The Fault Diagnosis And Privacy Protection Method Research For Distributed Industrial Systems Based On Federated Learning

Distributed industrial systems often require multiple users to jointly train a fault diagnosis model due to the data generated by multiple users and the small number of fault samples for a single user.Traditional centralized learning requires the aggregation of data from each client,which raises concerns about data privacy and security,resulting in the problem of "data island".The emerging federated learning technology transfers the model training phase to the client,which effectively alleviates the "data island" phenomenon.However,the data of industrial systems are characterized by high dimensionality and few labeled samples,which leads to many challenges for federated learning applications.In addition,federated learning is a privacy protection approach for multi-party participants.Correlational studies have shown malicious attackers can intercept data during the transmission of model parameters,which puts forward higher requirements for privacy protection of federated learning.In this paper,we focus on fault diagnosis and privacy protection using federated learning for distributed industrial systems.In the framework of federated learning,the achievements are as follows:(1)A fault diagnosis method based on information gain and autoencoder is proposed.First,the information gain was used to preprocess the data set.Second,the stacked autoencoder network is used as the global model of federated learning for unsupervised learning.Third,the trained autoencoder network is used to extract the features from the labeled data,which are used as the training samples of the classifier for fault diagnosis.Finally,the experiments demonstrate that the proposed method can effectively solve the fault diagnosis problem of federated learning under a little labeled sample environment.(2)A privacy preserving fault diagnosis method based on noise fragments is proposed.First,non-Gaussian noise slices are randomly added to the local model that is trained by the client via a selection matrix.Second,a high-order filter for nonGaussian noise is established on the client side to remove the added noise.Finally,the experiments demonstrate that the proposed method can protect the privacy of the data.(3)A method of stochastic upload and distribution mechanism for preventing interception of information and an adaptive compensation method are proposed.First,the model parameter information is uploaded via random triggers during the data upload phase.Second,the randomly arriving local models are aggregated asynchronously using the Sequential Kalman filter in the cloud center and then sent to each client via random trigger.Third,a delay parameter aggregation and compensation method based on adaptive compensation filter is designed on the client side.Finally,the effectiveness of the method is verified by experiments.