Genetic algorithm for automatic construction, optimization and testing of multi-input forked convolutional neural networks used for diagnostics of rotating agricultural machines

The third decade of the 21st century can undoubtedly be called the “smart” decade. Manufacturers of all devices try to integrate the functions of their products with methods based on artificial intelligence as much as possible. This trend has also reached industrial and agricultural machine manufacturers, where neural network diagnostics is no longer just a thing of the future and is becoming something every day. As developing AI methods requires a high degree of expert knowledge and is time-consuming, we propose a method to automate the entire process. Thanks to the use of a three-step genetic algorithm, we managed to create a system that can create the appropriate AI architecture based on neural networks, optimise it, and test it completely autonomously. We confirmed its effectiveness by using our method to develop a multi-input, forked convolutional network capable of recognizing nine different states (healthy and eight different damages) of an agricultural rotary tedder with an accuracy of approximately 98%.