AI-based 5g active mimo beamforming intelligent optimization method

Abstract

The application relates to the technical field of communication, and discloses an AI-based 5G active MIMO beamforming intelligent optimization method.The method comprises the following steps: acquiring real-time radio frequency amplitude and phase data of each antenna unit; inverting downlink channel state information; collecting obstacle distribution information; combining a benchmark model to generate an antenna health state confidence score; fusing multi-source information to construct a feature tensor; inputting a deep reinforcement learning network, taking edge throughput and main lobe precision as optimization objectives, and outputting a complex weighting coefficient; and reconstructing a beamforming matrix according to the complex weighting coefficient.The application realizes dynamic adaptive optimization of beams under abnormal working conditions, and improves link robustness and spectral efficiency.

Description

AI-based intelligent optimization method for 5G active MIMO beamforming Technical Field

[0001] This invention belongs to the field of communication technology, specifically relating to an AI-based intelligent optimization method for 5G active MIMO beamforming. Background Technology

[0002] With the large-scale deployment of fifth-generation mobile communication (5G) technology, massive MIMO (Multiple-Input Multiple-Output) beamforming has become a means to improve system capacity, spectral efficiency, and coverage quality. This technology integrates a large number of independently controllable active antenna elements on the base station side, dynamically adjusting the amplitude and phase of each channel to synthesize a high-gain, narrow-main-lobe directional beam, thereby achieving precise spatial focusing on user terminals. However, active MIMO systems are highly dependent on the physical integrity and coordination of each radio frequency channel in the array, and their performance is extremely sensitive to hardware conditions.

[0003] Beamforming technology based on active antenna arrays requires all radiating elements to maintain stable electrical performance parameters during long-term operation. However, in complex deployment environments, antenna elements may experience localized failures due to factors such as component aging, power fluctuations, lightning strikes, or physical obstructions, leading to abnormal output power, phase shifts, or complete failure. While such failures do not affect the overall power supply and control logic of the base station, they distort the array's composite radiation pattern, causing unintended radiation characteristics such as main lobe shift, side lobe rise, and null filling, severely weakening beam pointing accuracy and interference suppression capabilities.

[0004] Existing technologies typically rely on periodic manual inspections or simple alarm mechanisms based on fixed thresholds to identify hardware anomalies, lacking real-time perception and intelligent diagnostic capabilities for the "vital signs" of RF channels. Once a fault occurs, the system often continues to perform beamforming calculations based on the original complete array model, leading to a significant deviation between the actual radiation pattern and expectations, causing link interruptions, drastic throughput drops, and even service unavailability. Even with the introduction of redundant designs or coarse channel shielding strategies in some solutions, it is difficult to perform refined modeling and dynamic compensation based on the fault type and location, making it impossible to restore optimal beamforming performance within millisecond-level latency.

[0005] Therefore, in the context of 5G high-reliability, low-latency communication, there is an urgent need for an intelligent beamforming optimization mechanism that can achieve fault self-sensing, self-location, self-isolation, and self-repair to ensure the robustness and quality of service of active MIMO systems under non-ideal hardware conditions. Summary of the Invention

[0006] This invention provides an AI-based intelligent optimization method for 5G active MIMO beamforming, aiming to solve the problem of severe discrepancies between actual radiation patterns and expected designs, and drastic performance degradation, caused by hardware failures or physical obstructions in active antenna array elements. This method constructs a closed-loop optimization system integrating real-time channel state awareness, antenna element health status diagnosis, and an AI-driven beam reconfiguration mechanism. This enables dynamic adaptive adjustment of beamforming parameters under abnormal operating conditions, ensuring the stability and spectral efficiency of the communication link.

[0007] This invention provides an AI-based intelligent optimization method for 5G active MIMO beamforming, comprising:

[0008] Acquire the real-time RF output signal amplitude and phase data of each antenna element in the active antenna array;

[0009] Collect uplink pilot signals received by the base station to invert downlink equivalent channel state information;

[0010] Simultaneously acquire information on the distribution of obstacles in front of the antenna panel detected by the environmental perception sensor;

[0011] Based on the amplitude and phase data of the radio frequency output signal, and combined with the preset normal working state benchmark model, a health status confidence score for each antenna unit is generated.

[0012] The channel state information, obstacle distribution information, and health status confidence score are spatiotemporally aligned to form a multi-source heterogeneous input feature tensor.

[0013] The multi-source heterogeneous input feature tensor is input into a pre-trained deep reinforcement learning policy network. The deep reinforcement learning policy network takes maximizing the throughput of cell edge users and the pointing accuracy of the main lobe as the joint optimization objective and outputs the complex weighting coefficients of each effective antenna element.

[0014] The digital precoding matrix of the beamformer is reconfigured based on the complex weighting coefficients to complete the online reconstruction of the beam pattern.

[0015] Preferably, acquiring the real-time RF output signal amplitude and phase data of each antenna element in the active antenna array includes:

[0016] A directional coupler and a quadrature demodulation module are integrated at the output of the power amplifier of each antenna unit. The coupled signal is digitized by an analog-to-digital converter at a sampling rate of not less than twice the Nyquist frequency to extract the in-phase and quadrature components, and then the instantaneous amplitude and phase values ​​are calculated.

[0017] Preferably, a health status confidence score is generated for each antenna element, including:

[0018] Establish an amplitude-phase response database for each antenna element under normal conditions across the entire frequency band. This amplitude-phase response database includes calibration curves under different temperatures, power supply voltages, and operating frequencies.

[0019] The deviation between the measured amplitude and phase values ​​and the calibration curve under the corresponding operating conditions is calculated. If the amplitude deviation is greater than 15% or the phase deviation is greater than 30 degrees, the antenna element is determined to be abnormal.

[0020] The confidence score is mapped to a range of 0 to 1 based on the degree of bias using an exponential decay function; the greater the bias, the lower the score.

[0021] Preferably, the distribution information of obstacles in front of the antenna panel detected by the environmental perception sensor is acquired simultaneously, including:

[0022] The space in front of the antenna is periodically scanned by a millimeter-wave radar array located around the antenna panel, within a 60-degree horizontal field of view and a 30-degree vertical field of view.

[0023] By using time-of-flight ranging and Doppler shift analysis, the distance, azimuth, pitch angle and relative velocity of stationary or moving obstacles are identified, and a three-dimensional point cloud obstacle map is generated.

[0024] After being voxelized, the 3D point cloud obstacle map is projected onto the far-field radiation coordinate system of the antenna array to form an obstacle occlusion mask matrix.

[0025] Preferably, the channel state information, obstacle distribution information, and health status confidence score are spatiotemporally aligned to form a multi-source heterogeneous input feature tensor, including:

[0026] The channel state information is represented as a complex matrix of the number of user equipments multiplied by the number of antenna elements, with the real part and the imaginary part serving as two channels respectively.

[0027] The obstacle occlusion mask matrix is ​​expanded into a single-channel binary tensor with the same dimension as the number of antenna elements;

[0028] The health status confidence score vector is copied and stacked into a weight matrix with the same number of rows as the channel matrix, which is used as the 4th channel;

[0029] The above four channels are concatenated along the feature dimension to form a four-dimensional input tensor, whose dimensions are the number of users, the number of antennas, the number of feature channels, and the time step.

[0030] Preferably, the multi-source heterogeneous input feature tensor is input into a pre-trained deep reinforcement learning policy network. This deep reinforcement learning policy network aims to maximize the throughput of users at the cell edge and the pointing accuracy of the main lobe as a joint optimization objective, and outputs complex weighted coefficients for each effective antenna element, including:

[0031] The deep reinforcement learning policy network adopts a two-stream attention architecture, including a channel-aware stream and a health state stream;

[0032] The channel-aware stream is composed of a multi-layer complex convolutional neural network, used to extract spatial correlation and inter-user interference features;

[0033] The health state flow consists of a fully connected layer and a gated loop unit, used to model the temporal evolution of antenna element failure;

[0034] The two streams of features are fused at a higher level through a cross-attention mechanism to generate an attention-weighted joint representation;

[0035] The joint representation is input to the strategy head, which consists of a complex fully connected layer and directly outputs the complex weighted coefficients of each antenna element. The magnitude of the coefficients is constrained by the confidence level of the health status, and the phase value meets the beam pointing requirements.

[0036] Preferably, the training process of the deep reinforcement learning policy network includes:

[0037] Construct a training set in a simulation environment that includes scenarios such as random antenna unit failure, dynamic obstacle occlusion, and multi-user movement.

[0038] The reward function is defined as the weighted sum of the average throughput of users at the cell edge and the main lobe direction error, where the main lobe direction error is determined by the angle between the desired beam pointing direction and the actual direction of maximum gain of the synthesized beam.

[0039] The network parameters are iteratively updated using a near-end policy optimization algorithm until the reward function converges to a stable threshold.

[0040] Preferably, reconfiguring the digital precoding matrix of the beamformer according to the complex weighting coefficients includes:

[0041] The output complex weighting coefficients are multiplied element-wise with the original precoding vector to obtain the corrected precoding vector;

[0042] The corrected precoding vector is then normalized and loaded into the beamforming register of the baseband processor;

[0043] The weighting coefficients of antenna elements with health status confidence scores below a threshold are forcibly set to zero to achieve physical isolation.

[0044] Preferably, the database of the normal operating state benchmark model is obtained by collecting the amplitude-phase response calibration curves of each antenna unit under different ambient temperatures, power supply voltages and operating frequencies through an automated testing platform before leaving the factory, and then compressing and storing them.

[0045] Preferably, the deep reinforcement learning policy network is deployed on the neural network accelerator of the base station baseband processing unit. The neural network accelerator supports a complex number operation instruction set and the single inference latency is no more than 5 milliseconds.

[0046] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0047] 1. This invention constructs a multi-dimensional sensing input system for abnormal operating conditions by deeply integrating antenna unit health status diagnosis, environmental obstacle perception and channel state information, breaking through the limitation of traditional beamforming that only relies on ideal channel assumptions.

[0048] 2. By using a deep reinforcement learning strategy network to jointly analyze multi-source heterogeneous features, dynamic reconstruction and optimal redistribution of beam patterns are achieved under the condition that some antenna elements fail or are blocked, thus suppressing sidelobe lifting and main lobe splitting phenomena.

[0049] 3. This method can automatically isolate faulty units and re-optimize beam weights without manual intervention, thereby improving the throughput and link robustness of users at the cell edge while ensuring the main lobe pointing accuracy. Attached Figure Description

[0050] Figure 1 is a schematic diagram of the overall technical solution architecture of the present invention;

[0051] Figure 2 is a schematic diagram of the principle framework of the deep reinforcement learning policy network in this invention;

[0052] Figure 3 is a logical flowchart of the construction and spatiotemporal alignment of multi-source heterogeneous input feature tensors in this invention;

[0053] Figure 4 is a logical flowchart of antenna unit health status diagnosis and confidence score generation in this invention.

[0054] Figure 5 is a flowchart illustrating the logical process of environmental obstacle perception and occlusion mask matrix generation in this invention. Detailed Implementation

[0055] Referring to Figures 1 to 5, this invention provides an AI-based intelligent optimization method for 5G active MIMO beamforming, aiming to solve the problem of severe discrepancies between the actual radiation pattern and the expected design, and a sharp deterioration in system performance, caused by hardware failures or physical obstructions in active antenna array elements. This method constructs a closed-loop optimization system integrating real-time channel state awareness, antenna element health status diagnosis, and an AI-driven beam reconfiguration mechanism. This enables dynamic adaptive adjustment of beamforming parameters under abnormal operating conditions, ensuring the stability and spectral efficiency of the communication link.

[0056] The method includes the following steps:

[0057] S1, acquire the real-time RF output signal amplitude and phase data of each antenna element in the active antenna array;

[0058] S2, collect the uplink pilot signal received by the base station to invert the downlink equivalent channel state information;

[0059] S3, synchronously acquires information on the distribution of obstacles in front of the antenna panel detected by the environmental perception sensor;

[0060] S4. Based on the amplitude and phase data of the real-time radio frequency output signal, and combined with the preset normal working state benchmark model, generate a health status confidence score for each antenna unit.

[0061] S5, spatiotemporally aligns the channel state information, obstacle distribution information and health status confidence score to form a multi-source heterogeneous input feature tensor;

[0062] S6, the multi-source heterogeneous input feature tensor is input into a pre-trained deep reinforcement learning policy network. The deep reinforcement learning policy network takes maximizing the throughput of cell edge users and the pointing accuracy of the main lobe as the joint optimization objective and outputs the complex weighting coefficients of each effective antenna element.

[0063] S7. The digital precoding matrix of the beamformer is reconfigured according to the complex weighting coefficients to complete the online reconstruction of the beam pattern.

[0064] In step S1, the specific implementation of obtaining the real-time RF output signal amplitude and phase data of each antenna element in the active antenna array is as follows: A directional coupler and a quadrature demodulation module are integrated at the output of the power amplifier of each antenna element. The coupling degree of the directional coupler is set to 20 dB, which is used to extract a small portion of RF energy from the main signal path as a monitoring signal. This monitoring signal is sent to the quadrature demodulation module, where it is mixed with in-phase and quadrature carriers generated by a local oscillator to generate in-phase and quadrature components, respectively.

[0065] These two analog components are digitized by a high-speed analog-to-digital converter (ADC). The ADC has a sampling rate of at least 500 MHz and a resolution of at least 12 bits to meet the sampling requirement of at least twice the Nyquist frequency. The digitized in-phase component... Orthogonal components It is sent to the baseband processing unit, where the instantaneous amplitude is calculated. With instantaneous phase This process obtains the amplitude and phase values ​​of the RF output signal for each antenna element at the current moment. It is executed continuously every 20 milliseconds to ensure high temporal resolution monitoring of the antenna element's operating status.

[0066] In step S2, the specific implementation of acquiring the uplink pilot signal received by the base station to invert the downlink equivalent channel state information is as follows: The base station receives known pilot sequences transmitted by multiple user equipment on specified time-frequency resources. These pilot sequences have good autocorrelation and cross-correlation characteristics in the time and frequency domains, which facilitates channel estimation. After the baseband processor performs synchronization, denoising, and equalization processing on the received signal, it extracts the uplink channel impulse response from the received signal using the minimum mean square error algorithm or compressed sensing technology.

[0067] Based on the channel reciprocity principle of TDD systems, uplink channel state information is directly considered as downlink equivalent channel state information. The channel state information is represented as a complex matrix with dimension O(n). , For the number of user devices, This represents the total number of antenna elements in the active antenna array. Each element of the complex matrix... Indicates the first The user and the first The complex channel gain between antenna elements, with its real and imaginary parts reflecting the in-phase and quadrature components of the signal, respectively.

[0068] In step S3, the specific implementation of synchronously acquiring the obstacle distribution information detected by the environmental perception sensor in front of the antenna panel is as follows: A millimeter-wave radar array is uniformly arranged around the antenna panel. This millimeter-wave radar array operates in the 77 GHz frequency band and has a scanning capability of a 60-degree horizontal field of view and a 30-degree vertical field of view. The millimeter-wave radar transmits frequency-modulated continuous wave signals with a period of 50 milliseconds and receives the echo signals reflected by objects in front of it. By analyzing the frequency difference between the transmitted signal and the echo signal, the distance to the target object is calculated; by analyzing the phase difference of the multi-channel receiving antenna, the azimuth and elevation angles of the target are calculated; and by analyzing the Doppler frequency shift, the relative velocity of the target is determined.

[0069] All detected target points are integrated into 3D point cloud data. This 3D point cloud data undergoes voxelization, which divides the space into a cubic grid with sides of 10 cm. Points within each grid are marked as obstacles. Subsequently, this voxelized map is projected onto the far-field radiation coordinate system of the antenna array, forming a binary vector corresponding to the number of antenna elements. , Indicates the first The radiation path of each antenna element is blocked by an obstacle. This indicates no occlusion. This binary vector forms the basis of the obstacle occlusion mask matrix.

[0070] In step S4, based on the real-time RF output signal amplitude and phase data, and combined with a preset normal operating state benchmark model, the specific implementation method for generating the health status confidence score of each antenna element is as follows: A normal operating state benchmark model database is pre-stored in the base station storage unit. This database contains amplitude-phase response calibration curves for each antenna element under different ambient temperature ranges (-40 degrees Celsius to +70 degrees Celsius), power supply voltage ranges (4.5 volts to 5.5 volts), and operating frequency ranges (3.3 GHz to 3.6 GHz). These calibration curves are collected by an automated testing platform before the equipment leaves the factory and compressed using compressed sensing technology to save storage space.

[0071] During operation, the system reads the current ambient temperature, power supply voltage, and operating frequency in real time, and retrieves the calibration amplitude for the corresponding operating conditions from the database. Phase with calibration .Will and respectively with , Perform deviation calculation to obtain the amplitude deviation. Phase deviation .like or If so, the antenna element is determined to be abnormal. Health status confidence score. Generated by mapping through an exponential decay function: , and These are weighting coefficients, set to values ​​of 5 and 0.05 respectively, to ensure that the impact of amplitude deviation on the scoring is greater than that of phase deviation. Ultimately, Normalized to the interval between 0 and 1 The closer the value is to 1, the better the health status of the antenna element.

[0072] In step S5, the channel state information, obstacle distribution information, and health status confidence score are spatiotemporally aligned to form a multi-source heterogeneous input feature tensor. The specific implementation method is as follows: First, the channel state information matrix... Split into real part matrix With the imaginary part matrix These two constitute two independent feature channels. Secondly, the obstacle occlusion mask vector is expanded to... matrix Each row is a copy of the obstacle occlusion mask vector, making the dimension the same as the previous row. Consistency forms the third feature channel. Furthermore, the health status confidence score vector is also expanded to... weight matrix Each row is a copy of s, forming the fourth feature channel.

[0073] The above four channels are spliced ​​along the third dimension in the feature dimension to form a three-dimensional tensor. To support time series modeling, the system retains the most recent... X at time steps are stacked to form a four-dimensional input tensor. This tensor is the multi-source heterogeneous input feature tensor, which is used for subsequent deep reinforcement learning policy network inference.

[0074] In step S6, the multi-source heterogeneous input feature tensor is input into a pre-trained deep reinforcement learning policy network, and the complex weighted coefficients of each effective antenna unit are output. The specific implementation is as follows: The deep reinforcement learning policy network is deployed on a dedicated neural network accelerator of the base station baseband processing unit. This dedicated neural network accelerator supports a complex operation instruction set, and the single inference latency is no greater than 5 milliseconds. The network adopts a dual-stream attention architecture, including a channel-aware stream and a health state stream. The channel-aware stream consists of three layers of complex convolutional neural networks, each layer containing a complex convolutional kernel, complex batch normalization, and a complex activation function, used to... and Features such as spatial correlation, inter-user interference, and beamforming potential are extracted. The health state stream consists of two fully connected layers and one gated cyclic unit, used to model the temporal evolution of antenna element health states and obstacle masks, and to capture the dynamic characteristics of fault propagation or obstruction movement.

[0075] The two streams of features are fused at a higher level through a cross-attention mechanism: the output of the channel-aware stream is used as the query vector, and the output of the health state stream is used as the key-value vector. After calculating the attention weights, they are weighted and summed to generate an attention-weighted joint representation. The joint representation Z is fed into the policy head, which consists of two complex fully connected layers and directly outputs... A complex vector of dimension 1. During the output process, a constraint is forcibly imposed: the first... The magnitude of the complex weighting coefficients of each antenna element That is, the magnitude of the weighting coefficient must not exceed the health status confidence score of the antenna element, thereby avoiding the allocation of too much energy to faulty elements.

[0076] The training process of the deep reinforcement learning policy network was completed in an offline simulation environment. The simulation environment constructed diverse scenarios including random antenna unit failures (failure rates ranging from 5% to 30%), dynamic obstacle occlusion (movement speeds from 0 to 5 meters per second), and multi-user movement (speeds from 0 to 60 kilometers per hour). Reward function... Defined as: , The average throughput of users at the cell edge The main lobe direction error (i.e., the angle between the desired beam direction and the actual direction of maximum gain of the synthesized beam, in degrees). The balancing factor is set to 0.7. A proximal policy optimization algorithm is used to iteratively update the network parameters. Each training round contains 1 million interaction samples until the moving average of the reward function R remains stable above the threshold for 100 consecutive rounds, at which point training terminates. After training, the network parameters are fixed and loaded into the neural network accelerator at the base station.

[0077] In step S7, the specific implementation of reconfiguring the digital precoding matrix of the beamformer based on the complex weighting coefficients is as follows: The original digital precoding matrix of the beamformer is generated by the base station scheduler according to the user scheduling results, and each column is the precoding vector of the corresponding user. For the current user, its corresponding precoding vector is... Complex weighted coefficient vector output by the policy network Element-wise multiplication yields the corrected precoding vector. , This indicates the Hadamah. Subsequently, regarding... L2 norm normalization is performed to ensure the total transmit power remains constant. The normalized v' is loaded into the beamforming register of the baseband processor. Simultaneously, the system checks the health status confidence score vector; if the health status confidence score vector is less than a threshold (set to 0.3), the system forcibly... The complex weighting coefficients of each antenna element are set to zero to achieve physical isolation of that antenna element and prevent it from participating in beamforming. The entire beam reconfiguration process is completed within 20 milliseconds, synchronized with the subframe structure of 5G NR, ensuring real-time performance.

[0078] The aforementioned method constructs a multi-dimensional sensing input system for abnormal operating conditions by deeply fusing information on antenna element health status, environmental obstacle distribution, and channel state. Within this system, a deep reinforcement learning policy network achieves dynamic reconstruction and optimal reallocation of the beam pattern, suppressing sidelobe rise and main lobe splitting. Furthermore, by using health status confidence as a weighted constraint, energy waste on faulty elements is avoided, system power consumption is reduced, and the lifespan of the active antenna array is extended.

[0079] In one embodiment of the present invention, the antenna unit RF monitoring module includes a directional coupler, a quadrature demodulator, and a high-speed analog-to-digital converter integrated at the output of the power amplifier of each antenna unit. The directional coupler adopts a microstrip line structure with a coupling degree of 20 dB and a directivity greater than 25 dB, ensuring the purity of the monitoring signal. The quadrature demodulator consists of a mixer, a low-pass filter, and a voltage-controlled oscillator, with the local oscillator frequency strictly synchronized with the RF carrier. The analog-to-digital converter adopts a pipelined architecture with a sampling rate of 500 MHz, an effective bit depth of 12 bits, and a signal-to-noise ratio of not less than 70 dB.

[0080] In one embodiment of the present invention, the environmental obstacle perception module includes a millimeter-wave radar array disposed around the antenna panel. This millimeter-wave radar array consists of 16 transmitting antennas and 16 receiving antennas, employing MIMO virtual aperture technology, achieving an equivalent antenna count of 256 and an angular resolution of 1 degree. The radar signal processing unit uses FMCW waveforms with a bandwidth of 4 GHz and a range resolution of 10 cm. The point cloud generation algorithm employs DBSCAN clustering to filter out noise points.

[0081] In one embodiment of the present invention, the antenna health status assessment module embeds a normal operating condition benchmark model database. This database is stored in the form of lookup tables, with each antenna element corresponding to a three-dimensional array, the dimensions of which are temperature, voltage, and frequency. Calibration data is collected in a temperature-controlled test chamber using a vector network analyzer, covering the entire operating condition range. Data compression employs principal component analysis, retaining 99% of the variance information, achieving a compression ratio of 10:1.

[0082] In one embodiment of the present invention, the intelligent beam decision module is deployed on a dedicated neural network accelerator of the base station baseband processing unit. This dedicated neural network accelerator adopts an in-memory computing architecture, supports mixed-precision INT8 and FP16 operations, and has a peak computing power of 10 trillion operations per second. Complex number operations are implemented through parallel processing of the real and imaginary parts, with latency controlled within 5 milliseconds.

[0083] In one embodiment of the present invention, the beam reconfiguration execution module is connected to the beamforming register of the baseband processor via a high-speed serial interface. The interface adopts the JESD204B protocol, with a line rate of 12.5 gigabits per second, and supports eight-channel parallel transmission. Data updates are triggered at a period of 20 milliseconds, aligned with the TTI of 5G NR, ensuring the timely effectiveness of beamforming parameters.

[0084] In summary, this embodiment fully discloses an AI-based intelligent optimization method for 5G active MIMO beamforming and its supporting system components. Through strict step division, parameter visualization, sub-module refinement, and anomaly handling mechanism, robust beamforming control under antenna element anomalies or environmental obstruction conditions is achieved, meeting the requirements of full disclosure under patent law.

[0085] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0086] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI-based intelligent optimization method for 5G active MIMO beamforming, characterized in that, include: Acquire the real-time RF output signal amplitude and phase data of each antenna element in the active antenna array; collect the uplink pilot signal received by the base station to invert the downlink equivalent channel state information; Simultaneously acquire information on the distribution of obstacles in front of the antenna panel detected by the environmental perception sensor; Based on the amplitude and phase data of the real-time radio frequency output signal, and combined with the preset normal working state benchmark model, a health status confidence score for each antenna unit is generated; the channel state information, obstacle distribution information, and health status confidence score are spatiotemporally aligned to form a multi-source heterogeneous input feature tensor. The multi-source heterogeneous input feature tensor is input into a pre-trained deep reinforcement learning policy network. This deep reinforcement learning policy network takes maximizing the throughput of cell edge users and the pointing accuracy of the main lobe as the joint optimization objective and outputs the complex weighting coefficients of each effective antenna element. The digital precoding matrix of the beamformer is reconfigured according to the complex weighting coefficients to complete the online reconstruction of the beam pattern.

2. The AI-based intelligent optimization method for 5G active MIMO beamforming according to claim 1, characterized in that, To acquire the real-time RF output signal amplitude and phase data of each antenna element in an active antenna array, the following steps are taken: integrating a directional coupler and a quadrature demodulation module at the output of the power amplifier of each antenna element; digitizing the coupled signal using an analog-to-digital converter at a sampling rate of not less than twice the Nyquist frequency; extracting the in-phase and quadrature components; and then calculating the instantaneous amplitude and phase values.

3. The AI-based intelligent optimization method for 5G active MIMO beamforming according to claim 2, characterized in that, Generate a health status confidence score for each antenna element, including: establishing an amplitude-phase response database for each antenna element under normal conditions across the entire frequency band, which includes calibration curves under different temperatures, supply voltages, and operating frequencies; calculating the deviation between the currently measured amplitude and phase values ​​and the corresponding calibration curves; if the amplitude deviation is greater than 15% or the phase deviation is greater than 30 degrees, the antenna element is determined to be abnormal; and using an exponential decay function to map the deviation to a confidence score between 0 and 1, with a higher score for a larger deviation.

4. The AI-based intelligent optimization method for 5G active MIMO beamforming according to claim 3, characterized in that, The system synchronously acquires obstacle distribution information detected by environmental perception sensors in front of the antenna panel, including: periodically scanning the space within a 60-degree horizontal field of view and a 30-degree vertical field of view in front of the antenna using millimeter-wave radar arrays set around the antenna panel; identifying the distance, azimuth, pitch angle, and relative velocity of stationary or moving obstacles through time-of-flight ranging and Doppler frequency shift analysis, and generating a three-dimensional point cloud obstacle map; after voxelization, the three-dimensional point cloud obstacle map is projected onto the far-field radiation coordinate system of the antenna array to form an obstacle occlusion mask matrix.

5. The AI-based intelligent optimization method for 5G active MIMO beamforming according to claim 4, characterized in that, The channel state information, obstacle distribution information, and health status confidence score are spatiotemporally aligned to form a multi-source heterogeneous input feature tensor. This includes: representing the channel state information as a complex matrix of the number of user equipments multiplied by the number of antenna elements, with its real and imaginary parts serving as two channels; expanding the obstacle occlusion mask matrix into a single-channel binary tensor with the same dimension as the number of antenna elements; copying and stacking the health status confidence score vector into a weight matrix with the same number of rows as the channel matrix, serving as the fourth channel; and concatenating these four channels along the feature dimension to form a four-dimensional input tensor, with dimensions of the number of users, the number of antennas, the number of feature channels, and the time step.

6. The AI-based intelligent optimization method for 5G active MIMO beamforming according to claim 5, characterized in that, The multi-source heterogeneous input feature tensor is input into a pre-trained deep reinforcement learning policy network. This deep reinforcement learning policy network aims to maximize the throughput of users at the cell edge and the main lobe pointing accuracy as a joint optimization objective, and outputs the complex weighted coefficients of each effective antenna element. The deep reinforcement learning policy network adopts a dual-stream attention architecture, including a channel-aware stream and a health state stream. The channel-aware stream is composed of multi-layer complex convolutional neural networks and is used to extract spatial correlation and inter-user interference features. The health state stream is composed of fully connected layers and gated recurrent units and is used to model the temporal evolution of antenna element failure. The two stream features are fused at a higher layer through a cross-attention mechanism to generate a joint representation after attention weighting. This joint representation is input to the policy head, which is composed of complex fully connected layers and directly outputs the complex weighted coefficients of each antenna element. The magnitude of the coefficients is constrained by the health state confidence, and the phase value meets the beam pointing requirements.

7. The AI-based intelligent optimization method for 5G active MIMO beamforming according to claim 6, characterized in that, The training process of the deep reinforcement learning policy network includes: constructing a training set in a simulation environment that includes random antenna unit failure, dynamic obstacle occlusion, and multi-user movement scenarios; defining the reward function as the weighted sum of the average throughput of users at the cell edge and the main lobe direction error, wherein the main lobe direction error is determined by the angle between the desired beam direction and the actual synthetic beam maximum gain direction; and using a near-end policy optimization algorithm to iteratively update the network parameters until the reward function converges to a stable threshold.

8. The AI-based intelligent optimization method for 5G active MIMO beamforming according to claim 7, characterized in that, The digital precoding matrix of the beamformer is reconfigured based on the complex weighting coefficients, including: multiplying the output complex weighting coefficients element-wise with the original precoding vector to obtain the corrected precoding vector; loading the corrected precoding vector into the beamforming register of the baseband processor after normalization; and forcibly setting the weighting coefficients corresponding to antenna elements with health status confidence scores less than a threshold to zero to achieve physical isolation.

9. The AI-based intelligent optimization method for 5G active MIMO beamforming according to claim 8, characterized in that, The database of the normal operating state benchmark model is collected by the automated testing platform before delivery, which collects the amplitude-phase response calibration curves of each antenna unit under different ambient temperatures, power supply voltages and operating frequencies, and then compresses and stores them.

10. The AI-based intelligent optimization method for 5G active MIMO beamforming according to claim 9, characterized in that, The deep reinforcement learning policy network is deployed on the neural network accelerator of the base station baseband processing unit. The neural network accelerator supports complex number operation instruction set and the single inference latency is no more than 5 milliseconds.

Images