Diagnostica dei guasti a circuito aperto negli switch di convertitori di potenza Cascaded H-bridge con rete neurale di deep learning Bi-directional LSTM

The use of Deep Learning (DL) techniques in power converter fault diagnostics introduces a technique that incorporates data-driven approaches to effectively detect and localize faults in a very dynamic system. The Bidirectional Long-Short Term Memory Neural Network (Bi-LSTM NN), a specialized DL architecture known for its strength in processing sequential and time-series data will be explored in this work. The Cascaded H-Bridge (CHB) power converter is commonly used in most motor drives and the topology for such a system utilizes multiple transistor switches to provide power to electric machines. CHB converters are complex and fast-changing, making accurate fault detection essential for reliable operations. The complex topologies and fast dynamics in system operation require effective incorporation of light, fast-responding, accurate diagnostic tools. This work explores how Bi-LSTM networks can identify faults with high accuracy, using feature engineering techniques to process and analyze system data. The Deep Neural Network (DNN) architecture will be explored, highlighting its performance efficacies. Experimental results demonstrate the effectiveness of Bi-LSTM networks in diagnosing faults in CHB converters, achieving an accuracy of 99.89%, a precision of 0.9463, a recall of 0.9994, and an F1-Score of 0.9714. Additionally, the three-layer Bi-LSTM networks provide a fast prediction time of 46812.83 ms and are efficient in terms of memory size, requiring only 0.81 MB for the 2-class classifier with 30 time-samples and 128 mini-batches. These findings show that Bi-LSTM networks offer a reliable and efficient solution for improving system performance and reducing downtime. Furthermore, the demonstration of feature engineering techniques and comparisons for such a dynamic system will be discussed and analyzed. This paper provides a comprehensive overview of methods utilizing DL, specifically Bi-LSTM, for the diagnosis and localization of Open Switch faults in 3-level, 3-phase CHB inverter architecture.