PaperlayerView exact revision on arXiv

Source-linked AI summary

SMCC-Empowered Digital Twins for Sensorless Monitoring in Large-Scale AI-Driven IoT Systems

Vincenzo Sammartino

arXiv:2609.09161v1cs.DCcs.CRcs.NI

TL;DR

Large-scale Digital Twins require high-fidelity synchronization, while dense sensor deployments impose hardware and network burdens. SMCC-DT combines ISAC-based sensorless monitoring with joint sensing, memory, communication, and computation allocation, using PPO-based DRL for the resulting NP-hard problem. On a 500-node industrial IoT testbed, it reduces synchronization latency and energy consumption versus orthogonal and compute-only baselines while maintaining sensing and inference performance.

  • Problem

    Dense sensor deployments for large-scale Digital Twins impose hardware, energy, and network-bandwidth burdens that complicate continuous synchronization.

  • Method

    SMCC-DT uses ISAC waveforms for sensorless monitoring and jointly allocates transmit power, beamforming, memory partitions, and CPU frequency through the PPO-based DRL agent SMCCAGENT.

  • Results

    38.7% lower synchronization latency and 27.4% lower energy consumption were achieved on a 500-node industrial IoT testbed versus orthogonal and compute-only baselines.

  • Takeaways & Limitations

    SMCC-DT maintains sensing accuracy above 95% while jointly coordinating the SMCC resource layers for Digital Twin maintenance.

Abstract

from arXiv · show

The deployment of AI-driven Digital Twins (DTs) in large-scale Internet-of-Things (IoT) ecosystems demands continuous, high-fidelity synchronization between the physical environment and its virtual replica. Conventional approaches rely on dense sensor deployments, which introduce prohibitive costs in terms of hardware, energy, and network bandwidth. In this paper, we propose SMCC-DT, an integrated Sensing-Memory-Communication-Computation (SMCC) framework that enables sensorless monitoring of physical assets by exploiting Integrated Sensing and Communication (ISAC) waveforms at the 6G Edge. Under the SMCC-DT paradigm, a single radio signal simultaneously extracts environmental telemetry (Sensing) and delivers it to an Edge server (Communication), where a large-scale AI model is loaded into constrained memory (Memory) and executed (Computation) to update the DT state. We formulate the DT synchronization problem as a cross-layer optimization that jointly allocates transmit power, beamforming vectors, memory partitions, and CPU frequency to minimize the end-to-end synchronization latency subject to sensing accuracy, throughput, memory capacity, and computational budget constraints. Because the resulting mixed-integer nonlinear program is NP-hard, we design a Proximal Policy Optimization (PPO)-based Deep Reinforcement Learning (DRL) agent, termed SMCCAGENT, that learns near-optimal resource allocation policies online. Extensive simulations over a 500-node industrial IoT testbed demonstrate that SMCC-DT reduces DT synchronization latency by 38.7% and total energy consumption by 27.4% compared to state-of-the-art orthogonal and compute-only baselines, while sustaining sensing accuracy above 95% and model inference throughput above 30 frames per second.

I. INTRODUCTION

Large-scale Digital Twins need continuous synchronization, but dense dedicated sensors create hardware, bandwidth, and deployment limitations. SMCC-DT addresses this gap by combining ISAC-based sensorless monitoring with cross-layer resource optimization and PPO-based DRL.

  • Motivation: Dense industrial sensor deployments incur hardware costs that scale with monitored assets and generate high-frequency traffic that can saturate wireless uplinks.These limitations make large-scale Digital Twin deployment difficult under conventional sensing architectures.
  • Motivation: ISAC uses one waveform for radar-like sensing and data communication, reducing dependence on dedicated sensor hardware while still requiring coordinated telemetry delivery.The same transmission extracts environmental properties and carries payloads toward the communication endpoint.
  • Architecture: SMCC-DT replaces dedicated sensor networks with ISAC-based sensorless monitoring and integrates sensing, memory, communication, and computation at the 6G Edge.The architecture is designed to maintain a continuously synchronized Digital Twin for large-scale IoT systems.
  • Optimization: The paper jointly allocates ISAC transmit power, beamforming, Edge memory, and CPU frequency to minimize synchronization latency under sensing, throughput, memory, and computation constraints.This formulation captures cross-layer dependencies rather than optimizing isolated resource layers.
  • Optimization: SMCCAGENT uses PPO-based DRL to learn near-optimal resource-allocation policies for the NP-hard cross-layer optimization problem.The agent uses a hierarchical action space and domain-specific reward shaping to converge within practical time budgets.
  • Evaluation: 38.7% lower synchronization latency, 27.4% lower energy consumption, and 52.1% lower memory fragmentation were achieved on a 500-node industrial IoT testbed.The comparison covers orthogonal allocation, compute-only optimization, and heuristic baselines.

A. Digital Twins for IoT Systems

Prior work established Digital Twins, ISAC, and Edge Intelligence as complementary foundations for cyber-physical monitoring. The paper instantiates these ideas in an industrial IoT architecture whose BS and Edge server jointly sense assets and process telemetry.

  • Digital Twin Foundations: Digital Twin research spans model-driven and data-driven architectures, five-dimensional data-fusion models, Edge-based systems, and cybersecurity-oriented virtual replicas.Security Digital Twins can reflect vulnerability states, patch levels, and access-control configurations from passive observations.
  • Integrated Sensing and Communication: ISAC research covers dual-function waveform design, 6G perceptive networks, sensing–communication bounds, multiuser beamforming, and vehicular or millimeter-wave implementations.The literature frames sensing accuracy and communication capacity as coupled design objectives.
  • Edge Intelligence and SMCC: Edge Intelligence deploys deep-learning models on resource-constrained Edge servers, motivating tighter integration of sensing, communication, and computing.Related work also addresses task-oriented integration, over-the-air Digital Twin updates, semantic compression, and IoT resource allocation.
  • System Architecture: The proposed industrial IoT setting contains physical assets monitored by a Digital Twin hosted on an Edge server co-located with a 6G base station.The base station transmits ISAC waveforms that simultaneously sense the environment and communicate payloads to IoT devices.
  • Sensing Model: The sensing waveform uses separate sensing and communication beamforming vectors, with total transmit power constrained by Pmax.The sensing model represents each asset through its reflected echo, position-dependent delay, steering vector, radar cross-section, and noise.
  • Sensing Model: Sensing accuracy is characterized for angle, delay, and Doppler estimation through a CRLB based on the Fisher Information Matrix and sensing-beam power.The requirement ϵk(ρs) ≤ ϵmax imposes a minimum estimation-accuracy condition.

B. Communication Model

The communication model carries sensed telemetry from the base station to the Edge server over a wireless channel while accounting for residual interference from the sensing component. It imposes a minimum achievable-rate requirement for timely Digital Twin updates.

  • Telemetry Delivery: The communication component transports sensed telemetry ˆθk from the base station to the Edge server over channel H.This establishes the communication link needed after environmental parameters are extracted.
  • Signal Model: The received Edge signal combines the communication term Hwcsc(t) with sensing interference from Hwsx(t).The model explicitly represents the sensing waveform as interference within the communication reception.
  • Rate Constraint: After interference cancellation, the achievable rate is constrained by R ≥ Rmin to ensure sufficient throughput for Digital Twin telemetry delivery.The rate expression uses communication bandwidth Bc and residual interference factor γ after sensing-signal cancellation.
  • Rate Constraint: The residual interference factor γ lies in [0, 1] and quantifies interference remaining after sensing-signal cancellation.Lower residual interference supports a less interference-limited communication link under the stated model.

C. Memory Model

The Edge server partitions limited memory between AI model storage and buffered ISAC telemetry, creating a model-quality versus data-retention trade-off.

  • C. Memory Model: Edge memory is divided between AI model weights and the incoming telemetry data buffer.The total capacity Mmax must accommodate both demands.
  • C. Memory Model: Larger models can improve inference accuracy but leave less buffer capacity, potentially causing data drops.
  • C. Memory Model: The architecture routes ISAC-derived telemetry to the Edge server, where memory supports model execution and data buffering.

D. Computation Model

The computation model represents Edge inference as CPU-cycle processing whose latency and energy depend on model size, asset count, and processor frequency within a sequential synchronization pipeline.

  • D. Computation Model: Inference processing uses CPU frequency fcpu within [fmin, fmax], with required cycles determined by model size, asset count, and architecture-dependent factor κ.
  • D. Computation Model: Increasing CPU frequency reduces computation time but is associated with energy consumption under a cubic DVFS model.The energy model uses effective processor capacitance coefficient ξ.
  • E. DT Synchronization Latency: DT synchronization latency consists of four sequential stages aligned with Sensing, Communication, Memory, and Computation.
  • E. DT Synchronization Latency: Model loading time is determined by model memory footprint divided by storage-to-memory bandwidth Bmem.

IV. CROSS-LAYER OPTIMIZATION PROBLEM

The paper formulates DT synchronization as a joint resource-allocation problem and establishes that the resulting optimization is NP-hard.

  • IV. CROSS-LAYER OPTIMIZATION PROBLEM: The optimization jointly selects sensing power ρs, communication beamformer wc, memory partitions, and CPU frequency to minimize synchronization latency.
  • IV. CROSS-LAYER OPTIMIZATION PROBLEM: The formulation enforces sensing accuracy, communication throughput, power, memory, model-size, CPU-frequency, energy, and integer model-variant constraints.
  • IV. CROSS-LAYER OPTIMIZATION PROBLEM: Problem P1 is NP-hard.The proof reduces a simplified model-selection and memory-allocation instance to 0-1 Knapsack.

V. DRL-BASED SMCC RESOURCE ALLOCATION

SMCCAGENT formulates SMCC resource allocation as an MDP and learns resource-allocation policies with PPO under constraint-aware rewards.

  • V. DRL-BASED SMCC RESOURCE ALLOCATION: SMCCAGENT maps real-time system observations to resource-allocation actions using PPO because the underlying optimization is NP-hard.The learned policy is denoted πϕ(at|ot), and PPO is selected for stability with continuous actions and compatibility with reward-based constraint handling.
  • V. DRL-BASED SMCC RESOURCE ALLOCATION: The agent observes channel state, sensing errors, available memory, CPU load, and buffer occupancy as its system state.These observations define the state information used for online allocation decisions.
  • V. DRL-BASED SMCC RESOURCE ALLOCATION: Its hierarchical action space controls sensing power, communication beamforming, memory reallocation, CPU frequency, and discrete model selection.The model variant selector ω(t) is discrete, while ΔM(t) represents the memory reallocation increment.
  • V. DRL-BASED SMCC RESOURCE ALLOCATION: The training loop samples actions, executes them in the environment, stores transitions, computes GAE advantages, and updates the PPO policy.Algorithm 1 repeats this process across episodes and epochs using trajectory data stored in buffer B.
  • V. DRL-BASED SMCC RESOURCE ALLOCATION: A composite reward balances synchronization-latency minimization with penalties for constraint violations.The violation magnitudes gi(xt) are weighted by positive coefficients λ1, λ2, and λ3.

B. Network Architecture

SMCCAGENT uses shared feature extraction with separate policy heads for continuous resource controls, discrete model selection, and value estimation.

  • B. Network Architecture: A shared encoder maps observations to a latent representation zt that feeds the actor and critic components.The encoder comprises three fully connected layers with 256–128–64 neurons and ReLU activations.
  • B. Network Architecture: The actor uses a Gaussian policy for continuous actions and a categorical distribution for the discrete model-variant action.Continuous controls include sensing power, beamforming, memory reallocation, and CPU frequency.
  • B. Network Architecture: The critic head estimates the state-value function Vψ(ot) from the shared representation.This value estimate supports PPO policy optimization.

C. Training Procedure

The training and evaluation setup combines PPO trajectory optimization with Monte Carlo simulations in a large industrial IoT environment.

  • C. Training Procedure: SMCCAGENT trains by collecting environment trajectories and updating its policy with PPO’s clipped surrogate objective.The update uses Generalized Advantage Estimates, with clipping parameter ε = 0.2.
  • C. Training Procedure: The simulations model 500 assets in a 200 × 200 × 50 m3 smart factory served by a 64-antenna 6G base station.The system operates at 28 GHz with 100 MHz sensing and communication bandwidths.
  • C. Training Procedure: The dual-head architecture combines Gaussian continuous-action control, categorical model selection, and a critic value estimate.A shared encoder produces the latent representation used by all heads.
  • C. Training Procedure: The evaluation includes small, medium, and large AI models ranging from 7M to 1.3B parameters and 14 MB to 2.6 GB footprints.The Edge server provides 32 GB of memory and a CPU capped at 4.0 GHz.

B. Baselines

SMCC-DT is evaluated against orthogonal, compute-only, greedy, and random allocation baselines, with reported gains in latency, energy, and convergence behavior.

  • B. Baselines: The SMCC-DT comparison covers orthogonal allocation, compute-only optimization, greedy heuristics, and random allocation.These baselines provide distinct references for spectral efficiency, computation-only control, rule-based allocation, and lower-bound behavior.
  • B. Baselines: OA separates sensing and communication across frequency bands, whereas CO fixes their power split and memory partition while optimizing computation.GH prioritizes sensing accuracy, throughput, and computation sequentially, and RA samples feasible decisions uniformly at random.
  • C. Convergence Analysis: SMCCAGENT’s average reward converges after approximately 3,500 episodes, while constraint violations fall below 1% after 2,000 episodes.These learning-curve results indicate stabilization of the allocation policy under the shaped reward.
  • D. Synchronization Latency: At K = 500, SMCC-DT reaches Tsync = 12.3 ms, reducing latency 38.7% versus OA and 27.2% versus CO.The passage attributes the improvement to joint ISAC power splitting and memory-computation co-allocation.
  • E. Energy Efficiency: SMCC-DT maintains sensing accuracy above 95% while consuming 27.4% less energy than OA and 19.8% less than CO.The reported energy advantage is associated primarily with adaptive CPU-frequency scaling and memory buffering.

F. Impact of AI Model Size

SMCC-DT remains feasible as AI model size grows, with smaller latency degradation than compute-only optimization and continued operation at high asset counts. Its cross-layer design also depends materially on jointly optimizing sensing, memory, communication, and computation.

  • F. Impact of AI Model Size: 31% latency growth from medium to large models for SMCC-DT versus 67% for CO demonstrates more graceful degradation under increasing model size.At K = 500, SMCC-DT remains feasible, while OA violates the latency constraint for large models when K > 300 assets.
  • F. Impact of AI Model Size: Removing memory co-optimization increases latency by 21.5%, while removing sensing optimization increases it by 14.2%.The ablation is conducted at K = 500 with a medium model.
  • F. Impact of AI Model Size: Removing computation optimization increases energy consumption by 35.8% without improving latency, showing that fixed maximum CPU frequency is inefficient.The ablation results support the contribution of all four SMCC dimensions to overall performance.
  • F. Impact of AI Model Size: The full SMCC-DT design reduces synchronization latency by 38.7% and energy consumption by 27.4% versus orthogonal and compute-only baselines.The conclusion reports sensing accuracy above 95% alongside these gains.
Loading 2609.09161v1…