Predictive maintenance


Discrete manufacturing use case

Without Falkonry

A semiconductor manufacturer operates complex equipment that executes many different types of step-based operations in the course of a day. These machines are expensive, and optimizing utilization through predictive maintenance is a high priority. The machines are instrumented to collect operational execution data every second in the form of sensor readings, control parameters, and other settings – use of this data to improve maintenance and operational efficiency; however, has been very limited. Maintenance is driven by periodic inspection and response to faults presented to operators.

Applying Falkonry ready-to-use machine learning

A Falkonry server is installed on-premises, and is connected to the manufacturer’s operational datastore. This datastore captures the following information:

  • Trace data from machine execution – approximately 100 different signals
  • Quality measurements
  • Inspection logs
  • Operator logs

A maintenance engineer identifies a step type that is projected to be a good indicator of machine health. The engineer then executes the following tasks with Falkonry:

  1. Configured a datastream for the chosen step type.
    >  The datastream has 15 signals and a groupID field to identify with step execution event each measurement is associated with
  2. Loaded a 4 month history of trace data for model development
  3. Used the inspection logs, operator logs, and quality measurements to label a small set of known conditions for selected step execution events.
    >  Periods of confirmed normality
    >  Two known abnormality conditions
    >  One known maintenance condition
  4. Created a model based on the first 2 months of the history data
  5. Tested the model on the second 2 months of history data
  6. Through iterative examination of results & log data and refinement of models through small additions of Fact labels the engineer identifies the following additional conditions:
    >  A maintenance condition associated with periodic calibration
    >  A condition that indicates a need for maintenance intervention.
    >  Note: this condition consistently appears prior to periods of frequent maintenance interventions
  7. Additional history data is loaded and the latest model is tested to confirm consistent predictions of maintenance intervention need
  8. The datastream is made live and the assessment result stream is fed back into the operational datastore
  9. The monitoring application is augmented to alert when the condition needs maintenance is seen from the Falkonry assessment stream

The availability of advance warning provided by the Falkonry assessment stream supports early intervention by the maintenance team and results in an improvement in uptime and Overall Equipment Effectiveness (OEE).

Ready to discuss your use case?