Smartphone Architecture for Edge-Centric IoT Analytics

Bockarie Daniel Marah, Zilong Jing, Tinghuai Ma, Raeed Alsabri, Raphael Anaadumba, Abdullah Al-Dhelaan, Mohammed Al-DhelaanSensors 2020 · DOI

TL;DR

This paper introduces SMIoT, a novel edge-centric IoT architecture that leverages smartphones for real-time data analytics. By enabling on-device model deployment, evaluation, and updating, SMIoT significantly reduces latency and communication overhead compared to traditional cloud-centric approaches, making IoT solutions more efficient and responsive.

What Problem Does This Paper Address?

The Internet of Things (IoT) generates vast amounts of data, traditionally processed in the cloud. However, this cloud-centric approach faces challenges like high latency, increased communication costs, and potential privacy concerns. Edge computing aims to address these by moving computation closer to the data source. This paper focuses on utilizing smartphones, ubiquitous and powerful devices, as edge computing nodes for IoT analytics. The core problem addressed is how to effectively design and implement an edge-centric IoT architecture on smartphones that can perform complex tasks like data capture, preprocessing, model deployment, and real-time evaluation and updating, while maintaining efficient communication with a central core.

Methods

The proposed architecture, SMIoT (Smartphone IoT Architecture), adopts a "default to fitted" model styled approach. This means an initial, generalized model is deployed on the smartphone, which is then continuously evaluated and updated in real-time using local sensor data.

Key methodological components include:

  • Edge-Centric Data Processing: Smartphones are equipped to handle data capture, preprocessing (univariate analysis, missing data treatment, outlier detection), and model evaluation directly at the edge.
  • Artificial Neural Networks (ANNs) for Regression: The core component for analytics involves training ANN models to learn relationships between sensor inputs and outputs for regression tasks.
  • Model Deployment and Caching: Trained models, along with their associated normalizers and evaluators, are deployed and cached on the smartphone for fast, local inference.
  • Real-time Model Evaluation: A novel evaluation scheme using confidence intervals (CI) is employed on the smartphone to assess the performance of the deployed model against incoming data. This scheme raises flags when predictions fall outside the acceptable CI range.
  • Selective Data Forwarding: Only when the model's performance degrades (indicated by evaluation flags) is relevant data sent back to the core for retraining. This significantly reduces communication overhead.
  • Core-Side Model Creation and Updating: The core is responsible for initial model creation using pre-gathered datasets and for retraining models based on the data received from smartphones.
  • Sustainable Machine Learning: The concept of sustainable machine learning is adopted, ensuring that models can adapt to changing data subspaces in real-time.

Datasets

  • Training Dataset: A dataset of light intensity data collected from an Android smartphone (OPPO A83t) in a university lab was used. This dataset contained light intensity values (in Lux) and timestamps over an 8-hour period with 1-second granularity, totaling 3066 instances. This dataset was pre-processed to remove outliers and normalize values.
  • Test Dataset: A reserved dataset for the 18th hour (6 pm) of the day from the same location was used for testing the default model.
  • New Location Data: Data collected from a smartphone moved to a new location (a park) was used to demonstrate the model's adaptation to environmental changes.

The paper does not introduce a new benchmark dataset but utilizes real-world data collected by a smartphone.

Key Insights and Ideas

  • Smartphones as Powerful Edge Nodes: The research demonstrates that smartphones, with their integrated sensors and processing capabilities, are highly suitable for performing sophisticated edge analytics, moving beyond simple data collection.
  • "Default to Fitted" for Adaptability: The "default to fitted" model strategy is crucial for maintaining model relevance in dynamic IoT environments. It allows models to adapt to changing conditions without constant manual intervention.
  • Confidence Intervals for Effective Evaluation: Using confidence intervals as a primary metric for model evaluation at the edge provides a robust and statistically sound method for detecting performance degradation.
  • Selective Data Forwarding is Key to Efficiency: Transmitting only necessary data for retraining, rather than all raw data, drastically reduces communication overhead and energy consumption, a critical aspect for mobile edge devices.
  • Reduced Latency for Real-time Applications: By performing inference and evaluation locally on the smartphone, the system achieves significantly lower latency, enabling more responsive IoT applications.

Key Results

  • High Performance Score: The default ANN model achieved a performance score of 98.248% on the test dataset, indicating that most predictions fell within the 95% CI band.
  • Effective Adaptation to Environmental Changes: When a smartphone was moved to a new location (park), the model quickly adapted to the significantly different light intensity range (from 145-195 Lux to 8000-21,000 Lux). This adaptation was confirmed by subsequent testing.
  • Reduced Communication Overhead: The evaluation algorithm resulted in only 52% of the total data being transmitted to the core during the experiment, demonstrating the effectiveness of selective data forwarding.
  • Low Latency: The edge-centric approach significantly reduces latency, as queries are evaluated locally on the smartphone.

Limitations and Future Work

Limitations:

  • Sensor Availability and Accuracy: Not all smartphones have the same sensors, and their accuracy can vary.
  • Smartphone Portability: Smartphones are often carried in various positions, which might not be ideal for all applications unless dedicated.
  • External Sensor Integration: The current architecture relies on built-in smartphone sensors, which may have limitations in certain scenarios.

Future Work:

  • Integration with External Sensors: The authors propose linking the architecture with external sensors to collect data in a more coordinated manner, overcoming the limitations of built-in smartphone sensors.
  • Broader Application Scope: The proposed architecture is demonstrated with a light intensity prediction use case but is intended to be applicable to a wide range of IoT solutions.

Summary updated

BibTeX
@article{marah2020smartphone,
  author = {Bockarie Daniel Marah and Zilong Jing and Tinghuai Ma and Raeed Alsabri and Raphael Anaadumba and Abdullah Al-Dhelaan and Mohammed Al-Dhelaan},
  title = {{Smartphone Architecture for Edge-Centric IoT Analytics}},
  journal = {Sensors},
  year = {2020},
  doi = {10.3390/s20030892},
}
Figure