Source-linked AI summary
Learning-Aided Short Code Design for ISAC based on MIMO-OFDM
Mingcheng Nie, Shuangyang Li, Geng Wang, Peng Cheng, Shenghong Li, Chang Liu, Giuseppe Caire, Yonghui Li
TL;DR
Communication and sensing favor conflicting waveform properties, motivating joint ISAC waveform design. This paper uses a DL-based MIMO-OFDM framework to map heterogeneous inputs into transmit waveforms and balance communication and delay-sensing objectives. Its learned designs offer BER–MCRB trade-offs, with a balanced operating point close to communication- and sensing-oriented baselines.
Problem
Communication’s need for information-bearing randomness conflicts with sensing’s preference for deterministic signals, making joint waveform design challenging.
Source: §I · p. 1Method
A DL-based MIMO-OFDM framework maps bits, communication channels, and sensing angles into waveforms using a weighted communication-and-sensing training objective.
Source: §I · p. 1, A. Neural Network Architecture · p. 3, §IV · p. 5Results
At λ = 0.4, BER is close to the CC-QPSK baseline while delay MCRB approaches the sensing-oriented baseline; all learned designs have lower MCRB than both communication benchmarks.
Source: B. Comparisons · p. 6Takeaways & Limitations
The balanced waveform has an intermediate structure between communication-oriented energy spread across delay bins and sensing-oriented impulse-like energy concentration.
Source: §V · pp. 6–7
Abstract
from arXiv · showhide
This paper proposes a deep learning (DL)-based coded waveform design for integrated sensing and communications (ISAC), enabling flexible trade-offs between communication reliability and ranging accuracy in short-block transmissions. The proposed scheme is built upon a practical multiple-input multiple-output orthogonal frequency-division multiplexing (MIMO-OFDM) architecture, where the communication channel state information and the angles of the static targets are assumed available at the transmitter. A transformer-based transmitter encodes input information bits directly into ISAC transmit waveforms to jointly optimize the bit error rate (BER) performance and the delay modified Cramer-Rao bound (MCRB). A corresponding transformer-based receiver is adopted at the communication side to recover the transmitted information bits. We further examine the learned codewords for communication-oriented and sensing-oriented designs, revealing that a balanced ISAC waveform naturally exhibits an intermediate structure between these two extremes. Numerical results illustrate these codeword structures and demonstrate that the proposed design provides substantial trade-off gains over conventional schemes based on standard channel coding and modulation.
I. Introduction
ISAC must reconcile communication’s need for informative, random signals with sensing’s preference for stable, deterministic signals. The paper proposes a DL-based coded waveform framework for MIMO-OFDM and positions it beyond earlier single-antenna or simplified OFDM designs.
I. Introduction
- Communication and sensing impose conflicting waveform requirements: information-bearing randomness versus stable, deterministic signals for sensing.ISAC jointly designs waveforms to support both functions while optimizing spectrum use and reducing deployment and operational cost.Source: §I · p. 1
- Earlier learned ISAC codes interpolate between sensing-oriented PSK-like structures and communication-oriented, more dispersed constellations,.Prior work also formulated sensing loss through codeword autocorrelation sidelobes [7].Source: §I · p. 1
- The paper distinguishes its approach from prior single-antenna or simplified OFDM settings by jointly coding across subcarriers and handling heterogeneous inputs.The transmitter uses communication CSI and target-angle information alongside the information bits.Source: §I · pp. 1–2
- The proposed DL-based coded waveform framework targets practical MIMO-OFDM, where spatial-frequency degrees of freedom support joint communication and sensing design.The paper focuses on the post-probing stage, with target angles available at the transmitter and delay estimation as the primary sensing task.Source: §I · pp. 1–2
II. System Model
The system model describes a MIMO-OFDM link used for communication and monostatic sensing, with separate channel models for the two functions. The design focuses on the post-probing phase, when target angles are assumed known.
II. System Model
- A primary base station transmits ISAC signals through independent communication and sensing channels, with a secondary base station receiving the communication signal and the primary processing target echoes.For each subcarrier, an input bit sequence is mapped to a transmit sequence subject to a per-subcarrier power constraint.Source: §II · p. 2
- The communication and sensing receivers observe complex additive white Gaussian noise, with the corresponding noise variances specified separately.The model distinguishes resolvable communication paths from sensing targets and associates channel and target parameters with their respective paths.Source: §II · p. 2
- The paper studies the second ISAC phase, after initial probing and angle estimation, and assumes target angles are exactly known for balancing delay sensing and communication error performance.Before angles are known, the described first phase uses communication-oriented SVD precoding to produce a statistically near-omnidirectional beampattern.Source: §II · p. 3
A. Neural Network Architecture
The transmitter projects information bits, communication-channel matrices, and sensing angles into tokens, then uses a transformer to jointly model them across subcarriers. At the receiver, a neural decoder uses received-signal and channel features to recover the bits without sensing angles.
A. Neural Network Architecture
- Dedicated projections map the heterogeneous bit, channel, and angle inputs into feature tokens in a shared latent space for joint transformer processing.The encoder learns the space-frequency transmit signal from information bits, communication CSI, and target angles.Source: A. Neural Network Architecture · p. 3
- Bit sequences are projected as length-C blocks, reducing information-bit tokens from MC to M while preserving correlations across blocks through transformer processing,.Communication-channel matrices are also projected into M channel tokens after their complex entries are separated into real and imaginary components.Source: 1) Projection for Information Bits c · p. 3, 2) Projection for Communication… · p. 3
- Sensing angles are encoded as sine-cosine pairs to avoid angular discontinuities, and receive no positional embeddings because they form an unordered set of physical paths.The sine-cosine representation is consistent with the array-response model’s dependence on sin(θq).Source: 3) Projection for Sensing Angles · p. 3
- After token fusion, only the M information-bit tokens generate the transmit waveform; channel and angle tokens enrich the learned representations as auxiliary inputs.The transmitter applies power normalization to the generated sequence.Source: 3) Projection for Sensing Angles · p. 3
- The communication receiver uses a neural decoder because maximum-likelihood decoding is infeasible when sensing angles are unavailable, relying instead on received signals and communication-channel information.Its input features include channel-matched observations and channel Gram matrices.Source: 3) Projection for Sensing Angles · p. 3
B. Design of Performance Metric
The section defines BER and delay MCRB as communication and sensing performance measures, then combines their losses to control the trade-off.
B. Design of Performance Metric
- BER is evaluated from hard decisions, while training uses binary cross-entropy on soft probabilities stacked across subcarriers.The joint-loss weight λ controls the communication–sensing trade-off.Source: B. Design of Performance Metric · pp. 4–5
- The CRB lower-bounds unbiased delay-estimation error covariance, with target angles treated as known and fading coefficients as nuisance parameters.The modified CRB averages the delay Fisher information over the random nuisance parameters; its matrix is diagonal under the stated independence and zero-mean assumptions.Source: B. Design of Performance Metric · p. 4
- Training minimizes the logarithm of the average MCRB across sensing targets rather than minimizing MCRBs directly.The logarithm preserves optimization direction while compressing dynamic range, preventing exceptionally large MCRBs from dominating training and improving stability.Source: B. Design of Performance Metric · pp. 4–5
C. Training Procedure
The transmitter encoder and communication decoder are trained in alternating updates, with different objectives and more decoder cycles per epoch.
C. Training Procedure
- Training alternates between updating the transmitter encoder and the communication decoder, following a strategy similar to.The overall procedure is summarized as Algorithm 1.Source: C. Training Procedure · p. 5
- The encoder minimizes the overall loss while the decoder is fixed, whereas the decoder minimizes only communication loss with the encoder fixed.This separates the encoder's joint objective from the decoder's communication-only objective.Source: C. Training Procedure · p. 5
- Each epoch uses more decoder training cycles than encoder cycles to mitigate decoder performance lag and reduce the risk of poor local optima.The algorithm listing includes steps that fix the decoder during encoder updates and fix the encoder during decoder updates.Source: C. Training Procedure · p. 5
IV. Numerical Results
The numerical-results setup specifies channel, system, and training conditions and identifies BER comparisons across weighting factors and baselines.
IV. Numerical Results
- The setup uses C = 6, M = 12, Nt = Nr = 9, P = 5, and Q = 3, with communication-path delays uniform over [0, 5] and target angles drawn from a 3° grid spanning −60° to 60°.Both Transformers use hidden dimension 128, four attention heads, four layers, and zero dropout.Source: §IV · p. 5
- Fig. 3 presents BER comparisons across different λ values and baselines.The stated caption identifies the comparison scope but does not provide numerical outcomes.Source: Figure 3 · p. 5
- Training uses batch size 256 for 240 epochs, with one transmitter cycle and six receiver cycles per epoch and separate models for λ ∈ {0, 0.2, 0.4, 0.6, 0.8, 1}.Communication SNR is sampled uniformly from [10, 26] dB; sensing SNR is sampled from {−15, −10, −5, 0, 5, 10} dB.Source: §IV · p. 5
A. Baselines
The section describes coded communication benchmarks and a known-angle sensing-oriented waveform used as a sensing reference.
A. Baselines
- The communication baselines are convolutionally coded QPSK and 8-PSK with SVD precoding, soft-decision Viterbi decoding, and effective payload rates of 3/5 and 1/2, respectively.They use five and four strongest singular modes per subcarrier, respectively.Source: A. Baselines · p. 5
- The sensing-oriented waveform carries no information bits and is evaluated only as a sensing reference.Its construction uses known sensing angles.Source: A. Baselines · pp. 5–6
- Fig. 4 compares MCRB across different λ values and baselines.The caption specifies comparison scope, not numerical outcomes.Source: Figure 4 · p. 6
- The sensing-oriented waveform concentrates transmit energy in the angular subspace spanned by the target steering vectors, motivated by the delay FIM expression.It is formed from the Q dominant eigenvectors of the angular covariance matrix, with beam coefficients varying across subcarriers.Source: A. Baselines · p. 6
B. Comparisons
The learned designs span BER–MCRB trade-offs: λ = 0.4 is an appealing balance, while other weighting values favor different sensing and communication outcomes.
B. Comparisons
- Increasing λ generally degrades BER, while λ = 0 outperforms CC-8PSK over most evaluated SNRs and λ = 0, 0.2, and 0.4 outperform CC-QPSK.CC-8PSK's use of four stronger modes gives it better communication performance than CC-QPSK, which uses an additional weaker fifth mode.Source: B. Comparisons · p. 6
- Communication-oriented waveforms generally have relatively large delay MCRB because their SVD precoding concentrates energy on communication eigenmodes that may not align with target directions.CC-QPSK nevertheless has lower MCRB than CC-8PSK; its additional spatial mode may couple effectively to sensing targets despite weak communication gain.Source: B. Comparisons · p. 6
- λ = 0.4 achieves BER close to CC-QPSK and delay MCRB approaching the sensing-oriented baseline.All learned designs have lower MCRB than both communication benchmarks; λ = 0.8 and 1 also outperform the sensing-oriented baseline.Source: B. Comparisons · p. 6
C. Analysis and Insights
The learned codeword structure changes with the communication–sensing weighting: communication-oriented waveforms spread energy, sensing-oriented waveforms concentrate it, and λ = 0.4 is intermediate.
C. Analysis and Insights
- The visualization compares 50 randomly sampled codewords using the same communication channel and target angles across λ values.The target angles are −51°, −48°, and 6° for the angular power comparison.Source: C. Analysis and Insights · p. 6
- At λ = 0, codewords are dispersed within delay bins; at λ = 1, energy concentrates in the first bin, forming an impulse-like waveform.The sensing-oriented coefficients remain nearly unchanged across codewords because sensing alone does not require large codeword distances for reliable message discrimination.Source: C. Analysis and Insights · p. 6
- In the angle domain, λ = 1 concentrates power near target directions, λ = 0 distributes it more evenly, and λ = 0.4 is intermediate.The λ = 0.4 angular distribution is shaped by both communication and sensing objectives.Source: C. Analysis and Insights · p. 6
V. Conclusion
The proposed framework yields an attractive communication–sensing trade-off, with learned codewords spanning communication-oriented and sensing-oriented structures. The results identify a balanced operating point close to both specialized baselines.
V. Conclusion
- The learned coded waveforms significantly outperform conventional CC-QPSK baselines in MCRB without compromising BER.An operating point is close to both communication-oriented and sensing-oriented baselines simultaneously.Source: §V · pp. 6–7
- Weighting the joint loss produces different trade-offs: communication-oriented codewords spread energy across delay bins, sensing-oriented codewords concentrate power near target directions, and balanced waveforms are intermediate.Figs. 5 and 6 visualize the learned codewords in the delay and angle domains, respectively.Source: §V · p. 6, Figure 5 · p. 7, Figure 6 · p. 7
- The designed neural network handles heterogeneous, high-dimensional inputs to generate transmit ISAC waveforms.The framework is evaluated in MIMO-OFDM systems.Source: §V · p. 6