Optimizing execution of multiple machine learning models over a single edge device

A non-transitory computer-readable medium, storing instructions that, when executed by one or more processors of an edge gateway managing plurality of sensors, the plurality of sensors observing a plurality of manufacturing process steps sequentially performed by at least one physical manufacturing asset cause the one or more processors to:

• receive sensor data generated by the plurality of sensors;
• select, based on available computing resources of the edge gateway, a subset of the plurality of manufacturing process steps as first selected steps for observation by a subset of the plurality of sensors, wherein (i) each manufacturing process step of the first selected steps is observed by a respective sensor of the subset of the plurality of sensors and (ii) a number of the first selected steps is less than a total number of the plurality of manufacturing process steps;
• execute anomaly detection on the sensor data from the subset of the plurality of sensors using at least one machine learning model executed by a machine-learning inference module;
• in response to detection of an anomaly from a sensor of the subset of the plurality of sensors:
• select each manufacturing process step in which the anomaly is detected and all unobserved previous manufacturing process steps between the each manufacturing process step in which the anomaly is detected and a previously observed manufacturing process step as second selected steps;
• estimate a probability of anomaly occurrence for the second selected steps;
• and
• in response to the estimated probability of anomaly occurrence meeting a predetermined criterion based on a statistical distance between a reference probability distribution and the estimated probability of anomaly occurrence, select at least one of the second selected steps for observation at a second instance;
• and
• output a control signal to the at least one physical manufacturing asset associated with the at least one of the second selected steps to modify operation of the at least one physical manufacturing asset by discarding a part associated with the detected anomaly or routing the part to a secondary inspection line.

An apparatus for managing a system comprising an edge gateway, a plurality of sensors, and at least one physical manufacturing asset that performs a plurality of manufacturing process steps, the apparatus comprising:

• a memory configured to store instructions;
• and
• one or more processors configured to execute the instructions to;
• receive sensor data generated by the plurality of sensors;
• select, based on available computing resources of the edge gateway, a subset of the plurality of manufacturing process steps as first selected steps for observation by a subset of the plurality of sensors, wherein (i) each manufacturing process step of the first selected steps is observed by a respective sensor of the subset of the plurality of sensors and (ii) a number of the first selected steps is less than a total number of the plurality of manufacturing process steps;
• execute anomaly detection on the sensor data from the subset of the plurality of sensors using at least one machine learning model executed by a machine-learning inference module;
• in response to detection of an anomaly from a sensor of the subset of the plurality of sensors, wherein the detected anomaly may have originated from an unobserved previous manufacturing process step rather than the currently observed step:
• select each manufacturing process step in which the anomaly is detected and all unobserved previous manufacturing process steps between the each manufacturing process steps in which the anomaly is detected and a previously observed manufacturing process step as second selected steps;
• estimate a probability of anomaly occurrence for the second selected steps;
• and
• in response to the estimated probability of anomaly occurrence meeting a predetermined criterion based on a statistical distance between a reference probability distribution and the estimated probability of anomaly occurrence, select at least one of the second selected steps for observation at a second instance;
• and
• generate a control signal to the physical manufacturing asset associated with the at least one of the second selected steps to modify operation of the physical manufacturing asset by discarding a part associated with the detected anomaly or routing the part to a secondary inspection line.