not be the right data, but a brief analysis of the process capability and stability over time with the measurements that are available might provide some insights into how much variance and mean shifts are happening. Perhaps there are data between similar production units and these data can be mined to look for differences between operating crews, production lines, or plants in the aggregate. Regression models might be possible for process variables versus product quality data. The key might be to find a problem or anomaly that sparks greater investigation. Ideally these explorations would be guided somewhat by known or perceived problems from operations, such as high waste or downtime areas.
Explore the Data You Should Have Based On Engineering Models
If the results of regression models reveal you are not explaining a sufficient amount of the variability in your Key Performance Indicator (KPI), that’s a sign you are not measuring potentially the right things. One way to check is to compare your data model to first principles model and computer simulations of the physics and chemistry involved in your process. For example, for a web converting process, it is known that web tension/strain control is critical to avoiding wrinkles, mistracking, and registration errors and there have been published reports from Oklahoma State University’s Web Handling Research Center and other researchers on the dynamics of dancers, accumulators and driven rollers and the effects on web tension and strain, so these models can be used to provide an expected response from the production process and provide ideas for the critical variables to measure.
For a given Engineering model and a production process with data based on the variables guided by the model, the process data may be compared with the model prediction and the residuals charted to understand where the process deviates from the expected behavior. If it isn’t known for a new process what should be measured, even basic first principles models based on conservation of mass and energy can provide insights into what to measure. Sometimes a variable can’t be measured directly because the sensor to do it is too expensive or the method is too difficult or time-consuming and so-called “soft sensors”, that are combinations of other measured variables can be used, and models can guide the selection of these other sensors and their required discrimination power.
Be sure to include the knowledge from experienced operators as data
Finally, since all models are representations of reality and all data models are incomplete, another source of data to understanding your process are those who work with it every day. Operators may not always know why something behaves the way it does, but they see lots of effects and they run mini-experiments on an on-going basis as conditions degrade over time. Interviewing them and spending time with them operating a process can be invaluable in augmenting the theory and the data analytics with other clues for key variables to add or include in the hybrid model of data/engineering model/human experience.