Metric Learning-Based Neural Network Model for Electromagnetic Compatibility Fault Diagnosis: An Application Study

Abstract

With the growing prevalence of electronic equipment and the increasing severity of the electromagnetic environment, the likelihood of electromagnetic compatibility failures is on the rise. As a result, the difficulty of diagnosing EMC faults is also increasing. However, by employing neural networks in deep learning for EMC fault diagnosis, we can simplify and streamline the process of feature extraction and similarity analysis. Compared to traditional artificial feature extraction methods, neural networks can learn to measure the similarity between features more efficiently, resulting in more accurate diagnoses. To train the model, we obtain response data from each port of the electronic equipment system in a high radio frequency environment and pair it with the corresponding equipment fault status. However, due to the limited availability of labeled data, conventional neural networks are susceptible to overfitting. Therefore, we use a neural network model that is well-suited for few-shot learning, which is based on a metric learning approach. This approach enables the model to learn from a small amount of labeled data, making it more effective in diagnosing EMC faults.