Agriculture Rotary Tedder Fault Diagnosis Based on Evolutionary Convolutional Neural Network with Genetic Algorithm Optimization

Since the mechanization of agriculture, agriculture itself inherited the same problems, as an industry has: the necessity of machines maintenance and machines failures which causes downtime and may make unanticipated costs. To counteract such problems, we propose a method of monitoring agricultural rotary machinery, which should allow to predict when the machine would break down, and to plan repair time in advance, to eliminate downtime to minimum. Our approach utilizes machines’ vibration signal, which after being turned into a spectrogram, is fed as input data to the convolutional neural network (CNN). The convolutional part of a neural network is the feature extractor, and the dense part is the feature selector and classifier. Since machines are usually not created equal, CNNs architecture is permanently defined, but rather is adjusted to the machine’s and system’s needs by using the genetic algorithm (GA), which adds an additional layer of flexibility. Experimental results showed, that our approach was able to reach 99.7% of classification accuracy.