Thank you for the interesting question. The model used for the virtual factory blends attributes from discrete events simulation as well as agent-based modelling. The digital twin is an analytic model in MATLAB. As for the connection, we deliberately used rudimentary data transfer techniques between the virtual factory and its digital twin to replicate the potential sensing and communication issues that would be observed in a small to medium enterprise. I would classify this as an exploratory study where the theoretical boundaries of the potential benefits are characterised. So it is somewhat beyond concept definition but not quite at the industrial implementation level.
Dear Aydin, thank you for your very interesting paper presentation on Digital Twins predictive maintenance of machines used for manufacturing CFRP bicycle frames. I should be interested in knowing what type of sensors were considered to provide signals to the sensorial data processing algorithm based on machine learning. Are these sensors already available in the machines or are they “external” sensors that need to be bought and installed on the existing machines in the factory?
Thank you for the interesting question. The short answer is a mix of both.
To provide more detail: data acquisition was initiated by first breaking down the factory into processes. The machines required for each process were found; processes that did not require machines were eliminated. For each machine the dominant types of failure were found. A criticality matrix was used to compare the failure rate to the severity of the fault and impacts of no action; failures with low criticality were eliminated. The data used to identity this fault type was then found. To assess the suitability of sensing this data for PDM, fault types were chosen that satisfied two main characteristics: ‘Is it appropriate to sense this data through technology over using manual inspection?’ and ‘Will a sensor be able to predict the failure in an appropriate time scale?’. For the fault types that satisfied this, it was decided if sensors needed to be built into the machine or if the information was already being sensed in the machine. The chosen sensed data was then taken forward into the factory design.
The results were as follows:
CNC Machine: Misalignment of cutting head – camera/image recognition, dulling of cutting blade – motor load sensor, bearing issues – low impedance accelerometers, loss of lubrication – low impedance accelerometer, loss of hydraulic pressure – pressure sensor.
Professor Nassehi, a very complex case study is analyzed well. Do you implement hybrid modelling in your case or is it a concept definition?
Thank you for the interesting question. The model used for the virtual factory blends attributes from discrete events simulation as well as agent-based modelling. The digital twin is an analytic model in MATLAB. As for the connection, we deliberately used rudimentary data transfer techniques between the virtual factory and its digital twin to replicate the potential sensing and communication issues that would be observed in a small to medium enterprise. I would classify this as an exploratory study where the theoretical boundaries of the potential benefits are characterised. So it is somewhat beyond concept definition but not quite at the industrial implementation level.
Dear Aydin, thank you for your very interesting paper presentation on Digital Twins predictive maintenance of machines used for manufacturing CFRP bicycle frames. I should be interested in knowing what type of sensors were considered to provide signals to the sensorial data processing algorithm based on machine learning. Are these sensors already available in the machines or are they “external” sensors that need to be bought and installed on the existing machines in the factory?
Hello Alessandra
Thank you for the interesting question. The short answer is a mix of both.
To provide more detail: data acquisition was initiated by first breaking down the factory into processes. The machines required for each process were found; processes that did not require machines were eliminated. For each machine the dominant types of failure were found. A criticality matrix was used to compare the failure rate to the severity of the fault and impacts of no action; failures with low criticality were eliminated. The data used to identity this fault type was then found. To assess the suitability of sensing this data for PDM, fault types were chosen that satisfied two main characteristics: ‘Is it appropriate to sense this data through technology over using manual inspection?’ and ‘Will a sensor be able to predict the failure in an appropriate time scale?’. For the fault types that satisfied this, it was decided if sensors needed to be built into the machine or if the information was already being sensed in the machine. The chosen sensed data was then taken forward into the factory design.
The results were as follows:
CNC Machine: Misalignment of cutting head – camera/image recognition, dulling of cutting blade – motor load sensor, bearing issues – low impedance accelerometers, loss of lubrication – low impedance accelerometer, loss of hydraulic pressure – pressure sensor.
Curing oven: Temperature – Thermocouple
Debulker: Vaccum pressure – pressure sensor
Robotic arm: Bearing issues – low impedance accelerometer
Robotic arm (cutting) : tool misalignment – camera/image recognition, wear on tool heads – camera/image recognition
Robotic arm (painting) : paint causing wear/restricting motion – motor load