Optimising Measurement Processes Using Automated Planning S. Parkinson, A. Crampton, A. P. Longstaff Department of Informatics, School of Computing and Engineering University of Huddersfield s.parkinson@hud.ac.uk 1
Introduction Measurement processes often contain multiple measurements. E.g. calibration of precision machine tools. The processes estimated uncertainty of measurement can be affected by the sequence of measurements due to order dependencies. Dynamic Environment can further affect variation. E.g. Environmental temperature variation. 2
Introduction Choice of Equipment and Measurement Method More than one way to measure something, but they could have different affects on uncertainty. E.g. Squareness of two perpendicular axes of a machine tool. Renishaw XL-80 Renishaw Ballbar Granite Square/DTI 3
Introduction Each different instrumentation/method can have different temporal properties. Absence of method to reduce estimated uncertainty for entire measurement processes. GUM and PUMA do not provide the entire solution. All these factors make planning measurement sequences to reduce time and/or uncertainty challenging. Expert knowledge is only part of the solution. 4
Automated Planning as a Solution Automated planning and scheduling is a branch of Artificial Intelligence that is concerned with the realisation of strategies or action sequences. Two well known implementations: Mars Rover Drone 5
Automated Planning as a Solution Planning knowledge can be encoded into a domain model and a problem definition. Planning Domain Definition Language (PDDL) developed by the AI planning community. Domain encodes actions. Actions encode events that change the current state (e.g. performing a measurement). Problem specifies initial and goal state as well as the optimising metric. The domain and problem can then be interpreted by state-ofthe-art planners. Solutions found to satisfy the goal state, and/or minimise a specified metric. 6
Encoding Knowledge Knowledge engineering is the abstraction and formulation of the domain model and problem definition. The level of abstraction depends on the complexity of the problem. Requires AI planning knowledge. Graphical tools are available to aid with the process. Software engineering knowledge is required (UML). Current tools fall short of allowing access to AI planning for all. We are working on tools to expose AP to metrologists. 7
Benefits to Metrology Measurement plans can be produced by non-experts. Once domain knowledge has been encoded, plans can be generated by anyone and computationally validated. KE tools are making it easier to extract domain knowledge with the appropriate level of abstraction. Plans can be optimised based on a specified metric. Individual. Multi-objective. Quick plan generation Plans can be regenerated during the measurement process. 8
Example: Machine Tool Calibration A machine tool is a mechanically powered device used for manufacturing components. Machine tools are typically designed for optimum manufacturing of a small range of components. Hence the variety of different machines. Optimised in terms of size, material and power. 9
Example: Machine Tool Calibration All machine tools have motion errors. Machine tool calibration is the process of examining a machine s behaviour using a standard set of tests. From this, we can establish the machine s predictability. Modern emphasis is on improving machine tool accuracy for manufacturing to small tolerances. 10
Example: Machine Tool Calibration Temporal Strong desire on reducing machine down-time. Measurement quality Machine s configuration, environment and use determine the motion errors and their significance. Available measurement instrumentation and test methods. 11
Example: Domain Modelling PDDL 2.1 domain Five actions that represent: 1. Setting up the equipment. 2. Adjusting the equipment to measure a different error. 3. Measure an error that is not influenced by any others. 4. Measure an error which is influenced by other errors. 5. Remove the measurement setup. 12
Example: Temporal Optimisation Selecting the quickest equipment. Providing it can satisfy the requirements (E.g. accuracy, resolution.). Sequence measurements that use the same equipment. If possible, perform concurrent measurements. Providing that measurement characteristics agree (E.g. Number of samples, feedrate, etc.). 13
Example: Temporal Optimisation Temporal model (PDDL2.1) optimisation based on: Selection of instrumentation for each error component. Scheduling measurements to minimise instrumentation setup time. Resulted in a 9% reduction in downtime when compared to expert generated plans. (12:30 to 11:18). Some measurements are still performed sequentially, but where possible concurrent measurements have been planned. Grouped by axis. Optimisation has reduced instrumentation setup. 14 Parkinson, S.; Longstaff, A. P.; Crampton, A.; and Gregory,. The application of automated planning to machine tool calibration. Proceedings of the Twenty-Second International Conference on Automated Planning and Scheduling, ICAPS 2012. Sao Paulo, Brazil
Temperature Example: Uncertainty Optimisation The temporal reduction model was then extended to encode the necessary knowledge for reducing the estimated uncertainty of measurement. This involves considering: 1. Selecting measurement equipment with the lowest uncertainty. 2. Schedule the measurement where the effect of environmental temperature is minimised. 3. Minimising the change in environmental temperature when considering interrelated measurements. Long measurement Interrelated measurements 15 Short measurement Time
Example: Uncertainty Optimisation 1) Temperature variation 0.3 C. 2) Increased downtime of 1 hour, but a 31% reduction in the uncertainty of measurement from 235µm to 161µm. 16
Example: Multi-objective Uncertainty due to plan schedule: 183µm Duration: 82:20 Temporal Measurements that can use the same instrumentation are scheduled sequentially. Estimated Uncertainty Three interrelated measurements are scheduled, but the change in environmental temperature is not optimal. 17
Example: Multi-objective Graph shows average metric values for all 12 problem instances. Results show a good compromise when multi-objective optimisation is shown. 18
Example: Conclusion We have established that it is possible to: Create valid calibration plans without the requirement of a domain expert. Reduce machine tool downtime (9% over expert generated plans). Reduce the estimated uncertainty of measurement (31% reduction from the worse-know solution). Perform multi-objective optimisation (on average 9% worst than the best-known solution). 19
Future Challenges Tools to enable metrologists to utilise AP. Suitable knowledge engineering tools. Development and integration of state-of-the-art planning technology. Suitable guides for reference. A wider variety of metrology planning applications. We need more applications to develop more powerful tools. Better validation of the proposed technology. Better understanding of the impact on metrology processes. 20
Thanks for listening Any Questions? s.parkinson@hud.ac.uk 21