Organizations: School of Engineering, Institute of Science Tokyo, 2-12-1 Ookayama, Tokyo, 152-8550, Japan · Wireless Systems R&D Department, Infrastructure Systems R&D Center, Corporate Laboratory, Toshiba Corporation, 1, Komukai Toshiba-cho, Saiwai-ku, Kawasaki 212-8582, JAPAN
We propose a WiFi Channel State Information (CSI) sensing framework for multi-station deployments that addresses two fundamental challenges in practical CSI sensing: station-wise feature missingness and limited labeled data. Feature missingness is commonly handled by resampling unevenly spaced CSI measurements or by reconstructing missing samples, while label scarcity is mitigated by data augmentation or self-supervised representation learning. However, these techniques are typically developed in isolation and do not jointly address long-term, structured station unavailability together with label scarcity. To bridge this gap, we explicitly incorporate station unavailability into both representation learning and downstream model training. Specifically, we adapt cross-modal self-supervised learning (CroSSL), a representation learning framework originally designed for time-series sensory data, to multi-station CSI sensing in order to learn representations that are inherently invariant to station-wise feature missingness from unlabeled data. Furthermore, we introduce Station-wise Masking Augmentation (SMA) during downstream model training, which exposes the model to realistic station unavailability patterns under limited labeled data. Our experiments show that neither missingness-invariant pre-training nor station-wise augmentation alone is sufficient; their combination is essential to achieve robust performance under both station-wise feature missingness and label scarcity. The proposed framework provides a practical and robust foundation for multi-station WiFi CSI sensing in real-world deployments.
Figures & tables
Method
Approaches
#Stations
Handles Station-wise Missingness
Works with Uncontrolled Transmissions
Label-efficient
RFboost [ 22 ]
DA
-
-
-
✓
RR-scheduled CSI recovery with inpainting [ 11 ]
DA
Multi
✓
✗
✗
CSI-BERT [ 23 ]
DA
Single
✗
✓
✗
AutoFi [ 18 ]
Pre-training
Single
✗
✗
✓
Multi-device SSL [ 19 ]
Pre-training
Multi
✗
✗
✓
Ours
Pre-training + DA
Multi
✓
✓
✓
TABLE I : Comparison of related CSI sensing methods in terms of missingness handling and learning assumptions.
Fig. 1 : Multi-stations CSI sensing setup and visualization of timestamp alignment. Multiple stations transmit frames to the AP. A camera provides images for ground-truth of downstream model training. As illustrated in the upper timeline, WiFi frames from different stations and camera images are acquired asynchronously and at heterogeneous intervals.
Fig. 2 : Illustration of labeled and unlabeled sample generation from asynchronous CSI and label streams. Both labeled and unlabeled samples are constructed by applying a fixed-length window around reference timestamps. Due to asynchronous sensing and heterogeneous transmission intervals, some windows may contain no CSI from particular stations, resulting in station-wise feature missingness.
Fig. 3 : Overview of the proposed framework. The framework learns station-missingness-invariant representations from multi-station CSI and applies them to downstream tasks. CroSSL is used to learn robust global embeddings from unlabeled CSI by masking intermediate station embeddings during representation learning. For downstream model training, SMA masks entire station inputs to simulate realistic station unavailability. The same encoders and aggregator are shared across both phases, enabling robust inference under station-wise feature missingness and limited labeled data.
Fig. 4 : Experimental setup of office-like environment.
office-like environment
factory-like environment
AP
ESP32
NETGEAR Nighthawk X 10
Station
ESP32
Raspberry Pi 4 model B
Wireless LAN standard
802.11n
802.11ac
Channel
6
36
Bandwidth
20 MHz
80 MHz
CSI sensor firmware
ESP32-CSI-Tool [ 39 ]
Nexmon CSI [ 40 , 41 ]
TABLE II : Experimental configuration
Hyperparameter
office-like env
factory-like env
Downstream Model
learning rate
-
2.0e-5
Pre-training: AutoFi
learning rate
6.0e-4
2.4e-4
noise scale
4.5e-2
9.0e-5
probability size
16
32
TABLE III: Optimized hyperparameters
Fig. 5 : Examples of one-dimensional position estimation. The white line denotes the ground-truth label, while the blue, green, and red lines indicate the predictions obtained with 1, 4, and 8 stations, respectively. RMSExstations represents the RMSE of using x stations.
Fig. 6 : Performance comparison under different numbers of available stations on the office-like environment dataset. The proposed method is shown as the purple solid line. For methods whose worst-case RMSE exceeds 0.21, the standard deviation bars are omitted for clarity. While most baselines degrade as stations become unavailable, the proposed method maintains low RMSE, demonstrating robustness to station-wise feature missingness.
Fig. 7 : Performance comparison under different amounts of labeled data used for downstream model training on the office-like environment dataset. While the performance of most baselines degrades as the amount of labeled data decreases, the proposed method maintains stable performance and remains competitive across all label ratios, demonstrating robustness to limited labeled data.
Label usage ratio(%)
0.1
10
Available stations
1
4
1
4
ProposedMethod
0.0924
0.0424
0.0867
0.0353
Constant
0.1977
0.1977
0.1977
0.1977
NaiveSupervised
0.1888
0.1488
0.1910
0.1177
OutputEnsemble
0.1233
0.0894
0.0845
0.0563
AutoFi
0.2415
0.2190
0.2522
0.2254
TABLE IV : RMSE comparison under different numbers of available stations and label ratios on the office-like environment dataset. For each condition, the reported RMSE is averaged over all combinations of station-missingness patterns. Best results for each condition are shown in bold.
Fig. 8 : PCA-transformed scatter plots of raw inputs and embeddings under different station-wise feature missingness conditions. Points colored by the ground-truth position. While the raw inputs become increasingly irregular as missingness increases, the embeddings maintain a smooth and consistent manifold, indicating robustness to station-wise feature missingness.
Label usage ratio(%)
0.1
10
100
Available stations
1
4
8
1
4
8
1
4
8
ProposedMethod
0.0924
0.0424
0.0368
0.0867
0.0353
0.0276
0.0863
0.0334
0.0251
AutoFi × SMA
0.2126
0.1964
0.0977
0.1594
0.1080
0.0323
0.1158
0.0818
0.0289
DAE × SMA
0.2394
0.2182
0.1142
0.1709
0.1438
0.0512
0.1087
0.0825
0.0545
CroSSL × RE
0.1083
0.0670
0.0619
0.0940
0.0391
0.0285
0.0926
0.0388
0.0267
TABLE V : RMSE comparison of different combinations of pre-training and data augmentation strategies on the office-like environment dataset under varying numbers of available stations and label ratios. For each condition, the reported RMSE is averaged over all combinations of station-missingness patterns. Best results for each condition are in bold. The results show that the proposed method consistently achieves the lowest error.
Fig. 9 : Each heatmap shows the RMSE for combinations of CroSSL and SMA pmask values under different numbers of available stations on the office-like environment dataset. Across all conditions, the combination pmask=0.5 provides the most stable performance, suggesting that this masking rate offers an appropriate training difficulty for the office-like environment dataset.
Fig. 10 : Experimental setup of factory-like environment.
Fig. 11 : Ground-truth and predicted images obtained by the proposed method using all labeled data for training, under different numbers of available stations for inference.
Fig. 12 : Performance comparison under different numbers of available stations on the factory-like environment dataset. The proposed method is shown as the purple solid line. For methods whose worst-case RMSE exceeds 0.175, the standard deviation bars are omitted for clarity. The proposed method maintains consistently lower RMSE than the baselines as stations become unavailable, demonstrating robustness to station-wise feature missingness in the factory-like environment.
Fig. 13 : Performance comparison under different amounts of labeled data used for downstream model training on the factory-like environment dataset. For methods whose worst-case RMSE exceeds 0.175, the standard deviation bars are omitted for clarity. While the performance of most baselines degrades as the amount of labeled data decreases, the proposed method maintains stable performance and remains competitive across all label usage ratios, demonstrating robustness to limited labeled data.
Label usage ratio(%)
30
50
Available stations
1
4
1
4
ProposedMethod
0.1310
0.1235
0.1287
0.1202
Constant
0.1370
0.1370
0.1370
0.1370
NaiveSupervised
0.2522
0.2309
0.2596
0.2340
OutputEnsemble
0.1746
0.1592
0.1375
0.1317
AutoFi
0.1973
0.1793
0.3438
0.2364
TABLE VI : RMSE comparison on the factory-like environment dataset, evaluating combined robustness to label scarcity and station-wise feature missingness. Best results for each condition are in bold. For all conditions, proposed method show the best performance.
Label usage ratio(%)
30
50
100
Available stations
1
4
9
1
4
9
1
4
9
ProposedMethod
0.1310
0.1235
0.1145
0.1287
0.1202
0.1109
0.1260
0.1149
0.1043
AutoFi × SMA
0.1310
0.1230
0.1165
0.1279
0.1177
0.1097
0.1252
0.1125
0.1037
DAE × SMA
0.2532
0.2554
0.2364
0.1847
0.1818
0.1685
0.1789
0.1741
0.1582
CroSSL × RE
0.1545
0.1383
0.1149
0.1583
0.1387
0.1106
0.1613
0.1390
0.1061
TABLE VII : RMSE comparison of different combinations of pre-training and data augmentation strategies on the factory-like environment dataset under varying numbers of available stations and label usage ratios. For each condition, the reported RMSE is averaged over all combinations of station-missingness patterns. Best results for each condition are shown in bold. The results demonstrate that the proposed method consistently achieves low error across different station availability and label usage ratios.
Fig. 14 : Each heatmap shows the RMSE for combinations of CroSSL and SMA pmask values under different numbers of available stations on the factory-like environment dataset. Similar to the office-like environment, the combination pmask=0.5 provides the most stable performance across all conditions.