PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Spectrum Cartography: Completing Radio Map Tensors Using Learned Neural Models

Sagar Shrestha, Xiao Fu, Mingyi Hong

arXiv:2105.00177v2eess.SPcs.LG

TL;DR

Spectrum cartography must reconstruct full radio maps from sparse sensor measurements, a challenging high-dimensional completion problem. The paper models individual emitters and uses NMF-assisted and iterative completion methods. In indoor measurements, the proposed methods show satisfactory performance, with reported SRE values for three methods.

  • Problem

    Spectrum cartography reconstructs a full radio-map tensor from sparse sensor fibers, posing a challenging high-dimensional inverse problem.

    Source: §I · p. 1, A. Problem Setup and Signal Model · p. 3

  • Method

    The framework models individual emitters’ radio maps with deep networks and proposes NMF-assisted two-stage completion and a generative-model-regularized iterative method.

    Source: §I · p. 2

  • Results

    With 8 sensor locations, the proposed methods yielded satisfactory performance; with 166 ground-truth measurements, SRE was 0.6142 for DowJons, 0.6315 for Nasdac, and 0.6337 for DeepComp.

    Source: 3) Result · p. 14

  • Takeaways & Limitations

    Emitter-level modeling reduces offline training cost and online completion error, while the NMF-based two-stage approach is fast and lightweight at prediction time.

    Source: §VI · p. 14

    The NMF-based two-stage approach works only when emitter spectral occupancy is relatively sparse and suffers from error propagation.

    Source: §VI · p. 14

Abstract

from arXiv · show

The spectrum cartography (SC) technique constructs multi-domain (e.g., frequency, space, and time) radio frequency (RF) maps from limited measurements, which can be viewed as an ill-posed tensor completion problem. Model-based cartography techniques often rely on handcrafted priors (e.g., sparsity, smoothness and low-rank structures) for the completion task. Such priors may be inadequate to capture the essence of complex wireless environments -- especially when severe shadowing happens. To circumvent such challenges, offline-trained deep neural models of radio maps were considered for SC, as deep neural networks (DNNs) are able to "learn" intricate underlying structures from data. However, such deep learning (DL)-based SC approaches encounter serious challenges in both off-line model learning (training) and completion (generalization), possibly because the latent state space for generating the radio maps is prohibitively large. In this work, an emitter radio map disaggregation-based approach is proposed, under which only individual emitters' radio maps are modeled by DNNs. This way, the learning and generalization challenges can both be substantially alleviated. Using the learned DNNs, a fast nonnegative matrix factorization-based two-stage SC method and a performance-enhanced iterative optimization algorithm are proposed. Theoretical aspects -- such as recoverability of the radio tensor, sample complexity, and noise robustness -- under the proposed framework are characterized, and such theoretical properties have been elusive in the context of DL-based radio tensor completion. Experiments using synthetic and real-data from indoor and heavily shadowed environments are employed to showcase the effectiveness of the proposed methods.

I. INTRODUCTION

Spectrum cartography reconstructs multi-domain radio maps from sparse measurements, but handcrafted priors and direct deep learning face challenges in complex environments. The paper models individual emitters and develops completion methods for this setting.

II. PROBLEM STATEMENT AND BACKGROUND

  • Handcrafted sparsity, smoothness, and low-rank priors may inadequately represent complex radio maps, while severe shadowing in indoor or urban settings challenges smoothness-based approaches.Existing approaches may also depend on environmental parameters or emitter spectra that are difficult to obtain in fast-changing environments.Source: §I · pp. 1–2

I. INTRODUCTION

  • Direct deep-learning completion faces high training cost and generalization challenges because radio-map state space grows with emitters, their frequencies, locations, and combinations.Even large training sets may miss cases such as additional emitters entering the scene.Source: §I · p. 2
  • Emitter-disaggregated neural modeling represents each emitter’s spatio-spectral radio map, whose state space is smaller than that of aggregated maps, and addresses out-of-distribution completion cases.The paper also presents synthetic and indoor real-data validation, and reports that the framework offers promising performance in challenging scenarios.Source: §I · p. 2
  • The paper proposes a two-stage greedy NMF and deep-network method for sparse spectral occupancy, plus a more computationally demanding generative-model-regularized iterative method for dense occupancy.It characterizes recoverability, noise robustness, and sample complexity; this journal version adds the dense-occupancy algorithm, analyses, broader simulations, and a real-data experiment.Source: §I · p. 2

A. Problem Setup and Signal Model

  • Spectrum cartography estimates an I × J × K tensor of received power spectral density from fibers measured at a sparse set of sensor locations.The PSD represents average spectrum usage over a period when signals are stationary random processes.Source: A. Problem Setup and Signal Model · p. 3
  • For a sufficiently narrow band, the radio map is modeled as a sum of emitter-specific outer products between spatial loss functions and power spectral densities.The stated condition is that each emitter’s spatial propagation pattern is identical across frequencies up to scaling differences caused by its PSD.Source: A. Problem Setup and Signal Model · p. 3

B. Prior Art

Prior-art methods use structural priors or learned completion networks, but shadowing can undermine structural assumptions and the large state space challenges deep learning. The proposed approach instead learns individual emitters’ spatial loss fields (SLFs), with the aim of easing training and generalization.

B. Prior Art

  • Under severe shadowing, the first 10 principal components capture less than 70% of an SLF’s energy, versus more than 90% with little shadowing.This weakens the low-rank assumption used in prior-art completion methods.Source: B. Prior Art · pp. 3–4

C. Challenges of Deep Learning-Based SC

  • Existing deep-learning SC methods train completion networks on aggregated radio maps, but their training samples may not cover the large state space and test maps may fall outside the training distribution.The state space includes combinations of emitter counts, locations, and transmission power levels.Source: B. Prior Art · p. 4, C. Challenges of Deep Learning-Based SC · p. 4

III. PROPOSED APPROACH

  • The proposed model-assisted approach keeps the aggregation model and uses deep networks to learn individual emitters’ SLFs rather than the aggregated radio map.The authors state this can circumvent pure-DL training and generalization challenges while retaining deep networks’ expressive power.Source: §III · p. 4, A. Training DNN for Completing… · p. 5

A. Training DNN for Completing Individual SLFs

  • For training, the method teaches a completion network to recover a single emitter’s complete SLF from randomly masked observations.The paper uses an autoencoder structure for this network and notes that other generative models could also be used.Source: A. Training DNN for Completing… · p. 5

B. Emitter Disaggregation-Assisted Deep Completion

The method disaggregates observed measurements into emitter components with nonnegative matrix factorization (NMF), then completes each emitter’s SLF with a learned network. Under a spectral-separability assumption, it gives a recoverability guarantee for the NMF stage and assembles the completed SLFs into the radio map.

B. Emitter Disaggregation-Assisted Deep Completion

  • The observed measurements form a nonnegative matrix factorization problem, whose estimated emitter components and SLFs are used to recover the radio map.The method forms G from measured columns and estimates its nonnegative factors C and H.Source: B. Emitter Disaggregation-Assisted… · pp. 5–6
  • When emitters use sparse spectral occupancy, the method assumes each emitter has a frequency where its component is positive and all other emitters’ components are zero.The paper identifies this as the separability condition for NMF.Source: B. Emitter Disaggregation-Assisted… · p. 6
  • With that assumption and at least R sensors, a polynomial-time algorithm recovers the NMF factors up to permutation and scaling ambiguities.The paper uses the successive projection algorithm for efficiency; the ambiguities cancel in the final assembly when the estimated factors correspond.Source: B. Emitter Disaggregation-Assisted… · p. 6
  • Nasdac completes each emitter’s SLF with the trained network after NMF estimates the emitter components, then combines the results to reconstruct the radio map.The method assumes the number of emitters R is known or estimated; for nonsparse spectral usage, the paper proposes a more computationally burdensome iterative criterion.Source: B. Emitter Disaggregation-Assisted… · p. 6

IV. PERFORMANCE-ENHANCED APPROACH

The performance-enhanced approach models each emitter’s spatial loss function with a learned generator and jointly estimates emitter components from sensor measurements, avoiding Nasdac’s two-stage procedure. Its recoverability guarantee tolerates sensing noise and imperfect representation, with error governed by noise, sampling generalization, and representation error.

IV. PERFORMANCE-ENHANCED APPROACH

  • DowJons estimates emitter components through a generative-model-regularized completion criterion that does not require the sparse spectral-occupancy assumption used by Nasdac.The generator maps each emitter’s latent vector to its spatial loss function; the recovered emitter components are aggregated into a radio map.Source: §IV · p. 6

A. Recoverability Analysis

  • The analysis treats the ground-truth radio map as deterministic and the sensor sampling locations as random.The recoverability theorem applies under its stated network and solution-set assumptions and holds with probability at least 1 − δ.Source: A. Recoverability Analysis · pp. 7–8
  • The recoverability bound accounts for sensing noise, sampling-induced generalization error, and generative-model representation error.Increasing network complexity may reduce representation error while increasing the covering number and sampling-induced error; the latter approaches zero as sampling becomes dense under the stated conditions.Source: A. Recoverability Analysis · p. 8
  • Overestimating the emitter count adds redundant components that increase model covering number and therefore require more samples for the same recoverability accuracy.If the count is underestimated, the energy of missed components can instead be absorbed into the noise term.Source: A. Recoverability Analysis · p. 8

B. Algorithm Design

DowJons alternates between estimating emitter coefficients and optimizing the latent vectors that generate emitter spatial loss functions. It offers a recoverability advantage over lightweight Nasdac at higher computational cost, and the experiments evaluate the proposed methods on synthetic and real data.

B. Algorithm Design

  • DowJons alternates between coefficient estimation and latent-vector optimization, using nonnegative least squares for the coefficient step and back-propagation-based gradient descent for the latent step.The latent-vector block is nonconvex; the authors report that few inner iterations can work well and, under some conditions, inexact alternating optimization converges to the vicinity of a stationary point.Source: B. Algorithm Design · p. 8
  • Nasdac is lightweight but can propagate errors between its two stages, whereas DowJons offers a recoverability bound at higher computational cost.The comparison concerns the methods’ guarantees when their respective optimization procedures are used.Source: B. Algorithm Design · p. 8

V. EXPERIMENTS

  • The experiments use synthetic-data and real-data evaluations to showcase the proposed methods.The supplied experiment material introduces this evaluation scope but does not provide results here.Source: §V · p. 8

A. Synthetic-Data Experiments - Settings

The synthetic experiments compare the proposed methods with interpolation, tensor-completion, and direct deep-completion baselines. They generate shadowed emitter maps and train emitter-level models, then assess reconstruction and component-estimation accuracy.

1) Baselines:

  • The benchmarks include TPS, LL1, and DeepComp; LL1 replaces learned deep completion with TPS interpolation of disaggregated emitter spatial loss functions, serving as an ablation of learned priors.DeepComp instead learns the aggregated radio map directly.Source: 1) Baselines · p. 9

2) Data Generation:

  • Synthetic data use a joint path-loss and log-normal-shadowing model, with stronger shadowing associated with smaller decorrelation distance and larger shadowing variance.The simulated region is 50 × 50 m² on a 50 × 50 grid, with K = 64 frequency bins unless otherwise specified.Source: 2) Data Generation · p. 9

3) Deep Network Configuration and Training Sample Generation:

  • The emitter-level autoencoder is trained on 500,000 spatial loss functions, with parameters sampled to cover moderate-to-severe shadowing.DeepComp is trained on 500,000 aggregate map slices, with emitter count sampled from {3, 4, 5, 6}.Source: 3) Deep Network Configuration and… · p. 9

4) Performance Metric:

  • Radio-map accuracy is measured by squared reconstruction error, while emitter coefficient accuracy is measured by normalized absolute error after column-permutation removal.The captioned illustration compares ground-truth and estimated PSDs at ρ = 10%, R = 2, η = 5, and dcorr = 50.Source: 4) Performance Metric · pp. 9–10, Figure 8 · p. 10

5) Hyperparameters of Algorithms:

Under sparse spectral occupancy, Nasdac and DowJons generally outperform the baselines, with DowJons more robust to noise and difficult shadowing while Nasdac is faster. Their emitter-wise modeling also avoids DeepComp’s difficulty when test scenes exceed its training emitter range.

5) Hyperparameters of Algorithms:

  • DowJons stops at a relative cost change below 0.003 or 10 outer iterations, and results are averaged over 30 Monte Carlo trials.LL1 uses SLF rank up to 4, 10^-3 regularization, and stops at relative loss change below 10^-3 or 50 iterations.Source: 5) Hyperparameters of Algorithms · p. 10

1) Sparse Spectral Occupancy:

2) Dense Spectral Occupancy:

DowJons performs well when spectral occupancy is dense and under training–testing mismatch, and it leads the reported real-data comparisons. The paper also notes limits of Nasdac in dense occupancy and its two-stage workflow, while concluding that the one-shot method has broader recovery guarantees.

2) Dense Spectral Occupancy:

4) Training-Testing Mismatch:

1) Data Description:

  • The real-data map covers a 14 × 34m2 indoor area on 1 × 1m2 grids, with measurements from 166 cells roughly covering the hallway.The dataset has nine frequency bands and more than 400 observations for each frequency-location pair, reduced to the median.Source: 1) Data Description · p. 13, Figure 16 · p. 14

2) Training Sample Generation:

  • Training samples use η ∈[8, 12] and dcorr ∈[5, 50], with emitters allowed across a 200 × 200m2 region to account for sources outside the mapped area.The mapped training region is 14 × 34m2, and emitter-induced SLFs within it are extracted as examples.Source: 2) Training Sample Generation · p. 13

3) Result:

  • With eight random measurement locations, the DL-based methods outperform LL1; DowJons, Nasdac, and DeepComp yield SREs of 0.6142, 0.6315, and 0.6337 using 166 measurements.The 166-measurement comparison evaluates the DL-based methods.Source: 3) Result · pp. 13–14, Figure 17 · p. 14
  • DowJons performs consistently better than the other DL-based methods across all tested ρ values on real data.The SRE results are averaged over 30 random trials using estimated bR = 7.Source: 3) Result · p. 14, Table X · p. 14

VI. CONCLUSION

  • The conclusion distinguishes fast, lightweight Nasdac, limited to relatively sparse spectral occupancy and vulnerable to error propagation, from one-shot optimization with more relaxed recovery conditions.The one-shot criterion is reported to guarantee recoverability even with noise; synthetic and real-data experiments corroborate the framework and recovery theory.Source: §VI · p. 14

APPENDIX A NEURAL NETWORK SETTING

The appendix specifies the autoencoder used in the experiments, establishes identifiability under Assumption 1, and gives convergence conditions for DowJons.

APPENDIX A NEURAL NETWORK SETTING

  • The experimental autoencoder architecture uses SELU activations except for a sigmoid final activation, with padding chosen to obtain the desired output dimensions.Table XI and Fig. 18 give the network-setting details.Source: Table XI · p. 14, Appendix A · p. 14, Figure 18 · p. 15

APPENDIX B PROOF OF THEOREM 1

  • Under Assumption 1, C has rank R and satisfies separability, allowing standard separable-NMF identifiability analysis to support the theorem’s conclusion.The proof notes that a scaled permutation matrix is a submatrix of C.Source: Appendix B · p. 15

APPENDIX C CONVERGENCE PROPERTIES OF DOWJON S

  • DowJons converges to an ε-stationary point under the stated boundedness and iteration assumptions.The proof also uses strong convexity of the C subproblem when H(k) is full rank.Source: Appendix C · p. 17

APPENDIX D PROOF OF LEMMA 2

The proof bounds the covering number of the tensor class by covering the neural-model outputs and coefficient factors, then combining those approximations.

APPENDIX D PROOF OF LEMMA 2

  • An ε/P-net of the latent Euclidean ball induces an ε-net of the generated-model outputs through the network’s Lipschitz bound.The bound uses P, the product of layer Lipschitz and operator-norm factors.Source: Appendix D · p. 17
  • The proof separately covers the generated source matrices and coefficient matrices, then combines their approximations to bound matrix-factorization error.The error bound decomposes into terms for approximating each factor.Source: Appendix D · pp. 17–18
  • The resulting product construction gives an ε-net for the tensor class with a bounded covering number.The bound follows from the covering-number bounds for the factor sets.Source: Appendix D · p. 18

APPENDIX E PROOF OF LEMMA 1

The proof uses concentration for sampling without replacement to relate the sampled empirical loss to the loss over all entries.

APPENDIX E PROOF OF LEMMA 1

  • A sampling-without-replacement Hoeffding extension bounds the difference between empirical loss on sensed entries and actual loss over all entries.The proof bounds the sampled variables between zero and ξ before applying the inequality.Source: Appendix E · pp. 18–19

APPENDIX F PROOF OF THEOREM 2

The appendix proof uses empirical-loss optimality, then reviews SPA and its use of separability to identify NMF factors up to scaling and permutation.

APPENDIX F PROOF OF THEOREM 2

  • The empirical-loss optimizer minimizes the sensed-data residual among candidates, supporting the proof’s comparison with any other feasible model.The argument applies this optimality to candidates in the generative-model class.Source: Appendix F · p. 19

APPENDIX G THE SPA ALGORITHM FOR NMF

  • Under separability, scaled columns of A appear among the columns of F, reducing factor recovery to identifying the corresponding column indices.For Nasdac, the appendix relates separability to Assumption 1.Source: Appendix G · p. 20
  • SPA identifies columns by selecting a data column with maximal residual norm, then projecting the data onto the selected columns’ orthogonal complement before repeating.The appendix states recovery is up to column permutation and scaling ambiguities.Source: Appendix G · p. 20

APPENDIX H SUPPLEMENTARY EXPERIMENT RESULTS

In additional synthetic comparisons, deep learning-based methods outperform the interpolation baselines. In the real-data scenario, reconstruction is closer to ground truth with more measurements, while LL1 still struggles.

A. Synthetic Data: Additional Baselines

B. Real Data: Additional Scenario

Related papers

Loading 2105.00177v2…