PaperlayerView exact revision on arXiv

Source-linked AI summary

Machine learning and data analytics for the IoT

Erwin Adi, Adnan Anwar, Zubair Baig, Sherali Zeadally

arXiv:2007.04093v1eess.SPcs.LG

TL;DR

IoT applications generate large data volumes, but heterogeneous infrastructures and communication protocols complicate intelligent, adaptive analytics. The paper reviews machine learning for IoT, proposes a framework for cross-application learning and knowledge discovery, and illustrates it through a literature-based case study.

  • Problem

    Heterogeneous IoT infrastructures and application-layer protocols hinder intelligent applications from adaptively learning from other IoT applications.

  • Method

    The paper reviews machine-learning data processing for IoT and proposes a framework using knowledge discovery, ontologies, inference, and shared databases.

  • Results

    The case study shows ontology-based inference turning speculative relationships into asserted knowledge from collected IoT data samples.

  • Takeaways & Limitations

    The framework supports adaptive learning across IoT applications locally at the edge, through fog, or in the cloud.

  • Takeaways & Limitations

    Current network systems are designed for specific applications, constraining cross-domain convergence among community-network nodes.

Abstract

from arXiv · show

The Internet of Things (IoT) applications have grown in exorbitant numbers, generating a large amount of data required for intelligent data processing. However, the varying IoT infrastructures (i.e., cloud, edge, fog) and the limitations of the IoT application layer protocols in transmitting/receiving messages become the barriers in creating intelligent IoT applications. These barriers prevent current intelligent IoT applications to adaptively learn from other IoT applications. In this paper, we critically review how IoT-generated data are processed for machine learning analysis and highlight the current challenges in furthering intelligent solutions in the IoT environment. Furthermore, we propose a framework to enable IoT applications to adaptively learn from other IoT applications and present a case study in how the framework can be applied to the real studies in the literature. Finally, we discuss the key factors that have an impact on future intelligent applications for the IoT.

1 Introduction

IoT generates substantial data across heterogeneous infrastructures and protocols, creating a need for machine-learning-based analytics. The paper reviews this role and proposes a framework for IoT applications to learn adaptively from heterogeneous sources.

  • Motivation: IoT devices generate large data volumes that require processing and analysis to produce knowledge.Data may be locally processed in limited ways before transfer to centralized computing or cloud storage.
  • IoT communication constraints: MQTT and AMQP support longer messages but consume more power, whereas CoAP targets resource-constrained devices and HTTP suits higher-end devices.Protocol suitability depends on device power, computation, bandwidth, storage, and message requirements.
  • Computational infrastructure: Cloud, fog, and edge computing provide alternative infrastructures for collecting, storing, processing, and analyzing IoT data.Their integration with IoT can provide real-time insights and support machine-learning-based intelligent analytics.
  • Paper scope: The paper analyzes machine learning for IoT data analytics, reviews its integration with IoT, and proposes a framework for heterogeneous data sources.It also presents future research directions and concludes with implications for machine-learning applications in IoT.

2 The convergence of machine learning and IoT

The convergence of machine learning and IoT supports analytics over large, heterogeneous, and uncertain data streams, but requires balancing scalability, speed, and accuracy. The section reviews analytical classes and explains how these techniques process IoT data into insights and recommendations.

  • Convergence: Machine learning and IoT can improve communication, computation, controllability, decision making, efficiency, accuracy, productivity, and cost savings.The paper presents their complementarity as important for addressing complex real-world problems under computation and communication constraints.
  • Drivers: 18 billion connected IoT devices were forecast globally by 2022, increasing the need for intelligent analytics of distributed-device data.Analytics can identify and predict future states by mining large data volumes.
  • Drivers: Heterogeneous IoT sources produce data with differing features, formats, attributes, and domain-specific characteristics.This variability complicates unified processing across application areas.
  • Drivers: IoT data streams contain uncertainty from device or communication failures, missing data, gross errors, and possible cyber intrusions.Advanced preprocessing is required to support accurate decision making.
  • Drivers: IoT analytics must balance scalability with speed and accuracy when many devices generate data for time-critical decisions.Cloud transfer can introduce delay, making the trade-off especially important in connected-vehicle environments.
  • Analytical classes: Descriptive, predictive, prescriptive, and adaptive analytics respectively summarize data, forecast outcomes, recommend responses, and adjust outcomes using recent data.These classes span historical summarization, future prediction, recommended action, and ongoing optimization.

2.2 Classification of IoT data analytics based on Technological Infrastructure

IoT data analytics is distributed across cloud, edge, and fog infrastructures with different locations, capabilities, and trade-offs. Cloud supports centralized and complex analysis, edge supports low-latency local processing, and fog bridges the two.

  • Cloud computing: Cloud computing centrally stores, gathers, manages, and processes IoT data through Internet-based services.Cloud architectures may be public, private, hybrid, or community-based.
  • Edge computing: Edge computing processes IoT data near the source network, reducing transfer bandwidth, cost, power use, and single-point-of-failure risk.It is suited to low-latency real-time operations rather than delay-tolerant complex analysis.
  • Fog computing: Fog computing lies between cloud and edge and acts as a bridge between cloud and edge resources.It provides networking, computing, and storage capabilities between end devices and the cloud for latency-sensitive applications.
  • Analytics placement: Most computationally intensive analytical approaches suit cloud servers, while lightweight online algorithms better fit edge computing.This reflects the differing computational resources and processing capabilities across the infrastructures.

2.3 Applications of IoT Data Analytics

IoT data analytics supports decision-making across vehicles, healthcare, agriculture, and energy systems, but applications must handle heterogeneous data, distributed infrastructures, and security requirements.

  • Smart vehicles: Connected-vehicle analytics addresses high-volume broadcasting data, intelligent management, security, and decision-making needs.Centralized approaches may be infeasible because they are more prone to a single point of failure.
  • Smart vehicles: Machine learning and analytical frameworks support vehicle fault detection, warranty analysis, parking prediction, crime detection, and traffic management.Reported applications include low-false-alarm fault detection, Bayes and Logistic Regression models, occupancy prediction, fog-based crime analytics, and periodic pattern mining.
  • Smart vehicles: Transportation architectures combine sensing and communication layers with application-layer optimization and machine learning for subway, bus, and shared-taxi decisions.The three-tier framework is applied to scheduling and traffic management in connected public transportation.
  • Smart healthcare: Healthcare analytics uses IoT data for real-time diagnosis, patient modeling, sleep monitoring, nutrition monitoring, dyslexia detection, and emotion detection.Studies combine fog or edge preprocessing, cloud processing, Bayesian or neural models, distributed ledgers, and signal-processing methods.
  • Agriculture and energy systems: Smart agriculture and energy systems use IoT-generated data with advanced analytics, while machine learning is identified as relevant to cyber-attack detection in smart grids.Smart-farming architectures include physical, acquisition, processing, and analytics components; three of the four major components are data-related.

3 A critical review on data processing and knowledge discovery for IoT

IoT knowledge discovery depends on transforming raw sensor data through suitable processing architectures and application-layer protocols. The review identifies resource, interoperability, and adaptive-learning limitations across these designs.

  • IoT consumers receive meaningful knowledge when AI algorithms such as machine learning process raw data for machines or humans.
  • Application-layer protocols: CoAP suits constrained edge devices, while MQTT and AMQP support higher-end devices and HTTP is the most resource-intensive protocol.CoAP-enabled applications generally support prescriptive and predictive AI, whereas MQTT and AMQP scenarios can also support adaptive AI.
  • Knowledge-creation architectures: Knowledge-creation architectures vary with available computing frameworks and network bandwidth, using simplex remote processing, duplex feedback, or local and edge processing.Simplex architectures cannot return feedback to constrained devices; duplex and resource-rich edge architectures can support adaptive learning closer to the device.
  • Critical review: Current studies justify data-processing locations and protocol choices, but application protocols still provide limited support for IoT devices to adaptively learn.
  • Deriving knowledge from data: Raw IoT sensor streams combine insights with noise, requiring processing that derives meaningful attributes and relationships for consumers.Ontologies label data and represent object properties and relationships, but current work still lacks holistic output sharing between machines.
  • Smart objects: Ontology-based systems remain difficult to reuse across domains because semantic service identification disintegrates when constructed ontologies are applied elsewhere.

4 A framework for machine learning and knowledge discovery for IoT

The paper proposes a framework for machines to exchange knowledge, create new knowledge, and adaptively learn across different IoT domains or case studies. It adapts human learning and classification to autonomous Smart Objects.

  • The proposed framework enables machines to autonomously exchange knowledge, create new knowledge, and adaptively learn across different domains or case studies.

4.1 A review of intelligent big data analytics for IoT

The reviewed IoT analytics studies span smart city, production, building-management, and health applications. They use machine learning to predict, classify, detect, and recognize events from sensor data.

  • The reviewed case studies cover smart city, production, building management, and health domains.Examples include traffic routing, farming, manufacturing maintenance, occupancy prediction, and health-risk detection.
  • Production-farming: 91 percent accuracy enabled Random Forest to detect dairy-cow lameness one day before some visual signs appeared.
  • Production-farming: 81 percent accuracy enabled k-Nearest Neighbor to detect dairy-cow lameness three days before symptoms were visually noticeable.
  • Production-manufacturing: 100 percent accuracy enabled Neural Networks to identify faults in factory motors, with confidence levels between 80 percent and 99 percent.
  • Building management: 95 percent accuracy enabled Random Forest to predict room occupancy from building temperature, CO2, air-volume, and air-conditioning data.
  • Health: Deep learning achieved 94 percent accuracy in real-time detection of anomalous health risks from wearable-device data.

4.2 Learning from human-computer interaction

The paper derives two knowledge-sharing principles from human-computer interaction: machines should use human-understandable names and semantically structured information. It proposes ontologies to let machines share knowledge and learn adaptively from one another.

  • The proposed principles are presented as two characteristics of knowledge sharing drawn from human-computer interaction.
  • Human-understandable naming should replace machine-only abbreviations, codes, or binary mappings for entities, properties, relationships, and services.
  • Semantic organization should let users derive meaning by following the structure of information.
  • Ontologies provide the semantic model for machines to autonomously share knowledge and adaptively learn from other machines.

4.3 Epistemology

The paper classifies machine knowledge using Primary, Secondary, and Invented categories adapted from epistemological distinctions. These categories can change as sensor observations are generalized or hypotheses are scientifically verified.

  • Primary knowledge consists of definitions and classifications built into ontologies.The paper illustrates this with the rule “If the soil is wet then crops flourish”.
  • Secondary knowledge is obtained from IoT sensor observations.The paper illustrates this with sensor observations about soil conditions.
  • Invented knowledge consists of hypotheses created through abductive inference.The paper describes speculative indicators such as sunlight and fertilizers for crop flourishing.
  • Secondary knowledge can become Primary when generalized, while scientifically verified Invented knowledge becomes Secondary.

4.4 A framework of machine knowledge

The proposed framework lets Smart Objects exchange and classify knowledge across Ontology, Parameters, and Hypotheses databases. It combines ontology sharing, NLP, induction, and abduction to support adaptive learning between machines.

  • The framework contains Ontology, Parameters, and Hypotheses databases that store inference rules, name-value data, and non-verified rules.
  • Smart Objects advertise and search for services while tagging exchanged knowledge as Primary, Secondary, or Invented.
  • Inductive learning can promote verified Secondary knowledge to Primary knowledge using rule learning, classification, or Bayesian inference.
  • Abduction creates speculative rules when observed values fail to establish a class, such as inferring “apple” from observing “red”.
  • NLP supports human-language naming by identifying synonyms, homonyms, and word categories when speculative rules are generated.
  • Ontology exchange structures inference rules so that machine-shared relationships convey meaning in human-understandable form.

4.5 Case Study

The case study applies the framework to knowledge transfer between Smart Objects using fall detection and dairy-cow lameness studies. It shows how transferred rules can generate hypotheses, which later become assertions after relevant human data are collected.

  • 4.5 Case Study: The case study populates SO1 and SO2 with ontologies and parameters derived from fall-detection and dairy-cow lameness studies.
  • 4.5.1 From Primary knowledge sharing: SO1 queries SO2 for Primary knowledge and receives event labels Normal, Active, Dormant plus sensor concepts Lying time, Swaps/hour, and Step count.
  • 4.5.2 From Secondary knowledge sharing: Transferred Secondary values may not immediately apply because cow measurements and relevant indicators differ from human measurements and tasks.
  • 4.5.3 From Secondary knowledge becoming Primary knowledge: After collecting enough human data, SO1 can verify whether lying-time values correspond to events and accept the relationship as Primary knowledge.
  • 4.5.4 From Primary knowledge becoming Invented knowledge: Through abductive reasoning over ontology relationships, SO1 speculates that its User is a Person and stores the resulting relationship in the Hypotheses database.
  • 4.5.5 From Invented knowledge becoming Secondary knowledge: Activated relationships are asserted after data samples support them, including the previously speculative equivalence between Person and User.

4.6 Discussion

The framework enables IoT applications to exchange and build knowledge across domains through ontologies, human-language naming, and NLP. However, it does not solve adaptive learning under severe resource constraints and adopts a weak definition of intelligence.

  • The framework applies to smart home, farming, smart city, and health domains for autonomous knowledge exchange and construction.
  • The framework does not address adaptive learning when IoT computing and network resources are highly constrained, instead relying on cloud or high-performance edge computing.
  • Natural Language Processing for machine-to-machine communications: It supports ontology integration by mandating human-language names for entities and relationships, with NLP adapting meanings across ontologies.
  • Natural Language Processing for machine-to-machine communications: Three databases—the Ontology, Parameters, and Hypotheses databases—support machine-to-machine communication and autonomous knowledge expansion through scientific abduction.
  • Drawbacks: The case study uses a weak definition of intelligent systems and gives little attention to whether strong AI characteristics would be appropriate for IoT.

5 Future research opportunities

Future research focuses on securing intelligent IoT systems, moving analytics toward the edge, improving scalability and distributed learning, and addressing hyper-convergence and regulatory concerns. These directions are motivated by heterogeneous devices, constrained resources, cyber threats, and growing data volumes.

  • Improved cyber security: Machine-to-machine NLP introduces a cyber-security risk because homonyms and lexical-semantic variation can enable malicious nodes to spread deceptive knowledge.
  • Improved cyber security: IoT security research should improve security by design, cryptography, key management, authentication, PKI, and lightweight intrusion detection across devices, edge, and cloud.
  • Machine learning at the edge: Edge analytics can reduce the time and bandwidth required for cloud transmission, although edge devices generally provide less computing capacity than cloud data centers.
  • Machine learning at the edge: Future work includes federating edge devices to share knowledge, delegate processing, communicate with the cloud, and vote on which device transforms data into knowledge.
  • Scalability: Distributed machine learning with edge computing is proposed to address scalability by placing computing decisions closer to IoT devices.
  • Hyper convergence: Hyper-convergence can combine distributed storage performance and shared Ontology, Parameters, and Hypotheses databases, but seamless node addition, faults, heterogeneous technologies, and increasing data remain unresolved.
  • Future research should examine regulation because distributed intelligent systems may consume substantial power and compete with human basic needs.

6 Conclusion

The paper concludes that resource constraints and cross-domain integration hinder adaptive machine-learning solutions for IoT. Its framework supports adaptive learning locally, at the edge, through fog, or in the cloud, while future work must address security, resource distribution, scalability, and edge federation.

  • IoT devices remain constrained in computation and communication resources, limiting adaptive intelligent solutions and making cross-domain integration difficult.
  • The proposed framework provides pathways for integrating adaptive learning locally, at the edge, through fog, or in the cloud.
  • Future work includes mitigating malicious-machine influence, seamlessly distributing resources, scaling connected devices, and studying federated edge devices.
Loading 2007.04093v1…