Autonomy for industrial control systems

Machine calibration was a manual, time intensive process, often performed by an out-of-town expert, that requires machine downtime.

A manual process requiring expert human operators averaging 20–25 iterative steps over 2+ hours.

Achieved 2 micron precision at an average of 4–5 iterative steps over 13 seconds. System achieved superhuman precision (<1 micron) in ~10 iterative steps.

Silicon epitaxy is grown in thermal reactors which must be tightly controlled for uniform heating. Different phases of the process require different heat lamp settings to maintain optimal conditions in the reactor (called a recipe).

PID controller provides error correction in order to manage consistent temperature. PID singularly focused on heat distribution and doesn’t consider optimization goals of speed or energy use.

Minimized time and cost of growing wafer substrate above a target quality consistency.

Some repetitive robotic tasks require complex learned behaviors (such as grasping and stacking) which are extremely difficult to program and automate.

Extensive programming and testing applied to limited scenarios that require specialized skills and time-consuming process to automate only subset of tasks.

Established a new benchmark for programming industrial control systems. The Project Bonsai machine teaching technique imported existing controls, then trained learned sub-concepts separately, and finally trained the overall “grasp and stack” task that performed 45x faster.

A consumer-packaged goods manufacturer produces food using >50 extrusion lines. Each extruder is different and extruder behavior and process inputs change over time.

Operators manually control the extruder. Changes in the feed composition and moisture content make it very difficult for human operators or existing control systems to extrude puffs of consistent quality.

Maximized the throughput of product produced while maintaining the quality standard.

Time required to achieve target polymer quality varies dramatically based on operator experience and dynamic ambient conditions.

Lengthy equipment calibration process required for each batch, and extensive experience required to control highly dynamic reaction.

Reduced the amount of time required to calibrate the simulation from several months to a few weeks.

Reduced the loss of non-prime grade polymer by 5%.

Inefficient storage and retrieval decisions can have severe negative financial impacts.

Human operators and rules-based algorithms dictate storage location upon loading, and retrieval policies upon shipment. Manual methods are static and do not adapt to changing customer and market demands.

Balance warehouse fill density and labor efficiency. Balance fill density with long term throughput. Maximize financial return by choosing optimal actions in varying warehouse states and inventory storage patterns.

Many to many relationships between distribution centers and local supply locations force complex allocation decisions. Poor regional replenishment decisions lead to local stock out conditions and lost sales.

Linear optimization decisions using SAP APO. Struggles to adapt to forecast inaccuracies and rapidly changing local demand.

Decompose inventory allocation options into product and distribution center combinations. Simulate and train AI against dramatically different and changing demand profiles. Combine different inventory allocation alternatives into master optimization strategy.

Maximize rate and volume of oil extraction from subsurface asset via horizontal drilling.

Human operators follow and iteratively adjust the prescribed well plan based on static geophysical models. Inaccuracies in geophysical models require constant adjustments based on imprecise sensor data, reducing percentage of time drill is extracting oil on plan.

Eclipsed human-like performance within one week.  Drill bit never deviated more than 2’ from optimal path. Reduced time to complete drilling operation more than 20%.

Maximize energy output of wind farm by controlling operating parameters of individual turbines.

Human operators supported by custom optimization algorithms. Human operators struggle to comprehend voluminous sensor input, and algorithms that model dynamic operating conditions are still emerging.

Trained BRAIN to match controls benchmarks in 3 hours and extended approach to other five control parameters combining them across the entire wind farm.

Inconsistent production yield of land or offshore asset because of complexities in choke valve control tactics.

Production engineers and operators make control decisions several times per day based on pressures and flows. Extensive experience required for veteran production engineers to understand and optimize complex interactions, resulting in dramatic output differences.

Automated recommendations 24x7 that produce more output than Offshore Control Room Operator recommendations. Automatically operate choke valves for an oil field from control room to maximize oil production.

While conference rooms are unoccupied about half of the time, HVAC systems are predominantly operated by simple PI thermostats, consuming energy when unnecessary, and sometimes violating CO2 safety levels.

PID corrections are reactive, while MPC controllers consider future state but still struggle to find global cost optimization behavior. Both approaches require time consuming commissioning process to tune controller behavior.

The autonomous system learned to control HVAC for improved comfort and air quality with ~23% lower energy consumption.

Sub-optimal control of the SAG mill grinding process increases energy cost, long term machine cost, and potential disruptions in the downstream mining process.

Humans monitor operational process, control machine set-points, and episodically add steel balls to the grinder. Human operators struggle to optimize processing of varying particle size distributions as well as hardness of input ore.

Autonomously control input feed tonnage, water addition, mill rotation speed, material recirculation rates and recommend episodic ball introduction actions.

Sub-optimal control of the initial rock crushing process increases energy cost, long term machine cost, and potential disruptions in the downstream mining process.

Humans monitor operational process, periodically changing material feed speeds and machine set-points, struggling to optimize processing of varying particle size distributions as well as hardness of input ore.

Autonomously control conveyor speed, material feed speed, and throat crusher gap.

An operator must tune a PID controller for each piece of equipment & cutting material. Each new piece of equipment must be tuned separately. This is a repetitive and time intensive process.

Currently dozer blade lift and lower is controlled by 5 PD parameters. The operator may need to increase or decrease the parameters on-site if the default settings for the PID controller are not suitable. PD controllers take time to tune, requiring different tuning settings for each dozer model and for some cutting materials.

Achieved a waviness number (from the cut) that exceeded the PID benchmark across multiple dozer models at multiple speeds.