Sensor Fusion is an interesting topic for automation and manufacturing. The concept is rooted in the 1960s and is now reaching a new level. Data fusion is closely connected to the discussion around a sensor grid, which describes a network of (wireless) sensors, the internet of things, or big data.
- Why this could be useful?
- How is the data combined?
- Where to fuse the data?
- Do we need to store all this data? If yes, for how long?
Why Sensor Fusion could be useful for manufacturing?
Imaging that on object or batch is recognized, identified, and controlled by various machines and in various steps of production. At a later point in the production process, a large amount of information will become available, but because much of this does not pass the controller of a given application/machine, it is not shared. If the information were available, production could adapt to it, inspection systems could be streamlined and become cheaper, and waste could be reduced.
- Central Limit Theorem: Here, the final error of the result is the linear combination of the different measurements weighted by their noise variance. Obviously, the higher the variance, the lower the weight.
- Kalman Filter: The Kalman Filter weights different results also by their weights, but includes all other results, as soon as one measuremet is noise-free.
- Baysian Network
ARC’s views inspection and quality control as the main application of sensor fusion in a plant or factory. As much of the data will be stored and handled, we recommend simple decision making, which favors the Kalman Filter. Also, the Kalman filter has been prove to be mathematically optimal.
The process of sensor fusion includes different levels:
- Level 0: Data alignment – Collects the data and brings it into a standardized form to be processes and combined. This is done at or near the sensor level, but should be done at the controller layer, as the sensor often only deliver an on/off or analog signal, which cannot be quantified or combined.
- Level 1: Entity Assessment – Includes new inputs about an object and aligns them with existing information. This is important to keep the dataset small and to reduce redundancies. Information conflicts are resolved. This could be done at the line-controller or HMI layer, depending on which information is added.
- Level 2: Situation Assessment – At this level, data is aggregated and knowledge of the situation and the process is taken into account. Depending on the purpose of sensor fusion, this could be done at the controller or HMI/CPM layer.
- Level 3: Impact Assessment – This is the first level, where no data is included, or aggregated, but is clearly “cognitive”, so algorithms and calculations derive information for decision making from the data. Depending on the purpose of sensor fusion, this could be done at the controller or HMI/CPM layer. Level 4 and 5 are part of the setup or the HMI/CPM layer and are not part of daily manufacturing routine.
- Level 4: Process Refinement – This level looks at the fusion process in order to evaluate it in terms of accuracy, timeliness, etc.
- Level 5: User refinement – This level brings the information to the user/operator.
As all data are collected from heterogeneous and remote sources, this system requires a certain degree of openness, so cyber security issues need to be taken into account.
Where to Fuse?
For OEMs and automation suppliers, a central question is, on which layer should the fusion be performed? Either the clients (sensor) fuses all data remotely, or the data is fused centrally in the controller (PLC, DCS, or IPC). An important aspect is that sensor fusion typically also includes historical data, which means that data from historians need to be linked to current information. So next to the controller layer, the HMI layer is also a potential layer to fuse the data. When the data also include product quality data, the may need to be stored for a longer time and linked to the CPM layer.
Looking at the calculations and data necessary to do sensor fusion, the controller or HMI/CPM layer are the most important layers, as the majority of sensors in a factory is not capable of doing it. In case the sensors have a built-in DA converter, are I/O Link slaves, or network enabled, sensors could contribute to Level 0.
Do we need to store all this data? If yes for how long?
Probably yes and probably for a longer time. For example, the food beverage industry needs to take into account FDA regulations. Sensor Fusion can contribute to process analytics technology (PAT), which can help turn batch manufacturing into continuous processes. In general, quality control, predictive maintenance, energy management, or condition monitoring require a long time storage of data.
The standard process of sensor fusion also includes the integration of existing data, and historical information used can improve the overall accuracy of data and decision making.