Privacy-Aware Resource Sharing in Cross-Device Federated Model Training for Collaborative Predictive Maintenance

Abstract

The proliferation of Industry 4.0 has made modern industrial assets a rich source of data that can be leveraged to optimise operations, ensure efficiency, and minimise maintenance costs. The availability of data is advantageous for asset management, however, attempts to maximise the value of this data often fall short due to additional constraints, such as privacy concerns and data stored in distributed silos that is difficult to access and share. Federated Learning (FL) has been explored to address these challenges and has been demonstrated to provide a mechanism that allows highly distributed data to be mined in a privacy-preserving manner and offering new opportunities for a collaborative approach to asset management. Despite the benefits, FL has some challenges that need to be overcome to make it fully compatible for asset management or more specifically predictive maintenance applications. FL requires a set of clients that participate in the model training process, however, orchestration, device heterogeneity and scalability can hinder the speed and accuracy in the context of collaborative predictive maintenance. To address this challenge, this work proposes a split-learning-based framework (SplitPred) that enables FL clients to maximise available resources within their local network without compromising the benefits of a FL approach (i.e., privacy and shared learning). Experiments performed on the benchmark C-MAPSS dataset demonstrate the advantage of applying SplitPred in the FL process in terms of efficient use of resources, i.e., model convergence time, accuracy, and network load.

Blockchain-Based Decentralized Model Aggregation for Cross-Silo Federated Learning in Industry 4.0

“…The lifetime of an engine is estimated in terms of the number of operation cycles before the engine runs to failure. The pre-processing of data is performed in a manner similar to what is discussed in [3]. A simple 2-hidden layers MLP model (Table 1) is used as the global ML model.…”

“…FL deployment strategies are categorized as either (1) crosssilo or (2) cross-device. While clients in cross-silo settings are limited in numbers (2-100) of organisations, cross-device clients are generally referred to as a large number (>100,000) of resource-constrained IoT edge devices installed at industry premises [3]. Traditionally, data across the clients share a feature space for different sample IDs (known as Horizontal FL), however, the data across clients can also differ in terms of features for the same sample IDs (known as Vertical FL).…”

Data mining in predictive maintenance systems: A taxonomy and systematic review

“…In the case of de Carvalho Chrysostomo et al (2020), the forecasting of the critical variables is part of a complete framework for decision support. In Bharti and McGibney (2021), MLP is used as a baseline model to implement a federated learning framework that enables several clients to train local predictive models that are updated by the other clients preserving their privacy. Beyond MLP, other types of ANN have been explored in regression task for PdM.…”

CoRoL: A Reliable Framework for Computation Offloading in Collaborative Robots

“…Generally, approaches around how to offload are classified into system level, method level and application level offloading. In the context of modern cobot scenarios and data privacy requirements, this decision is mainly governed by privacy preserving offloading techniques [13]. This paper focuses on the decision of whom to offload which involves the selection of a reliable offloading server which can collaboratively execute a complex ML task.…”

“…The values of α is calculated as 250 J, whereas the value of β is calculated in terms of maximum number of tasks a cobot can execute in a day (kept as 2). The value of γ (100 MB) is adopted from the empirical results presented in [13] about the memory expenditure in split-learning. Another algorithm parameter i.e., cobot's transmission range's value (25m) is assumed to be half of the factory-floor area.…”