With the increasing use of the Internet of Things, this technology has permeated various industries, leading to the emergence of the concept of the smart industry. The smart industry has provided extensive opportunities for collecting and processing production data from industrial devices, enabling comprehensive data analysis through cloud computing. However, certain requirements and limitations, such as the reluctance of some organizations and users to send and store data outside their internal network, hinder the full utilization of cloud computing for processing. Conversely, many industrial devices often have significant unused processing power that can be harnessed as edge devices within a cloud system. This research aims to introduce a framework for collecting data from sensors and processing it in real-time on edge systems, as well as collaboratively training artificial intelligence models on these edge systems. Given that the most computationally intensive tasks in the smart industry typically involve AI models, particularly their training, utilizing collaborative learning at the edge leverages existing capacities and eliminates the need for more powerful systems. Training models at the edge also addresses the concern of not transferring data outside the organization’s network. This research proposes a platform based on the collaboration and participation of existing edge systems within industries, encompassing initial data collection, pre-processing, model training, and ultimately, model deployment. The proposed method has been eva‎luated using the HAI Security Dataset. The results demonstrate that collaborative learning, utilizing nine nodes, can increase training speed by more than 80% while maintaining AI model performance. The simulated experiment in this research focuses on anomaly detection in the smart industry using time series data. To ensure more reliable results, we employed the eTaPR library and methodology for model eva‎luation. Additionally, the Cassandra database was used in a distributed manner for data collection in the simulation scenario. The distribution feature in this research facilitates the movement and coordination of information between different nodes and enables the implementation of horizontal and vertical cooperative learning regardless of the information collection method.

زينب زالي , مسعودرضا هاشمي

الهام محمودزاده , علي فانيان