PaperlayerView exact revision on arXiv

Source-linked AI summary

Orientation-Aware Deep Learning Framework for RSSI-Based Indoor Localization in RIS-Aided mmWave Systems

Varun S. Advani, Ahmed Nasser, Mohamed Y. Selim, Ahmed M. Eltawil

arXiv:2610.21043v1cs.NIeess.SP

TL;DR

Indoor RSSI localization faces NLoS propagation challenges, and device orientation changes the fingerprints used for positioning. O-DSG infers orientation from RSSI and uses it to condition localization processing. In measured 28 GHz NLoS data, it achieves a 0.141 m median error, while its weights remain tied to the surveyed site.

Abstract

from arXiv · show

Reconfigurable intelligent surface (RIS)-aided millimeter wave (mmWave) systems provide directional measurements of received power for indoor localization, but device orientation changes the fingerprints. This work proposes an orientation-aware dual-stream gated (O-DSG) deep learning framework for received signal strength indicator (RSSI)-based indoor localization. An orientation classifier estimates device orientation from RSSI, so inference needs only received power. An embedding of the predicted angle and posterior entropy controls feature modulation and soft expert gating. Parallel convolutional and statistical streams are fused through bidirectional long short-term memory (BiLSTM) processing and temporal attention. Single-user measurements in an indoor laboratory at 28 GHz under non-line-of-sight (NLoS) conditions yield a median localization error of 0.141 m, a 25.8% improvement over the closest evaluated baseline (supplied with the measured orientation), and 92.3% orientation classification accuracy across four discrete classes. In spatial regions withheld from training, O-DSG has an average median error of 0.247 m, against 0.392 m for a validation-selected sequence model. A separate offline evaluation constructs non-orthogonal multiple access (NOMA)-style power mixtures from the single-user measurements and measures localization accuracy and computational scaling with up to eight emulated users whose power contributions overlap, under the specified mixture model.

I. INTRODUCTION

Indoor RSSI localization must contend with attenuation, multipath, and NLoS conditions, while device orientation changes the measured fingerprints. The paper introduces O-DSG to infer orientation from RSSI and use it to condition spatial and temporal processing.

I. INTRODUCTION

  • Indoor RF localization faces strong attenuation, multipath propagation, and frequent NLoS conditions; mmWave directional measurements and RIS-shaped reflected paths provide spatial information for localization.RSSI sweeps can be collected without explicit CSI acquisition, though fluctuations and multipath complicate the mapping from power to position.Source: §I · p. 1

A. Related Works

  • Prior RIS-localization work spans geometry-based, fingerprinting, and learned approaches, while adaptive RIS-control studies learn communication policies rather than position and orientation from prescribed RSSI sweeps.The paper contrasts its approach with prior DSOG, which used orientation supplied to the localization model.Source: A. Related Works · pp. 1–2

B. Motivation and Contributions

  • O-DSG combines inferred orientation, spatial RSSI structure, and temporal context; a CNN derives expected angle and posterior entropy from RSSI to condition localization without supplied orientation at inference.Orientation labels supervise training; FiLM modulates spatial feature channels, while the orientation representation also informs soft expert gating.Source: B. Motivation and Contributions · p. 2
  • The evaluation uses 28 GHz NLoS measurements and separately tests localization accuracy and computational load on offline power mixtures of single-user fingerprints as emulated contributors increase.The mixture experiment is conducted under the stated synthesis assumptions.Source: B. Motivation and Contributions · p. 3

C. Organization and Notation

II. SYSTEM MODEL AND PROBLEM FORMULATION

  • The system model has a BS, RIS, and UEs in an indoor NLoS environment; UEs measure downlink RSSI during prescribed beam and RIS sweeps, and the BS estimates position and orientation.In the experiment, one mobile UE acquires separate fingerprints at measured locations, and the multi-user extension combines them offline.Source: §II · p. 3

A. Signal Model

The signal model describes RSSI fingerprints formed through beam and RIS sweeps, including the channel and codebook setup. Its multi-user interpretation relies on shared precoding and RIS states and a specified mean-power mixture assumption.

A. Signal Model

  • The model includes a direct BS–UE path for completeness, treated as negligible when an obstruction blocks the direct path.The BS–RIS channel is modeled geometrically, and the RIS–UE channel contains multipath components whose arrival angles vary with device orientation.Source: A. Signal Model · pp. 3–4
  • The BS uses a fixed precoder toward the RIS, while common codebooks specify the RIS configurations and UE receive beams swept for measurement.The measured fingerprints retain the testbed’s physical amplitude and phase response, including possible coupling.Source: A. Signal Model · p. 4
  • Power-allocation design is outside the work’s scope; the localization framework neither estimates nor optimizes the allocation fractions.For offline evaluation, relative mixture powers use a normalized distance-based rule recomputed from ground-truth positions.Source: A. Signal Model · p. 4
  • Single-user measurements can represent mean fingerprints under simultaneous transmissions when streams are uncorrelated, share the precoder and RIS state, and do not alter propagation channels.With user-specific precoders, each fingerprint would depend on other users’ precoders and power fractions.Source: A. Signal Model · p. 4
  • The offline NOMA-style construction combines independently measured matrices in linear power units, with a common composite and noise contribution from the measured matrices.The term NOMA-style refers specifically to this offline construction.Source: A. Signal Model · p. 4

1) Deployment Motivation and Scope:

The deployment scope distinguishes per-device RSSI reports from offline mixtures in which multiple contributors share one observation. The learning task estimates position and orientation from RSSI, with orientation inferred internally at inference.

1) Deployment Motivation and Scope:

  • A shared composite requires the estimator to localize each requested contributor from overlapping received-power contributions, unlike a per-UE fingerprint.The offline construction approximates mean power at a common receiver; concurrent measurements would be needed to quantify synchronization, blockage, receiver, and access effects on accuracy and latency.Source: 1) Deployment Motivation and Scope · p. 5

B. Learning Tasks and Problem Formulation

  • At inference, the BS processes an RSSI sequence and target index to estimate position and orientation, with orientation inferred internally.The fixed sweep operates independently of unknown position; ground-truth positions and orientations provide training labels.Source: B. Learning Tasks and Problem… · p. 5
  • Training combines class-weighted orientation cross-entropy and Huber position regression, then jointly fine-tunes both networks after orientation-module pretraining.The Huber loss is used to reduce sensitivity to large position errors.Source: B. Learning Tasks and Problem… · p. 5

III. PROPOSED O-DSG DEEP LEARNING FRAMEWORK

O-DSG maps beam–RIS power-grid sequences to position coordinates by combining orientation context with spatial features and temporal processing. Its orientation representation conditions feature modulation and expert weighting before attention aggregates recurrent features across the observation window.

III. PROPOSED O-DSG DEEP LEARNING FRAMEWORK

  • The expected angle and posterior entropy summarize predicted orientation, conditioning feature modulation and soft expert weighting alongside spatial features.FiLM adapts features to orientation, while the expert gate combines distinct spatial representations.Source: §III · pp. 5–6
  • Temporal self-attention aggregates recurrent features over the observation window, within an architecture integrating orientation prediction, dual-stream feature extraction, and temporal processing.The target user index denotes the contributor slot, and statistical descriptors summarize normalized grid entries.Source: §III · p. 5, Figure 2 · p. 6

A. Input Representation and Normalization

The framework represents input as successive RSSI grids collected across receive beams and RIS configurations. An orientation branch uses the latest normalized snapshot to supply an angle-and-uncertainty context for the window.

A. Input Representation and Normalization

  • Each RSSI snapshot sweeps all receive beams and RIS configurations, and training-set mean and standard deviation normalize the grids before they are fixed for use.A window contains T snapshots; advancing it by one snapshot reuses T−1 prior snapshots and requires one new sweep.Source: A. Input Representation and… · p. 6

B. Integrated Orientation Prediction Module

  • A lightweight CNN orientation branch predicts a posterior over orientation classes from the latest normalized snapshot, whose expected angle and entropy form the context embedding.The context and FiLM coefficients are fixed within a window, while expert weights can vary by time step because the gate also receives snapshot features.Source: B. Integrated Orientation Prediction… · p. 6

1) Conditioning for Emulated Multi-User Scenarios:

The framework conditions spatial processing on orientation and, in emulated multi-user settings, on a target-contributor index. Multi-scale residual features are attended, modulated, and combined through soft expert gating.

1) Conditioning for Emulated Multi-User Scenarios:

  • The target-index embedding identifies the requested contributor in an offline mixture, while single-user processing uses a fixed slot; deployment with new devices requires consistent index association.The index represents contributor rank rather than device identity, and indices are trained for mixtures of up to eight contributors.Source: 1) Conditioning for Emulated… · pp. 6–7

1) Multi-Scale Residual Processing:

  • Parallel convolutional branches apply different kernel sizes to normalized RSSI and fuse residual-block outputs while retaining the receive-beam and RIS-configuration dimensions.The multi-scale design targets both sharp specular-path peaks and diffuse scattering clusters.Source: 1) Multi-Scale Residual Processing · p. 7

2) Spatial Attention Mechanism:

  • Spatial attention produces one sigmoid weight per beam–RIS grid position and broadcasts it across feature channels to highlight informative regions.The module is intended to mitigate noise and deep fading.Source: 2) Spatial Attention Mechanism · p. 7

D. Feature-wise Linear Modulation and Mixture of Experts

  • FiLM uses the joint embedding to rescale and shift attended feature channels before a soft gate combines outputs from distinct expert networks.The gate depends on the joint embedding and pooled features, allowing expert combinations to vary with orientation and observed features.Source: D. Feature-wise Linear Modulation and… · p. 7

E. Statistical Aggregation Branch

A statistical branch compresses each normalized power grid into descriptors, while the temporal head processes fused spatial, statistical, and context features to predict position.

E. Statistical Aggregation Branch

  • The statistical branch feeds dimensionless summaries of the normalized power grid to a residual multilayer perceptron, producing a statistical feature vector.The summaries include total power, dynamic range, and sums across RIS configurations and beams.Source: E. Statistical Aggregation Branch · p. 7

F. Temporal Fusion and Localization Head

  • The temporal head concatenates spatial, statistical, and joint-context features per snapshot, then processes each acquired window with a BiLSTM and temporal self-attention.The BiLSTM passes remain within the acquired window; attention layers combine multi-head self-attention, feedforward processing, and layer normalization.Source: F. Temporal Fusion and Localization Head · pp. 7–8
  • Temporal mean pooling and a linear output head map the processed window representation to the two-dimensional endpoint position.The output projection uses a 2 × dh matrix and a two-element bias.Source: F. Temporal Fusion and Localization Head · p. 8

G. Multi-Task Loss and Training

Training jointly optimizes the networks after orientation pretraining, progressing from single-user inputs to offline mixtures. The complexity discussion accounts for spatial, recurrent, attention, and contributor-conditioned computation.

G. Multi-Task Loss and Training

  • After orientation pretraining, joint optimization begins with single-user inputs and then adds offline mixtures with increasing contributor counts.The training objective is the composite loss defined in Section II.Source: G. Multi-Task Loss and Training · p. 8

H. Computational Complexity

  • Per-sequence computation includes orientation inference and spatial, statistical, and conditioning operations, with convolutional cost summed across parallel scales and all soft-gated experts.The soft MoE evaluates every expert.Source: H. Computational Complexity · p. 8
  • For a batch of M sequences, the stated attention-stack operation terms are Nh(Tdhdk + T^2dk) and Tdhdff, alongside recurrent computation.The attention terms cover projections, pairwise snapshot interactions, and feedforward processing in each layer.Source: H. Computational Complexity · p. 8
  • The embedding table stores O(Umaxde) parameters, and each contributor requires a conditioned prediction; batching can use GPU parallelism, with latency depending on batch size, hardware, and memory.The deployment latency budget also includes measurement reporting and communication when estimates are returned to devices.Source: H. Computational Complexity · pp. 8–9

IV. EXPERIMENTAL SETUP AND PERFORMANCE EVALUATION

The evaluation uses measured RSSI from a 28 GHz indoor NLoS laboratory testbed and four UE orientations. It also describes held-out evaluation procedures and offline mixtures synthesized from single-user measurements.

1) Experimental Testbed and Environment Description:

  • The testbed operates at 28 GHz under enforced NLoS conditions, using identical 4×4 arrays at the base station and mobile UE and a programmable 20 × 20-element RIS.The RIS codebook has 15 configurations; the receiver uses 60 fixed beam directions.Source: 1) Experimental Testbed and… · p. 9

2) Data Acquisition and Dataset Characteristics:

  • The dataset contains 1,534 RSSI-matrix samples from 45 rover trajectories and four UE orientations, with position labels derived from LiDAR annotations.Each matrix covers 60 UE beam patterns and 15 RIS configurations; trajectories span 3 m to 7 m from the RIS.Source: 2) Data Acquisition and Dataset… · pp. 9–10

3) Dataset Preprocessing and O-DSG Implementation:

  • Complete trajectories or acquisition groups are split into five outer folds before normalization, windowing, or mixture construction, with training-only statistics and orientation pretraining.Sequences contain ten chronological snapshots with 50% overlap, and no test group contributes to fitting, normalization, or model selection.Source: 3) Dataset Preprocessing and O-DSG… · p. 10

4) Emulated NOMA-style Multi-User Data Generation:

  • Offline mixtures pair complete windows from distinct raw records, combine their measurements in linear power, and apply fixed training normalization.The synthesis enforces at least 1.5 m separation, uses four measured orientation classes and a path-loss exponent of 2.5, and assigns each composite window one target per contributor.Source: 4) Emulated NOMA-style Multi-User Data… · p. 11
  • Offline training samples comprise 60% single-user, 25% two-user, 10% four-user, and 5% eight-user windows, all derived from the original single-user measurements.Approximately 4,500 single-user and composite training windows are drawn per fold.Source: 4) Emulated NOMA-style Multi-User Data… · p. 11

5) Additional Single-User Evaluation Protocol:

Across the additional single-user evaluation, O-DSG outperforms the evaluated baselines and sequence models, while ablations show substantial error increases when its temporal and orientation-aware components are removed.

5) Additional Single-User Evaluation Protocol:

  • The additional experiments use five outer folds of 45 acquisition groups, with ten-record windows at stride five yielding 212 test endpoints; O-DSG and sequence baselines receive RSSI at inference.Spatial tests hold out five regions and orientation tests hold out four angles; reported statistics average across three seeds and folds.Source: 5) Additional Single-User Evaluation… · p. 11

1) Overall Localization Performance:

2) Comparison with Sequence Models:

  • O-DSG attains the lowest average median and 95th-percentile errors among the sequence models, with 7.09 ms inference time versus 3.70–5.81 ms for the others.The independent sequence baselines receive the same ten-snapshot histories; this compares the complete O-DSG architecture with separately tuned models.Source: 2) Comparison with Sequence Models · p. 12, Table VI · p. 12

3) Ablation Study of the O-DSG:

  • Removing temporal processing causes the largest degradation: a 387% increase to 0.687 m; removing the CNN and statistical streams raises median error 272% to 0.524 m and 223% to 0.456 m.Removing orientation prediction, FiLM, and MoE raises median error by 192% to 0.412 m, 160% to 0.367 m, and 101% to 0.283 m, respectively.Source: 3) Ablation Study of the O-DSG · p. 12, Figure 8 · p. 12

4) Emulated Multi-User Evaluation and Computational Scaling:

Offline mixtures of single-user measurements test localization and computational scaling as emulated contributors increase. Additional analyses examine distance and orientation effects, generalization to held-out regions and angles, and measured inference costs.

4) Emulated Multi-User Evaluation and Computational Scaling:

  • 0.141 m median error for one user rises to approximately 0.53 m for eight emulated contributors; throughput grows from approximately 140 to 500 localizations per second.Processing latency increases from approximately 7 ms to 45 ms, while combined peak host and GPU memory rises from 2.5 GB to 27 GB.Source: 4) Emulated Multi-User Evaluation and… · pp. 12–13, Figure 9 · p. 13

5) Distance, Orientation, and Prediction Analysis:

  • Median error increases from 0.147 m at 3.25 m to 0.235 m at 6.75 m, while orientation classification reaches 92.3% accuracy and a weighted F1-score of 0.924.The lowest median error by orientation is 0.128 m at 0°; it rises to 0.178 m at 63.4°.Source: 5) Distance, Orientation, and… · p. 13, Table VII · p. 13

6) Spatial and Orientation Generalization:

  • Across five held-out spatial regions, O-DSG's average region median is 0.247 m versus 0.392 m for the selected sequence comparator; at interior omitted angles, its position medians are 0.297 m and 0.423 m.The comparator's corresponding interior-angle medians are 0.314 m and 0.555 m; across all four angle holdouts, the models average 0.482 m each.Source: 6) Spatial and Orientation… · pp. 13–14, Table VIII · p. 14

7) Statistical Significance and Computational Efficiency:

  • Paired O-DSG–DSOG group comparisons give p = 0.00145 and a 95% bootstrap interval of [0.019, 0.075] m for median-error reduction; O-DSG uses 2.884M parameters and 7.1 ms inference.A deployment timing example assumes 10 snapshots, 300 ms updates, and estimates 2.98 s plus processing time for the first estimate.Source: 7) Statistical Significance and… · p. 14, Table IX · p. 14

8) Measurement Grid and Sequence Length:

The section examines acquisition and deployment trade-offs, orientation effects, and scope limits for O-DSG, then situates its supervised estimator alongside adaptive RIS learning and offline power-mixture evaluation.

8) Measurement Grid and Sequence Length:

9) Discussion, Implications, and Limitations:

  • With each orientation omitted from training, conditioning lowers error at three of four holdouts, while both models average 0.482 m; between trained angles, expected orientation has angular MAE near 10°.Source: 9) Discussion, Implications, and… · p. 15
  • O-DSG weights are site-specific, so a new site requires a labeled survey and retraining; within the surveyed area, it has lower median error than the selected sequence comparator in every held-out spatial region.Source: 9) Discussion, Implications, and… · p. 15
  • Each BS update recomputes the ten-snapshot window at 4.283 GMACs, or about 14.3 GMAC/s per UE at τupd = 300 ms; motion during sweeps must also remain small relative to snapshot spacing.Source: 9) Discussion, Implications, and… · p. 16

10) Relationship to Adaptive RIS Learning:

  • Unlike supervised localization from labeled data, adaptive RIS control learns through action–reward interactions; combining the approaches requires a localization-tied reward and an estimator compatible with the chosen configurations.Source: 10) Relationship to Adaptive RIS… · p. 16

V. CONCLUSION

  • O-DSG reaches 0.141 m median error from received power alone; in held-out spatial regions, its average median error is 0.247 m versus 0.392 m for the selected sequence model.Source: §V · p. 16
  • Offline mixtures of separately measured fingerprints test target-conditioned localization from shared composites and computational scaling as the number of contributors grows.Source: §V · p. 16

APPENDIX NETWORK IMPLEMENTATION AND MODEL SELECTION

The appendix specifies O-DSG’s orientation and expert layers, statistical and temporal processing, baseline configurations, and arithmetic-counting procedure.

A. Orientation Network and Expert Layers

B. Statistical and Temporal Layers

C. Original Baselines

  • The original Random Forest, ridge-regression, MLP, CNN, and NA baselines use specified model searches, with Huber loss and Adam for the neural models.Source: C. Original Baselines · p. 17

D. Additional Sequence Baselines

  • Additional CNN–BiLSTM, CNN–Transformer, and CNN–TCN baselines use searched sequence configurations, with selected parameter counts ranging from 0.391M to 2.462M.Source: D. Additional Sequence Baselines · p. 17

E. Parameter and Operation Counts

  • The operation counts include convolutional, dense, recurrent, and attention matrix products; O-DSG’s full ten-snapshot, four-expert configuration totals 4.283 GMACs.Source: E. Parameter and Operation Counts · p. 17

Related papers

Loading 2610.21043v1…