Pervasive and Mobile Computing最新文献

In on-device training of machine learning models on microcontrollers a neural network is trained on the device. A specific approach for collaborative on-device training is federated learning. In this paper, we propose embedded federated learning on microcontroller boards using the communication capacity of a LoRa mesh network. We apply a dual board design: The machine learning application that contains a neural network is trained for a keyword spotting task on the Arduino Portenta H7. For the networking of the federated learning process, the Portenta is connected to a TTGO LORA32 board that operates as a router within a LoRa mesh network. We experiment the federated learning application on the LoRa mesh network and analyze the network, system, and application level performance. The results from our experimentation suggest the feasibility of the proposed system and exemplify an implementation of a distributed application with re-trainable compute nodes, interconnected over LoRa, entirely deployed at the tiny edge.

This paper presents SmartSPEC, an approach to generate customizable synthetic smart space datasets using sensorized spaces in which people and events are embedded. Smart space datasets are critical to design, deploy and evaluate systems and applications under issues of heterogeneity, scalability and robustness, leading to cost-effective operation which improves the safety, comfort and convenience experienced by space occupants. However, many challenges exist in obtaining realistic smart space datasets for testing and validation, from a lack of fine-grained sensing to privacy/security concerns. SmartSPEC is a smart space simulator and data generator that leverages a semantic model augmented with user-defined constraints to represent important attributes, relationships, and external domain knowledge for a smart space. We employ machine learning (ML) approaches to extract relevant patterns from a sensorized space, which are used in an event-driven simulation strategy to generate realistic simulated data about the space (events, trajectories, sensor observation datasets, etc.). To evaluate the realism of the generated data, we develop a structured methodology and metrics to assess various aspects of smart space datasets, including trajectories of people and occupancy of spaces. Our experimental study looks at two real-world settings/datasets: an instrumented smart campus building and a city-wide GPS dataset. Our results show the realism of trajectories produced by SmartSPEC ($1.4x$ to $4.4x$ more realistic than the best synthetic data baseline when compared to real-world data, depending on the scenario and configuration), as well as sensor data derived from such trajectories which adhere to the underlying semantics of the smart space as compared to synthetic sensor data baselines, even under hypothetical changes.

In this paper, we propose an architecture, design and build a prototype of a novel IoT system with intelligence, distributed at multiple tiers including the network edge. Our proposed architecture hosts a modular, three-tier IoT system including the edge, gateway (fog) and cloud tiers. The proposed system relies on data acquired by edge devices to realize a distributed machine learning model and achieve timely response at the edge using a lightweight machine learning model. In addition, it employs more sophisticated machine learning models at the higher fog and cloud tiers for wider-scope, long-term decision making. One of the prime objectives of the proposed system is reducing the volume of data transferred across tiers. This is attained through intelligent data filtering at the edge/gateway tiers to distill key events that avail the most relevant data points to higher-tier machine learning models at the gateway and cloud. This, in turn, reduces the outliers and the redundant data that may impact the gateway and cloud models and reduces the inter-tier communications overhead. To demonstrate the merits of our proposed system, we build a proof-of-concept prototype hosting the three tiers, using COTS components and supporting networking technologies. We demonstrate through extensive experiments the merits of the proposed system. A major finding is that our system is capable of achieving prediction performance comparable to the centralized machine learning baseline model, while reducing the inter-tier communications overhead by up to 80%.

Sensor-based human activity recognition (HAR), with the ability to recognise human activities from wearable or embedded sensors, has been playing an important role in many applications including personal health monitoring, smart home, and manufacturing. The real-world, long-term deployment of these HAR systems drives a critical research question: how to evolve the HAR model automatically over time to accommodate changes in an environment or activity patterns. This paper presents an online continual learning (OCL) scenario for HAR, where sensor data arrives in a streaming manner which contains unlabelled samples from already learnt activities or new activities. We propose a technique, OCL-HAR, making a real-time prediction on the streaming sensor data while at the same time discovering and learning new activities. We have empirically evaluated OCL-HAR on four third-party, publicly available HAR datasets. Our results have shown that this OCL scenario is challenging to state-of-the-art continual learning techniques that have significantly underperformed. Our technique OCL-HAR has consistently outperformed them in all experiment setups, leading up to 0.17 and 0.23 improvements in micro and macro F1 scores.