The Next Phase of Enterprise Automation: Bridging the Edge MLOps Gap in Physical Infrastructure
Today’s manufacturing and logistics environments are full of data. Executives can see live wall dashboards showing all facilities across the world in a green light state. Hundreds of thousands of connected sensors monitor every tiny change in temperatures, vibrations and pressures within a facility. However, when an issue arises this becomes very slow and expensive.
This is called the Dashboard Blindspot. Internet of Things platforms are great for monitoring and highlighting anomalies in data. But once there is an issue with a vital asset, there is no automation. An alert will be highlighted but someone will need to manually check where the asset is, check inventory records and then decide how risky it is to action.
Leading manufacturing companies are moving away from dashboards to Physical AI to turn this data into real, tangible value.
Four Layers of Automation
To turn continuous data into deterministic actions we need to add four different engines to our stack.
- Identification & Location
We need to have an accurate understanding of what asset the alert relates to and where it is. Our Identification Engine will accurately map out all assets in real-time. Using RFID tags (active and passive), Ultra Wideband RTLS, Bluetooth low energy tags and visual markers we can understand exactly what the data relates to.
- Sensing
Now we know what asset is experiencing an issue we can start gathering data. Our Sensing Engine will gather high frequency data around the asset. This could be sound waves, vibrations, temperature, power usage and pressure. We can process and normalise this data at the edge to reduce latency of sending data to the cloud and reduce local noise.
- Reasoning
Our normalised data will now go through our local MLOps models. Instead of using simple threshold values, our AI Decision Engine will look at how patterns have changed across multiple variables. We will get an accurate score around how risky the issue is to operations.
- Physical AI Action
Finally we can take action. Our Physical AI Action Engine will take our high confidence decisions and turn them into physical actions. Based on our pre-validated safety parameters we will create top priority tickets, change the duty cycle of the machine or trigger safety interlocks.
Human in the Loop for Heavy Industry
The main issue with taking action physically is risk. If you are controlling heavy machinery, robots or high pressure systems without any limitations it can cause issues.
Within enterprise Physical AI we have set limitations on how much we can do based on our risk scores.
Low Risk Physical Action – Fully automated actions for low risk items. For example adding condition based inspections to your log, changing location tags or reordering components.
Medium Risk Physical Action – Recommendations to change something, however they will need to be manually confirmed by a one click action. For example changing the speed of a motor drive or changing the maintenance schedule.
High Risk Interlock – Hardcoded interlocks to ensure we are doing a deterministic action. If there is an issue with a key parameter we will automatically put the asset into a safe state and immediately alert the supervisors.
By splitting our intelligence into the engines we can remove the dashboard friction and have closed loop control. If you’re an engineer or founder looking to understand modern Physical AI and venture building models you can find research from Aperture Venture Studio.