End-to-end sensing integrated cooperative codebook design method and system based on mutual information estimation

Abstract

The invention discloses an end-to-end sensing integrated cooperative codebook design method and system based on mutual information estimation, and aims to solve the problems that statistical characteristics of communication and sensing signals are inconsistent, and performance indexes are difficult to optimize uniformly. According to the method, communication mutual information and perception mutual information are adopted as unified optimization indexes, and a mutual information estimator based on comparative learning is constructed to respectively obtain differentiable lower bound estimation of two types of mutual information. By setting a sensing tradeoff coefficient, the weighted sum of communication and sensing mutual information is maximized in end-to-end training, and meanwhile, power constraint and phase dispersion punishment are introduced to meet actual emission requirements. And after training, enumeration is carried out to generate a joint sensing codebook, the joint sensing codebook is issued and deployed, rapid signal transmitting and receiving processing can be realized through index table lookup in an online stage, and flexible switching and performance balancing in different service modes are supported.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and in particular to a method and system for designing an integrated sensing codebook based on mutual information estimation and end-to-end collaborative learning in a sensing integration scenario. Background Technology

[0002] Integrated communication and sensing, by sharing spectrum and RF front-end, enables the same transmitted signal to simultaneously undertake information transmission and environmental sensing tasks, significantly improving resource utilization efficiency in scenarios such as connected vehicles and industrial IoT. However, unified signal design faces the challenge of a deterministic-random tradeoff: communication prefers random signals to enhance information carrying capacity, while sensing prefers deterministic or structured signals to improve echo discernibility and parameter estimation accuracy. This conflict makes it difficult for traditional codebook designs to balance the sensing tradeoffs in different scenarios. Furthermore, existing sensing systems exhibit diverse forms and inconsistent measurement scales for communication and sensing performance indicators, lacking a unified performance evaluation and optimization framework, further increasing the difficulty of joint design. Mutual information can characterize the information transmission capacity of the communication link and the discernibility of the sensing link to environmental states under a unified information theory scale, establishing a connection between information theory and estimation theory. Therefore, it is considered an ideal unified indicator for joint sensing optimization. However, under conditions of continuous high-dimensional variables and fast-fading channels, mutual information typically lacks a usable closed-form expression, making direct calculation and gradient-based optimization difficult. Although neural mutual information estimation methods can be made differentiable through variational lower bounds, existing estimators still face problems such as large gradient variance and insufficient training stability in practical applications. For these reasons, there is an urgent need for a method that can stably estimate communication mutual information and sensing mutual information in fast fading environments, and thereby enable the design and deployment of an integrated communication and sensing codebook. Summary of the Invention

[0003] This invention provides an end-to-end integrated sensing cooperative codebook design method and system based on mutual information estimation. Addressing the problems of significant conflicts between communication and sensing objectives, difficulty in measuring performance indicators on a unified scale, and challenges in end-to-end joint optimization in existing integrated sensing signal designs, this invention introduces mutual information as a unified performance indicator to simultaneously characterize the information carrying capacity of the communication link and the identifiability of the sensing link. To solve the problem of the difficulty in directly calculating mutual information under continuous high-dimensional variables and fast-fading channel conditions, this invention constructs differentiable lower bound estimates for both communication and sensing mutual information, and uses these estimates as a unified optimization basis to drive the integrated sensing codebook design. Specifically, this invention introduces a mutual information estimator, which achieves low-variance and stable lower bound estimates of mutual information through a comparative learning method using positive and negative samples, enabling the system to obtain reliable optimization feedback even in complex channel environments such as fast fading. Simultaneously, an integrated sensing encoder is constructed to generate unified sensing codewords, and under power and phase structure constraints, the weighted sum of the estimated values ​​of communication and sensing mutual information is maximized, thereby achieving adaptive codeword generation under different sensing trade-offs. Furthermore, this invention employs a collaborative alternating training mechanism between a mutual information estimator and an integrated sensing encoder, improving overall training convergence and robustness while ensuring the stability of mutual information estimation. After training, this invention enumerates the learned codewords and generates a joint sensing codebook for deployment. This allows for the selection of codebooks with different sensing trade-off points according to business needs during the online phase, and the transmission and reception of unified sensing signals are processed through an index lookup table, thereby achieving a flexible balance between communication and sensing performance. This invention overcomes the limitations of traditional analytical mutual information or fixed waveform design methods, which struggle to adapt to time-varying channel environments and multi-objective sensing trade-offs, thus improving the deployability and generalization capability of the integrated sensing system.

[0004] This invention provides an end-to-end integrated sensing cooperative codebook design method and system based on mutual information estimation, comprising the following steps:

[0005] Step 1: Obtain the length as The information bit sequence is generated by the integrated sensor encoder with a length of [length missing]. The complex digital word vectors are obtained, and the communication received signal and the sensing echo signal are obtained through the fast fading communication channel and the fast fading sensing channel, respectively. The communication mutual information and the sensing mutual information are constructed as a unified optimization index.

[0006] Step 2: Construct a communication mutual information estimator and a sensing mutual information estimator, and make them output variational lower bound estimates of communication mutual information and sensing mutual information respectively based on positive and negative sample comparison learning. Establish a training loss function for the integrated synesthesia encoder, so that the encoder maximizes the weighted sum of the communication mutual information estimate and the sensing mutual information estimate under power and phase constraints.

[0007] Step 3: Employ end-to-end collaborative alternating training: While freezing the parameters of the integrated synesthetic encoder, train the communication mutual information estimator and the perception mutual information estimator to obtain a stable mutual information estimate; while freezing the parameters of the communication mutual information estimator and the perception mutual information estimator, train the integrated synesthetic encoder to minimize the training loss function based on the synesthetic mutual information weighting, and iterate alternately until convergence;

[0008] Step 4: Based on the converged integrated sensor encoder, enumerate all information sequences to generate a joint sensor codebook and distribute it for deployment; during online operation, the transmitter selects the codeword to transmit according to the index corresponding to the bit to be transmitted, the communication end completes the detection and decoding based on the communication received signal, and the sensing end performs parameter estimation based on the sensing echo signal under the condition of known integrated sensor signal.

[0009] Preferably, step 1 specifically includes:

[0010] Step 101: Obtain the length as bit sequence The bit sequence is mapped to a sequence index. ,in Belongs to set ;

[0011] Step 102: Index the sequence Input inductive encoder The output length is Complex digital vectors The codeword vector has both communication and sensing functions;

[0012] Step 103: Define the fast fading channel coefficients in the downlink communication link. In addition to noise, establish a user received signal model. ;

[0013] Step 104: Define the equivalent sensing channel coefficients in the echo sensing link. To eliminate noise, a sensing echo model is established after the sensing signal is reflected by the target. ;

[0014] Step 105: Based on the communication signals and sensing signals described in steps 103 and 104, using mutual information as a unified performance index, give the specific forms of communication mutual information and sensing mutual information.

[0015] Preferably, step 2 specifically includes:

[0016] Step 201: Based on the communication and sensing channel model described in Step 1, construct a communication mutual information estimator with parameters based on contrastive learning. With perceptual mutual information estimator The input to the communication mutual information estimator is The input to the perceptual mutual information estimator is And establish output scoring functions respectively. and ;

[0017] Step 202: For sample sizes of Communication sample set Construct positive and negative samples, where the positive samples are... Negative samples are fixed Different instances within the same batch of samples Mismatch ;

[0018] Step 203: Train the communication mutual information estimator using contrastive loss. By updating parameters through comparison of positive and negative samples within a batch, the lower bound estimate of the communication mutual information is adjusted based on the sample size. Loss compared to communication To be determined jointly;

[0019] Step 204: For sample sizes of Perceptual sample set Construct positive and negative samples, where the positive samples are... Negative samples are fixed Different instances within the same batch of samples Mismatch ;

[0020] Step 205: Train the perceptual mutual information estimator using contrastive loss. By updating parameters through comparison of positive and negative samples within a batch, the lower bound estimate of the perceptual mutual information is adjusted based on the sample size. Loss compared to communication To be determined jointly;

[0021] Step 206: Establish a training loss function for the integrated sensory encoder, which is used to jointly optimize codewords under the sensory trade-off coefficient and physical constraints. The training loss function maximizes the weighted sum of communication sensing mutual information after satisfying the power and phase penalty constraints of the generated codewords.

[0022] Preferably, step 3 specifically includes:

[0023] Step 301: Initialize the parameters of the integrated sensor encoder, the communication mutual information estimator, and the sensing mutual information estimator, and set the sample size. Learning rate, alternating update steps ;

[0024] Step 302: Freeze the parameters of the integrated inductive encoder ,implement The estimator update loop includes generating codewords from the current integrated sensor encoder, obtaining observation data of the sensor channel and signal through the communication link and the sensing link respectively, constructing positive and negative samples according to the schemes described in steps 202 and 204, minimizing the loss function, and obtaining the estimated value of the sensor mutual information.

[0025] Step 303: Freeze the parameters of the communication mutual information estimator and the perception mutual information estimator, minimize the training loss function described in step 206, and update the parameters of the synesthetic encoder.

[0026] Step 304: Repeat steps 302 and 303 until the convergence condition is met. The convergence condition is that the change of the loss function or mutual information estimate in several consecutive iterations is less than a preset threshold, or the maximum number of iterations is reached.

[0027] Preferably, step 4 specifically includes:

[0028] Step 401: After training converges, enumerate all sequence indices. Each sequence index is input into the integrated sensor encoder to generate the corresponding codeword. And form a joint synesthesia codebook ;

[0029] Step 402: Transfer the codebook It is transmitted to both the transmitter and receiver, and the inductive tradeoff coefficient can be changed ( Train or export codebook families with different synesthetic working points for business switching;

[0030] Step 403: In the online transmission phase, the transmitter sends the bit sequence to be transmitted. Mapped to sequence index and from the codebook Select the corresponding codeword Launch;

[0031] Step 404: In the online signal reception stage, the communication end analyzes the codebook based on the received communication signal. The detection and decoding are performed, and the sensing end outputs the estimated target parameters based on the echo signal under the condition of known integrated sensing signal.

[0032] The present invention also provides a sensor-integrated collaborative codebook design system for implementing the method, comprising: a control unit, a transmitter, a communication receiver, and a sensing receiver;

[0033] The control unit, deployed during the training phase, includes:

[0034] The integrated sensor encoder module is used to map information indexes into complex digital characters;

[0035] The channel simulation module is used to simulate communication links and sensing links, and generate received signals;

[0036] The mutual information estimator module includes a communication mutual information estimation submodule and a perceptual mutual information estimation submodule based on contrastive learning, which are used to provide differentiable performance evaluation;

[0037] The training engine module is used to execute the alternating training mechanism and coordinate the optimization process of the encoder module and the estimator module.

[0038] The codebook generation and deployment module is used to generate a joint synesthesia codebook after training is completed and to distribute it.

[0039] The transmitting end stores the joint sensing codebook and, during online operation, retrieves and transmits the corresponding codeword based on the input information bit sequence.

[0040] The communication receiver stores the joint sensing codebook and uses the codebook to detect and decode the received signals during online operation.

[0041] The sensing receiver stores the joint sensing codebook, and uses this codebook as prior knowledge to estimate the target parameters of the received echo signal during online operation.

[0042] Furthermore, the communication mutual information estimation submodule and the perception mutual information estimation submodule in the mutual information estimator module are structured as deep neural networks, and the input is a real number vector formed by concatenating the real and imaginary parts of the complex signal after decomposition.

[0043] Furthermore, the training engine module controls the optimization objective of the training loss function by configuring different synesthesia tradeoff coefficients, thereby driving the synesthesia integrated encoder module to generate codebooks with different characteristics, which are then organized into a switchable codebook family by the codebook generation and deployment module.

[0044] Furthermore, the transmitting end is equipped with multiple antennas, all of which transmit the same complex symbol determined by the joint sensing codebook at the same time; the sensing receiving end is equipped with multiple receiving antennas for receiving the echo signal reflected by the target.

[0045] Beneficial effects: The end-to-end integrated sensing cooperative codebook design method and system based on mutual information estimation in the embodiments of the present invention address the technical challenges in unified symbol design, such as the trade-off between communication and sensing objectives, the difficulty in directly solving mutual information, and the instability of the learning process. It forms a comprehensive solution that is trainable, deployable, and adjustable. This scheme first characterizes the mutual information performance of the communication and sensing links simultaneously within a unified information metric framework, avoiding inconsistencies and trade-offs caused by separate optimization of communication and sensing. Second, it introduces a mutual information estimator based on contrastive learning, obtaining low-variance, stable variational lower bound estimates of communication and sensing mutual information through batch positive and negative sample construction and contrastive loss, providing continuously differentiable feedback metrics for end-to-end optimization. Simultaneously, it introduces power and phase structure constraints to the integrated sensory encoder, allowing for rapid derivation of integrated sensory codebooks for different needs by adjusting the sensory trade-off coefficients. Finally, it employs a collaborative alternation mechanism between the mutual information estimator and the encoder, enabling mutual promotion between mutual information estimation and codeword generation module training. This achieves the learning of a joint integrated sensory codebook in fast-fading channels and time-varying target environments, ensuring the output codebook meets the joint objectives of communication and sensing, supporting service mode switching and scenario migration. This scheme overcomes the dependence of traditional analytical codebook design on ideal models and static environments, improving the feasibility and overall performance of the integrated sensory system in complex dynamic scenarios. Attached Figure Description

[0046] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0047] Figure 1 A flowchart illustrating an end-to-end integrated synesthesia cooperative codebook design method and system based on mutual information estimation according to an embodiment of the present invention;

[0048] Figure 2 This is a comparison diagram of communication mutual information results according to an embodiment of the present invention, used to compare the communication mutual information performance of the proposed method and the baseline method under different synesthesia trade-off coefficients;

[0049] Figure 3 This diagram illustrates the comparison results of perceptual mutual information according to an embodiment of the present invention, used to compare the perceptual mutual information performance of the proposed method and the baseline method under different synesthesia tradeoff coefficients. Detailed Implementation

[0050] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0051] Based on the problems raised in the background, existing integrated sensing symbol design exhibits a clear deterministic-stochastic trade-off between communication and sensing performance, and the dimensions of related performance indicators are inconsistent, making joint measurement and optimization on the same scale difficult. Furthermore, mutual information, as a unified metric for sensing, is difficult to directly calculate under continuous high-dimensional conditions, resulting in a lack of stable and differentiable training feedback during end-to-end joint codebook optimization. This makes the training process prone to non-convergence or getting trapped in suboptimal sensing trade-off solutions. Therefore, this invention proposes an end-to-end integrated sensing cooperative codebook design method and system based on mutual information estimation. By constructing a contrastive learning mechanism of positive and negative samples, it stably estimates the variational lower bounds of communication mutual information and sensing mutual information, and drives the integrated sensing encoder to learn and generate a joint codebook that satisfies physical emission constraints and sensing trade-off requirements, thereby solving the aforementioned technical problems.

[0052] Figure 1 This is a flowchart illustrating an end-to-end integrated sensing cooperative codebook design method based on mutual information estimation, according to an embodiment of the present invention.

[0053] like Figure 1 As shown, the combined codebook design scheme includes the following steps:

[0054] Step 1: Establish a unified sensing signal generation and reception model to describe the signal model under fast fading communication and sensing channels, and define communication mutual information and sensing mutual information as unified optimization indicators.

[0055] In one embodiment of the present invention, step 1 specifically includes:

[0056] Step 101: Establish the overall architecture of the integrated sensing system, which includes a transmitter, a communication receiver, and a sensing receiver; wherein, the transmitter is configured with... One transmitting antenna is used to generate and transmit a unified sensing signal. The communication receiving end is a single-antenna user terminal used to complete information reception and recovery. The sensing receiving end is equipped with... A receiving antenna is used to receive the echo. The transmitter and sensing receiver are connected via a backhaul link, enabling echo processing and parameter estimation based on shared information. Training and deployment can be completed offline by the control unit: end-to-end collaborative training of the encoder and mutual information estimator is performed by synthesizing or acquiring samples. After training, the codebook or model parameters are distributed for online operation.

[0057] Step 102, obtain the length as bit sequence And input it into the integrated sensor encoder Generate unified synesthesia transmit codeword vector In this embodiment, the bit sequence Corresponding to a set of sequences Sequence index in and will As The equivalent input, the mapping satisfies and This is a one-to-one mapping. The encoder output is defined as... .in Indicates the first The re-emission symbol at each symbol time, This represents the dimension of a unified synsensory symbol codeword, and all antennas reuse this transmitted signal. To meet physical transmission constraints, [the following can be done]: Apply an average power constraint, i.e., set the power constraint for a single codeword as follows: The aforementioned As a signal that integrates communication and sensing, it is used for both communication information transmission and sensing detection. Therefore, its statistical characteristics will directly affect the information carrying capacity of the communication link and the parameter estimation capability of the sensing link.

[0058] Step 103: In the downlink communication link, the transmitter adopts a unified symbol transmission mechanism, that is, indexing any symbol... Each antenna transmits the same complex symbol. Let the first antenna transmit the second complex symbol. The fast fading channel vector from the transmitter to the single-antenna communication user at each symbol time is: Additive white Gaussian noise is Then the user receives the signal as

[0059]

[0060] Among them, in the first The equivalent communication channel at each symbol time is represented as follows: ,in For length is A column vector of all 1s. The communication receive vector is obtained by stacking the symbols at each time point. Thus obtain

[0061]

[0062] Among them, the definition and diagonalize to obtain , .

[0063] Step 104: In the echo sensing link, the sensing receiver receives the sensing echo signal. Let the first... The dual-base station sensing response matrix at each symbol time is: It is determined by the target state parameters, and the sensing receiver uses the receiving vector. In this embodiment, a synesthetic symbol is considered. At the sensing receiver, it is assumed to be known to support coherent sensing processing and subsequent parameter estimation. Let the noise vector received at the sensing end be... The sensed signal after receiving vector combining is:

[0064]

[0065] Wherein, the equivalent sensing channel coefficient is The combined equivalent noise is ,Will The sensing and receiving vector is obtained by stacking the symbols at each time step. Thus obtain

[0066]

[0067] Among them, the definition and diagonalize to obtain , .

[0068] Step 105: To evaluate the information carrying capacity of the communication link, the communication mutual information is defined as... The communication mutual information is used to measure the state of a given communication channel. Under the condition of transmitting codeword With communication received signals The amount of information shared between them is used as a unified indicator of communication performance. More specifically, communication mutual information can be expressed as the expected form of the log-likelihood ratio.

[0069]

[0070] Step 106: To characterize the sensing link's ability to estimate sensing parameters, the sensing mutual information is defined as... The perceptual mutual information is used to measure the perceptual integration signal given a given synesthetic signal. Under these conditions, sense echo signal Includes information about the sensing channel state The amount of information, and used as a unified indicator of perception performance. Perceptual mutual information can be expressed as...

[0071]

[0072] In subsequent steps of this invention, the mutual information will be estimated using a differentiable variational lower bound based on a contrastive learning mutual information estimator, in order to drive end-to-end learning of the synesthetic codeword.

[0073] Step 2 involves constructing a communication and sensing mutual information estimator to generate a stable and differentiable variational lower bound estimate of the mutual information; simultaneously, a synesthetic encoder is constructed, and a joint synesthetic training loss function with physical constraints is established. Specifically, this includes:

[0074] Step 201, Construct a communication mutual information estimator and perceptual mutual information estimator They are used respectively for communication mutual information and perceptual mutual information A differentiable variational lower bound estimate is provided. The communication mutual information estimator... The input is Perceptual mutual information estimator Input By optimizing the parameters of the communication mutual information estimator and the perception mutual information estimator respectively. and Output the scoring functions for both. and This is used to measure the degree of matching between input samples and the joint distribution. The scoring function... and It can be implemented using a multilayer perceptron, residual network, or other differentiable functions. To facilitate the processing of complex vectors by neural networks, in a preferred embodiment, the complex vector is expanded by its real and imaginary parts and concatenated to form a real vector. , , , , Then, the variables are concatenated to obtain the estimator input vector. and .

[0075] Step 202: Construct a training sample set during the training phase to achieve mutual information estimation. Let the sample size of a batch be... , No. Each sample corresponds to a bit sequence to be transmitted. or index via encoder Generate codewords And generate channel state under the communication link. and received signals This results in a batch of communication sample sets. Within this batch, construct the following set of positive and negative samples. The positive samples are the joint samples, indexed... The data constituted as positive samples It approximately follows a joint distribution. Negative samples, also known as mismatched samples, are found in each batch of communication samples. Keeping the combination unchanged, the synesthetic codewords of other sample instances within the same batch will be used. To construct a negative sample, the data is mismatched. Its approximate product follows the marginal distribution. In the specific implementation, to efficiently generate all negative samples, one can... Obtained by random permutation along the sample dimension and index each sample use and Form negative samples.

[0076] Step 203: Based on the positive and negative samples constructed in Step 202, define a contrastive scoring function. Define the loss function as the negative of the scoring function, i.e.

[0077]

[0078] Among them, the numerator corresponds to the positive sample. The scores, with the denominator corresponding to the same batch of samples. All code words The scoring system causes the model to tend to increase the scores of positive samples and decrease the scores of negative samples, with the temperature coefficient being a factor. The softmax strength is used to adjust the contrastive learning algorithm. During training, this is achieved by minimizing the loss function. Update the parameters of the communication estimator ,Right now ,in Let be the learning rate. Finally, the differentiable variational lower bound estimate of the communication mutual information can be obtained from this loss function.

[0079]

[0080] Step 204: Construct a training sample set during the training phase to achieve perceptual mutual information estimation. For the... Each sample corresponds to a bit sequence or index via encoder Generate codewords And generate the sensing channel state under the sensing link. With echo received signal This results in a batch of perceptual sample sets. Within this batch, the following positive and negative sample sets are constructed. The positive samples are the joint samples, defined by the index. The data constituted as positive samples It approximately follows a joint distribution. Negative samples are mismatched samples, which are present in each batch of perceived samples. The combination remains unchanged, and the perception channels of other instances within the same batch are changed. A mismatch with the sample is used to construct a negative sample. It approximately follows the product of marginal distributions. In practical implementation, to efficiently generate negative samples, one can... Obtained by random permutation at the batch dimension and index each sample use and Form negative samples.

[0081] Step 205: Based on the positive and negative samples constructed in Step 204, define the contrast scoring function. Define the loss function as the negative of the scoring function, i.e.

[0082]

[0083] Among them, the numerator corresponds to the positive sample. The scores, with the denominator corresponding to the same batch of samples. All sensing channels The scoring system causes the model to tend to increase the scores of positive samples and decrease the scores of negative samples, with the temperature coefficient being a factor. The softmax strength is used to adjust the contrastive learning algorithm. During training, this is achieved by minimizing the loss function. Update the parameters of the perception estimator ,Right now ,in Let be the learning rate. Finally, the differentiable variational lower bound estimate of the communication mutual information can be obtained from this loss function.

[0084]

[0085] Step 206: Establish an integrated inductive encoder Bit sequence Mapped to a length of Complex digital vectors To achieve an adjustable trade-off between communication and sensing performance, and to ensure that the generated codewords meet engineering reproducibility and structural constraints, a training loss function for the encoder is defined to jointly optimize the codewords under the influence of the sensing trade-off coefficients and physical constraints. The training loss function is defined as follows:

[0086]

[0087] Among them, the mutual information weights of the synesthesia are defined as follows: and And satisfy . The lower bound estimate of the communication mutual information obtained by the method described in steps 202-203 is... This is the lower bound estimate of the perceptual mutual information obtained by the method described in steps 204-205. Power penalty term. The average transmit power is constrained to not exceed Its definition ,in, This is the power penalty coefficient. Phase penalty term. Used to enhance codeword phase dispersion, it is defined as follows: ,in This is the phase penalty coefficient. During training, the encoder parameters are adjusted. Perform gradient updates; the update rule can be written as follows: ,in This represents the encoder learning rate.

[0088] Step 3 employs a collaborative alternating training mechanism. First, the encoder is frozen while training the mutual information estimator to stabilize the estimation. Then, the estimator is frozen while training the encoder to minimize the joint loss. This process is iterated until convergence. Under this mechanism, the mutual information estimator provides a differentiable variational lower bound of the mutual information as a training signal, and the encoder learns to generate a better set of joint synesthetic codewords under the combined influence of this training signal and physical constraints. Specifically, this includes:

[0089] Step 301: Initialize the integrated sensor encoder described in step 206 parameter The communication mutual information estimator described in step 201 parameter and perceptual mutual information estimator parameter In a preferred embodiment, an initialization method based on a Gaussian distribution is used. The sample batch size is then set. Communication estimator learning rate Perceptual estimator learning rate Learning rate of the integrated sensor encoder Set temperature coefficient and and synesthesia tradeoff coefficient , Set power penalty coefficient With phase penalty coefficient and power threshold Set to alternate step count updates This is used to control the number of times the mutual information estimator is updated before each encoder update. To ensure training reproducibility, the random seed can be fixed and the above hyperparameter configuration can be recorded.

[0090] Step 302: Freeze the parameters of the integrated inductive encoder Performing a synesthetic mutual information estimator Training rounds. In each update, the random sampling batch size is... bit sequence Codewords are generated by a fixed encoder. Communication samples were obtained based on steps 103 and 104, respectively. With perceived samples A training dataset for communication and perception is constructed. Then, based on the positive and negative sample construction and contrast loss form described in steps 202-205, the loss functions of the synesthetic mutual information estimator described in steps 203 and 205 are calculated respectively. and The lower bound estimate of the mutual information of the current batch is calculated using the loss function and the number of samples in the current batch, and used for training and monitoring. Finally, the estimator parameters are updated through backpropagation. and .

[0091] Step 303: Freeze the parameters of the communication mutual information estimator With perceptual mutual information estimator parameters Train the synesthetic encoder. The batch size for resampling samples is [missing information]. bit sequence And generate codewords Communication samples were obtained based on steps 103 and 104, respectively. With perceived samples A training dataset for communication and perception is constructed, and based on the mutual information estimation scheme described in steps 203 and 205, the sample input is frozen. and Obtain the variational lower bound estimate of the communication-sensing mutual information. and As one of the encoder training metrics, the encoder training loss is then constructed based on the method described in step 206. , and according to Update parameters .

[0092] Step 304: Repeat the training method described in steps 302 and 303 to form a collaborative alternating training process of the synesthetic mutual information estimator and the synesthetic integrated encoder until the preset convergence condition is met or the maximum number of iterations is reached. The convergence condition can be set to any one or a combination of the following: the loss function value of the synesthetic encoder continuously... The change in each update round is less than the threshold. Or, the estimated value of communication and sensing mutual information in continuous The improvement in the round is less than the threshold. After training, save the final network parameters, such as the parameters of the synesthetic encoder. Communication mutual information estimator parameters and perceptual mutual information estimator The parameters are used for subsequent analysis and online codebook deployment; in actual deployment, only encoder parameters are usually needed.

[0093] Step 4: Based on the converged synesthetic encoder trained through collaborative alternation, a joint synesthetic codebook covering all message indices is enumerated and deployed. During the online operation phase, the transmitter selects the corresponding codeword from the codebook according to the message index for transmission. Specifically, this includes:

[0094] Step 401: After the collaborative alternating training in Step 3 meets the convergence condition, the synesthetic encoder is used. final parameters information set Perform mapping. Input the encoder to obtain the corresponding codeword It performs emission constraint processing consistent with the training phase to ensure that codewords satisfy reproducibility and consistency, thereby forming a joint synesthetic codebook. In a preferred embodiment, the codebook is organized into a lookup table by index, and the mapping relationship between the index and the bit sequence is retained for online retrieval. The enumeration generation process ensures that the codebook covers all sequence indices, so that codewords can be obtained directly through index retrieval without neural network inference in the online stage, thereby reducing the running complexity and improving latency determinism.

[0095] Step 402: Transfer the codebook Deployed at both the transmitting and sensing receiving ends. Furthermore, to adapt to different business needs and scenario preferences, the sensing trade-off coefficient can be adjusted. Repeat steps 2 and 3 to derive codebook families for different synesthetic operating points. Used for service switching; during online switching, only the current codebook number needs to be indicated to complete the codebook selection. During this distribution process, the transmitting and receiving ends can synchronously store the codebook index range, codebook number, and its corresponding weighting coefficient configuration, so that subsequent service switching only involves codebook selection without affecting the consistency of the index mapping rules.

[0096] Step 403: In the online transmission phase, the transmitter sends the bit sequence to be transmitted. Converted into a sequence index according to a pre-agreed mapping rule. When the system enables a codebook family, the transmitter first selects a codebook number based on business requirements and determines the currently used codebook. Then from the codebook Select the corresponding codeword The signal is transmitted as a unified synesthetic signal, and necessary identification information is sent along with the frame structure for consistent processing at the receiving end. This online process allows the transmitting end to generate the transmitted signal simply by performing index calculations and codeword lookups, thus solidifying the synesthetic tradeoff capabilities acquired during the training phase into a codebook for online operation.

[0097] Step 404: In the online receiving phase, the communication receiver and the sensing receiver respectively receive the corresponding link signals. and and in the shared codebook The integrated sensing and communication processing is completed when relevant prior information is available. The communication end outputs the estimation result of the transmitted bit sequence, and the sensing end outputs the parameter estimation result related to the target state.

[0098] The following specific embodiment illustrates the performance of the integrated synesthetic codebook design method and system based on contrastive learning variational mutual information estimation proposed in this invention.

[0099] The simulation uses a fast-fading Rayleigh channel to verify the performance of the proposed scheme. The length of the bits to be transmitted is taken as... Code word length Corresponding bitrate The channel power gain for communication and sensing is normalized to 1. Training uses the Adam optimizer, with a sample batch size of [value missing]. and set To ensure that the mutual information estimator provides a relatively stable lower bound feedback on mutual information before each encoder update, the integrated sensor encoder and mutual information estimator are both implemented using a multilayer perceptron. It consists of an input layer, an embedding layer, several fully connected hidden layers, and an output layer. The width of the hidden layers is taken as... The hidden layer activation function uses ELU, and the embedding dimension is taken as... The output layer uses a linear mapping; the mutual information estimator consists of an input layer, several fully connected hidden layers, and an output layer, with the hidden unit size taking [value missing]. The hidden layer activation function is Leaky ReLU. To enhance training stability, Gaussian noise is introduced after the embedding layer of the synesthetic encoder, with a noise variance of [value missing]. The transmit power constraint adopts a soft penalty approach, with the power upper limit set to... And set the penalty coefficient to and The learning rate employs a segmented strategy, with the encoder learning rate set to... The mutual information estimator learning rate is taken Regarding the trade-off coefficients, by changing... To achieve training and comparative analysis of different synesthetic working points.

[0100] Figure 2 and Figure 3 The mutual information performance of the proposed scheme, the communication baseline scheme, and the sensing baseline scheme are compared at different transmit powers. Figure 2 For the comparison of communication mutual information performance, Figure 3This study compares the performance of mutual information sensing. The comparison baselines include a communication-centric baseline coding and modulation scheme (convolutional code + QPSK) and a sensing-centric baseline scheme (normal-mode waveforms), along with theoretical upper bound reference curves to calibrate the performance gap. Figure 2 It is evident that as the transmission power increases, the communication mutual information of each scheme also improves; At that time, the communication-optimal codebook learned by this invention can closely approximate the upper bound of communication mutual information, demonstrating the effective approximation capability of end-to-end learning to the optimal performance of the communication link. While the convolutional code + QPSK scheme can also improve communication mutual information with increasing power, its overall performance is still slightly lower than the upper bound of communication mutual information, indicating that traditional schemes have shortcomings under this unified optimization metric. from Gradually reduce to , Until At this time, the mutual information exchanged between systems shows a decreasing trend, reflecting that the system is gradually shifting its optimization focus from communication to perception, thus forming a controllable transfer of communication performance. Figure 3 As can be seen, the present invention is in Right now Under optimal perceptual settings, its perceptual mutual information curve remains consistent with or very close to the baseline of the norm-modulated waveform and the upper bound of the perceptual mutual information, indicating that the learned codeword structure can achieve the high mutual information level expected by the perceptual side; while the traditional cumulative code + QPSK scheme achieves lower perceptual mutual information. As the transmit power increases, the perceived mutual information still improves, but it gradually decreases relative to the optimal perception curve, reflecting the loss of perceived mutual information caused by communication enhancement. It is worth noting that... and At the intermediate operating point, the perceptual mutual information can be maintained at a high level, while the communication mutual information is also significantly better than that of pure perception-oriented systems. The situation illustrates that the present invention can achieve a compromise region where communication and sensing can benefit simultaneously by weighing coefficients; the comparison results jointly verify the effectiveness and adjustability of the joint codebook design of the present invention from two dimensions: communication mutual information and sensing mutual information.

[0101] In summary, this invention addresses the technical challenge of simultaneously balancing communication information carrying capacity and sensory discriminability in the design of joint codebooks for integrated sensing systems. It proposes an end-to-end integrated sensing joint codebook optimization method and system based on mutual information estimation. This invention uses communication mutual information and sensory mutual information as unified performance metrics to construct mutual information estimators for both communication and sensing, providing a stable and differentiable variational lower bound for mutual information for the encoder. Through a cooperative alternating training mechanism, a closed-loop optimization is formed between the stable mutual information metric of the estimator and the encoder's adaptive learning of the codeword structure, thereby achieving codebook generation and online deployment under different sensing trade-offs. Simulation results show that this invention can learn a codebook design covering communication-optimal to sensing-optimal levels and maintains a stable performance trend under different transmit powers. Furthermore, compared to traditional baseline schemes with communication and sensing centers, this invention can approximate the upper bound of mutual information on the corresponding side targets and achieves a flexible trade-off between communication and sensing performance at intermediate operating points. Therefore, this invention provides an implementable, switchable, and scalable joint codebook design approach for next-generation integrated sensing systems, possessing significant engineering application value.

[0102] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0103] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature.

[0104] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

Claims

1. An end-to-end integrated sensing cooperative codebook design method based on mutual information estimation, characterized in that, Includes the following steps: Step 1: Map the information bit sequence to be transmitted to a sequence index, input it into the integrated encoder, and generate a complex digital word vector that is used for both communication and sensing; the complex digital word vector is passed through the communication link model and the sensing link model respectively to obtain the simulated communication received signal and sensing echo signal; communication mutual information and sensing mutual information are used as a unified performance optimization index. Step 2: Construct a communication mutual information estimator and a perception mutual information estimator, both based on a contrastive learning mechanism; using the codeword vector generated by the integrated sensory encoder and the signal output by the link model described in Step 1, construct communication training samples and perception training samples respectively, and train the two estimators to output differentiable variational lower bound estimates of communication mutual information and perception mutual information. Based on the two variational lower bound estimates, a training loss function for the integrated sensory encoder is established. This function maximizes the weighted sum of the communication mutual information estimate and the sensing mutual information estimate under power constraint penalty and phase dispersion penalty. Step 3: Optimize the system using an alternating training mechanism: First, freeze the parameters of the integrated sensory encoder, and train the communication mutual information estimator and the perception mutual information estimator using its current output and the link model results; then freeze the parameters of the two estimators, and train the integrated sensory encoder using the training loss function; repeat this alternating process until the model converges. Step 4: Based on the converged integrated sensor encoder from Step 3, generate corresponding codewords for all possible information sequence indices, form a joint sensor codebook, and deploy it to the transmitter and receiver. During online operation, the transmitter searches for the corresponding codeword in the codebook according to the information index and transmits the signal. The communication receiver uses the codebook for signal detection and decoding. The sensing receiver estimates the target parameters under the condition of knowing the transmitted codebook.

2. The method according to claim 1, characterized in that, In step 1, the communication link model is a fast fading channel model, and the sensing link model is a target reflection echo channel model; the communication mutual information is defined as the amount of information between the transmitted codeword and the received signal under given channel conditions, and the sensing mutual information is defined as the amount of information between the sensing echo signal and the target state under given transmitted signal conditions.

3. The method according to claim 1, characterized in that, In step 2, the method for constructing communication training samples is as follows: for a batch of training data, the combination of codewords, channel state and received signal actually generated by the same information index is used as positive samples; codewords from different information indices are mismatched with the current channel state and received signal to form negative samples. The method for constructing sensing training samples is as follows: For a batch of training data, the combination of codewords actually generated by the same information index, sensing channel state and echo signal is used as positive samples; the channel state from different sensing scenarios is mismatched with the current codeword and echo signal to form negative samples.

4. The method according to claim 1, characterized in that, The training loss function in step 2 is specifically the negative of the weighted sum of the communication mutual information estimate and the perception mutual information estimate, plus a power penalty term and a phase penalty term; wherein the weighting weights are configurable communication-sensory tradeoff coefficients used to adjust the priority of communication and sensing performance.

5. The method according to claim 4, characterized in that, The synesthesia tradeoff coefficient includes a communication weight coefficient and a perception weight coefficient, the sum of which is one.

6. The method according to claim 4, characterized in that, By configuring different communication-sensory tradeoff coefficients and repeatedly executing the training process of steps 2 to 3, a series of codebooks corresponding to different communication-sensory tradeoff working points are generated, forming a codebook family to support dynamic switching of online service modes.

7. A sensor-integrated collaborative codebook design system for implementing the method as described in any one of claims 1 to 6, characterized in that, include: Control unit, transmitter, communication receiver, and sensing receiver; The control unit, deployed during the training phase, includes: The integrated sensor encoder module is used to map information indexes into complex digital characters; The channel simulation module is used to simulate communication links and sensing links, and generate received signals; The mutual information estimator module includes a communication mutual information estimation submodule and a perceptual mutual information estimation submodule based on contrastive learning, which are used to provide differentiable performance evaluation; The training engine module is used to execute the alternating training mechanism and coordinate the optimization process of the encoder module and the estimator module. The codebook generation and deployment module is used to generate a joint synesthesia codebook after training is completed and to distribute it. The transmitting end stores the joint sensing codebook and, during online operation, retrieves and transmits the corresponding codeword based on the input information bit sequence. The communication receiver stores the joint sensing codebook and uses the codebook to detect and decode the received signals during online operation. The sensing receiver stores the joint sensing codebook, and uses this codebook as prior knowledge to estimate the target parameters of the received echo signal during online operation.

8. The system according to claim 7, characterized in that, The communication mutual information estimation submodule and the perception mutual information estimation submodule in the mutual information estimator module are structured as deep neural networks, and the input is a real number vector formed by concatenating the real and imaginary parts of the complex signal after decomposition.

9. The system according to claim 7, characterized in that, The training engine module controls the optimization objective of the training loss function by configuring different synesthesia tradeoff coefficients, thereby driving the synesthesia integrated encoder module to generate codebooks with different characteristics, which are then organized into a switchable codebook family by the codebook generation and deployment module.

10. The system according to claim 7, characterized in that, The transmitting end is equipped with multiple antennas, all of which transmit the same complex symbol determined by the joint sensing codebook at the same time; the sensing receiving end is equipped with multiple receiving antennas for receiving echo signals reflected from the target.

Images