Open-Circuit Fault Diagnosis of NPC Inverter Based on Improved 1-D CNN Network

Abstract

With the increasing popularization of a three-level neutral point clamped (NPC) inverter, realizing an accurate and fast fault diagnosis is also required. However, most methods are either not accurate enough or require a long period of time to collect signals as input. To compensate for these drawbacks, a novel 1-D convolution neural network (CNN) with the improved stochastic gradient optimization method is proposed in this article. First, the input data are segmented to ensure that the fault diagnosis can be completed if the current of a period is collected at any time in a period. Then, the feature of the waveform is extracted by the proposed network. In the meantime, the improved Adamod (IAdamod) method is implemented to better adaptively adjust the learning rate during the training process. Specifically, the learning rate is bigger at the beginning and smaller at the end. Finally, the classifier softmax is used for classification. The experimental results demonstrate the effectiveness of the proposed method in detecting and locating the faulty devices, which has shown that it perfectly realizes the fault diagnosis with 100% accuracy and can be diagnosed online with the time of less than 0.25 ms, far less than one cycle. The proposed method is also compared the traditional methods, and shows a better performance both in speed and accuracy, the loss can be less than 0.5 in iteration 220 and achieve zero in 1000.