A distributed training framework for abnormal sound detection in industrial devices: aggregation strategies and performance analysis

Abstract In today's digital and networked industrial landscape, the detection of abnormal sounds in device has emerged as a vital aspect for guaranteeing the normal operation of industrial machinery. Nevertheless, traditional centralized training approaches demand substantial amounts of audio data, imposing considerable burdens on data storage and transmission, and concurrently presenting obstacles to data privacy and security. Federated learning, as a distributed machine learning paradigm, enables model training with local data from each client without sharing the original data, thereby effectively safeguarding data privacy. Hence, in this study, we propose a distributed training framework based on federated training, which enables multiple clients to collaboratively train an abnormal sound detection model, thereby mitigating the risk of data privacy exposure. In the distributed training framework, each client possesses data from different types of device or various machines within the same device type, posing significant challenges for distributed training. To overcome this, we devised two client device distribution scenarios and proposed aggregation strategies based on client sample size, model performance, and domain shift among clients. Additionally, we introduced a Sample-Performance-Shift (SPS) aggregation strategy to ensure robust model performance across diverse device scenarios in industrial settings.
The proposed methods were evaluated on the DCASE 2020 Challenge Task 2 dataset.
Experimental results demonstrate that the SPS aggregation strategy enhances the accuracy and reliability of abnormal sound detection for industrial device within the distributed training framework while simultaneously reducing the risk of data privacy leakage.