Application of TPE-Random Forest in Microwave Oven Fault Diagnosis Based on GAN and ADASYN Algorithm

The utilization of data-centric machine learning methodologies has gained significance within the domain of microwave oven fault diagnosis. Conventional fault diagnosis approaches that rely on mechanism analysis demonstrate limited adaptability to various fault scenarios and environmental conditions. Moreover, the distribution frequency of different faults in the data sets collected from factories is severely imbalanced. This paper introduces a TPE-Random Forest fault diagnosis approach, which integrates the Generative Adversarial Network (GAN) and Adaptive Synthetic Sampling (ADASYN) algorithm. The combination of GAN with ADASYN aims to mitigate biases present in data generated solely by ADASYN with improving the feasibility and variety of the data generated and effectively addressing imbalanced data issues within dataset. Specifically, the ADASYN-generated samples from minority classes are utilized to train the GAN network, replacing the random noise generator typically employed in GAN networks. Subsequently, a set number of samples are produced using the refined generator within the GAN framework to achieve class equilibrium. Ultimately, the created data is utilized for training of the random forest model, with optimization of its hyperparameters carried out through the TPE optimization algorithm. Experiments confirm the approach we proposed attains an F1 score of 92.59% on the microwave oven fault diagnosis dataset, surpassing alternative algorithms like SVM, MLP, KNN, and LGBM.