Federal learning algorithm for bearing fault diagnosis

Abstract

The invention relates to a federated learning algorithm for bearing fault diagnosis, which is characterized in that the algorithm runs on a plurality of local nodes and an aggregation node, and comprises the following steps of: 1, converging data of a sensor network by each local node, the data of the sensor are preprocessed in a time-sharing, partitioning, sampling and normalizing manner; 2, training the preprocessed data by adopting a convolutional neural network model; 3, after training is completed, whether the aggregation condition is met or not is judged according to the improved aggregation strategy, and if yes, the round of training is ended; then, calculating an F1 score of the local model; finally, performing homomorphic encryption on the model parameters, the F1 score and the total number of samples, and sending to an aggregation node; and 4, after receiving the information sent by all the local nodes, the aggregation node decrypts the information, then weights and aggregates all the local models according to an F1 score weighting strategy to obtain a new initial model, and sends the new initial model to the local nodes.

Description

technical field [0001] The invention belongs to the field of the Internet of Things, relates to data computing equipment of the Internet of Things, and belongs to the field of combining artificial intelligence and edge computing. Background technique [0002] Bearings are important components of industrial equipment. As the core of bearings, rolling bearings are the most prone to mechanical failures. Shock alternating loads, thermal fatigue and mechanical wear are common causes of bearing failure. The health status of the bearing directly affects the overall performance of the equipment, so the fault diagnosis method of rolling bearings has become the focus of scholars' research. Although the practicability of deep learning is undeniable, the training data it uses may involve serious privacy issues: factory production data and user usage data are collected centrally, and these data are permanently stored by large companies, and users can neither Deleting this data also gi...

AI Technical Summary

Problems solved by technology

Although the practicability of deep learning is undeniable, the training data it uses may involve serious privacy issues: factory production data and user usage data are collected centrally, and these data are permanently stored by large companies, and users can neither delete the data and have no control over how these companies use the data

Secondly, production data is likely to contain a large amount of sensitive information such as production capacity, work efficiency

If these data are transmitted to the cloud, many security issues will arise

Moreover, complex industrial data usually has the following problems: 1) In massive data, effective information is limited, and there is a lack of characteristic information related to faults

2) Different sampling rates or packet loss usually lead to the disappearance of observed data at a specific point in time

These issues often lead to uncertainty in the data space and diagnostic space, severely limiting the learning ability of deep learning algorithms

On the other hand, in many fields, especially those related to medicine, the law does not allow the sharing of personal data. Since the implementation of the "Network Security Law of the People's Republic of China" on June 1, 2017, information security has been listed as a priority The scope of protection, the country has put forward clear requirements for privacy clauses, and further improved the protection of information. Therefore, researchers in related industries can only analyze and mine data sets belonging to their own institutions

If the amount of data owned by a single organization is not very large, and the similarity is high, and the diversity is insufficient, then if machine learning is performed on such a data set, researchers may end up with models with poor scalability or prone to overfitting. Combined result

In this case, the limitations of data privacy and confidentiality clearly affect the effectiveness of machine learning

Images