Sheet-Metal Manufacturing: Sheet metal manufacturing is a versatile and efficient process that involves transforming flat metal sheets into various components and products. The figure below summarizes gives an overview of the main processing steps. The process typically begins with material selection, where the appropriate type and thickness of metal are chosen based on the desired end product. Next, the metal sheets are cut into specific shapes using methods like shearing, laser cutting, or punching. These shapes are then subjected to bending or forming operations, where the sheets are shaped into the desired contours and angles using techniques such as press braking or roll forming. Additional processes like welding, fastening, and surface finishing may be employed to join multiple components or enhance the appearance and durability of the finished product.
The geometries cut into the metal sheet during the first processing step are not subject to a regular pattern. This is because sheet-metal products are ubiquitous across industries: From household appliances and automotive components to aerospace parts and construction materials, sheet-metal plays a crucial role. Sheet-metal can be cut into intricate and complex shapes, allowing for the production of intricate designs and customized components. For instance, in the automotive industry, sheet-metal is used to create body panels with precisely cut contours and curves, while in the aerospace sector, it enables the fabrication of aerodynamic structures and aircraft components with intricate geometries. Furthermore, in architecture and interior design, sheet-metal is employed to craft decorative elements, façades, and furniture featuring unique and artistic cut patterns.
The figure below exemplifies the variety described above. Two exemplary sets of shapes cut into metal sheets are depicted using colors to indicate the product association of individual pieces.
Because of the high degree of variance and intricacy associated with the metal embedded geometries coupled with material properties, the extraction process can become a bottleneck: Shapes can be either large, heavy and hard to handle or, conversely, small, with complex contours which make them difficult to remove from the contained metal. As such, the extraction process can become a bottleneck during production.
In an effort to streamline the cutting process, the company that provided the data developed a (laser) cutting machine capable of automatically extracting geometries from metal-sheets. This is being done by means of a specialized pin bed in conjunction with suction cups. The suction cups are used to latch on to cut geometries from above, while the pins are used to prop up the geometry from below. After the cutting process finishes, the sheet is conveyed onto the pin bed. A heuristic algorithmic solution is then used to select the pins to activate along with the position of the suction cups. While quite reliable, the employed algorithmic solution cannot guarantee that a particular pin-suction cup configuration will lead to the successful extraction of the desired part.
The figure below seeks to visualize the described process. On the left a positioned suction cup is depicted along with an activate pin configuration. On the right, we see an example of a failed configuration along with a successful one. The suction cup placement is drawn in light gray, dark gray is used to mark the geometry and re marks the position of the activated pins. If the geometry cannot be lifted from the bed, a new configuration is computed. If the machine fails to extract a part eight times in a row, a human worker needs to extract the piece manually. Naturally, this situation is to be avoided at any cost.
Prediction Guided Extraction: Machine Learning can be used to minimize the probability of failed extraction as shown in the next figure. The geometry of the part to be extracted along with material properties, as well as the tentative pin and suction cup placement information can be fed into a model which must then predict whether the extraction will be successful or not given. Instead of the the variable size raw geometric and configuration information, fixed size descriptors (features) could be instead computed and passed to the model. Should the model indicate extraction failure, a new suction cup-pin configuration would be computed and fed back into the model. This loop is to continue until the model indicates a successful extraction configuration. At this point, the cups would be moved at the desired locations and activated along with the corresponding pins.
Data Set: To train and validate the extraction prediction model, the company constructed an anonymized data set of 16.000 geometry-extraction configuration combinations, marking whether the particular configuration was successful or not. For each data point, several experiments were run, such that the probability of extraction rather than a flag can be learned from the data. Aside from the extraction probability, every record in the data set consists of 74 features describing the material, the geometry, the pin configuration and the suction cup position.-Two such features are shown in the figure below. Since the exact geometric and material information represents sensitive customer data, feature name mangling and feature scaling was additionally applied to the data set.
Challenge: Help perfect their technology by training and evaluating the necessary predictive model.