A genetic-based convolutional neural networks optimization for fault diagnosis of rotary agriculture machine

Convolutional neural networks (CNNs) are an effective and accurate method for vibration analysis in fault detection. However, optimizing hyper-parameters and network topology is a time-consuming task. This paper focuses on two main points. First, it demonstrates that CNN architecture can be automatically optimized using a genetic algorithm (GA) to achieve the desired accuracy. Second, it tests these optimized CNNs for fault detection in rotary agriculture machines using vibration analysis, an ongoing challenge in the field. The methodology includes the population of various CNNs with the architecture optimized by GA for the spectrographic representation of vibration signals. Two most common faults were considered: 1) lack of oil in the main shift, and 2) broken tooth in the rotor. The results show that CNNs with GA optimization predict the lack of oil in the shift with 99% accuracy, whereas the broken tooth in the rotor was detected with 100% accuracy. The best results were obtained using metrics vector with MLP and SVM. The mean accuracy across 100 iterations was 0.995.