Transfer learning (TL) is a machine learning tool that focuses on storing knowledge gained while solving one problem and applying it to a different but related problem.
Example 1: Assembly process modeling in different cars
The car assembly is a stage-by-stage manufacturing process. Each assembly line constructs a complex (tree) network.
Traditional data-driven methods focus on learning the network (inter-relationship) individually for each type of car. The transfer learning can borrow information from the existing assembly lines to benefit the understanding of a new line, e.g., use information from an old-fashioned sedan line to help understand the newly designed truck line. See Figure 2, where the edges (absence/presence) are the inter-relationship information we want to infer. Note the nodes and edge absence/presence are different in each assembly line, which makes the edge information transfer a non-trivial task.
Figure 2. Information transfer in the car assembly.
Example 2: Ceramic stereolithography model calibration
The ceramic stereolithography (CSL) is an emerging additive manufacturing (AM) process, whose products can achieve high-density and flexible shapes. A key challenge for CSL process is a lack of capability in accurately controlling the microstructures of fabricated ceramics to desired levels. Most of traditional methods require very large amount of CSL data to calibrate the CSL microstructure features. However, the CSL data is time and cost-consuming to collect (one data point needs 2 days to prepare).
The basic idea to handle the CSL data limitation is to transfer rich and cheap data from traditional ceramic manufacturing processes, e.g., ceramic injection molding, to benefit the understanding of the CSL process. See the ideas in Figure 3. The challenge here is how to effectively differentiate the manufacturing data from different processes, at the same time extract the common information that can be beneficial for CSL.
Figure 3. Data transfer for CSL.