SigT: An Efficient End-to-End MIMO-OFDM Receiver Framework Based on Transformer

I. INTRODUCTION

• Conventional receivers use independently designed cascaded FFT, channel-estimation, detection, and demodulation blocks whose ideal channel assumptions can fail in complex wireless environments.Hardware nonlinearity and mismatched channel assumptions can deteriorate BER performance.Source: §I · p. 1
• End-to-end DL receivers optimize recovery globally but can treat the receiver as a black box and face exponentially increasing recovery difficulty as input length grows.Existing individual-block DL designs also require expert knowledge and handcrafted information between subnets.Source: §I · pp. 1–2
• SigT models each receiver antenna’s signal as a Transformer token, learning spatial correlations to mitigate the zero-shot problem.The zero-shot problem becomes more severe in MIMO because input digit sequences are longer, while SigT learns antenna correlations during recovery.Source: §I · p. 2
• SigT is proposed as an end-to-end MIMO-OFDM receiver framework and reference baseline for future learning-based receiver studies.The framework applies Transformer modeling to the complete receiver task rather than separately optimizing conventional receiver blocks.Source: §I · p. 2
• SigT recovers data accurately without foreknown pilots, freeing time-frequency resources for useful signals and potentially enlarging effective channel capacity.This distinguishes the proposed receiver from most existing DL-based receivers described by the authors.Source: §I · p. 2
• Experiments on a real dataset report that SigT outperforms existing DL-based end-to-end models by a large margin.The paper describes the evaluation as extensive and uses signal recovery accuracy as the comparison basis.Source: §I · p. 2

A. MIMO-OFDM System

• The conventional receiver performs CP removal, FFT, MIMO channel estimation, signal detection, and QAM demodulation to recover transmitted bits.Pilot symbols estimate the MIMO channel before data-symbol detection.Source: A. MIMO-OFDM System · p. 2

B. End-to-End MIMO-OFDM Receiver Design

• The end-to-end DL model substitutes channel estimation, signal detection, and QAM demodulation, mapping the post-FFT input Y directly to predicted labels X-hat.During testing, hard decisions produce the recovered bitstream X-tilde with the same dimensions as the transmitted bitstream.Source: B. End-to-End MIMO-OFDM Receiver Design · p. 2
• The receiver is evaluated with average accuracy (AACC), comparing each recovered test signal with its ground-truth signal.N denotes the number of test signals in the stated metric definition.Source: B. End-to-End MIMO-OFDM Receiver Design · p. 2
• Figure 1 contrasts the conventional cascaded receiver with the end-to-end DL receiver that performs these recovery functions as one model.The figure’s two panels correspond to the conventional system and the end-to-end DL-based design.Source: B. End-to-End MIMO-OFDM Receiver Design · p. 2, Figure 1 · p. 3
• Figure 2 reshapes the received signal across 256 subcarriers, flattens it for Transformer input, and splits rows into antenna-information tokens.The demonstration uses Nr = 16 receiving antennas and Nt = 4 transmitting antennas.Source: Figure 2 · p. 3
• Because the label space grows as 2^(Ns×Nt×2), finite datasets cannot cover all labels, requiring prediction of unseen labels and creating a zero-shot generalization problem.This problem can produce a large training–test accuracy gap identified as overfitting.Source: B. End-to-End MIMO-OFDM Receiver Design · p. 3

III. SIGNAL TRANSFORMER

• SigT is presented as a Transformer-based end-to-end MIMO-OFDM receiver built around multi-head self-attention.The section introduces the attention mechanism before describing SigT’s design details.Source: §III · p. 3

A. Multi-Head Self-Attention Mechanism

• Self-attention captures each vector's features and aggregates information across the input sequence without changing its number or dimensionality.Queries, keys, and values are obtained through linear transformations.Source: A. Multi-Head Self-Attention Mechanism · p. 3
• The attention mechanism uses the input tensor and the dimensions of query, key, and value vectors, with scaling applied when query and key dimensions match.The scaling factor is 1/√dK.Source: A. Multi-Head Self-Attention Mechanism · p. 3
• Multi-head attention concatenates the outputs of multiple attention heads and applies an output projection.Source: A. Multi-Head Self-Attention Mechanism · p. 3

B. Signal Transformer Model

• SigT learns correlations among receiver antennas with Transformer self-attention because each receiver antenna observes signals from all transmitting antennas.These correlations are used during signal recovery.Source: B. Signal Transformer Model · p. 3

1) Tokenization:

• Tokenization reshapes and permutes the received signal according to antenna and subcarrier structure, then splits it into Nr antenna-specific vectors for the Transformer Encoder.Each vector serves as one token in the input sequence.Source: 1) Tokenization · p. 3

2) Transformer Encoder:

• The Encoder captures relations between antenna tokens, helping complement missing information and enhance information shared across antennas.Its structure follows the original Transformer design.Source: 2) Transformer Encoder · pp. 3–4
• The model review shows the processing path from antenna-wise input vectors through the Encoder and aggregation layers before prediction.The Encoder output retains the input feature dimension before vector reduction.Source: Figure 3 · p. 4
• Multi-head self-attention can learn multiple relation types and generally performs better than single-head attention without increasing asymptotic complexity.This provides the Encoder with multiple views of token relationships.Source: 2) Transformer Encoder · p. 4

3) Aggregation:

• Pooling or convolutional layers aggregate information from nearby Encoder features and produce Nt output vectors from the Encoder feature sequence.The Encoder features are denoted fi, and the aggregated outputs pi.Source: 3) Aggregation · p. 4

4) Output and Loss Calculation:

• The vectors pi are concatenated into P and fed into a two-layer MLP, with Sigmoid applied to produce ˆX in RNs×Nt×2.Source: 4) Output and Loss Calculation · p. 4
• The sigmoid maps MLP outputs to [0,1], while Hardmax converts the predicted values into the test-time bitstream X-tilde.The hard decision is applied after prediction rather than during the continuous-output training stage.Source: 4) Output and Loss Calculation · p. 4
• Training updates the MLP, convolutional or pooling layers, and Transformer Encoder using mean squared error between X-hat and the ground-truth signal.The supplied passage introduces the MSE objective but does not provide its completed equation.Source: 4) Output and Loss Calculation · p. 4

A. Experiment Configurations

• Experiments use PCL channel data sampled at 3.5 GHz from 500 locations with 20 measurements each.Source: A. Experiment Configurations · p. 4

B. Performance Comparison with Benchmark Methods

• The MIMO-OFDM setup uses 256 subcarriers, 4 transmitting antennas, 16 receiving antennas, and one information symbol per antenna per subcarrier.Source: B. Performance Comparison with… · p. 4
• The dataset uses a test-to-training size ratio of 1/10, with default training and test sets of 25,600 and 2,560 signals.Source: B. Performance Comparison with… · p. 4
• Training uses minibatches of 640 signals and Adam with learning rate 0.0001, β1 = 0.9, β2 = 0.999, and ϵ = 10^-8 without weight decay.Source: B. Performance Comparison with… · p. 4

C. Benchmark DL-Based Models

• SigT outperforms FC-DNN and CSINet by a large margin across SNRs, with all models improving rapidly below 10 dB.Source: C. Benchmark DL-Based Models · p. 5, Figure 4 · p. 5
• At higher SNR, an error floor occurs mainly because end-to-end MIMO-OFDM models face generalization difficulty.Source: C. Benchmark DL-Based Models · p. 5
• At 10 dB SNR, SigT achieves 14.9% higher AACC than FC-DNN and 8.95% higher AACC than CSINet.Source: C. Benchmark DL-Based Models · p. 5, Figure 5 · p. 5
• FC-DNN shows overfitting, whereas CSINet exhibits underfitting; SigT reaches higher training and test AACC and grows faster than both.Source: C. Benchmark DL-Based Models · p. 5
• SigT test AACC rises rapidly with small training sets and saturates as NB increases, so enlarging the training set does not always improve learning.Source: C. Benchmark DL-Based Models · p. 5

D. Ablation Study

• SigT with convolutional layers achieves higher training and test AACC than SigT with pooling layers.Convolution captures local information, whereas pooling primarily reduces feature dimensionality.Source: D. Ablation Study · p. 5
• Applying dropout to SigT’s MLP head lowers performance for both convolution-based and pooling-based variants.The authors associate this with reduced representative power when hidden layers are not overly large.Source: D. Ablation Study · pp. 5–6
• Adam converges faster and reaches higher AACC than SGD, which converges more slowly and performs poorly.Source: D. Ablation Study · p. 6, Figure 7 · p. 6
• An LSTM-backbone variant learns little and has much lower training and test AACC than SigT.The comparison retains the remaining SigT components and uses convolutional layers, no dropout, and Adam for LSTM.Source: D. Ablation Study · p. 6

• SigT uses a transformer to learn correlations among receiving antennas and relieve the zero-shot problem in end-to-end learning models.Source: §V · p. 6
• SigT is designed as an end-to-end DL-based MIMO-OFDM receiver for detecting and demodulating received signals.Source: §V · p. 6
• SigT achieves the highest signal recovery accuracy compared with the other two benchmark end-to-end DL-based models.Source: §V · p. 6
• The authors propose SigT as a performance baseline for future learning-based MIMO-OFDM receiver design.Source: §V · p. 6
• Future work will study dynamic mechanisms that adapt the DL-based receiver to dynamic environments.Source: §V · p. 6