Pressing cables with cable shoes currently involves the use of high pressure to ensure successful crimping. However, this approach lacks the ability to detect when the pressing has been completed. Elpress intends to develop a handheld tool that can classify cables in real-time and stop the pressing before unnecessary energy is lost. This project is trying to solve this by creating a neural network on an embedded system, which will learn on the device.
In a previous project, transferring TensorFlow and Scikit-learn models to embedded systems proved ineffective due to the limited memory capacities of the embedded system and large machine-learning models. Therefore, the neural network was instead implemented by hand on the embedded system directly. The implemented neural network was compared with a Python implementation, as well as with two traditional methods: linear regression, and averaging. The comparison focused on performance and memory consumption. The neural network was designed with 12 input nodes, one hidden layer consisting of 12 nodes, and an output layer of 11 nodes.
The highest accuracy achieved by the neural network implemented in C was 80\%, which is also the lowest accuracy that the other methods achieved. The neural network in C does not achieve equal or better precision compared to the traditional methods. However, since the neural network implemented in Python achieves higher accuracy, it should in theory be possible for the neural network in C as well.