PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Learning for IoT Big Data and Streaming Analytics: A Survey

Mehdi Mohammadi, Ala Al-Fuqaha, Sameh Sorour, Mohsen Guizani

arXiv:1712.04301v2cs.NIcs.DBcs.LG

TL;DR

IoT produces heterogeneous big data and time-sensitive streams that require analytics for insight, prediction, and control. The survey synthesizes deep-learning architectures, applications, deployment strategies, and challenges across IoT data and infrastructure. It reports deep learning’s broad use in IoT while identifying limitations in evaluation realism, regression coverage, and theoretical understanding.

  • Problem

    IoT generates big and fast data streams whose complexity and timing requirements exceed traditional learning approaches, creating a need for suitable analytics.

  • Method

    The paper surveys IoT data characteristics, deep-learning architectures and algorithms, IoT applications, frameworks, and deployment from devices through fog and cloud.

  • Results

    The survey synthesizes IoT deep-learning approaches across big-data and streaming analytics, applications, smart devices, and fog- and cloud-centric deployment.

  • Takeaways & Limitations

    The survey provides a basis for selecting deep-learning approaches across foundational IoT services, vertical applications, and deployment levels.

  • Takeaways & Limitations

    Reported deep-learning improvements remain empirical, while realistic IoT evaluation data and regression-focused applications remain insufficiently addressed.

Abstract

from arXiv · show

In the era of the Internet of Things (IoT), an enormous amount of sensing devices collect and/or generate various sensory data over time for a wide range of fields and applications. Based on the nature of the application, these devices will result in big or fast/real-time data streams. Applying analytics over such data streams to discover new information, predict future insights, and make control decisions is a crucial process that makes IoT a worthy paradigm for businesses and a quality-of-life improving technology. In this paper, we provide a thorough overview on using a class of advanced machine learning techniques, namely Deep Learning (DL), to facilitate the analytics and learning in the IoT domain. We start by articulating IoT data characteristics and identifying two major treatments for IoT data from a machine learning perspective, namely IoT big data analytics and IoT streaming data analytics. We also discuss why DL is a promising approach to achieve the desired analytics in these types of data and applications. The potential of using emerging DL techniques for IoT data analytics are then discussed, and its promises and challenges are introduced. We present a comprehensive background on different DL architectures and algorithms. We also analyze and summarize major reported research attempts that leveraged DL in the IoT domain. The smart IoT devices that have incorporated DL in their intelligence background are also discussed. DL implementation approaches on the fog and cloud centers in support of IoT applications are also surveyed. Finally, we shed light on some challenges and potential directions for future research. At the end of each section, we highlight the lessons learned based on our experiments and review of the recent literature.

I. INTRODUCTION

IoT applications generate heterogeneous, noisy, large-scale data that may be accumulated as big data or arrive as fast streams requiring timely analytics. The survey positions deep learning as a suitable approach for extracting knowledge from these demanding data and reviews its IoT-specific use.

  • IoT data characteristics: IoT data come from diverse devices and can be heterogeneous, spatially and temporally correlated, and noisy.
  • Motivation for deep learning: Traditional inference and learning approaches are insufficient for extracting hidden knowledge from complex IoT big data, motivating new algorithms and infrastructures.
  • IoT data analytics: IoT data support two analytics treatments: accumulated big data and continuously generated streaming data requiring prompt analysis and fast decisions.
  • IoT streaming analytics: Time-sensitive applications such as autonomous driving, fire prediction, and posture recognition demand fast processing and near-real-time actions.
  • Deep learning for IoT: Deep learning reduces handcrafted feature engineering and improves accuracy, while its use in IoT spans diverse application domains and both big-data and real-time analytics.

C. Contributions

The survey organizes the intersection of IoT and deep learning around data characteristics, architectures, applications, deployment levels, and open challenges. It covers both big-data and streaming analytics, from constrained devices through fog and cloud platforms.

  • Scope: The survey identifies key IoT data characteristics and issues relevant to adopting deep learning in IoT ecosystems.
  • Methods: It reviews state-of-the-art deep learning methods for both IoT big-data and streaming-data analytics.
  • Applications: It surveys IoT applications using deep learning and compares DNN types while providing guidance for matching models to domains and applications.
  • Deployment: It reviews deployment approaches for deep learning across the IoT hierarchy, from resource-constrained devices to fog and cloud.
  • Challenges: The paper highlights challenges and future research directions for combining deep learning with IoT applications.

B. IoT Big data

IoT big data is characterized by multiple dimensions including volume, velocity, variety, veracity, variability, and value. The survey discusses the resulting analytics and deployment challenges alongside deep learning’s representation-learning capabilities and unresolved theoretical limitations.

  • 6V characteristics: IoT big data is defined through six characteristics: volume, velocity, variety, veracity, variability, and value.
  • 6V characteristics: IoT generates data in varied forms, at high and changing rates, while data quality and usefulness affect the reliability and value of analytics.
  • Analytics challenges: Big-data analytics must address scalability, data transfer, and cleansing as IoT sources expand across multiple datasets.
  • Deployment challenges: A major gap is the lack of deep-learning frameworks and algorithms deployable on fog or IoT devices, where model depth must be traded against performance.
  • Deep learning: Deep learning architectures learn hidden features from raw data through layered processing, but reported improvements remain empirical without a concrete analytical explanation.

A. Architectures

The survey introduces common deep learning architectures, their basic components and training process, and the data types or tasks they address in IoT applications.

  • A. Architectures: Deep learning models use multiple processing layers to learn hierarchical representations, and are broadly categorized as generative, discriminative, or hybrid.A DNN contains input, hidden, and output layers; neurons combine weighted inputs and bias through activation functions.
  • A. Architectures: Training passes inputs forward to produce predictions, computes loss against true values, and updates neuron weights with optimization algorithms such as SGD.
  • 1) Convolutional Neural Networks (CNNs): CNNs use local connectivity, convolutional filters, pooling, and ReLU to extract features while reducing parameters, computation, and overfitting risk.Pooling reduces spatial representation size, while ReLU supports faster training without substantially harming generalization.
  • 1) Convolutional Neural Networks (CNNs): CNNs support IoT vision applications including drone-based flood or landslide prediction, smartphone plant-disease detection, and vehicle traffic-sign recognition.
  • 2) Recurrent Neural Networks (RNNs): RNNs address sequential and time-series problems by modeling dependencies across inputs, and are trained using Backpropagation Through Time after unrolling over time spans.Examples include driver-behavior detection, movement-pattern identification, and household energy-consumption estimation.

3) Long Short Term Memory (LSTM):

The survey describes LSTM as a gated RNN extension for controlling memory over time, then reviews autoencoders, variational and adversarial generative models, and restricted Boltzmann machines.

  • 3) Long Short Term Memory (LSTM): LSTM extends RNNs with feedback and multiplicative forget, read, and write gates that control access to memory cells.The gates compute values between 0 and 1 and regulate stored information.
  • 3) Long Short Term Memory (LSTM): LSTM forget gates actively control cell states, helping prevent stored computations from degrading over time; BPTT commonly trains the network.
  • 3) Long Short Term Memory (LSTM): For long temporal dependencies, LSTM models perform better than RNN models in applications such as activity recognition, educational-performance prediction, and disaster prediction.
  • 4) Autoencoders (AEs): Autoencoders encode inputs into latent representations and decode them into reconstructions by minimizing reconstruction error, supporting unsupervised and transfer learning.
  • 5) Variational Autoencoders (VAEs): VAEs support generative and semi-supervised learning with relatively weak data-structure assumptions, fitting IoT settings with diverse data and scarce labels.
  • 6) Generative Adversarial Networks (GANs): GANs train generator and discriminator networks adversarially to produce synthetic data that can deceive discrimination between generated and real samples.IoT examples include path generation for localization and way-finding and services for visually impaired people.

8) Deep Belief Network (DBNs):

The section covers DBNs, ladder networks, streaming-oriented deep learning, real-time object detection, and deep learning’s role in nonlinear approximation and feature extraction.

  • 8) Deep Belief Network (DBNs): DBNs use a visible layer and multiple hidden layers to extract hierarchical representations, reconstruct inputs, and support prediction with a classifier layer.
  • 8) Deep Belief Network (DBNs): DBNs train layer by layer, treating each new layer as an RBM built on the previously trained layer, which makes training efficient and fast.
  • 8) Deep Belief Network (DBNs): DBNs can support industrial fault classification, security-threat identification, and emotional feature extraction from images.
  • 9) Ladder Networks: Ladder networks combine two encoders with a decoder for supervised and unsupervised learning, using denoising costs to reconstruct clean layer representations.
  • 9) Ladder Networks: Initial ladder-network evaluations focused on MNIST and CIFAR-10, while their IoT use remained limited despite potential for semi-supervised vision analytics.
  • B. Fast and Real-time DL Architectures: YOLO reached 45 fps and Fast YOLO 155 fps on resized images, whereas Faster R-CNN ran at 5–17 fps, making YOLO variants suitable for smart cameras.The cited work treats 30 fps or better as truly real-time for image processing.

2) Transfer Learning with Deep Models:

The section discusses transfer learning, deep learning frameworks, architecture selection for IoT data, and practical requirements for efficient, dynamically updated analytics.

  • 2) Transfer Learning with Deep Models: Transfer learning adapts knowledge representations learned in a related domain to improve learning in a new domain.For IoT localization, a model trained on one smartphone platform can transfer to another without recollecting training data.
  • 2) Transfer Learning with Deep Models: Deep models suit transfer learning because they learn both low-level and abstract representations, although real-world IoT applications had not been reported.
  • C. Online Learning: Cloud-based IoT streaming requires online learning because batch-trained models cannot evolve as incremental data arrive.
  • D. DL Frameworks: The survey compares DL frameworks by supported architectures, optimization algorithms, development ease, and deployment characteristics.It reports that Torch and Theano showed the best overall performance in most scenarios in one comparative study.
  • E. Lessons Learned: Architecture choice depends on input structure: RNNs suit serial data, LSTMs long dependencies, CNN variants multidimensional inputs, and RBM, DBN, or AE variants feature extraction.
  • F. Challenges: Fog and device deployment still lack adequate DL frameworks and algorithms, requiring a trade-off between network depth and performance.

A. Foundational Services

Foundational IoT services use DL across visual, localization, human-state, and security tasks, with models adapted to diverse sensor and multimedia inputs. These applications also expose practical concerns around privacy, resource constraints, and evaluation realism.

  • Image Recognition: IoT image applications commonly use cameras for recognition, classification, and object detection in homes, campuses, and manufacturing.
  • Image Recognition: Common benchmark datasets such as MNIST, VGG Face, and CIFAR do not represent difficult IoT conditions including nighttime, rain, or fog.
  • Indoor Localization: DeepFi uses DL over WiFi channel-state fingerprints in offline training and online localization, while other systems report accurate indoor positioning.
  • Indoor Localization: An LSTM-based soccer-robot localization system improved accuracy and efficiency over standard EKF and static particle-filtering baselines.
  • Physiological and Psychological State Detection: A CNN-LSTM framework for wearable-sensor activity recognition outperformed a convolution-only baseline by 4% on average.
  • Security and Privacy: IoT security applications use DL for attack, anomaly, malware, and privacy-preserving detection across smart grids, vehicles, and mobile devices.

B. Applications

DL applications span smart homes, smart cities, smart grids, transportation, and other IoT settings. Reported systems address forecasting, monitoring, classification, localization, and infrastructure management using varied DL architectures.

  • Smart Homes: Smart-home DL applications support appliance analytics, household supply prediction, and energy-use control.
  • Smart Homes: LSTM sequence-to-sequence forecasting predicted household electricity usage better than CNN and LSTM, and all three outperformed a conventional ANN.
  • Smart City: Smart-city systems analyze heterogeneous cross-domain data for resource management, service improvement, crowd movement, and station crowd-density prediction.
  • Smart Grid: Smart-grid DL applications use meter data to learn local consumption patterns, predict needs, and estimate building energy flexibility in real time.
  • Intelligent Transportation Systems: Transportation applications apply RBM, RNN, LSTM, and CNN models to congestion, mobility, traffic signs, and real-time detection tasks.
  • Intelligent Transportation Systems: A multi-column DNN traffic-sign recognizer achieved a 99.46% recognition rate, reported as 0.62% more accurate than humans.

5) Healthcare and Wellbeing:

DL supports healthcare, agriculture, education, and industrial IoT through image, speech, sensor, and time-series analytics. Reported applications include diagnosis, monitoring, crop management, learning analytics, and manufacturing inspection, with data and deployment constraints remaining important.

  • Healthcare and Wellbeing: Healthcare applications use CNNs and other DL models for dietary-image analysis, disease indicators, voice pathology, cardiovascular screening, and fall detection.
  • Healthcare and Wellbeing: Fall-detection evaluation found RBM more accurate than DBN, while limited datasets and offline detection restricted the method.
  • Healthcare and Wellbeing: LSTM models analyze multivariate ICU measurements from vital signs and laboratory tests for disease-pattern recognition and prediction.
  • Agriculture: Agricultural DL systems classify plant diseases from leaf images, including 13 disease categories with about 96% accuracy.
  • Agriculture: Remote-sensing CNNs support large-scale land and crop monitoring, while deep CNNs detect agricultural obstacles and help autonomous machines operate safely.
  • Education: Education applications use learner data for progress prediction, early identification of struggling MOOC students, and classroom occupancy estimation on embedded platforms.
  • Industry: Industrial IoT applies DL to smart manufacturing, visual inspection, rotating-device faults, and vehicle-component defect detection.

9) Government:

The survey links DL with IoT services across government, sports, retail, and infrastructure, while identifying uneven adoption across application types and models.

  • Government: DL supports government tasks including natural-disaster recognition and prediction and environmental monitoring.
  • Cross-domain applications: IoT and DL applications span sports analytics, visual shopping, customer-orientation estimation, anomaly detection, and network load balancing.
  • Foundational services: The survey identifies five foundational IoT service classes and summarizes their use across domains.
  • Lessons learned: Security and privacy services remain limited, and voice recognition is not widely used partly because domain-specific training datasets are scarce.
  • Model usage: About 43% of surveyed papers used CNNs, while RNNs and LSTMs together appeared in 30% of works.

1) Complexity vs. Performance:

The survey examines how DL model complexity affects accuracy, latency, memory, and energy, and reviews compression and hardware-oriented strategies for constrained IoT devices.

  • Complexity vs. Performance: Accuracy and inference time have a hyperbolic relationship, while operations count increases linearly with inference time.
  • Resource trade-offs: Energy constraints limit maximum achievable accuracy, and memory usage grows linearly with batch size after initial allocation.
  • Depth and accuracy: Increasing depth does not guarantee higher accuracy: localization performed best with three hidden layers, whereas deeper networks benefit many vision tasks.
  • Pitfalls and criticisms: DL models are black boxes with high predictability but low interpretability, and may be unsuitable for reasoning, long-term planning, or algorithmic data manipulation.
  • Network Compression: Network compression converts dense networks into sparse networks to reduce storage and computation, but applicability depends on the model and hardware.
  • Network Compression: AlexNet and VGG-16 were compressed at least 9 and 13 times, respectively, while largely preserving accuracy.
  • Network Compression: HashedNets group connection weights into hash buckets so shared parameters reduce representation size while retaining trainable models.

2) Approximate Computing:

Approximate computing and specialized acceleration are surveyed as ways to deploy DL on IoT devices while meeting acceptable output quality and resource constraints.

  • Approximate Computing: Approximate computing reduces implementation and energy costs by allowing predictions to remain within application-defined quality thresholds.
  • Hardware acceleration: Specialized hardware and circuits target improved DL energy efficiency, memory footprint, and inference time on IoT devices.
  • Software acceleration: Software acceleration can compress layers at runtime and decompose deep models across available processors for mobile inference.

4) Tinymotes:

Tinymotes and other embedded approaches bring DL inference closer to IoT data sources, but resource constraints still limit training and broad deployment. Compression, approximate computing, specialized hardware, and fog/cloud placement address different parts of this challenge.

  • Tinymotes: Resource-constrained IoT devices rarely host DL models because their processing, memory, and energy limitations complicate deployment.The paper frames embedded DL support as challenging even when devices only perform predictions.
  • Tinymotes: Tinymotes target onboard inference with available or pretrained DL models, while training directly on resource-constrained devices remains an open challenge.The paper identifies distributed, on-site training as desirable for scalable deployments such as smart security cameras.
  • Tinymotes: Network compression reduces model size but requires balancing compression rate against accuracy, especially for dynamic streaming models and time-series architectures.The paper notes that compression methods for RNNs and LSTMs remain insufficiently investigated.
  • Tinymotes: Approximate computing simplifies DL models by reducing computation bit length while preserving network structure, supporting more energy-efficient operation on constrained devices.Unlike network compression, approximate computing changes computation representations rather than manipulating network structure.
  • Tinymotes: Specialized small-form-factor hardware can support wearable, mobile, and IoT DL applications, but generality, interoperability, and compatibility remain unresolved.The paper presents these architectures as suitable for time-sensitive applications because of their reduced resource demands.
  • Fog and cloud support: Fog computing brings analytics closer to devices to reduce cloud-related latency and security drawbacks, while cloud platforms remain important for aggregation and complex analytics.The surveyed section contrasts fog proximity with cloud scalability and reviews enabling technologies for both.

A. Enabling Technologies and Platforms

Cloud and fog platforms provide complementary infrastructure for IoT DL analytics: cloud systems scale complex training and inference, while fog systems target proximity and responsiveness. The survey also identifies service integration and orchestration challenges across distributed fog nodes and application interfaces.

  • Cloud platforms: Project Adam trained a DNN with 2 billion connections at twice the baseline accuracy while using 30 times fewer machines.
  • Cloud platforms: Google’s TPU accelerates DNN inference 15 to 30 times versus contemporary GPUs or CPUs while consuming 30 to 80 times less energy per TeraOps/second.
  • DL models as a service: DL models need API-based access for IoT integration, but computational intensity and data transfer can starve cloud services or create bottlenecks.The survey describes DL models as a potential new cloud software-as-a-service category.
  • Fog platforms: Fog DL requires service discovery because geographically distributed nodes may provide different capabilities, such as CNN image detection or RNN time-series prediction.
  • Fog platforms: Fog systems must distribute DL models, tasks, and data streams across nodes while aggregating results and returning actions with low latency.
  • Mobile edge: Mobile-edge DL must account for devices joining or leaving dynamically and manage their energy accurately when analytic tasks are delegated.
  • DL interfaces: Current IoT DL integration is limited mainly to RESTful APIs, leaving integration with MQTT, CoAP, XMPP, and AMQP as an open need.

A. Challenges

IoT DL analytics face challenges from scarce real-world data, heterogeneous and high-velocity streams, resource constraints, security concerns, and model limitations. The survey highlights gaps in online learning, trustworthy data handling, regression, and deployment at the edge.

  • Data challenges: Large real-world IoT datasets are scarce, while copyrighted or privacy-sensitive datasets further constrain training and empirical validation.The paper links limited data availability to DL’s need for more data and reduced overfitting.
  • Data challenges: IoT data preprocessing is difficult because inputs arrive from different sources and require appropriate representations before DL models can produce good results.
  • Velocity: High data velocity requires high-speed processing, yet augmenting DL with online and sequential learning remains an open research need.
  • Veracity: IoT data must be validated for trustworthiness at each analytics level, particularly when online streams feed an analytic engine.
  • Deployment: Resource-constrained IoT devices make deploying DNNs difficult, and growing datasets and algorithms increase device designers’ requirements.
  • Model limitations: DL models can produce false confidence on images unrecognizable to humans, classifying fooling examples as familiar objects.
  • Model limitations: Many IoT applications require regression, but DL research has focused more heavily on classification and needs further investigation for regression capabilities.Examples include electricity-load and temperature forecasting.

B. Future Directions

Future IoT DL research emphasizes mobile data, multimodal context, adaptive resource provisioning, semi-supervised learning, system safety, and intelligent network management. The survey concludes that DL connects IoT-generated raw data with higher-level insights for service improvement.

  • IoT Mobile Data: Distributed learning over mobile big data, including iterative MapReduce on parallel Spark workers, is proposed for improved IoT services.
  • Semi-supervised learning: Semi-supervised deep reinforcement learning outperforms supervised learning in total rewards and closeness to the target at each epoch.The cited indoor-localization experiment used only 15% labeled data and benefited from unlabeled data.
  • Context fusion: Combining IoT sensor data with contextual information can improve environmental understanding and support faster reasoning through a bounded search space.
  • Adaptive provisioning: Fog and cloud provisioning needs algorithms that adapt to current streams because future data volumes cannot be known in advance.
  • Safety and resilience: DL can analyze CPS and IoT log traces to identify and predict weak points associated with attacks or functional defects.
  • Intelligent networking: DL architectures may support network self-configuration, self-optimization, self-healing, and self-load balancing for large IoT deployments.
  • Survey conclusions: The survey organizes IoT DL research around big-data and fast-streaming analytics, architectures, applications, implementation platforms, challenges, and future directions.
Loading 1712.04301v2…