Source-linked AI summary
Temporal vs Spectral Resolution in CSI Reporting for Wi-Fi ISAC
Abolfazl Zakeri, Sumit Roy, Nhan Thanh Nguyen, Markku Juntti
TL;DR
CSI feedback supports Wi-Fi HAR but consumes airtime that could serve data communications. The paper evaluates temporal and spectral reporting choices, including mean CSI payloads, against feedback cost on three public datasets. Temporal resolution is generally more cost-effective, and the proposed scheme attains near-full-CSI accuracy at at most 7.7% of full-CSI cost.
Problem
Frequent CSI feedback for HAR consumes airtime otherwise available for data transmission, while temporal and spectral reporting-cost tradeoffs have received limited investigation.
Source: §I · p. 1Method
The paper compares matched-rate temporal reporting and matched-size spectral reduction using a CNN–GRU with temporal-pyramid readout across three public datasets.
Source: §I · p. 1Results
Takeaways & Limitations
In the evaluated settings, full CSI is not Pareto-efficient, and standardized reporting mechanisms can provide substantial feedback savings.
Source: §V · p. 6
Abstract
from arXiv · showhide
CSI feedback is fundamental to proposed uses of Wi-Fi sensing such as human activity recognition (HAR); but it also constitutes significant overhead as resources must be allocated away from data communications. Wi-Fi CSI extraction can be explored along two dimensions in IEEE 802.11bf: reporting intensity, i.e., temporal resolution, and number of reported subcarriers, i.e., spectral resolution. We explore periodic and IEEE 802.11bf threshold-based reporting at matched rates, and uniform decimation and block-mean grouping at matched report sizes. We further propose reporting the mean CSI accumulated since the last report. Using the 802.11bf report size as the metric, we evaluate the accuracy-cost tradeoffs on three public datasets with our proposed model, a convolutional neural network combined with a gated recurrent unit and temporal pyramid pooling. The proposed periodic reporting with mean payload and block-mean grouping improves accuracy by up to 53 percentage points at equal CSI overhead cost. Temporal resolution is generally more cost-effective than spectral resolution. With only at most 7.7% of full-CSI feedback cost, the proposed scheme achieves 96-98% of full-CSI accuracy.
I. INTRODUCTION
The paper studies how temporal and spectral CSI reporting choices trade HAR accuracy against feedback cost. Across three datasets, temporal resolution is generally more cost-effective than spectral resolution, and the proposed scheme approaches full-CSI accuracy at a small fraction of its cost.
I. INTRODUCTION
- Frequent CSI feedback for HAR consumes airtime otherwise available for data transmission, making its overhead a central integration challenge.CSI report size can range from 44 to 64 624 octets, depending on antenna configuration, bandwidth, and subcarrier grouping.Source: §I · p. 1
- The study compares temporal reporting intensity and spectral subcarrier count using matched-rate reporting rules and matched-size spectral reduction.It evaluates periodic, random, and threshold-based reporting, instant or mean payloads, and uniform decimation or block-mean averaging.Source: §I · p. 1
- The proposed CNN–GRU with temporal-pyramid readout supports varying temporal lengths and subcarrier counts, and outperforms reported dataset models with full CSI.Its reported gains are 6.5, 2.7, and 5.5 percentage points over the SignFi CNN, SenseFi LeNet, and CSI-HAR 2D-CNN, respectively.Source: §I · pp. 1–2
- At at most 7.7% of full-CSI feedback cost, the proposed scheme reaches 95.7–98.3% of full-CSI accuracy.The results characterize accuracy–cost Pareto frontiers across three datasets.Source: §I · p. 2
II. SYSTEM MODEL AND PROBLEM FORMULATION
The system model treats trigger-based Wi-Fi sensing as a sequence of CSI measurements collected by a station and labeled for an end task. The formulation seeks to reduce feedback by controlling reporting time and the number of reported subcarriers.
II. SYSTEM MODEL AND PROBLEM FORMULATION
- Each sounding instance yields CSI X(t) ∈ C^(N×K), with N = N_TN_R spatial links and K subcarriers.N_T and N_R denote the numbers of transmit and receive antennas.Source: §II · p. 2
- The feedback-reduction problem has two dimensions: when the station reports and how many subcarriers each report contains.Both temporal and spectral mechanisms are supported by IEEE 802.11bf, while the passage describes prior investigation as limited.Source: §II · p. 2
- The sensing task assigns a label to a session after collecting T CSI samples, which need not be uniformly spaced in time.Sending every CSI instance to the access point incurs substantial feedback overhead.Source: §II · p. 2
A. Temporal Resolution Modeling
Temporal reporting is modeled through report instants, report payloads, and rules that determine when the station sends CSI. The reporting strategies are compared at matched average rates, with threshold control tuned using unlabeled CSI.
A. Temporal Resolution Modeling
- A report carries either the current CSI frame or the mean of frames accumulated since the previous report, with the same report size for either payload.The instant payload is the standard default; the mean payload is proposed.Source: A. Temporal Resolution Modeling · p. 2
- A session contains T sounding instances, indexed by sounding instance rather than exact CSI measurement timestamp.The sensing-window duration and sampling rate primarily determine the number of instances.Source: A. Temporal Resolution Modeling · p. 2
- Periodic reporting selects instances at a fixed interval, while random reporting samples independently after the initial report to match the mean rate.The random rule models report loss or contention without memory.Source: A. Temporal Resolution Modeling · p. 2
- Threshold reporting requests a full report when CSI variation reaches a threshold, after short variation reports at each instance.The standard leaves the variation measure to implementation; the described rule uses a threshold τ.Source: A. Temporal Resolution Modeling · p. 2
- The threshold is tuned on unlabeled CSI to meet the target average reporting rate, enabling rate-matched comparisons among rules.The standard quantizes τ to {0, 0.1, …, 1}, while the mapping from channel change to variation is implementation-dependent.Source: A. Temporal Resolution Modeling · p. 2
B. Spectral Resolution (Subcarrier Grouping)
Subcarrier grouping reduces spectral resolution by reporting one value for each group of adjacent subcarriers. The section compares uniform decimation with proposed block-mean averaging at equal report size.
B. Spectral Resolution (Subcarrier Grouping)
- Uniform decimation retains the first subcarrier in each group, whereas block mean averages the group’s subcarriers.Both grouping choices use the same report size; subcarriers that do not fill a complete group are dropped.Source: B. Spectral Resolution (Subcarrier… · p. 2
- Block mean averages each group of adjacent subcarriers, acting as a moving-average low-pass filter before down-sampling.For complex CSI, it is applied after removing the linear phase slope across subcarriers so timing offsets do not attenuate the average.Source: B. Spectral Resolution (Subcarrier… · p. 2
- The study uses fixed Sg rather than optimized subcarrier selection, which is outside its scope; its added computational complexity may not justify the limited performance gain.Source: B. Spectral Resolution (Subcarrier… · p. 2
C. CSI Feedback Cost Model
The feedback-cost model separates a component independent of reporting rate from one that scales with report count. Threshold-based reporting also incurs trigger-frame and short-report overhead, while sounding count is treated as a separate resource.
C. CSI Feedback Cost Model
- Feedback cost has a reporting-rate-independent component and a component that scales with the number of reports.The size of each CSI report depends on the number of reported subcarriers.Source: C. CSI Feedback Cost Model · p. 3
- Threshold-based reporting adds a trigger frame and a short report carrying a 4-bit CSI-variation field at each instant.Its feedback overhead includes a minimum Tcv term even when the channel rarely crosses the threshold.Source: C. CSI Feedback Cost Model · p. 3
- The model’s budgeted resource is CSI feedback overhead; the number of NDP soundings is separate and has its own cost.Table I reports both resources for each scheme.Source: C. CSI Feedback Cost Model · p. 3
D. Machine Learning Task and Pareto-frontier Reporting
The paper evaluates session-level classification across reporting scenarios by reconstructing CSI with zero-order hold and comparing accuracy against feedback cost. It defines Pareto-efficient scenarios as those not dominated in accuracy and cost.
D. Machine Learning Task and Pareto-frontier Reporting
- The AP reconstructs each CSI sequence with zero-order hold, retaining the most recently received report until the next one arrives.The first frame is assumed reported, so the reconstructed sequence is defined throughout the measurement window.Source: D. Machine Learning Task and… · p. 3
- Each reporting scenario combines a reporting policy, payload type, target reporting rate, grouping scheme, and grouping factor.The policies are periodic, random, or threshold-based; payloads are instantaneous or mean CSI, and grouping is uniform or block mean.Source: D. Machine Learning Task and… · p. 3
- For each scenario, a neural network performs session-level classification, with parameters trained by minimizing empirical cross-entropy and accuracy evaluated on sessions excluded from training.The training dataset and loss are specified in the evaluation definition.Source: D. Machine Learning Task and… · p. 3
- Scenarios differing only in payload type or grouping scheme have identical feedback costs, while the Pareto frontier gives the highest accuracy attained at each cost level.The cost depends on the reporting policy, target rate, and grouping factor; the frontier is evaluated for each dataset.Source: D. Machine Learning Task and… · p. 3
III. NEURAL NETWORK MODEL AND TRAINING
The model processes reconstructed CSI as a frequency–time input using a CNN-GRU with temporal-pyramid readout. Its fixed-size representation supports varying window lengths and subcarrier groupings, and the architecture is held constant across reporting settings.
III. NEURAL NETWORK MODEL AND TRAINING
- Inputs are reconstructed with zero-order hold and grouped, then represented as frequency–time maps with amplitude and, where available, sanitized phase channels.The input length remains T across reporting rates, so the same architecture serves all settings without padding or masking.Source: §III · pp. 3–4
- The CNN-GRU temporal-pyramid model uses convolutional blocks, a bidirectional GRU, and a readout that averages outputs over 1, 2, and 4 temporal segments.The readout concatenates these pooled representations into a 14H-dimensional vector, retaining coarse temporal order.Source: §III · p. 4
- The network accepts any window length and subcarrier grouping because its temporal-pyramid readout has fixed size.Adaptive average pooling collapses the frequency axis to D bands before the bidirectional GRU models temporal dynamics.Source: §III · p. 4
- A separate model is trained for each reporting configuration using validation-selected checkpoints and evaluated on held-out sessions.Accuracy is averaged across Monte Carlo runs and evaluation folds; the architecture remains fixed across configurations so accuracy differences reflect reporting settings rather than model tuning.Source: §III · p. 4
IV. SIMULATION RESULTS
This section evaluates CSI-reporting accuracy–cost tradeoffs, introducing the datasets and simulation setup before discussing results.
IV. SIMULATION RESULTS
- The evaluation first describes the datasets and simulation setup, then discusses the reporting schemes’ accuracy–cost tradeoffs.Source: §IV · p. 4
A. Datasets and Simulation Setup
The study emulates reporting schemes on three public CSI datasets using 200 sounding instances per session, with dataset-specific splits and threshold calibration.
A. Datasets and Simulation Setup
- Each scheme is applied to full-resolution CSI from three public datasets, treating each session as T = 200 sounding instances.NTU-Fi HAR sessions are decimated by four; CSI-HAR and NTU-Fi sequences are restricted to central instances where needed.Source: A. Datasets and Simulation Setup · p. 4
- Thresholds are calibrated by bisection to match the mean reporting rate using unlabeled CSI, with calibration grouped by user, subject, or NTU-Fi split.Source: A. Datasets and Simulation Setup · p. 4
- SignFi and CSI-HAR use leave-one-user or leave-one-subject-out folds, while NTU-Fi HAR uses its fixed split; scenario comparisons are paired across runs.Random CSI-HAR splits can inflate accuracy from about 61% to about 93%.Source: A. Datasets and Simulation Setup · p. 4
B. Performance Evaluations and Discussions
The reporting frontiers show dataset-dependent tradeoffs, with temporal reporting often a more effective use of budget than spectral resolution. The proposed frontier reaches full-CSI or near-full-CSI performance at low relative costs on all three datasets, though the benefit of mean payload varies by task.
B. Performance Evaluations and Discussions
- The table summarizes each budget’s highest accuracy, marking below-minimum-threshold settings with dashes.Source: Table III · p. 5
- Threshold reporting has minimum relative costs of 13.3%, 21.4%, and 4.0% on SignFi, CSI-HAR, and NTU-Fi HAR, respectively.Below those limits, only periodic schemes are viable.Source: B. Performance Evaluations and… · p. 5
- On SignFi, mean payload boosts accuracy by 23 percentage points at a 2% budget, while random reporting trails periodic reporting by 13.0 points at matched mean rate.Random reporting also trails periodic reporting by 2.3 and 0.5 points on CSI-HAR and NTU-Fi HAR, respectively.Source: B. Performance Evaluations and… · pp. 4–5
- On SignFi, the proposed scheme reaches 63.1% accuracy at 12.9% cost, matching the 62.6% full-CSI result.At 1.7% cost, it reaches 57.5%, compared with 34.5% for periodic instant-payload reporting with uniform grouping at the same cost.Source: Figure 2 · p. 5, B. Performance Evaluations and… · p. 5
- On CSI-HAR, periodic reporting reaches 40–42% accuracy at 1.9–2.1% cost and exceeds full-CSI accuracy from 7.7% cost onward.Mean payload provides no measurable gain beyond one standard error; the proposed frontier shifts early to grouping factors g = 2–4, then raises the report ratio.Source: B. Performance Evaluations and… · p. 5, Figure 3 · p. 6
- At 2.8% cost, the proposed scheme reaches 97.0% accuracy on NTU-Fi, while the standard scheme needs 5.4% cost to reach that accuracy.Above about 10% cost, NTU-Fi frontiers converge near the 98.6% full-CSI accuracy.Source: Figure 4 · p. 6, B. Performance Evaluations and… · p. 6
V. CONCLUSIONS
The evaluated results show that CSI feedback can be substantially reduced without a proportional loss in HAR accuracy. Temporal resolution provides more accuracy benefit per unit cost than spectral resolution, and the evaluated low-cost configurations approach full-CSI accuracy.
V. CONCLUSIONS
- At most 7.7% of full-CSI feedback cost achieves 95.7–98.3% of full-CSI accuracy, so full CSI is not Pareto-efficient in the evaluated settings.The conclusion reports these savings using standardized reporting mechanisms, without learned CSI compression or reconstruction.Source: §V · p. 6
- Temporal resolution provides greater accuracy benefit per unit feedback cost than spectral resolution, favoring increased reporting frequency with coarse subcarrier grouping over dense spectral information.Averaging CSI measurements between reports further improves use of the available feedback budget.Source: §V · p. 6
- The results suggest that substantial CSI feedback savings are possible with standardized reporting mechanisms, without learned CSI compression or reconstruction.Source: §V · p. 6