Fault Diagnosis of Cascaded Multilevel Inverter Using Multiscale Kernel Convolutional Neural Network

Over the last decade, multilevel inverter (MLI) has gained a global research paradigm in power electronics. The high-power application of multilevel inverters necessitates the safe and reliable operation of the inverter. Fault diagnosis of MLI is inevitable to enhance the reliability of the inverter. The real-time applications monitor the intricate switching process of the MLI using data from sensors like voltage, current, and temperature. Obtaining complementary fault features through conventional methods is difficult due to operating complexity and limited switch-fault samples. These methods heavily depend on prior knowledge. To deal with this problem, this paper proposes a new multiscale kernel convolution neural network (MKCNN) for fast fault diagnosis of cascaded MLI. Firstly, the proposed method necessitates using frequency domain samples instead of raw signals, which can leverage the power of a convolution neural network (CNN) to obtain hierarchical features from the images through a short-time Fourier transform (STFT). Secondly, the multiscale kernel convolution network model is constructed to capture and analyze low- to high-level fault features. This method improves the traditional CNN by extracting discriminative fault information through multiple convolution kernels of different scales with varying resolutions obtaining high diagnosis accuracy under single and multiple open circuit and short circuit faults. Finally, the softmax layer generates the output of the fault diagnosis results. Simulation results validate the effectiveness of the proposed model, demonstrating a high diagnosis accuracy rate of 98.3%. The model exhibits robustness in diagnosing single and multiple switch faults across different fault cases of an MLI. Comparison with other intelligent models further emphasizes the superiority of the proposed method.