Federated Learning Applications on Sensing Devices: Everything From Medical to Environmental

Table of Links

Highlights

Abstract & Introduction

Federated Learning: Collaborative and Privacy-Preserving Model Training

State of the Art Federated Learning Software Platforms and Testbeds

Sensors and Edge-Sensing Devices

Federated Learning Applications on Sensing Devices

Discussion

Conclusions, Acknowledgment, and References

5. Federated Learning Applications on Sensing Devices

Small sensing devices can capture location-based or person-based data related to specific environments or behaviours. Data collected from multiple devices can be combined according to the application’s requirements to gain a comprehensive understanding of situations or behaviours.

This section explores the diverse application areas where FL techniques have been utilized (see Figure 4). These areas include Human Activity Recognition (HAR), health monitoring, affect detection, abnormality detection, and other relevant areas where sensors on edge devices are utilized.

We are inspired by the study Nguyen et al. (2021) and extended it by adding different application areas. Due to the page limitations, we did not include all studies covering these application areas; instead, we provided examples of studies from these areas in the literature. In Table 1, we present an overview of the covered studies, which are explained in the following sections in detail.

5.1. Medical applications

In today’s world, computational-based techniques are widely used in the healthcare domain, especially with image recognition, e.g., the detection of diseases, the decision of cell types according to sizes and forms. The benefits of using these techniques are fast and accurate diagnosis and treatment Szegedi et al. (2019); Wang et al. (2023).

However, data protection regulations make it impossible to collect and analyze these data easily. In this regard, FL techniques are well suited to healthcare-related applications from the privacy perspective. In addition, intrinsically, FL is designed for a vast number of client participants.

• Disease Diagnosis: The study Ogbuabor et al. (2021) proposes an intelligent and privacy-oriented context-aware decision support system for cardiac health monitoring using physiological and activity data of the patient during rehabilitation, utilizing a federated machine learning approach for activity recognition while maintaining users’ privacy.

The system allows healthcare professionals to assess the health status of patients by considering multiple relevant parameters, such as heart rate, electrocardiogram (ECG) signals, and activity data.

It uses the data collected from Holter monitors and smartphones equipped with accelerometers. The federated machine learning approach enables the development of a collaborative model without sharing sensitive patient data, ensuring data privacy while achieving model generalization.

• Affect Detection: Affective computing Picard (2000) is an active research area for automatically monitoring a person’s mental and emotional state via physiological and physical signals. In Can and Ersoy (2021), the authors applied FL on stress data with the aim of privacy preservation.

The authors applied FL on heart activity data collected with smart bands, and the targeted devices include wrist-worn wearable devices (e.g., Samsung Gear S, Empatica E4) and smartphones, with various physiological sensors such as accelerometers, electrodermal activity, heart rate, etc.

The proposed federated deep learning algorithm achieves encouraging results in stress-level detection while ensuring privacy protection for health-related research conducted with widespread mobile unobtrusive devices.

Two use-case scenarios are considered: creating a general model from individual models while preserving user privacy and aggregating separately collected event data to develop an improved shared model. They applied deep learning algorithm in a federated manner and obtained better results compared to traditional learning methods.

• Human Activity Recognition: HAR aims to detect users’ activities, such as sitting, standing, and walking, mostly via motion sensors. The sensors used in HAR tasks typically include triaxial accelerometers and gyroscopes found in smartphones and smartwatches.

These sensors measure acceleration forces and rotational movements, allowing the models to recognize various activities such as walking, running, biking, and more. In Feng et al. (2020), FL models were applied for trajectory prediction. The framework utilizes various sensing devices and sensors present in mobile devices to gather location-based data.

The paper introduces a group optimization method for training on local devices to achieve a better trade-off between performance and privacy. In Sozinov et al. (2018), authors applied FL to detect basic human activities, where the goal is to recognize different human activities using sensor data from smartphones or smartwatches. They obtained good accuracies but slightly lower than a centralized one. In Ek et al. (2020), they focused on the HAR domain and applied several aggregation algorithms (FedAvg, FedPer, and FedMA) and compared their results.

5.2. Environmental applications

FL has emerged as a powerful approach to tackle environmental challenges like climate change, air pollution, and extreme weather events.

• Air Pollution Monitoring: FL is employed to predict air pollution using sensor networks in Nguyen and Zettsu (2021). The utilized sensing devices are environmental monitoring stations with various sensors, which are not specified in the paper.

Nevertheless, a combination of sensors is typically deployed to collect data on various environmental parameters such as particulate matter (PM), nitrogen dioxide (NO2), ozone (O3), carbon monoxide (CO), sulfur dioxide (SO2), meteorological data (temperature, humidity, wind speed, etc.), and more.

The paper focuses on predicting air pollution levels, particularly Oxidant warning levels, using Convolutional Recurrent Neural Networks (CRNN). The CRNN models are trained locally on data from different spatial areas, such as cities and prefectures, and then the common parts of these models are aggregated at the server to create a global model.

This federated learning framework allows for cooperative training among participants from different regions without exchanging raw data, ensuring data privacy and reducing data transmission latency.

• Weather Forecasting: FL is utilized for weather prediction using distributed weather stations Chen et al. (2023). The stated paper introduces a novel machine learning approach, MetePFL, for weather forecasting on a global scale. The proposed approach utilizes a foundation model (FM) pre-trained on extensive weather data and finetuned by regions to capture local weather patterns.

MetePFL employs prompt federated learning to train a Transformer-based FM collaboratively across participants with heterogeneous meteorological data. A spatiotemporal prompt learning mechanism is introduced to handle multivariate time series data efficiently, improving forecasting accuracy.

The experiments conducted on accurate weather forecasting datasets demonstrate the effectiveness and superiority of the MetePFL approach over traditional fine-tuning and other federated learning methods.

By utilizing only a fraction of the model’s parameters, MetePFL achieves excellent performance while reducing inter-device communication overhead, making it a promising solution for comprehensive weather forecasting on large-scale meteorological data.

5.3. Smart city applications

FL techniques are also applied to integrated city solutions related to food, water, and energy So et al. (2021). Applications use data collected from cameras, temperature sensors, and traffic sensors to understand human activities, population density, and pollution Jiang et al. (2020).

• Traffic Pattern Analysis: FL is employed for traffic flow prediction in smart cities Liu et al. (2020). The paper Liu et al. (2020) presents FedGRU, a novel privacy-preserving traffic flow prediction algorithm based on FL.

The goal is to develop a collaborative model that can accurately predict traffic patterns and congestion levels using data from various sources, such as traffic cameras and sensors. The proposed algorithm uses a secure parameter aggregation mechanism.

The paper introduces a joint-announcement protocol to handle large-scale scenarios, randomly selecting a subset of organizations for each training round, effectively mitigating communication challenges. Empirical case studies demonstrate that FedGRU achieves comparable prediction accuracy to centralized models while maintaining robust privacy protection.

An ensemble clustering-based approach is proposed to enhance model performance by leveraging spatiotemporal correlations among organizations’ data.

• Smart Transportation: Transportation is composed of two parts: vehicular traffic planning and resource management Nguyen et al. (2021). Traffic planning is essential for traffic prediction and congestion minimization. In Elbir et al. (2022), authors replaced centralized ML models with FL models for traffic prediction cases and ran the model on edge devices. As data, they utilized traffic flow, weather, and road geometry. In Liang et al. (2022), a traffic simulator is designed with FTL to guide Reinforcement learning agents’ (FTRL) driving. The proposed framework provides collision avoidance as well.

5.4. Industrial applications

Processing industrial data with AI techniques makes possible high-precision location detection, the creation of early warning systems, and the detection of abnormal behaviors. Thus, with smart security, pre-warnings, and postevent analysis can be performed. The manufacturing process comprises modeling, monitoring, prediction, and control stages Ge et al. (2017). The smart industry is a concept about integrating ML techniques into this process.

• Predictive Maintenance: FL is applied to develop models for predictive maintenance in industrial settings in Pruckovskaja et al. (2023). The paper evaluates FL aggregation methods for predictive maintenance and quality inspection in Industry 4.0, ensuring data privacy. Four strategies, FedAvg, FedProx, qFedAvg, and FedYogi, are explored. Data distributions significantly impact FL performance.

The authors introduce a real-world quality inspection dataset (FLADI) with diverse product variants as clients. FL relies on decentralized training, where clients train local models and share updates with a central server. Targeted sensing devices include machines, production units, sensors, and intelligent factories. FL’s suitability varies based on data distribution and feature heterogeneity among clients.

• Monitoring and Anomaly Detection: In Han et al. (2019); Liu et al. (2020), authors use FL for monitoring the environment collaboratively using various sensors to detect defects. In Zhai et al. (2021), authors developed an anomaly detection system using FL for IoT devices equipped with an ammeter, water meter, and camera sensors. In Mowla et al. (2019), it is used for attack detection for Unmanned Aerial Vehicles, which can be considered anomaly detection.

• Robotics and Industry 4.0: FL is used for allocating devices for realtime data processing in robotics. For instance, in Zhou et al. (2018), authors used FL to decrease the communication time, facilitating learning on the device.

In Liu et al. (2020, 2019), FL is applied for learning an imitation scheme. Each device runs the NN model independently and shares the parameters with the server, which increases the global knowledge knowledge leading to more accurate results.

5.5. Aerospace applications

The aerospace industry is equipped with cameras (for visual data), LiDAR (Light Detection and Ranging) sensors (for 3D mapping and obstacle detection), GPS (Global Positioning System) receivers (for location information), and inertial sensors (gyroscopes and accelerometers for attitude and motion tracking), is continually seeking innovative ways to enhance the safety, efficiency, and reliability of aircraft operations.

Key to achieving these objectives is the development of advanced technologies that enable real-time monitoring of aircraft health and optimizing critical resources Fu et al. (2023).

• Aircraft Health Monitoring: FL is employed for aircraft monitoring using distributed sensor networks in Li et al. (2022). The study addresses the challenge of intelligent fault diagnosis in aerospace equipment, focusing on bearings as critical components. Data-driven approaches have shown promise, but the scarcity of labelled data and data distribution discrepancies among different equipment hinder model training.

The authors propose a novel method using Clustering Federated Learning (CFL) with a self-attention mechanism to overcome this. Sensors collect vibration signals from various devices, and the CFL approach clusters clients with similar data distributions.

The proposed method achieved superior fault diagnosis results compared to other methods, demonstrating its effectiveness in leveraging distributed data for improved safety and maintenance of aerospace equipment.

• Power Allocation and Scheduling: FL is applied to optimize power and schedule in the aviation industry. The goal is to develop a shared model to improve fuel efficiency, reduce emissions, and enhance flight safety. Airlines or aviation companies train local models using their flight data, and the models’ updates are aggregated to create a global model.

By collaboratively learning from diverse flight patterns, FL enables optimized flight planning without sharing sensitive flight information. The paper Zeng et al. (2020) presents a novel framework for distributed FL in a swarm of unmanned aerial vehicles (UAVs).

Each UAV trains a local FL model based on its data, and a leading UAV aggregates these models to generate a global one, optimizing convergence using joint power allocation and scheduling. The study analyzes how wireless factors impact FL performance, highlighting the effectiveness of the proposed approach.

5.6. Agricultural applications

The agricultural applications exemplify how integrating smart sensors and FL revolutionizes the state of the art. By empowering farmers with automated disease detection, crop monitoring, and efficient resource management, these technologies play a pivotal role in maximizing agricultural productivity and sustaining livelihoods in the agricultural sector.

• Crop Disease Detection: FL is utilized for collaborative crop disease detection in Ebenezer et al. (2022). The method involves training machine learning algorithms across a decentralized network of edge devices or servers, each retaining its local data samples.

By utilizing image datasets of crop diseases and employing convolutional neural networks (CNNs), FL enables the creation of a powerful deep-learning model for disease identification. This approach eliminates the need for costly and continuous professional inspection, making disease detection more accessible and cost-effective for farmers, even in remote areas.

FL’s privacy-preserving nature ensures farmers can contribute their crop images without sharing raw data. Implementing FL for crop disease detection promises to enhance agricultural productivity, ensuring food security and economic growth for farming communities and the entire country.

• Agricultural Production: Another application area of FL is agricultural production Abu-Khadrah et al. (2023). In this study, authors utilize various smart sensors for agricultural monitoring and control. The specific sensors used include optical sensors, mechanical soil sensors, location sensors, electrochemical sensors, dielectric soil moisture sensors, airflow sensors, and electronic sensors.

These sensors are deployed in smart agricultural plots to sense crucial parameters such as temperature, pressure, humidity, soil moisture, crop health (infections and growth level), and climate conditions.

The target device for this application is the smart sensor system, which integrates with IoT technologies and cloud computing for decentralized data processing and global actuation. The sensor system, using FL, collaborates with the cloud server and other sensor devices to optimize sensor controls based on past accumulated data.

By modifying the sensor operations according to the validated and optimal data, the smart sensors enhance agricultural production and maximize productivity in various agricultural plots.

This paper is available on Arxiv under a CC 4.0 license.