Research On Fault Diagnosis Of Electric Drone Motor Based On Deep Learnin

Unmanned aerial vehicles(UAVs)are widely used in various fields,and their application scenarios and usage methods are rapidly expanding.Electric UAVs have more advantages than traditional fuel UAVs in terms of environmental friendliness,economy,adaptability,flexibility and intelligence.Motor failure can greatly affect the flight stability of electric UAVs,which may affect the flight quality or lead to loss of control of the UAVs.Therefore,the research on electric UAV motor fault diagnosis has important theoretical and engineering application value for improving the controllability and reliability of electric UAVs flight.With the rapid development of deep learning theory and methods,it has achieved remarkable results in the field of fault diagnosis.However,many challenges are still faced in the practical application of electric UAV motor fault diagnosis.For example,overfitting of the model under limited training samples,robustness under noise interference needs to be improved,degradation of inference performance under variable operating conditions,and poor generality of diagnostic signal acquisition.In this paper,we conduct research on deep learning-based electric UAV motor fault diagnosis in terms of deep neural network model improvement,regular loss construction,training method optimization,and inter-modal knowledge transfer to provide theoretical and technical guidance for electric UAV motor fault diagnosis.The main research work of this paper is as follows:(1)To address the problem of overfitting of deep learning models brought about by limited training samples,a new convolutional network,similar dilated convolution,is designed,in which the feature extraction capability of the convolutional network is retained as much as possible while reducing the model parameters through the method of sharing weights of multi-scale dilated convolutional kernels.A hybrid network model combining an auto-encoder with a similar dilated convolution backbone network is constructed,which enables the proposed model to have both supervised and unsupervised learning capabilities by sharing encoders between classifiers and decoders.The autoencoder can force the model to retain the useful features of the samples during the feature extraction and reduce the risk of overfitting.Meanwhile,inspired by the idea of manifold learning algorithm,a manifold regularization based on local linear embedding is designed to make full use of the local spatial features of training samples and improve the generalization ability of the model.Finally,the effectiveness of the proposed method is verified on a electric UAV motor fault dataset collected by ourself.(2)To address the problem that noise brings interference to the fault diagnosis model and leads to the degradation of the inference performance of the deep learning model,a RR-Dropout(Regularized Random Dropout)method is designed to improve the robustness of the model to noise perturbation by imposing interference on the model in the training phase through a random rate Dropout,while imposing consistency constraints on the features of different perturbed outputs on the output layer of the feature extractor.At the same time,the prototype matching network is introduced into the electric UAV motor fault diagnosis model to further reduce the impact of noise on the prediction performance of the model by using the idea of "averaging" of the prototype matching network.Finally,the effectiveness of the proposed method is verified on the electric UAV noisy motor fault dataset collected by ourself.(3)To address the problem of different distributions of training data and test data due to high-frequency changes of electric UAV motor operating conditions,a deep causal learning-based electric UAV motor fault diagnosis method is proposed to improve the cross-domain generalization ability of the deep learning model by extracting features that are more causally related to the prediction results.At the same time,the manifold mixup algorithm is introduced into the model training process to achieve a flatter feature representation to expand the classification decision boundary and further reduce the risk of performance degradation of the model in cross-domain tasks.Finally,multiple sets of multi-source domain cross-domain inference experiments are set up on an electric UAV motor fault dataset collected by ourself to verify the effectiveness of the proposed method in improving the model domain generalization capability.(4)To address the problems of high cost of vibration signal acquisition and weak mechanical fault characterization in current signal,a cross-modal knowledge transfer method from vibration signal model to current signal model is proposed in order to fully utilize the advantages of vibration signal and current signal in motor fault diagnosis.The fault diagnosis knowledge learned from the vibration signal model is transferred to the current signal model through feature alignment between different modal models and knowledge distillation at the output layer.The proposed knowledge transfer method is independent of the backbone network and has certain generality.In addition,a platform for simultaneous acquisition of multimodal signals of electric UAV motor faults is built,and the effectiveness of the proposed method is verified using different backbone networks on the multimodal fault dataset.(5)A cloud-edge combined electric UAV motor condition monitoring and fault diagnosis platform is designed and developed.Through the distributed structure of combining the central server in the cloud and the edge computing platform,the advantages of high arithmetic power of the central server in model training and good real-time status monitoring and fault diagnosis of the edge computing platform are fully utilized to realize the functions of motor status monitoring and fault diagnosis,data collection and visualization,and the electric UAV motor diagnosis method proposed in this paper is integrated into the prototype system for the deep learning-based electric UAV motor condition monitoring and fault diagnosis application based on deep learning to provide reference.