August 2023 Volume 5

FORGING RESEARCH

Description of the Candidate Forging Part and Closed Die Process The candidate part selected for the closed die ANN was an axisymmetric double-hub pulley forged from AA 6061 billet in the as-received condition. The part geometry and dimensions are shown in Figure 1. After the part and process definition was finalized, an H-13 tool set was machined by an FIA member company and then mounted on a 4-post Superior die base. The tool set was instrumented so that die temperature, transfer time, load-stroke data, and billet temperature could be measured and/or controlled.

Figure 1. Forged part geometry and dimensions used in the study. All dimensions are given in inches/degrees. To make the trial forging runs, AA 6061 billets were cut to specified dimensions and heated in a resistance type box furnace. Forging trials were conducted on a 25-ton hydraulic press (Figure 2) where high temperature vegetable oil was applied using a Rimrock 090 spray wand equipped with dual nozzles. Parts were forged according to a 3-level full-factorial experimental design with each set of process conditions replicated at least twice. Levels were determined using engineering judgment and simulation models to ensure that an appropriate percentage of failures (incomplete fill) would occur. A complete summary of the conditions that were used to make the trial forgings can be found in Table 1. A check fixture was also machined to confirm complete fill/incomplete fill conditions on actual parts that were forged in the press.

Figure 2. 25-ton laboratory hydraulic press and tool stack used to conduct forging trials. Table 1. Summary of the conditions used in producing experimental closed-die forging data for ANN use.

Artificial Neural Networks All ANNs were developed using the Python and MATLAB® programming languages, TensorFlow software, and run on a Windows-based computer. Rather than developing a single ANN program to predict complete fill/incomplete fill and detect surface defects, the team chose to employ separate ANN algorithms for simplicity and establish a baseline. Three different algorithms were developed and are discussed separately below. The first two algorithms were developed to predict complete/incomplete die fill while the third algorithm was intended to detect surface defects such as laps, scratches, and the amount of residual scale on AISI 1020 steel billets after a 10% reduction in height. ANN Fill/No-Fill Predictive Algorithm – The team considered that three different approaches could be used to predict die fill based on the type of input data supplied to the ANN. This included discrete

FIA MAGAZINE | AUGUST 2023 68