Beam-level fully decoupled network user identification method based on integrated sensing and communication

Abstract

This invention discloses a beam-level fully decoupled network user identification method based on integrated sensing and communication, belonging to the field of 6G wireless communication and sensing fusion technology. It combines target range, azimuth, and Doppler velocity parameters acquired by FMCW radar with the optimal index of the communication beam to construct a deep neural network inference framework. The network provides a confidence probability for each candidate target as a communication user, and finally selects the one with the highest probability as the identification conclusion. Evaluation using the DeepSense 6G public test platform shows an identification accuracy of 89.3%, which is significantly superior to the traditional angle mapping baseline scheme. The obtained identification results can be directly sent to the control base station to guide the optimal base station access in the uplink and downlink directions, thereby realizing perception-driven transmission and computing linkage under the "sensing-transmission-computing integration" framework. This solution is also applicable to fully decoupled heterogeneous network deployments in scenarios such as vehicle-to-everything (V2X) and industrial IoT.

Description

Technical Field

[0001] This invention belongs to the field of 6G wireless communication network technology, specifically involving a beam-level fully decoupled network user identification method based on integrated sensing and communication. It studies how to complete radar-assisted communication user identification in a fully decoupled heterogeneous network environment by means of integrated sensing, transmission and computing, especially for fully decoupled radio access network (FD-RAN) deployment scenarios that use millimeter-wave radar sensing capabilities to assist user access procedures. Background Technology

[0002] 6G mobile communication is moving towards a fully decoupled heterogeneous network architecture. In this architecture, base station functions are subdivided into three categories: control base stations, uplink transceiver base stations, and downlink transceiver base stations. User terminals can freely choose different base stations to access in both uplink and downlink directions, which theoretically can significantly improve spectrum utilization and enhance network flexibility. However, this functional decoupling brings new user access challenges: in dense network scenarios, radar sensors mounted on the base station side often detect multiple moving or stationary targets simultaneously (of which only one is a real communication user). The control base station must accurately locate the communication user from these candidate targets in order to allocate appropriate uplink and downlink access resources to it.

[0003] Current academic research on fully decoupled networks mainly focuses on load balancing strategies and spectral efficiency improvement, with insufficient attention paid to the crucial issue of "how to accurately map sensing information to the communication users to be served." The common approach is to simplify the relationship between the communication beam pointing angle and the radar detection angle into a linear model. However, in actual deployments, antenna installation deviations, multipath superposition, and detector noise all introduce varying degrees of nonlinear errors, ultimately leading to a low user identification hit rate and significantly reducing the reliability of access decisions.

[0004] Furthermore, "integrated sensing, transmission, and computing" is widely considered the core design paradigm for next-generation fully decoupled networks. Its essence lies in using environmental information obtained through sensing to drive transmission decisions and fully leveraging the computing resources of edge nodes to ensure real-time processing. Therefore, there is an urgent need for a lightweight approach that can implement the integrated sensing, transmission, and computing design concept in fully decoupled heterogeneous networks, accurately identifying the current communication users with lower computational overhead. Summary of the Invention

[0005] To address the aforementioned issues, this invention discloses a beam-level fully decoupled network user identification method based on integrated sensing and communication. The method uses communication beam number and radar measurement data as dual inputs, and relies on a deep neural network for joint inference to output the probability that each candidate target is a communication user. The identification results can be directly used by the control base station for uplink and downlink access resource allocation decisions, thereby realizing sensing-driven transmission and computing linkage under the "sensing-transmission-computing integration" framework.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] A beam-level fully decoupled network user identification method based on integrated sensing and communication includes:

[0008] Step S1: Construct an integrated sensing and communication system model of a fully decoupled heterogeneous network, including a communication model and a radar model.

[0009] Communication Model: Consider a single-user MISO downlink, with the base station equipped with... A uniform linear antenna array. The channel employs a geometric model:

[0010] ;

[0011] in and The first The complex coefficients and angle of arrival of the path, This is the array response vector. For ULA, The base station uses a beamforming codebook. The user receives the following signal:

[0012] ;

[0013] in For transmit power gain, It is additive white Gaussian noise. superscript This represents the Hermitian transpose of a channel matrix or vector. This refers to the beamforming vector used on the base station side. The optimal beam index for the communication data symbols (information) transmitted by the base station is obtained by maximizing the received power;

[0014] ;

[0015] Radar model: FMCW radar is used. Transmitting linear frequency modulated signals:

[0016] ;

[0017] in For carrier frequency, For frequency modulation slope, For bandwidth, For chirp duration, Represents a continuous time variable within the duration of a single linear frequency modulated (chirp) pulse; a radar frame contains 1 chirp, each chirp sampled There are [number] points, and the number of receiving antennas is [number]. Obtain radar data .right Range FFT, angle FFT, and Doppler FFT are performed sequentially to obtain a radar cube; high-power points are extracted using CFAR detection; points belonging to the same object are clustered using DBSCAN; and the average range of each cluster is output. ,angle Doppler velocity Form candidate object attribute vectors Ultimately obtained There are 10 candidate objects, which contain communication users.

[0018] Step S2: Define the user identification problem

[0019] Let the first The time-slot radar detected An object, whose attribute set is: The attribute vector of each candidate object All include distance ,angle and Doppler velocity Simultaneously, the communication system provides the optimal beam index. Let the actual index of the communication user be... Define indicator variables The goal is to learn a mapping function. Make The problem is transformed into supervised learning: finding parameters. This makes the neural network output Approaching Ultimately passed Output the predicted index.

[0020] Step S3: Design a deep neural network model

[0021] Transform the user identification problem into a supervised learning problem and design a deep neural network model. The network takes the attributes and optimal beam index of a single candidate object as input and outputs the probability that the object is a communication user. It calculates the probability independently for all candidate objects and selects the one with the highest probability as the identification result. .

[0022] The neural network structure is a 7-layer fully connected network:

[0023] (1) Input layer: 4-dimensional vector Among them, beam index It can be converted to one-hot encoding or directly input as an integer.

[0024] (2) First hidden layer: 64 neurons, ReLU activated.

[0025] (3) Second hidden layer: 32 neurons, ReLU activated.

[0026] (4) Third hidden layer: 16 neurons, ReLU activated.

[0027] (5) Fourth hidden layer: 8 neurons, ReLU activated.

[0028] (6) Fifth hidden layer: 4 neurons, ReLU activated.

[0029] (7) Output layer: 1 neuron, Sigmoid activation, output probability .

[0030] Since each candidate object passes through the same network independently, this structure naturally supports a variable number of object lists and is permutation invariant.

[0031] Step S4: Training and Deployment

[0032] Training was performed using a real-world dataset (e.g., DeepSense 6G Scene 35). The dataset contains 2158 samples, each including a list of radar objects, an optimal beam index, and manually labeled user tags. Training / testing was divided by independent driving trajectories (34 training trajectories, 9 testing trajectories). The loss function was the mean squared error.

[0033] ;

[0034] in, This represents the total number of data samples used for training (i.e., the total number of sampling times). This represents the probability prediction value output by the deep neural network (i.e., the confidence level that the object is the target communication user). The label is a true indicator variable representing the identity of the candidate object. ; (when the target) (Take 1 if it is a real communication user, otherwise take 0).

[0035] The optimizer was ADAM, with an initial learning rate of 0.001. This rate was decayed to one-tenth of its original value at the 50th and 80th epochs, for a total of 100 epochs of training. The batch size was set to 128. After model training, the parameters were fixed, and the model was deployed to a fully decoupled network control base station or nearby edge computing nodes. During online operation, in each decision time slot, the control base station collected the radar target list and the optimal beam number reported by the user, sending them to the deployed network to infer probabilities on a target-by-target basis. The target with the highest probability was selected as the communication user, and based on its measured angle and distance, uplink and downlink access base stations were selected from a pre-built base station coverage map.

[0036] From the perspective of integrated sensing, transmission, and computing, the design logic of this solution is as follows: the radar sensing stage is responsible for collecting and outputting the physical attributes of environmental targets; the communication transmission stage provides the optimal beam number for user feedback; the deep neural network inference stage completes user identification at the edge nodes; and the identification result is then fed back to the transmission stage to drive the execution of access decisions. The whole system forms a complete closed loop of "sensing-computing-transmission", which is highly consistent with the core connotation of the integrated sensing, transmission, and computing paradigm.

[0037] The beneficial effects of this invention are as follows:

[0038] This invention combines target range, azimuth, and Doppler velocity parameters acquired by FMCW radar with the optimal index of the communication beam to construct a deep neural network inference framework. The network provides a confidence probability for each candidate target as a communication user, and finally selects the one with the highest probability as the identification conclusion. Validated on a real-world dataset, the user identification accuracy reaches 89.3%, which is approximately 25 percentage points higher than the traditional angle-offset baseline (64.2%), effectively improving the reliability of the fully decoupled heterogeneous network in the user access stage. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the fully decoupled heterogeneous network architecture and the integrated sensing, transmission, and computing in an example of the present invention;

[0040] Figure 2 This is an overall flowchart of the user identification method in an example of the present invention.

[0041] Figure 3 This is a schematic diagram of the approximation algorithm flow in an example of the present invention. Detailed Implementation

[0042] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0043] As shown in the figure, the present invention provides a beam-level fully decoupled network user identification method based on integrated sensing and communication.

[0044] Step S1: System Model Construction (see...) Figure 1 )

[0045] The simulation scenario is set as a typical fully decoupled heterogeneous network: it includes a control base station responsible for issuing and managing global access policies, several uplink receiving base stations and downlink transmitting base stations each performing their own functions, and a communication user moving on the road (using an autonomous vehicle as a representative example). To acquire information about surrounding targets, an FMCW radar sensor is installed on the base station side, specifically for real-time perception of moving and stationary objects within the field of view.

[0046] Communication model: Base station equipped with A uniform linear antenna array, operating in the 28 GHz millimeter-wave band. The beamforming codebook uses a DFT codebook and contains... One beam, covering Range. During the initial access or periodic beam training phase, the user equipment measures the received signal power of each beam and indexes the optimal beam. Feedback is transmitted to the control base station via a control channel. The channel employs a geometric model and includes... One path (line-of-sight path and two reflection paths), angle of arrival for each path. and complex gain Randomly generated. Array response vector. The above formula is satisfied.

[0047] Radar Model: FMCW radar parameters are as follows: bandwidth MHz, chirp cycle FM slope MHz / Each frame contains 1 chirp, each chirp sampled Number of points, number of receiving antennas The radar has a maximum detection range of 249m and a maximum speed of 82km / h. After receiving data, a range FFT (512 points) is first performed to obtain a range-time map; moving target detection and clutter suppression are then performed; followed by an angle FFT (8 points) to obtain angular information; finally, a Doppler FFT (250 points) is performed to obtain velocity information, forming a radar cube (8×250×512). CFAR detection (constant false alarm rate, false alarm probability) is applied. High-power points are extracted by taking the coordinates of the detection points (distance, angle, Doppler) as input and using DBSCAN clustering (neighborhood radius). , the minimum number of points This clusters points belonging to the same physical object. Each cluster outputs the average distance. Average angle Mean Doppler velocity A candidate object list is formed. Because the environment contains communication user vehicles, pedestrians, and roadside parked vehicles, It usually varies between 1 and 8.

[0048] Step S2: User Identification Problem Definition

[0049] For each sampling time slot (Sampling interval approximately 111ms), control base station is known:

[0050] Candidate object list ;

[0051] The attribute vector of each candidate object Each contains the distance to the object. ,angle and Doppler velocity property;

[0052] And the optimal beam index based on user feedback ;

[0053] What is unknown is the index corresponding to the communication user. Due to the fixed installation deviation between radar and communication antennas (e.g., horizontal deviation) The radar's angular resolution is approximately 10°, and the simple angle matching error is relatively large. In addition, distance and velocity information also contain discriminative power, so a nonlinear mapping needs to be learned.

[0054] Step S3: Deep Neural Network Design (see...) Figure 2 , Figure 3 )

[0055] This embodiment transforms the user identification problem into a supervised learning problem and designs a deep neural network model. The network takes the attributes (distance, angle, velocity) of a single candidate object and the optimal beam index as joint input, and independently outputs the probability that the object is a communication user. By independently calculating the probability of all candidate objects, the network finally selects the one with the highest probability as the recognition result. .

[0056] The network employs a fully connected structure, with a 4-dimensional input as previously mentioned. It is worth noting the beam index. Normalized to Interval (divided by 64), distance Normalization to (Divided by the maximum distance of 249m), angle Normalization to (Mapped from) ),speed Normalization to (Mapped from) (km / h). Total number of network parameters: Layer 1: Layer 2: Layer 3: Layer 4: Layer 5: Output layer: With a total of approximately 3105 parameters, it is suitable for real-time inference on edge devices (approximately 0.1ms for a single forward propagation).

[0057] Step S4: Training and Deployment

[0058] The DeepSense 6G platform used a real parking lot as the data source for the experiment. The test location was a real parking lot with the base station installed on the second floor, facing a two-way four-lane main road. During the test, a vehicle equipped with 802.11ad communication equipment (60GHz band) and a GPS receiver repeatedly drove along the road, and the system simultaneously recorded the RSSI values ​​of the beams in each direction and the complete raw radar sampling data.

[0059] During the training phase, the loss function is defined as the mean squared error:

[0060] ;

[0061] in The total number of training samples, As an indicator variable. When This indicates the candidate object. If the target is the actual communication user, then set it to 0; otherwise, set it to 0. .

[0062] The data processing steps are as follows: (see...) Figure 3 (1) Extract the object list (distance, angle, speed) from the original radar data to obtain 3045 unlabeled samples. (2) By manually checking the radar Range-Doppler map and RGB camera images, and combining time continuity tracking, label the object index corresponding to the communication user to obtain 2158 labeled samples. (3) Divide the samples according to the independent "pass" of the vehicle: 34 passes were used for training (1581 samples), and 9 passes were used for testing (381 samples).

[0063] Training hyperparameter settings: The loss function was the mean squared error (MSE), and the optimizer was ADAM; the learning rate started at 0.001 and was multiplied by a reduction factor of 0.1 after the 50th and 80th training epochs, respectively; the mini-batch size was 128, and the cumulative iteration was 100 epochs. To evaluate training stability, 10-fold cross-validation was additionally introduced, ultimately achieving an accuracy of 89.3% on the held-out test set, with a standard deviation of 4.14% for each fold.

[0064] Deployment Phase: The trained model parameters are embedded into the control base station of the fully decoupled heterogeneous network. In each time slot, the control base station collects a list of objects detected by the radar. and user reports Calculate the probability for each object ,Pick Then, based on the object's angle... and distance Query the pre-established base station coverage map and select the uplink base station with the best signal quality. and downlink base stations (For example, selecting the nearest base with a load below a threshold) and notifying the user to access the site via a control channel.

[0065] Implementation effect

[0066] Evaluations on an independent test set showed that the user identification accuracy of the proposed method was 89.3%. In comparison, the accuracy rates of baseline methods were as follows: traditional angle offset method 64.2%, one-dimensional linear regression of angle approximately 63.8%, three-dimensional linear regression approximately 67.4%, lookup table method 64.7%. The deep learning method outperformed the best baseline by approximately 22 percentage points. Furthermore, by integrating the user identification results into the access control process, the system can save approximately 70% in beam training overhead, and the overall network throughput increases by approximately 18%. These performance levels demonstrate that the proposed method fully meets the engineering requirements for identification accuracy in practical radar-assisted communication systems.

[0067] Comparison scheme description: Traditional baseline methods rely on angle offset models ,in The beam pointing angle, Obtained through least squares estimation of the training data. Then select... This method does not utilize distance and velocity information and cannot handle nonlinear errors.

[0068] It should be noted that the above content merely illustrates the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, various improvements and modifications can be made without departing from the principle of the present invention, and all such improvements and modifications fall within the scope of protection of the claims of the present invention.

Claims

1. A beam-level fully decoupled network user identification method based on integrated sensing and communication, characterized in that, Includes the following steps: Step S1: Build an integrated sensing and communication system model for a fully decoupled heterogeneous network. This model includes two parts: a communication module and a radar module. The communication module describes the base station antenna array configuration, beamforming codebook structure, user terminal received signal expression, and the selection criteria for the optimal beam index. The radar module describes the FMCW waveform transmission signal model, the intermediate frequency signal form after down-conversion, the range-angle-Doppler three-dimensional data cube generation process, and the subsequent CFAR point detection, DBSCAN clustering segmentation and candidate target attribute vector extraction. Step S2: Define the user identification problem. Assume that the radar in the current time slot detects... There are candidate objects, and their attribute vector sets are: Each Includes distance, angle, and Doppler velocity; simultaneously obtains the optimal beam index selected by the communication system. The goal is to find the function ,in A true index for communication users; Step S3: Transform the user identification problem into a supervised learning problem. For cases where the number of candidate targets is not fixed, design a deep neural network model. The network takes the attributes and optimal beam index of a single candidate object as input and outputs the probability that the object is a communication user. By independently calculating the probability of all candidate objects, the network selects the one with the highest probability as the identification result. ; Step S4: Train the deep neural network using a dataset collected from a real-world environment, defining the loss function as mean squared error. ,in Represents objects A value of 1 indicates that object k is the target communication user; otherwise, it is set to 0. After the network converges, the trained model weights are deployed to the control base station of the fully decoupled heterogeneous network or the edge computing unit near the site to support the real-time decision-making of uplink and downlink base stations in the user access process.

2. The method according to claim 1, characterized in that, The specific form of the communication module in step S1 is as follows: The base station side is deployed with... A uniform linear antenna array using a geometric channel model. ,in and The first The complex coefficients and angle of arrival of the path, The array response vector; the base station uses a beamforming codebook. The user receives the signal as ,in For transmit power gain, The noise is additive white Gaussian noise; the optimal beam index is obtained by maximizing the received power. .

3. The method according to claim 1, characterized in that, The specific form of the radar module in step S1 is as follows: The system adopts an FMCW radar system, radiating linear frequency modulated pulse signals into space. ,in For carrier frequency, For frequency modulation slope, For bandwidth, The chirp duration; the received signal is mixed to obtain the intermediate frequency signal. If there is a single object in the environment at a distance of... The intermediate frequency signal A three-dimensional radar cube is formed by sequentially applying range-dimensional FFT, angle-dimensional FFT, and Doppler-dimensional FFT to a single frame of received radar data. After CFAR point detection and DBSCAN density clustering, attribute vectors are extracted for each candidate object. ,in For distance, For angle, This is the Doppler velocity.

4. The method according to claim 1, characterized in that, The topology of the deep neural network in step S3 is as follows: the input receives a feature vector consisting of four components. The network then passes through five fully connected hidden layers in sequence—with dimensions of 64, 32, 16, 8, and 4 respectively, all using ReLU as the activation function. Finally, the output layer compresses the 4-dimensional features into a 1-dimensional scalar and maps them to probability values ​​using the Sigmoid function. The entire network consists of 7 weighted layers, and the final output reflects the confidence probability that the candidate target is a communication user.

5. The method according to claim 1, characterized in that, The handling mechanism for the case where the number of candidate targets is not fixed in step S3 is as follows: in a certain time slot Detected For each object, assign an attribute vector to it. With the current optimal beam index Each input is fed into the same deep neural network, and the results are obtained. Group probability scoring; compare the scores one by one, and determine the object with the highest probability as the communication user, i.e. The process of independently performing probability inference and comparison for each candidate target is independent of the order in which multiple candidate targets detected by the radar sensor within a detection time slot are input into the neural network, and for any Both are applicable.

6. The method according to claim 1, characterized in that, The complete expression for the loss function in step S4 is: ,in The optimizer was ADAM, with an initial learning rate of 0.

001. The learning rate was reduced to one-tenth of its original value in the 50th and 80th training epochs, with a total of 100 training epochs and a batch size of 128 samples per iteration.

7. The method according to claim 1, characterized in that, The method further includes using the location information obtained from user identification in a fully decoupled network access control process: controlling the base station to read the azimuth angle of the target user based on the identification result. and distance Taking into account the current service load of each uplink and downlink base station, the optimal uplink access base station is selected for the user. and downlink access base station This enables the coordinated sensing-assisted transmission and computing in the integrated sensing, transmission, and computing system.

Images