Human Knowledge-based Compressed Federated Learning Model for Wind Turbine Blade Icing Detection

With the development of sensor technology, wind turbines are equipped with an increasing number of sensors, which are collecting more and more information. However, redundant features and the high cost of maintaining the internet connection between the server and the clients far from the urban area make it difficult when applying federated learning (FL) to wind turbine icing detection. To address these issues, first, this paper conducts a human knowledge-based feature selection technique to choose the informative features. Then, a neural network combines long short-term memory (LSTM) and a three-layer convolutional neural network (CNN) is built for wind turbine blade icing detection. In federated training, the server considers the data size effect and time effect when aggregating the updates of participating clients. Furthermore, a compression method is used on both the client-side and the server-side to reduce the size of packages exchanged between clients and the server. Experiments show that the human knowledge-based feature selection method helps neural networks become sensitive to icing class and when the compression method is used with a sparsity ratio set to 0.01, it saves 99.72% of bits of updates per client with an acceptable performance drop.