A Research Of Array Failure Diagnosis Method Based On Multi-label Deep Learning

The array antenna is composed of multiple radiating elements.If some of the elements fail,the performance of the array will decrease.In order to maintain the radiation performance of the array,periodic or real-time fault diagnosis is particularly important.The array fault diagnosis method based on radiation far-field information is taken as the research topic and a diagnosis model is established from the perspective of multi-label classification.Machine learning related technologies is also used to study the array fault diagnosis method based on multi-label deep learning and the planar array diagnosis network based on transfer learning in this paper.The main research content of this article mainly includes the following three parts.1.Most of the existing array fault diagnosis methods based on machine learning modeled the array fault diagnosis as a single classification problem.However,as the array size increases,the measurement cost increases rapidly and the diagnostic accuracy rate will decrease.Therefore,the array diagnosis is firstly as a multi-label classification problem in this paper and a lightweight multi-task neural network is trained as a diagnosis network using the far-field radiation information of the array as feature.In order to further increase the diagnostic accuracy of the lightweight network,a radiation compensation algorithm is proposed.Meanwhile,considering the measurement cost of feature data,a skeletonized sampling method is proposed.Finally,simulation experiments show that the proposed diagnosis method has a significant advantage in accuracy compared with the array fault diagnosis method based on classification algorithms.The measurement experiment shows that the proposed method can reduce the impact of the inherent error of the array compared with the traditional diagnosis method.2.Due to different application scenarios of arrays,their scale and shape are usually different.In order to make full use of the existing diagnosis network,this paper proposes a planar array diagnosis network based on transfer learning.The existing linear array data is taken as the source domain and the related data of the planar array under test is as the target domain in the proposed method.An end-to-end deep transfer network is built by transfering some layers of the basic diagnostic network.Then,the deep transfer network parameters are fine-tuneed through a small amount of plane array instances.The simulation results show that the network is easier to converge than ordinary neural network during training process and can achieve higher diagnostic accuracy under the condition of fewer training samples.3.In view of the difficulty of array pattern measurement in engineering applications,the high cost of phase measurement,the existence of measurement angles errors and other issues,a phased array fault diagnosis method based on far-field amplitude information is proposed.Considering that the current large-scale phased array usually contains many radiating elements,the array under test is firstly divided into multiple sub-arrays in the proposed method.According to the proposed excitation criterion,excitation is applied to each sub-array in turn and the probes at a fixed position are used to measure its far-field amplitude data.Then,the measured amplitude data is used as the feature to train a multilabel neural network to locate the failed array element.The simulation results show that the proposed method can achieve effective diagnosis for large-scale phased arrays.