How it works
Liveline Technologies uses a highly automated data science process to quickly train and deploy advanced AI-based factory controllers and to keep them up-to-date as plant conditions evolve. The AI controllers can be used alone or in conjunction with rules-based controls specified by your factory experts.
Let’s look at how the AI works.
Online inference and control
Figure 1: The automated data science process at Liveline Technologies
Consider the image in Figure 1. On the right hand side is your factory. In this case we’re making espresso, and the factory consists of two processing stages: A coffee bean grinder and a brewing machine. Both pieces of equipment are connected to a standard PLC device.
Sensors in the equipment, such a temperature and pressure, are read by the PLC and passed to our proprietary transport layer, the Liveline Digital Controls Platform (LDCP). When running production with our AI-based controls, setpoint commands for the equipment, such as a new temperature, are pushed down to the equipment through the PLC.
The trained AI-based controller is contained in a Liveline Controls Package (LCP). The LCP is a secure, encrypted object that contains extensive functionality, such as performance monitoring and knowledge about physical safety limits for equipment settings. The LCP continuously observes the state of the process and issues new equipment setpoints as needed.
Human operators can monitor the process and communicate with the LCP through an interface, which can be implemented on any device as a responsive React app (contact us to discuss other customizations). Normally, the LCP makes changes to equipment setpoints without human intervention, but LCP controls can be toggled on and off by operators.
The LCP and LDCP run on generic computers located inside your factory (a.k.a. edge devices). Liveline can ship pre-configured server blades to your facility, or configure hardware on-site. Specifications for edge devices will depend on your performance requirements. The LDCP is highly scalable and can be extended quickly by adding additional devices.
Offline modeling
In addition to running online controls, the Liveline Digital Controls Platform (LDCP) also handles transport of data to Liveline’s secure cloud service that is instantiated specifically for you, so all of your data is completely isolated from other customers.
Once sufficiently historical data has accumulated in the cloud, we automatically initiate a patent-pending 2-step data science process.
The first step is to train an Liveline Physics Package (LPP) for each “recipe” at your factory. A recipe is a combination of product SKU and production line or cell. The LPPs contains a physics model that can predict the future evolution of desired outputs, such as product quality characteristics, by observing an incoming stream of sensor data from the equipment. LPPs are encrypted and secure.
The second step is to use the LPPs as simulation engines to train an Liveline Controls Package (LCP) for each recipe. The correct LCP is loaded to the edge device at your factory when starting production. From there, online controls are implemented as described above.
LPPs and LCPs are automatically re-trained as needed based on how they perform over time. It’s natural for model performance to drift as conditions change inside the plant. Common changes include substitutions of raw materials, modifications to the product design, and changes to the equipment. Re-training can also be initiated manually if the plant is aware of significant changes that will impact controls.
Predictive alarming
Our AI-based controllers are model-based, which means they constantly predict how production lines will evolve and make proactive changes to ensure better outcomes. Because of this, controllers are able to generate highly sophisticated predictive alarms when they predict that line conditions will evolve into unfavorable states that cannot be automatically mitigated.
In addition, because our controllers are model-based, it is easy to detect when conditions in the plant have changed, because model predictions will begin to drift away from observed signals. This is an excellent way to detect changes in raw materials, equipment and tooling health, and other factors.
For more information, see our product descriptions.
Why is it so fast?
Our AI-based controllers are the key to unlocking affordable and effective APC applications at a wide variety of manufacturers because they are fast. For Liveline Technologies, “fast” is defined in two dimensions:
Fast to deploy and maintain
Our proprietary data science techniques have achieved world-record levels of data efficiency and training speed. For our customers, this means we can train controllers with a surprisingly small amount of historical data — often just a few days’ worth.
It also means that models train quickly (hours), allowing us to rapidly scale up to cover the full range of products and lines within a factory. Fast training also allows us to touch-up or retrain models as needed when conditions inside the plant change.
This is possible due to some deep technological advances made by our team:
New algorithms for training RL agents that break world records in AI.
Optimized implementation of new algorithms with hardcore computer science.
Optimized methods for real time signal processing and feature engineering.
Novel methods for auto-selection of rich data to include in training sets.
It’s also important to note that our controllers are relatively narrow in scope compared with something like full self-driving for a car. Controllers only need to handle conditions that will actually be observed in the production environment. This is another reason we don’t require copious amounts of historical data. To use a car analogy, we’re implementing highway cruise control, not full self-driving.
Fast to execute
Once AI controllers have been trained, they must be able to keep pace with production and execute setpoint changes on equipment in real time. Our entire approach was designed around this requirement, and this has influenced choices we made about feature engineering, modeling techniques, our streaming data platform, and overall stack architecture.
The result is that we can execute our AI control loops on the order of 1 second. Every second, our AI observes plant conditions and makes any changes necessary to ensure optimal production, all the time.
Your application may not require such frequent adjustments, but we’re ready to support a huge variety of industrial processes.
Working with rules
The Liveline Digital Controls Platform (LDCP) provides the capability for your process experts to create rules for controls. Rules can include If This Then That (ITTT) statements or classical PID controls.
The LDCP is built with a service-oriented architecture and our AI controls and rules are implemented as individual services. This allows an outstanding degree of flexibility and scalability. Rules can be used alone or in conjunction with AI controllers. In fact, any number of rules and controllers can be applied to individual production lines or cells.