November 2022 Volume 4

FORGING RESEARCH

Adam [4] was shuffled at the beginning of each epoch. The initial learning rate was 0.01, which relates to the rate at which the error (predicted minus actual) is used to adjust the weights of the ANN during training. The learning rate was reduced by 50% every 50 epochs. Figure 4 shows the results of the ANN predictions when it was applied to the testing data.

Figure 4. Comparison of out-of-sample predicted and actual upset workpiece diameter 0.1 and 0.2 inches from the top of the workpiece for AA6061. The data was generated using DEFORM® v12. Experimental Design and Prototype Setup The primary objectives in developing the prototype open die set-up were to physically replicate a buster operation and have the ability to control process conditions and record key process variables. The workpiece materials that were used included 6061 AA aluminum and AISI 1018 steel which were heated to 850°F and 2250°F prior to forging, respectively. After being manually transferred from the furnace using a pair of tongs, the samples were placed in the press and then upset to 10-13% reduction in height using heated H13 platens. The basis for the prototype system was a National Instruments data acquisition system used in conjunction with DAQExpress software to record the outputs provided by multiple sensors. The sensors employed were commercially available components. The hardware that was used in the prototype setup included two type K thermocouples that measured the upper and lower die surface temperatures, an infrared temperature detector that remotely measured spot temperature of the workpieces, a linear potentiometer that measured the die position, and four strain gauges mounted on the tie rods to monitor press load. Data collection was initiated from an automated remote trigger and was also recorded and stored for each physical run that was performed. Figure 5 illustrates the purpose-built hydraulic press (developed as part of a previous Forging Industry Education & Research Foundation project) capable of producing 20,000 lbs of force. The sensor package was mounted on a Superior 4-post die set. Temperature control was achieved by a system that monitored the temperature of the upper and lower die surfaces and controlled these to within ± 5 degrees F.

Figure 5. Photograph showing the experimental forging set-up used in the ANN/MIMO project. A purpose-built vision system was also implemented to record workpiece position and orientation. This setup was developed for a separate study involving robotic automation and future work to assess the feasibility of performing in-situ dimensional measurements and friction monitoring in a forging process and will be incorporated as part of the prototype open die in the near future. Results Once model training was complete, physical trials were run to assess the accuracy of the ANN model using various process conditions. The validation data was collected using the prototype set-up and measurements taken from each specimen. For each run, surface temperature and diametral dimensions (taken at mid-height, 0.1”, and 0.2” below the top surface) were used to compare the model accuracy and reliability. Results for AA6061 and AISI 1018 are shown below in Figure 6 and 7 for the diametral measurements of the compressed workpieces where it can be seen that the ANN predictions are very close to the actual data values.

FIA MAGAZINE | NOVEMBER 2022 80