August 2023 Volume 5

FORGING RESEARCH

data (individual process conditions), time series data (this consisted of the load-stroke information recorded for each process condition that was simulated ), and combination of both. As it was not clear which method would provide the best output results, the team developed two separate ANNs and compared the results. The discrete data consisted of the initial billet temperature and dimensions (height, diameter), die temperature, die friction, and maximum press force. The Design of Experiments (DoE) module of DEFORM v13® was used to generate a training dataset that contained 2,280 unique process conditions. One of the lessons learned from developing and testing the open die prototype process was that the training dataset needed to have an appropriate mix of good and defective parts. This was necessary for two reasons. The first is that a representative mix of conditions is needed to validate the ANN’s ability to correctly distinguish the conditions which result in good and bad parts. The second reason is to eliminate systemic bias. If the training data consists of predominantly good or bad parts, the ANN will skew predictions in that direction. The distinction between complete and incomplete fill was made using the Fill Constraints Tool in the DoE module which assessed the distance between the workpiece and die surfaces. If the distance at any point exceeded a specified value, the part was designated as an incomplete fill. While the exact details of the ANN code are beyond the scope of the article, a summary of the ANN system architecture is provided here. The 5-layer ANN architecture for the discrete inputs is shown in Figure 3 where the first and second hidden layers have six and two rectified linear unit (ReLU) nodes, respectively. The ReLU function is defined as f(x) = max(0, x) that is if the input is positive the output is the input otherwise the output is a zero. Both layers are fully connected, i.e., the outputs of the neurons in the first hidden layer are connected as inputs to all neurons in the second layer. The final output layer has a sigmoid neuron, which is an s-shaped function, that yields a probabilistic classification of whether a complete fill or an incomplete fill condition occurred.

Figure 4. Load-Stroke Input ANN Architecture used in the aluminum closed die forging process. ANN Surface Defect Algorithm Surface defect detection is an important quality control task in closed die forging processes. While it is possible to develop an ANN system that uses process conditions to predicts defects a priori, it must be considered that this would require that all relevant factors that would contribute to a defect be identified and incorporated in the training data. Considering that some factors are internal and others are external to the forging process, this approach was not considered to be worthwhile. An alternative approach is to use digital images of the surface as inputs to an ANN algorithm. While this represents a post facto approach, it significantly simplified the convolutional neural network (CNN) architecture and requirements and training data requirements. It was also considered that forging conditions and billet geometries are quite challenging for a vision system. Consequently the team also worked on developing some possible solutions that would enable satisfactory images of axisymmetric pieces to be taken in a forging environment. While numerous commercial systems are available, the team chose to develop a purpose built system in order to have full control over its operation and better understand how individual algorithms would perform under forging conditions. As metrology was not a primary objective, a single off-the-shelf GoPro camera capable of taking RGB (color) images was used to obtain images of part surfaces (front view) in a dedicated station outside of the die. Several algorithms were used and an optimum solution was found that did not require any external lighting or specific background. The CNN system processes the RGB images to recognize various types of surface defects, including scale, laps/folds, and surface scratches. Although not pursued in the present study, it should be noted that, the system also has the capability to assess surface temperature differences and generate thermographs for steel workpieces. This will be considered at a later stage of the project. The vision system, which can operate in ambient lighting, first divides an image of the workpiece surface into a matrix pattern composed of individual “patches” which are individually analyzed by the ANN. Each patch is evaluated to determine whether a particular defect is present or not. One advantage of this CNN system is that it was developed using only 70 training points. Another benefit is that it was trained using archival photos and previously collected information, making it possible for companies to use archival data for CNN training. Figure 5 shows the process of the CNN model. A steel billet which is partially covered with scale

Figure 3. ANN Architecture used in the aluminum closed die forging process. The ANN architecture for the load-stroke (time series) data was more complex and is shown in Figure 4. Since the load stroke curve is a sequence of values (time series), a recurrent node is used. A recurrent node creates a feedback loop by using its outputs as additional inputs. The specific node the team used was the so-called “long short-term memory” (LSTM) node [3]. After the LSTM layers are two fully connected layers.

FIA MAGAZINE | AUGUST 2023 69