基於機器學習之多天線空中測試裝置錯誤診斷

Abstract

於本論文中，吾人欲藉由多天線空中測試 (MIMO-OTA)之測試結果進行裝置錯誤之檢測。吾人在空中測試系統中利用空間相關性 (Spatial Correlation)並考慮功率頻譜之形狀提出通道仿真之最佳化問題，並解出每根天線的最佳能量權重以模擬實際空中測試環境。此後，吾人從測試結果中某些特性，以利後續錯誤檢測進行。因為特性與裝置錯誤之間的關係不容易直接找出對應關係，選擇基於機器學習的演算法去作錯誤的分類。在單一問題的情境下，吾人使用支援向量機 (Support Vector Machine, SVM)，而在多重問題的情境下，使用類神經網絡 (Neural Network)及多重支援向量機。經過模擬驗證，上述的方法皆可以高正確率的分出錯誤種類，進而佐證實際測試時可以利用空中測試之結果及機器學習演算法判斷裝置的錯誤種類。

In this thesis, we would like to realize device fault diagnosis by using the results of MIMO OTA testing. We first use spatial correlation and the shape of power azimuth spectrum (PAS) to emulate the channel by correctly allocating the power weights of probe antennas in the anechoic chamber, so as to simulate actual MIMO OTA testing. After that, we find some features to use in device fault diagnosis. Since it is hard to determine the correspondence between the testing result and faults, we use some machine learning based algorithms to resolve the classification problem. Support vector machine (SVM) is used for single-fault case, and neural network and multi-SVM are used for multi-fault case. By the simulations, these methods have good ability to classify different kinds of fault, and we can ascertain that it is possible to do device fault diagnosis by using MIMO OTA testing results.