Abstarct: Industry has begun to demand higher power ratings and multilevel inverter drives (MLIDs) have become a solution for high-power applications in recent years. Because of using these systems in the high power rating, the reliability of them is very important, so the uninterruptable working of the motor and its drive system is needed. If a fault (open or short circuit) occurs at a semiconductor power switch in a cell, it will cause an unbalanced output voltage and current, while the motor is operating. The unbalanced voltage and current may result in vital damage to the motor if the motor is run in this state for a long time. Generally, the passive protection devices will disconnect the power sources or gate drive signals from the multilevel inverter system whenever a fault occurs, stopping the operated process. Although a cascaded MLID has the ability to tolerate a fault for some cycles, thus the detection of fault is important in MLIDs. Research on fault diagnostic techniques using voltage FFT and wavelet transformation and motor supply current Park’s Vector along the classification methods such as Neural Network and SVM are done. In this thesis a level-shifted (LS-PWM) method is proposed that uses rotating carrier waveforms to balance switching pattern in all cells, at all modulation indices. Next a fault detection and reconfiguration method is proposed for a modular MLI by using the output phase voltages. The histogram analysis is used for feature extraction and these features have been used as input to the Neural Networks (NNs). After the detection of fault type and its location by the NNs, the faulty cell is bypassed and a reconfiguration method for this modulation strategy is used that guaranty continuous working. Simulation results are given for a cascade 7-level inverter at different modulation indices. These results show that the proposed diagnosis method is accurate for detection of fault types and their locations. This method works correctly under noisy condition and the classification performance for the noise with variance up to 1500 is 100%, whereas the other methods have not such accuracy. The proposed method is faster and less complicated because of using histogram analysis instead of using sophisticated methods such as wavelet.