A resource optimization method for an intelligent reflective surface-assisted integrated communication and sensing system

Abstract

This invention belongs to the field of wireless communication technology and discloses a resource optimization method for an integrated communication and sensing system assisted by intelligent reflectors. The method includes: obtaining the channel state set between all active intelligent reflectors and all communication users within the current time slot; inputting the channel state set into an association modeling model to obtain a user-intelligent reflector allocation probability matrix; adaptively adjusting the size of the candidate set and searching for candidate schemes from the user-intelligent reflector allocation probability matrix using a deterministic candidate strategy, a probability-guided sampling strategy, and a perturbation-enhanced exploration strategy; evaluating the performance of each candidate scheme in the candidate set, selecting the candidate scheme with the highest communication rate as the optimal allocation scheme for the current time slot, and storing it in an experience pool. This invention can improve the communication and sensing performance of the system under complex resource coupling conditions.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication technology, specifically relating to a resource optimization method for an intelligent reflective surface-assisted integrated communication and sensing system. Background Technology

[0002] In recent years, with the rapid development of wireless communication technology and smart terminals, application scenarios such as autonomous driving, augmented reality, smart security, and the Industrial Internet of Things have placed demands on wireless networks to simultaneously possess high data rates, high reliability, and high-precision sensing capabilities. Traditional cellular communication systems struggle to balance communication and sensing functions. Integrated communication and sensing technology, by simultaneously achieving information transmission and environmental perception on shared spectrum and hardware resources, provides a new technological path for building future intelligent networks.

[0003] Reconfigurable Intelligent Surfaces (RIS) improve wireless link quality and extend system coverage by programmably controlling the wireless propagation environment, achieving this at a lower cost and with lower energy consumption. To further compensate for cascaded link losses, active RIS incorporates a signal amplification module within the reflecting unit, thereby enhancing the reflected signal strength and making it more promising for medium- and long-range communication and sensing scenarios. However, the introduction of active RIS also introduces additional power consumption and hardware constraints, resulting in a more complex coupling between its parameter configuration and system performance.

[0004] In a multi-user, multi-intelligent reflector collaborative communication and sensing system, communication links and sensing links need to share spatial, power, and reflection resources. Channel conditions differ significantly between different users and different intelligent reflectors, and the interference and coupling relationships between links in the system are complex. The overall system performance largely depends on the association method between communication users and intelligent reflectors, as well as the reflection parameter configuration of the intelligent reflectors. However, these decisions are difficult to effectively determine through manual rules or static strategies in large-scale network and dynamic channel environments.

[0005] Furthermore, the base station beamforming design is highly coupled with the reflection parameter configuration of the active smart reflector. Constrained by transmit power limitations, reflector unit hardware characteristics, and sensing performance requirements, the related optimization problems exhibit significant non-convexity and multi-parameter coupling characteristics. Existing technologies mostly optimize for a single smart reflector or a single performance indicator, making it difficult to simultaneously balance communication rate and sensing performance in scenarios with multiple users and multiple smart reflectors. Meanwhile, traditional user-RIS association methods typically rely on fixed rules or exhaustive search, resulting in high computational complexity and failing to meet the real-time and adaptive requirements of large-scale networks.

[0006] Therefore, how to achieve efficient and flexible user-RIS association decision-making in a communication and sensing integrated system assisted by multiple active intelligent reflectors, and on this basis, jointly optimize base station beamforming and RIS reflection parameters, has become a key technical problem that urgently needs to be solved to improve the overall performance of the system. Summary of the Invention

[0007] To address the problems of complex user association decisions and difficulties in joint parameter optimization in existing communication and sensing integrated systems assisted by multiple active intelligent reflectors, the present invention aims to provide a resource optimization method for intelligent reflector-assisted communication and sensing integrated systems. This method is used to collaboratively optimize the association relationship between communication users and active intelligent reflectors, as well as the base station beamforming parameters and the reflection parameters of active intelligent reflectors in a multi-user, multi-active intelligent reflector collaborative communication and sensing system, thereby improving the communication and sensing performance of the system under complex resource coupling conditions.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] A resource optimization method for an integrated communication and sensing system assisted by intelligent reflective surfaces, the integrated communication and sensing system comprising a base station, multiple active intelligent reflective surfaces, and multiple communication users, includes the following steps:

[0010] Obtain the channel state set between all active smart reflectors and all communication users in the current time slot, and input the channel state set into the association modeling model. The association modeling model includes a user feature extraction network, a smart reflector feature extraction network, and a user-smart reflector interaction modeling module.

[0011] The user feature extraction network extracts features from the channel vectors corresponding to the communication users to obtain the user feature matrix.

[0012] The intelligent reflector feature extraction network extracts features from the channel vector corresponding to the active intelligent reflector to obtain the intelligent reflector feature matrix.

[0013] The user feature matrix and the intelligent reflective surface feature matrix are projected onto a unified feature space and input into the user-intelligent reflective surface interaction modeling module. The user-intelligent reflective surface relationship matrix is ​​output and normalized to obtain the user-intelligent reflective surface allocation probability matrix.

[0014] The size of the candidate set is adaptively adjusted based on the degree of difference between the current time slot's user-intelligent reflector allocation probability matrix and the optimal allocation scheme in the experience pool. A deterministic candidate strategy, a probability-guided sampling strategy, and a perturbation-enhanced exploration strategy are used to search for candidate schemes from the user-intelligent reflector allocation probability matrix.

[0015] The performance of each candidate scheme in the candidate set is evaluated, and the candidate scheme with the highest communication rate is selected as the optimal allocation scheme for the current time slot and stored in the experience pool.

[0016] Several alternative methods are provided below, but they are not intended as additional limitations on the overall solution above. They are merely further additions or optimizations. Provided there are no technical or logical contradictions, each alternative method can be combined individually with respect to the overall solution above, or multiple alternative methods can be combined with each other.

[0017] Preferably, both the user feature extraction network and the intelligent reflective surface feature extraction network adopt a multilayer perceptron structure, and the user-intelligent reflective surface interaction modeling module is an attention module based on the Transformer structure.

[0018] Preferably, the adaptive adjustment of the candidate set size based on the difference between the user-intelligent reflector allocation probability matrix of the current time slot and the optimal allocation scheme in the experience pool includes:

[0019] The degree of difference between the user-intelligent reflector allocation probability matrix of the current time slot and the optimal allocation scheme in the experience pool is calculated using cross-entropy.

[0020] The range of candidate set size is weighted by the ratio of the degree of difference to the upper bound of the difference, and the weighted value is superimposed with the minimum value of the candidate set size range to obtain the candidate set size after adaptive adjustment in the current time slot.

[0021] Preferably, the step of calculating the difference between the user-intelligent reflector allocation probability matrix of the current time slot and the optimal allocation scheme in the experience pool using cross-entropy includes:

[0022] The optimal allocation scheme in the experience pool is selected. If a communication user is allocated to an active intelligent reflector, the true label between the corresponding communication user and the active intelligent reflector is set to 1; otherwise, the true label is set to 0, thereby obtaining the true allocation probability matrix between all communication users and all active intelligent reflectors.

[0023] Based on the true allocation probability matrix and the user-intelligent reflector allocation probability matrix of the current time slot, the discrete classification cross-entropy is calculated to obtain the degree of difference.

[0024] Preferably, the deterministic candidate strategy outputs a candidate solution, and the probability-guided sampling strategy outputs... The perturbation enhancement exploration strategy outputs a candidate scheme. The candidate schemes, the The size of the candidate set.

[0025] Preferably, the deterministic candidate strategy performs the following operations:

[0026] Based on the maximum probability corresponding to each communication user in the user-intelligent reflector allocation probability matrix, the active intelligent reflector corresponding to the maximum probability is selected for the communication user, and a deterministic allocation matrix is ​​obtained as a candidate scheme.

[0027] Preferably, the probability-guided sampling strategy performs the following operations:

[0028] Based on the user-intelligent reflector allocation probability matrix, multiple random samples are taken from each communication user to generate multiple candidate schemes with differences.

[0029] Preferably, the perturbation enhancement exploration strategy performs the following operations:

[0030] A random perturbation is introduced into the user-intelligent reflector allocation probability matrix and normalized. Based on the perturbation-enhanced probability distribution, a random sampling operation is performed to generate multiple candidate schemes for perturbation enhancement.

[0031] As a preferred option, the following operations are performed during the training phase:

[0032] Several training samples are randomly selected from the experience pool. Each training sample contains a set of channel states, an optimal allocation scheme, and the corresponding communication rate.

[0033] The channel state set in the training samples is normalized and subjected to random noise perturbation to generate enhanced samples. The training samples and enhanced samples are used as training data.

[0034] For each sample in the training data, the set of channel states of the sample is input into the correlation modeling model to obtain the currently predicted user-smart reflector allocation probability matrix;

[0035] The classification loss is calculated by combining the currently predicted user-intelligent reflector assignment probability matrix with the optimal assignment scheme of the sample with the highest communication rate in the training data.

[0036] An L2 regularization term is introduced, and the parameters of the association modeling model are updated in conjunction with the classification loss.

[0037] The training process is repeated iteratively until the training termination condition is met, and the optimal correlation model is output.

[0038] This invention addresses the characteristics of high-dimensional, strongly coupled optimization problems in multi-active intelligent reflector collaborative scenarios, providing a resource optimization method for an integrated intelligent reflector-assisted communication and sensing system. First, it constructs a user-intelligent reflector association modeling method based on a deep neural network, transforming the original exponential search-based combinatorial decision-making problem into a single forward inference process, significantly reducing complexity from exponential to polynomial. Building upon this, a candidate allocation mechanism based on probability distribution is further introduced. Through deterministic selection, random sampling, and perturbation enhancement, a candidate set much smaller than the full combinatorial space is generated. This effectively compresses the search range while maintaining solution space coverage, allowing subsequent joint optimization to be performed only on a finite candidate set.

[0039] Furthermore, under the condition of fixed candidate allocation relationships, this invention decomposes the original high-dimensional joint optimization problem of multiple active RIS cooperative coupling into a hierarchical optimization structure. Specifically, the outer layer performs user-intelligent reflector association selection, while the inner layer performs cooperative optimization of the base station beamforming matrix and the reflection matrices of each active intelligent reflector for a given association relationship. This decomposition effectively reduces the coupling dimension between variables, transforming the problem, which originally needed to be solved in a global high-dimensional space, into a multiple optimization problem in a low-dimensional subspace, thereby reducing the overall complexity. During this process, the cooperative relationships between multiple active RIS are preserved through candidate selection and inner-layer joint optimization, thus avoiding significant performance degradation while reducing complexity. Attached Figure Description

[0040] Figure 1 This is a flowchart of the resource optimization method for the intelligent reflective surface-assisted communication and sensing integrated system of the present invention;

[0041] Figure 2 This is a graph showing the normalized total system rate variation curves over 8000 time blocks using different methods in the experiment of this invention.

[0042] Figure 3 This is a comparison chart of the total system rate performance under different numbers of communication users in the experiment of this invention;

[0043] Figure 4 This is a comparison chart of the system's total rate performance under different total transmit powers in the experiments of this invention;

[0044] Figure 5 This is a comparison chart of the overall system rate performance under different base station and active smart reflector distances in the experiment of this invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention.

[0047] In a multi-user, multi-active intelligent reflector-assisted communication and sensing integrated system, because the system simultaneously introduces multiple intelligent reflectors with amplification capabilities, each reflector not only needs to regulate the amplitude and phase of its own reflective unit but also needs to meet amplification power constraints. Its output signal will have a cascading effect on other links in the system, thus causing the system to exhibit significant cross-reflector coupling characteristics. Specifically, for systems containing... In a system with multiple active intelligent reflectors, the reflection matrices are interconnected through signal superposition and interference propagation, resulting in a high-dimensional, strong coupling between the base station beamforming matrix and all reflection matrices. Simultaneously, different users share spatial resources through different reflection paths, further impacting the overall interference structure and system performance through user-intelligent reflector association decisions. Therefore, compared to single or passive RIS systems, the optimization problem in multi-active RIS collaborative scenarios not only has a higher dimensionality of variables but also more complex coupling relationships, exhibiting characteristics of high dimensionality, strong coupling, and non-convexity superposition. Its solution complexity increases exponentially with the number of users and reflectors.

[0048] Traditional methods typically require joint optimization under all possible user-intelligent reflective surface association methods, resulting in a combined scale of... Furthermore, a high-dimensional optimization problem involving all active RIS reflection parameters and base station beamforming parameters is solved under each allocation method. The overall computational complexity can be expressed as: ,in This represents the computational cost of a single joint optimization. This represents the total number of active intelligent reflective surfaces. The total number of communication users, This is asymptotically an upper bound. Due to the significant coupling between multiple active RIS, the complexity of the single optimization problem itself increases with the number of reflectors, further amplifying the overall computational burden and making such methods difficult to deploy in practical systems.

[0049] To address the high complexity issue arising from multi-active RIS collaborative optimization, this invention first constructs a user-intelligent reflective surface association modeling method based on deep neural networks. This transforms the original combinatorial decision-making problem, which required exponential search, into a single forward inference process, thus reducing the complexity from exponential to high level. to polynomial complexity The problem is significantly reduced. Building upon this, a candidate allocation mechanism based on probability distribution is further introduced. Through deterministic selection, random sampling, and perturbation enhancement, a candidate set with a size much smaller than the full combinatorial space is generated. This effectively compresses the search range while ensuring solution space coverage, thus requiring subsequent joint optimization to be performed only on a finite candidate set.

[0050] Furthermore, under the condition of fixed candidate allocation relationships, this invention decomposes the original high-dimensional joint optimization problem of multiple active RIS cooperative coupling into a hierarchical optimization structure. Specifically, the outer layer performs user-intelligent reflector association selection, while the inner layer performs cooperative optimization of the base station beamforming matrix and the reflection matrices of each active intelligent reflector for a given association relationship. This decomposition effectively reduces the coupling dimension between variables, transforming the problem, which originally needed to be solved in a global high-dimensional space, into a multiple optimization problem in a low-dimensional subspace, thereby reducing the overall complexity from... Reduced to ,in , The total number of user-intelligent reflector allocation methods is determined. During this process, the collaborative relationships between multiple active RIS are preserved through candidate selection and inner-layer joint optimization, thereby reducing complexity while avoiding significant performance degradation.

[0051] Furthermore, in the inner-layer joint optimization process, considering the high-order coupling relationship between the base station beamforming matrix and the reflection matrix of multiple active intelligent reflectors, this application employs a combination of fractional programming, alternating optimization, and principal-secondary optimization to solve the problem. These methods are existing optimization techniques, and this application applies them to high-dimensional, strongly coupled optimization problems in a multi-active intelligent reflector collaborative scenario.

[0052] Specifically, by introducing auxiliary variables to equivalently reconstruct the original fractional objective function, the high-order coupled expression involving multivariate multiplication is transformed into a decomposable form. Based on this, an alternating optimization method is used to solve different variables in blocks, and a principal-secondary method is employed to construct convex upper bounds for non-convex terms. This transforms the original, difficult-to-solve non-convex high-dimensional optimization problem into a series of well-structured convex optimization subproblems. Since there are significant cooperative coupling relationships among multiple active intelligent reflectors, directly solving this type of problem has extremely high computational complexity. However, the above optimization strategy effectively reduces problem complexity while ensuring solution accuracy and improving the stability and convergence of the optimization process, thus giving the proposed resource optimization method good engineering feasibility.

[0053] In summary, this application addresses the characteristics of high-dimensional, strongly coupled optimization problems in multi-active intelligent reflector cooperative scenarios. It focuses on a continuous variable solution method combining deep learning modeling, candidate set compression, and hierarchical optimization structures to synergistically reduce system complexity from both discrete decision-making and continuous optimization perspectives. This effectively solves the key problems of combinatorial explosion and high coupling complexity in multi-RIS cooperative optimization. While significantly reducing computational complexity, it achieves near-global optimal performance, demonstrating stronger scalability, stability, and practical application value compared to existing single-RIS or simple extension methods.

[0054] In an integrated communication and sensing system, a base station with dual communication and sensing functions is deployed. The base station is equipped with multiple antennas to simultaneously transmit communication and sensing signals. Multiple active intelligent reflectors are deployed within the system. Each active intelligent reflector consists of multiple tunable reflector elements, capable of modulating the amplitude and phase of the incident signal to enhance the transmission performance of both the communication and sensing links. The system includes multiple communication users, with each active intelligent reflector covering its corresponding local area. It provides reflection-assisted links for communication users within its coverage area and reflection support for sensing targets within the corresponding area. During system operation, the base station dynamically determines the correspondence between communication users and active intelligent reflectors based on the channel state information between each active intelligent reflector and the communication user, indicating which active intelligent reflector provides the primary reflection enhancement service for each communication user.

[0055] like Figure 1 As shown in this embodiment, a resource optimization method for an intelligent reflective surface-assisted communication and sensing integrated system includes the following steps:

[0056] Step 1: Obtain the channel state set between all active smart reflectors and all communication users in the current time slot, input the channel state set into the correlation modeling model, and output the user-smart reflector relationship matrix.

[0057] During system operation, the base station first acquires a set of channel state information between all active smart reflectors and all communication users within the current time slot. This channel state information can be estimated using pilot signals. The channel state set is represented as... ,in Indicates the first The active intelligent reflector and the first Channel coefficients between individual communication users This represents the total number of active intelligent reflective surfaces. The total number of communication users. The base station obtains the set of channel states between all active smart reflectors and all communication users within the current time slot, and sends this set of channel states as input to a preset association modeling model. The association modeling model in this embodiment is a deep neural network model, including a user feature extraction network, a smart reflector feature extraction network, and a user-smart reflector interaction modeling module.

[0058] (1) User feature extraction network: The user feature extraction network adopts a multilayer perceptron structure to extract the channel vector corresponding to each communication user. Feature extraction is performed to obtain the first... The user feature vectors of each communication user are aggregated to obtain the user feature matrix.

[0059] In one embodiment, the user feature extraction network includes an input layer, at least two hidden layers, and an output layer, wherein the input layer has a dimension of [missing information]. The hidden layers are fully connected, consisting of three layers with 128, 128, and 64 neurons respectively. Interlayer connections are achieved through linear transformations and ReLU activation functions. Each hidden layer is followed by a batch normalization layer and a Dropout layer with a dropout rate of 0.2–0.5. The output layer generates the user feature vector with a feature dimension of [missing information]. This yields a user feature matrix containing the user feature vectors of all communication users. .

[0060] (2) Intelligent reflector feature extraction network: The intelligent reflector feature extraction network also adopts a multilayer perceptron structure, which is used to extract the channel vector corresponding to each active intelligent reflector. Perform feature encoding to obtain the first The feature vectors of each active intelligent reflector are aggregated to obtain the intelligent reflector feature matrix.

[0061] In one embodiment, the intelligent reflective surface feature extraction network includes an input layer, three hidden layers, and an output layer, wherein the input layer has a dimension of [missing information]. The number of neurons in the hidden layers are 128, 128, and 64, respectively. Linear transformations are used between layers, and ReLU activation functions are connected. Batch normalization layers and Dropout layers are added after the hidden layers, with a Dropout rate of 0.2–0.5. The output layer generates a smart reflective surface feature vector with a feature dimension of [missing information]. This yields the intelligent reflector feature matrix, which contains the feature vectors of all active intelligent reflectors. .

[0062] (3) User-Intelligent Reflector Interaction Modeling Module: The user feature matrix and the intelligent reflector feature matrix are projected onto a unified feature space through a linear mapping layer, and the feature dimensions are... The data is input into the user-intelligent reflective surface interaction modeling module, which outputs the user-intelligent reflective surface relationship matrix.

[0063] In one embodiment, the user-intelligent reflective surface interaction modeling module employs a Transformer-based attention module. This attention module includes a multi-head self-attention sublayer and a feedforward neural network sublayer, wherein the number of attention heads is set to 8, and the feature dimension of each attention head is [missing information]. Attention calculation employs a scaled dot product attention mechanism; the feedforward neural network is a two-layer fully connected structure with a hidden layer dimension of 256 and uses the ReLU activation function; residual connections and layer normalization are set in each sub-layer to enhance model stability. The user-intelligent reflector relationship matrix is ​​output through the attention module. .

[0064] This embodiment uses the aforementioned dual-feature extraction and interactive modeling structure to achieve explicit modeling of the collaborative relationship between users and multiple active intelligent reflective surfaces, thereby effectively characterizing the complex coupling characteristics in multi-RIS scenarios and improving the accuracy of user-intelligent reflective surface association decisions.

[0065] Step 2: Normalize the user-intelligent reflector relationship matrix to obtain the user-intelligent reflector allocation probability matrix.

[0066] The base station will use the user-intelligent reflector relationship matrix. Input the Softmax layer to generate the user-smart reflector assignment probability matrix. ,in Indicates communication user Assigned to the The probability of each active intelligent reflector, i.e., each element in the user-intelligent reflector allocation probability matrix, represents the probability that a communication user is assigned to an active intelligent reflector. This probability is used to characterize the preference distribution of the deep neural network regarding the user-intelligent reflector association. Based on the model prediction preference information reflected by this probability distribution, the generation process of candidate solutions can be further guided.

[0067] Step 3: Adaptively adjust the size of the candidate set based on the difference between the user-intelligent reflector allocation probability matrix of the current time slot and the optimal allocation scheme in the experience pool, and search for candidate schemes from the user-intelligent reflector allocation probability matrix using a deterministic candidate strategy, a probability-guided sampling strategy, and a perturbation-enhanced exploration strategy.

[0068] Based on step 2, this embodiment designs an efficient candidate generation method based on the output space of a deep neural network. This method uses the probability distribution of user-intelligent reflector allocation output by the deep neural network as prior information. By guiding sampling near the high preference region of the probability distribution, a set of user-intelligent reflector allocation candidate schemes with both diversity and high quality are generated.

[0069] Specifically, this embodiment does not blindly search the entire discrete solution space. Instead, it utilizes a deep neural network to predict the preferences of the user-intelligent reflective surface relationship, constraining the candidate generation process to the vicinity of high-probability regions. This reduces search complexity while ensuring the quality of candidate solutions. Furthermore, by introducing appropriate perturbations within high-preference regions, the generated candidate solutions maintain consistency with model predictions while possessing a certain exploratory capability, thereby effectively mining potential high-quality solutions. The specific construction of candidate solutions is achieved through the following three strategies:

[0070] (1) Deterministic Candidate Strategy: Based on the maximum probability term corresponding to each communication user in the user-intelligent reflector allocation probability matrix, a corresponding active intelligent reflector is selected for that communication user, thereby generating a deterministic allocation matrix as a candidate scheme, denoted as . This strategy directly utilizes the optimal prediction results of deep neural networks, enabling the rapid acquisition of high-confidence solutions to ensure the basic performance of the candidate set.

[0071] (2) Probability-guided sampling strategy: Based on the user-intelligent reflector probability matrix, multiple random samplings are performed on each communication user. Randomness is introduced while adhering to the prediction preferences of the deep neural network, thereby generating multiple candidate schemes with differences to improve the coverage of the solution space. Each random sampling yields the sampling results of each communication user, and the sampling results of all communication users are combined into a candidate scheme. The candidate set generated by this strategy is denoted as:

[0072]

[0073] (3) Perturbation-enhanced exploration strategy: A random perturbation is introduced into the user-intelligent reflector allocation probability matrix and normalized. Then, multiple random sampling operations are performed based on the perturbated probability distribution, thereby expanding the search range and enabling the candidate generation process to escape local optima, enhancing the model's ability to explore the global optimum. The candidate set generated by this strategy is denoted as:

[0074]

[0075] Through the synergistic effect of the three strategies described above, the candidate generation process can focus on the high-preference regions of the deep neural network output while also considering the diversity and exploration capabilities of the solution space, thus achieving effective coverage of high-quality candidate solutions. Specifically, the three strategies supplement and optimize the candidate generation process from different perspectives: the deterministic candidate strategy provides high-confidence solutions based on the current model's predictions, ensuring a high initial performance level for the candidate set; the probabilistic guided sampling strategy enhances the diversity of candidate solutions, thereby strengthening the coverage of the solution space; and the perturbation-enhanced exploration strategy further expands the search range by applying perturbations to the probability distribution, enabling the candidate generation process to escape local optima and thus improving the ability to explore potential global optima. Therefore, the three strategies complement each other in terms of high-quality solution acquisition and solution space exploration capabilities.

[0076] Based on this, to enable the aforementioned candidate generation strategy to adaptively adjust according to the model's prediction accuracy, this embodiment unifies the candidate generation process into an adaptive candidate size adjustment mechanism based on the difference between prediction preferences and empirical optimal solutions. Specifically, by measuring the degree of difference between the current prediction allocation result of the deep neural network and the optimal allocation scheme in the empirical pool, the size of the candidate set is adjusted. Dynamic adjustments are made. The degree of difference is calculated using cross-entropy, and its expression is:

[0077]

[0078] in, The degree of difference characterizes the degree of deviation between the model's predicted distribution and the optimal allocation. Indicator variables representing the optimal allocation scheme in the experience pool (when communication users) Select the The value is 1 when there is an active intelligent reflective surface, and 0 otherwise, which is used as the real label.

[0079] Based on the degree of difference, the size of the candidate set Adaptive adjustment is performed. This embodiment sets a minimum value for the candidate size range. With the maximum value And according to the degree of difference A normalized mapping is performed to determine the number of candidates.

[0080]

[0081] in, This represents the upper bound or preset threshold of the difference. Using the above method, when the difference is large, the number of candidates... near This expands the search scope to enhance the ability to explore the solution space, thereby discovering potentially better allocation schemes; when the difference is small, it indicates that the model prediction is relatively accurate and the number of candidates is small. near This reduces computational complexity and improves solution efficiency. Under the control of this adaptive adjustment mechanism, the three candidate generation strategies mentioned above can achieve dynamic and coordinated adjustment between the exploration range and candidate quality, thereby enabling the model to maintain stable performance and high optimization efficiency under different channel environments, and further achieving an effective balance between exploration and utilization in the candidate generation process.

[0082] Step 4: Evaluate the performance of each candidate scheme in the candidate set, select the candidate scheme with the highest communication rate as the optimal allocation scheme for the current time slot, and store it in the experience pool.

[0083] After obtaining the set of candidate user-intelligent reflector allocation matrices, for each user-RIS allocation matrix in the candidate set... , Execute the base station beamforming matrix respectively With each active intelligent reflector's reflection matrix The joint optimization process. During the joint optimization process, with the current candidate assignment matrix fixed... Under these conditions, the base station beamforming matrix and reflection matrix are designed collaboratively. The base station beamforming matrix controls the energy distribution of the transmitted signal in space, while the reflection matrices of each active intelligent reflector are used to adjust the amplitude and phase of the reflected signal, thereby reconstructing the wireless propagation environment. By jointly optimizing the signal modulation parameters of the transmitting and reflecting ends, the received signal power of users is improved and multi-user interference is suppressed, thereby increasing the overall system rate. The corresponding optimization objective can be expressed as:

[0084]

[0085] Through the aforementioned joint optimization process, the system can obtain the optimal total system rate corresponding to each candidate user-RIS allocation matrix under the current channel state. The system performs the joint optimization and performance evaluation process on all user-RIS allocation matrices in the candidate set, compares the optimal communication rates of each candidate scheme, and selects the user-RIS allocation matrix with the highest communication rate as the optimal allocation scheme under the current channel state. Finally, the optimal user-RIS allocation scheme, its corresponding channel state set, and communication rate are stored in an experience pool for subsequent training and parameter updates of the deep neural network model.

[0086] During the training of the association modeling model, several training samples are randomly selected from the experience pool. Each training sample contains a set of channel states, an optimal allocation scheme, and the corresponding communication rate. The user feature extraction network, the intelligent reflector feature extraction network, and the user-intelligent reflector association modeling module are jointly updated. For each set of training samples, the corresponding channel state information is input into the association modeling model to obtain the currently predicted user-intelligent reflector allocation probability matrix. This matrix is ​​then compared with the optimal user-intelligent reflector allocation result stored in the samples. The deviation between the predicted result and the target result is calculated using a preset loss function as the classification loss.

[0087] In this embodiment, the cross-entropy loss function is preferably used. Backpropagation is performed based on the loss value to update the model parameters, gradually bringing the model closer to the optimal allocation strategy in the experience pool. The Adam optimization algorithm is preferably used during the optimization process, with a learning rate set to... to The learning rate decay strategy is used for dynamic adjustment. During training, the batch size is preferably 32-128, the number of training rounds is preferably 100-500, and an early stopping strategy is adopted based on the performance of the validation set. At the same time, an L2 regularization term is introduced to reduce the risk of model overfitting, and the parameters of the association model are updated in combination with the classification loss. The training process is repeated iteratively until the training termination condition is met, and the optimal association model is output.

[0088] Furthermore, during training, the input channel data is normalized and random noise perturbation is introduced to generate enhanced samples. These enhanced samples and training samples are used as training data to enhance the model's adaptability to channel changes, thereby improving its robustness in dynamic channel environments. As training continues, the model can output more stable and accurate user-intelligent reflector allocation results under different channel conditions.

[0089] To verify the effectiveness of the method of this invention, a typical simulation scenario was constructed for performance evaluation. The link between the base station and the intelligent reflector in the system is modeled as a Rician fading channel, and an α-β-γ path loss model is used to describe the large-scale fading characteristics, where α is the distance attenuation coefficient, β is the baseline loss offset, and γ is the environmental attenuation factor. Communication users and sensing targets are randomly distributed within an area centered on the base station, with each communication user assisted by a single intelligent reflector. In this simulation environment, the method of this invention is compared with the Direct-MLP (Direct Multi-Layer Perceptron) method, the CNN (Convolutional Neural Network) method, the randomized RIS combined with RZF (Regularized Zero-Forcing) method, and the passive RIS method (i.e., replacing the active intelligent reflector in the method of this invention with a passive intelligent reflector). The system total rate within the time block, the system total rate with different numbers of communication users, the system total rate with different total transmit power, and the system total rate with different distances between the base station and the intelligent reflector are monitored and compared for each scheme.

[0090] according to Figures 2-5 Simulation results show that, in terms of convergence performance, the proposed method converges quickly within fewer training iterations, and the final normalized total system rate approaches the optimal performance of exhaustive search, significantly outperforming Direct-MLP and CNN methods, demonstrating stronger feature modeling capabilities and stability. Regarding total system rate performance, as the number of communication users increases, the proposed method consistently outperforms various baseline schemes, maintaining high throughput even under high load conditions, indicating good scalability in multi-user scenarios. In terms of joint optimization performance, the proposed active RIS-assisted scheme significantly outperforms passive and random RIS schemes under different transmit power, user scale, and propagation distance conditions, with particularly pronounced advantages in low-power and long-distance scenarios. Regarding adaptability to complex environments, as the distance between the base station and the intelligent reflector increases or the system scale expands, the performance of traditional schemes declines significantly, while the proposed method maintains a high communication rate, demonstrating stronger robustness.

[0091] In summary, the method of this invention can achieve near-optimal system performance while ensuring low computational complexity in multi-active intelligent reflective surface collaborative scenarios, and has good engineering application value.

[0092] In another embodiment, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of a resource optimization method for an integrated intelligent reflective surface-assisted communication and sensing system.

[0093] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory.

[0094] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this invention.

[0095] The above embodiments merely illustrate several implementation methods of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of the present invention should be determined by the appended claims.

Claims

1. A resource optimization method for an integrated communication and sensing system assisted by intelligent reflective surfaces, wherein the integrated communication and sensing system comprises a base station, multiple active intelligent reflective surfaces, and multiple communication users, characterized in that, Includes the following steps: Obtain the channel state set between all active smart reflectors and all communication users in the current time slot, and input the channel state set into the association modeling model. The association modeling model includes a user feature extraction network, a smart reflector feature extraction network, and a user-smart reflector interaction modeling module. The user feature extraction network extracts features from the channel vectors corresponding to the communication users to obtain the user feature matrix. The intelligent reflector feature extraction network extracts features from the channel vector corresponding to the active intelligent reflector to obtain the intelligent reflector feature matrix. The user feature matrix and the intelligent reflective surface feature matrix are projected onto a unified feature space and input into the user-intelligent reflective surface interaction modeling module. The user-intelligent reflective surface relationship matrix is ​​output and normalized to obtain the user-intelligent reflective surface allocation probability matrix. The size of the candidate set is adaptively adjusted based on the degree of difference between the current time slot's user-intelligent reflector allocation probability matrix and the optimal allocation scheme in the experience pool. A deterministic candidate strategy, a probability-guided sampling strategy, and a perturbation-enhanced exploration strategy are used to search for candidate schemes from the user-intelligent reflector allocation probability matrix. The performance of each candidate scheme in the candidate set is evaluated, and the candidate scheme with the highest communication rate is selected as the optimal allocation scheme for the current time slot and stored in the experience pool.

2. The resource optimization method for the intelligent reflective surface-assisted communication and sensing integrated system according to claim 1, characterized in that, Both the user feature extraction network and the intelligent reflective surface feature extraction network adopt a multilayer perceptron structure, and the user-intelligent reflective surface interaction modeling module is an attention module based on the Transformer structure.

3. The resource optimization method for the intelligent reflective surface-assisted communication and sensing integrated system according to claim 1, characterized in that, The adaptive adjustment of the candidate set size based on the difference between the user-intelligent reflector allocation probability matrix of the current time slot and the optimal allocation scheme in the experience pool includes: The degree of difference between the user-intelligent reflector allocation probability matrix of the current time slot and the optimal allocation scheme in the experience pool is calculated using cross-entropy. The range of candidate set size is weighted by the ratio of the degree of difference to the upper bound of the difference, and the weighted value is superimposed with the minimum value of the candidate set size range to obtain the candidate set size after adaptive adjustment in the current time slot.

4. The resource optimization method for the intelligent reflective surface-assisted communication and sensing integrated system according to claim 3, characterized in that, The degree of difference between the user-intelligent reflector allocation probability matrix for the current time slot calculated using cross-entropy and the optimal allocation scheme in the experience pool includes: The optimal allocation scheme in the experience pool is selected. If a communication user is allocated to an active intelligent reflector, the true label between the corresponding communication user and the active intelligent reflector is set to 1; otherwise, the true label is set to 0, thereby obtaining the true allocation probability matrix between all communication users and all active intelligent reflectors. Based on the true allocation probability matrix and the user-intelligent reflector allocation probability matrix of the current time slot, the discrete classification cross-entropy is calculated to obtain the degree of difference.

5. The resource optimization method for the intelligent reflective surface-assisted communication and sensing integrated system according to claim 1, characterized in that, The deterministic candidate strategy outputs a candidate solution, and the probability-guided sampling strategy outputs... The perturbation enhancement exploration strategy outputs a candidate scheme. The candidate schemes, the The size of the candidate set.

6. The resource optimization method for the intelligent reflective surface-assisted communication and sensing integrated system according to claim 1, characterized in that, The deterministic candidate strategy performs the following operations: Based on the maximum probability corresponding to each communication user in the user-intelligent reflector allocation probability matrix, the active intelligent reflector corresponding to the maximum probability is selected for the communication user, and a deterministic allocation matrix is ​​obtained as a candidate scheme.

7. The resource optimization method for the intelligent reflective surface-assisted communication and sensing integrated system according to claim 1, characterized in that, The probability-guided sampling strategy performs the following operations: Based on the user-intelligent reflector allocation probability matrix, multiple random samples are taken from each communication user to generate multiple candidate schemes with differences.

8. The resource optimization method for the intelligent reflective surface-assisted communication and sensing integrated system according to claim 1, characterized in that, The perturbation enhancement exploration strategy performs the following operations: A random perturbation is introduced into the user-intelligent reflector allocation probability matrix and normalized. Based on the perturbation-enhanced probability distribution, a random sampling operation is performed to generate multiple candidate schemes for perturbation enhancement.

9. The resource optimization method for the intelligent reflective surface-assisted communication and sensing integrated system according to claim 1, characterized in that, During the training phase, perform the following operations: Several training samples are randomly selected from the experience pool. Each training sample contains a set of channel states, an optimal allocation scheme, and the corresponding communication rate. The channel state set in the training samples is normalized and subjected to random noise perturbation to generate enhanced samples. The training samples and enhanced samples are used as training data. For each sample in the training data, the set of channel states of the sample is input into the correlation modeling model to obtain the currently predicted user-smart reflector allocation probability matrix; The classification loss is calculated by combining the currently predicted user-intelligent reflector assignment probability matrix with the optimal assignment scheme of the sample with the highest communication rate in the training data. An L2 regularization term is introduced, and the parameters of the association modeling model are updated in conjunction with the classification loss. The training process is repeated iteratively until the training termination condition is met, and the optimal correlation model is output.

Images