A signal transmission method for a DSP-based antenna system

Abstract

The application relates to the technical field of interference cancellation, and discloses an antenna system signal transmission method based on a DSP, which comprises the following steps: establishing a parameter acquisition network of a millimeter wave antenna array, the parameter acquisition network being used for collecting target factors in real time; generating a target vector based on the target factors, inputting the target vector into a pre-trained interference prediction model, and outputting an interference probability in a future t time length; when the interference probability is greater than a preset interference probability threshold, activating a DSP operation unit integrated in a radio frequency front end, and executing a preset interference cancellation step to complete interference cancellation. The application guarantees the effect of interference cancellation and improves the reliability of an antenna system.

Description

Technical Field

[0001] This invention relates to the field of interference cancellation technology, and more specifically to a signal transmission method for an antenna system based on DSP. Background Technology

[0002] DSP, or Digital Signal Processing, is a technology and discipline that uses digital computing methods to analyze, transform, filter, and reconstruct discrete signals. It samples and quantizes analog signals (such as sound, images, and voltages collected by sensors), converts them into digital form, and then processes them in the digital domain through algorithms to achieve functions such as signal enhancement, compression, recognition, and synthesis.

[0003] In radio frequency systems, interference refers to the coupling between different signal channels or traces, which can cause signal distortion, energy crosstalk, or phase / amplitude errors. In millimeter-wave antenna arrays, when multiple antennas and radio frequency traces are arranged close together on the top layer, signals can crosstalk into or out between adjacent traces through electromagnetic fields.

[0004] A Chinese patent application with publication number CN114784467A discloses a signal transmission device and antenna system that transmits signals through multiple layers. This design utilizes a SIW (Substrate Integrated Waveguide) module on a multi-layer high-frequency PCB to transmit the antenna signal from the top layer down to the (m+1)th layer and back up to the top layer, effectively avoiding interference between the top-layer RF traces. However, this design typically requires controlling the board layer thickness to within ±10μm; the via diameter (e.g., φ0.3mm) and spacing (e.g., 1mm) error to be ≤±0.05mm; and precise matching of the coupling aperture and trapezoidal patch size. Otherwise, impedance mismatch and a sharp drop in coupling efficiency will affect performance and yield. Therefore, a DSP-based antenna system signal transmission method is proposed to ensure interference elimination. Summary of the Invention

[0005] The purpose of this invention is to provide a signal transmission method for an antenna system based on DSP, thereby solving the above-mentioned technical problems.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A signal transmission method for a DSP-based antenna system includes the following steps:

[0008] A parameter acquisition network for a millimeter-wave antenna array is established. The parameter acquisition network is used to acquire target factors in real time, which are factors that affect the interference of the millimeter-wave antenna array.

[0009] Based on the target factors, a target vector is generated, and the target vector is input into a pre-trained interference prediction model to output the interference probability within a future time period t, where t is a preset prediction step size.

[0010] When the interference probability is greater than a preset interference probability threshold, the DSP processing unit integrated in the radio frequency front end is activated, and the DSP processing unit executes preset interference elimination steps to complete the interference elimination.

[0011] As a further aspect of the present invention, the process of generating a target vector based on the target factors specifically includes:

[0012] The 360° horizontal scanning range is divided into several intervals, with each interval corresponding to an angle of 30°. The average reflection intensity of the laser echo point within each interval is calculated, where A1 represents the average reflection intensity corresponding to the first interval.

[0013] The proportion of the number of laser echo points in a single interval to the total number of laser echo points is recorded as the reference proportion, and B1 represents the reference proportion corresponding to the first interval.

[0014] Acquire the temperature change rate C1, humidity fluctuation amplitude C2, and the three-dimensional acceleration of the obstacle C3;

[0015] Generate the first vector (A1, A2, ..., A 12 B1, B2, ..., B 12 (C1, C2, C3);

[0016] The values ​​of the first vector in each dimension are standardized to eliminate dimensions, thus obtaining the target vector.

[0017] As a further aspect of the present invention, the process of pre-training the interferometric prediction model includes:

[0018] A database is established, which stores training samples labeled with interference probabilities. Each training sample is obtained by stacking the target vectors of four adjacent time points in chronological order.

[0019] An interferometric prediction model is constructed based on a spatiotemporal convolutional neural network. The interferometric prediction model is then trained and validated based on the database to obtain a pre-trained interferometric prediction model.

[0020] As a further aspect of the present invention: the interference elimination step specifically includes:

[0021] Step 1: Obtain the orientation of the obstacle relative to the millimeter-wave antenna array;

[0022] The difference in propagation path length from a single antenna element to the obstacle is determined based on the obstacle's location in space.

[0023] The phase offset value is determined based on the difference in propagation path length, and the phase offset value is sent to the corresponding antenna element. The phase offset value is used to ensure that the reflected wave cancels out the main beam.

[0024] Step 2: Divide the preset frequency range into m sub-bands at equal intervals, where m is the preset number. Send a short pulse signal at the center frequency of each sub-band. The receiving end measures the quality of the received short pulse signal and calculates the signal-to-noise ratio.

[0025] Obtain the sub-frequency band Z1 corresponding to the highest signal-to-noise ratio, and smoothly tune the local oscillator to the sub-frequency band Z1.

[0026] As a further aspect of the present invention: the training samples in the database are based on manually labeled interference probabilities.

[0027] As a further aspect of the present invention: the interference prediction model is trained using the backpropagation algorithm, and the interference prediction model is verified using five-fold cross-validation.

[0028] As a further aspect of the present invention: the process of training and validating the interferometric prediction model based on the database further includes the following steps:

[0029] The performance metrics of the validated interferometric prediction model are obtained, including mean square error, Brier score, and ECE. The performance score of the interferometric prediction model is calculated based on the superior-inferior solution distance method and the performance metrics. When the performance score is less than a preset performance score threshold, the hyperparameters of the interferometric prediction model are optimized, and the model is trained and validated again.

[0030] As a further aspect of the present invention, the methods for optimizing the hyperparameters of the interferometric prediction model include Bayesian optimization, grid search, and random search.

[0031] As a further aspect of the present invention: inputting the target vector into a pre-trained interference prediction model and outputting the interference probability within a future time period t specifically includes:

[0032] The target vectors at four consecutive time points are stacked according to the channel-space-time dimension to obtain a three-dimensional tensor;

[0033] The three-dimensional tensor is sequentially passed through H spatiotemporal convolution units, where H is the preset number of spatiotemporal convolution units. Each spatiotemporal convolution unit performs spatial convolution and temporal convolution operations.

[0034] A downsampling layer is set after any 0.3H spatiotemporal convolutional units to expand the network's ability to perceive spatiotemporal information over a wider range by progressively reducing the temporal and spatial resolution.

[0035] Residual connections are introduced between pre-defined deep convolutional units to alleviate gradient vanishing, accelerate convergence, and fuse shallow details with deep abstract features.

[0036] Global pooling is performed on the final spatiotemporal feature map to obtain a compact feature vector, which is then mapped through one or more fully connected layers and activation functions to predict the probability of interference within a future time period t.

[0037] As a further aspect of the present invention: after the DSP arithmetic unit performs the preset interference elimination steps to complete the interference elimination, the following steps are also included:

[0038] Interference probabilities greater than the aforementioned interference probability threshold are marked as abnormal probabilities. Within a preset monitoring period, the proportion of the abnormal probabilities to the total interference probabilities is calculated. If the proportion is greater than a preset proportion threshold, an early warning message is sent to alert the user.

[0039] The beneficial effects of this invention compared to the prior art are as follows:

[0040] 1) This invention constructs a parameter acquisition network on a millimeter-wave antenna array, enabling real-time acquisition of multi-dimensional environmental information such as laser echo intensity, echo quantity ratio, temperature change rate, humidity fluctuation amplitude, and obstacle three-dimensional acceleration. This information is then standardized and input into a spatiotemporal convolutional neural network model. Compared to traditional methods that rely solely on signal post-processing, this invention provides interference probability prediction before signal interference occurs, achieving "prediction-prevention" closed-loop control. The improved real-time perception and prediction capabilities not only make the system more robust to interference in complex and changing environments but also significantly reduce the time overhead of blind adjustments and trial-and-error, thereby enhancing the overall reliability and real-time performance of the communication system.

[0041] 2) When the interference probability output by the prediction model exceeds the threshold, the system can activate the DSP processing unit integrated in the RF front-end to simultaneously perform interference cancellation according to the preset phase cancellation and sub-band optimization strategy. By accurately calculating the path difference from each antenna element to the obstacle and applying a phase offset, the reflected wave and the main beam are canceled out. Combined with sub-band short pulse scanning and signal-to-noise ratio evaluation, the optimal transmission frequency band is quickly locked. This dual interference cancellation mechanism can not only finely adjust the phase and frequency, but also complete dynamic compensation at the millisecond level, effectively reducing interference residues and significantly improving signal quality and transmission efficiency.

[0042] 3) This invention employs a large-scale labeled database and constructs training samples based on target vectors from four adjacent time points. Backpropagation is used for model optimization, supplemented by five-fold cross-validation to ensure the model's generalization ability and stability. The database covers various typical environments and interference scenarios, and manually labeled interference probabilities provide high-quality supervision signals for the model. Five-fold cross-validation effectively prevents overfitting, ensuring the model maintains high prediction accuracy even in unknown environments. This comprehensive training and validation process enables the system to continuously and stably provide reliable interference predictions in actual deployment, laying a solid foundation for subsequent interference elimination.

[0043] 4) The parameter acquisition, DSP computation, and antenna phase control modules used in this solution can all be integrated into existing millimeter-wave antenna arrays and RF front-end hardware platforms without large-scale modifications to the antenna structure or the addition of expensive dedicated sensors. At the software level, only lightweight data standardization and spatiotemporal convolutional neural network inference modules need to be deployed. Combined with the DSP unit, real-time interference prediction and compensation can be completed. The overall solution takes into account both hardware and software synergy and is easy to deploy and scale up in various millimeter-wave communication, radar, and autonomous driving application scenarios, greatly reducing system upgrade and maintenance costs. Attached Figure Description

[0044] The invention will now be further described with reference to the accompanying drawings.

[0045] Figure 1 This is a flowchart illustrating a signal transmission method for an antenna system based on DSP according to the present invention. Detailed Implementation

[0046] 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.

[0047] Please see Figure 1 As shown, this invention is a signal transmission method for an antenna system based on DSP, comprising the following steps:

[0048] A parameter acquisition network for a millimeter-wave antenna array is established. The parameter acquisition network is used to acquire target factors in real time, which are factors that affect the interference of the millimeter-wave antenna array.

[0049] It is worth noting that multi-modal sensing nodes integrating temperature and humidity, micro-vibration and short-range ranging are deployed on each subarray or key element of the millimeter-wave antenna array. Low-power MCUs or small FPGAs are used to collect environmental parameters and perform filtering and calibration. Data from each node is transmitted to the central edge computing platform with millisecond-level delay through industrial Ethernet (such as TSN) or CAN-FD bus combined with IEEE1588 precise clock synchronization.

[0050] Based on the target factors, a target vector is generated, and the target vector is input into a pre-trained interference prediction model to output the interference probability within a future time period t, where t is a preset prediction step size.

[0051] Understandably, a target vector is generated from real-time collected and standardized multi-dimensional environmental parameters. This vector includes features such as the intensity and proportion of laser echoes in each zone, the rate of change of temperature and humidity, and the acceleration of obstacles. The target vectors from the current time step and the previous k-1 time steps are concatenated using a sliding window method to form a temporal input. The pre-trained interference prediction model adopts a spatiotemporal convolutional neural network structure, extracting deep features in both temporal and spatial dimensions. Convolutional layers and residual blocks capture the patterns of environmental changes. Finally, a fully connected layer and a sigmoid activation function output an interference probability between 0 and 1, which represents the confidence level of signal interference occurring within the next t time step. Model inference runs in real-time on the DSP computing unit with a controllable latency in the millisecond range. It can dynamically trigger interference cancellation strategies based on preset thresholds, thereby achieving accurate prediction and rapid response to interference.

[0052] In a preferred embodiment of the present invention, the process of generating a target vector based on the target factors specifically includes:

[0053] The 360° horizontal scanning range is divided into several intervals, with each interval corresponding to an angle of 30°. The average reflection intensity of the laser echo point within each interval is calculated, where A1 represents the average reflection intensity corresponding to the first interval.

[0054] The proportion of the number of laser echo points in a single interval to the total number of laser echo points is recorded as the reference proportion, and B1 represents the reference proportion corresponding to the first interval.

[0055] Acquire the temperature change rate C1, humidity fluctuation amplitude C2, and the three-dimensional acceleration of the obstacle C3;

[0056] Generate the first vector (A1, A2, ..., A 12 B1, B2, ..., B 12 (C1, C2, C3);

[0057] The values ​​of the first vector in each dimension are standardized to eliminate dimensions, thus obtaining the target vector;

[0058] Understandably, the 360° horizontal scanning range of the millimeter-wave antenna array is divided into 12 equal 30° intervals, and laser echo point cloud data is collected in each interval. The average reflection intensity A of the echo points within the interval is then statistically analyzed. i (i = 1…12) and the proportion B of the number of echo points in this interval to the total number of echo points. i This forms a 24-dimensional spatial feature. Simultaneously, the system also acquires environmental parameters such as the temperature change rate C1, humidity fluctuation amplitude C2, and the acceleration of obstacles in three-dimensional space C3, totaling 27 original feature vectors. The temperature change rate represents the rate of change of ambient temperature over time, reflecting the dynamic changes in the atmospheric temperature gradient. When the temperature rises or falls rapidly, it changes the refractive index and dielectric constant of the air, thus affecting the propagation path and phase stability of millimeter-wave signals. The humidity fluctuation amplitude refers to the maximum fluctuation range of ambient relative humidity. The specific calculation method is to take the difference between the maximum and minimum values ​​of the humidity sensor readings within a window period. Humidity changes affect the water vapor content in the air, thereby altering the signal absorption and attenuation characteristics, especially in millimeter-wave signals. The effect is more pronounced in the meter-wave band. The three-dimensional acceleration of an obstacle represents the magnitude of its acceleration vector in three-dimensional space. Acceleration data is acquired by deploying Doppler radar or inertial measurement units (IMUs) on moving obstacles (such as vehicles, drones, or leaves blown by the wind) or within the array's field of view, and the three-axis resultant acceleration is calculated. This acceleration reflects the intensity of the obstacle's motion, thus affecting the time-varying Doppler frequency shift and scattering characteristics of the reflected wave, playing a crucial role in predicting the timing and intensity of signal interference. Next, this first vector is standardized in each dimension using zero-mean, unit variance, or maximum / minimum normalization to eliminate dimensional differences, obtaining a target vector with a uniform scale and distribution. This target vector accurately reflects the current environment and the obstacle's motion state, providing a high-quality, easily learnable feature representation for the subsequent interferometric prediction model, thereby significantly improving prediction accuracy and system stability.

[0059] In a preferred embodiment, the process of pre-training the interferometric prediction model includes:

[0060] A database is established, which stores training samples labeled with interference probabilities. Each training sample is obtained by stacking the target vectors of four adjacent time points in chronological order.

[0061] An interferometric prediction model is constructed based on a spatiotemporal convolutional neural network. The interferometric prediction model is then trained and validated based on the database to obtain a pre-trained interferometric prediction model.

[0062] It is understood that the training samples in the database are based on manually labeled interference probabilities;

[0063] In another preferred embodiment, the interference prediction model is trained using the backpropagation algorithm and validated using five-fold cross-validation.

[0064] It should be noted that, firstly, a database containing various typical interference scenarios is established. Each training sample in this database consists of manually labeled interference probabilities and corresponding target vector sequences. The target vector sequence is formed by concatenating target vectors from four adjacent time points in chronological order to fully reflect the dynamic changes of the environment and obstacles. Based on this database, a spatiotemporal convolutional neural network model is constructed. In the time domain, the model captures the temporal correlation of the target vector sequence through a one-dimensional convolutional layer, and in the spatial domain, it learns the distribution characteristics of each target factor in different intervals and dimensions through a three-dimensional convolutional layer. During training, the backpropagation algorithm is used to optimize the network parameters, and the model's generalization ability is evaluated through five-fold cross-validation. Based on the accuracy, recall, and F1 score of the validation set, the hyperparameters are adjusted, ultimately obtaining a high-performance and robust pre-trained interference prediction model. This model can efficiently predict the signal interference probability within the next t time period during the real-time inference stage and can be seamlessly deployed on DSP computing units, achieving millisecond-level inference latency and low-power operation.

[0065] Understandably, spatiotemporal convolutional neural networks can extract features simultaneously in both time and space dimensions. By combining convolutional operations, they can capture the dynamic patterns of target factors evolving over time and the spatial correlations between different intervals. Compared to purely temporal or purely spatial models, they can more comprehensively model the multidimensional dependencies in complex scenarios in interferometric prediction tasks, thereby significantly improving prediction accuracy and system robustness.

[0066] It is worth noting that the process of training and validating the interferometric prediction model based on the database also includes the following steps:

[0067] The performance metrics of the verified interferometric prediction model are obtained, including mean square error, Brier score and ECE. The performance score of the interferometric prediction model is calculated based on the superior-inferior solution distance method and the performance metrics. When the performance score is less than a preset performance score threshold, the hyperparameters of the interferometric prediction model are optimized, and the model is trained and verified again.

[0068] It should be noted that methods for optimizing the hyperparameters of the interferometric prediction model include Bayesian optimization, grid search, and random search.

[0069] It is worth noting that by introducing performance metrics such as mean squared error, Brier score, and ECE during training and validation, and combining them with the best-or-worst solution distance method to calculate the comprehensive performance score, the model's performance on different evaluation dimensions can be quantified and compared with ideal and negative ideal solutions. This allows for a more comprehensive and objective evaluation of the interferometric prediction model's performance. When the comprehensive score falls below a preset threshold, the model's hyperparameters are systematically adjusted based on various hyperparameter optimization strategies such as Bayesian optimization, grid search, and random search, and the model is retrained and validated. This avoids the model getting trapped in local optima or overfitting, effectively improving the model's generalization ability and robustness. Simultaneously, it can quickly converge to near-optimal solutions under limited computational resources, ensuring stable and efficient interferometric prediction performance in real-world deployment environments. Furthermore, through dynamic monitoring and threshold-triggered automatic optimization, continuous iterative upgrades of the model can be achieved, reducing manual parameter tuning costs and shortening the deployment cycle, thus meeting the system's requirements for real-time performance and reliability.

[0070] When the interference probability is greater than a preset interference probability threshold, the DSP processing unit integrated in the radio frequency front end is activated, and the DSP processing unit executes preset interference cancellation steps to complete the interference cancellation;

[0071] In a preferred embodiment of the present invention, the target vector is input into a pre-trained interferometry prediction model, and the output of the interferometry probability within a future time period t specifically includes:

[0072] The target vectors at four consecutive time points are stacked according to the channel-space-time dimension to obtain a three-dimensional tensor;

[0073] The three-dimensional tensor is sequentially passed through H spatiotemporal convolution units, where H is the preset number of spatiotemporal convolution units. Each spatiotemporal convolution unit performs spatial convolution and temporal convolution operations.

[0074] A downsampling layer is set after any 0.3H spatiotemporal convolutional units to expand the network's ability to perceive spatiotemporal information over a wider range by progressively reducing the temporal and spatial resolution.

[0075] Residual connections are introduced between pre-defined deep convolutional units to alleviate gradient vanishing, accelerate convergence, and fuse shallow details with deep abstract features.

[0076] Global pooling is performed on the final spatiotemporal feature map to obtain a compact feature vector, which is then mapped through one or more fully connected layers and activation functions to predict the probability of interference within a future time period t.

[0077] It should be noted that by stacking the target vectors from four consecutive time points into a tensor in the channel-space-temporal three dimensions, and then sequentially passing them through H spatiotemporal convolutional units, the spatial convolution within each unit is used to mine the spatial correlation between different angular sectors, while the temporal convolution is used to capture the dynamic patterns of feature evolution over time, enabling the network to jointly model spatiotemporal coupled information. A downsampling layer is set after any 0.3H layer to expand the receptive field, achieving layer-by-layer aggregation of multi-scale information. Residual connections are introduced between preset deep convolutional units to alleviate gradient vanishing, accelerate convergence, and promote the fusion of shallow details and deep abstract features. Global pooling is performed on the final output spatiotemporal feature map to obtain a compact vector, which is then passed through one or more fully connected layers and activation mappings to output the interference probability within the next time period t. This balances the model's expressive power and computational efficiency, improving the accuracy, robustness, and real-time performance of interference prediction. The setting of H can be flexibly adjusted according to the complexity of the interference scenario, and the 0.3H downsampling strategy can take into account both local details and global trends, meeting the recognition needs of diverse interference patterns. Furthermore, this method can be deployed in real time on RF front-ends with limited DSP resources, and further reduces storage and computational overhead through model compression and quantization, achieving low-latency and high-reliability prediction of interference risks.

[0078] In another preferred embodiment of the present invention, the interference elimination step specifically includes:

[0079] Step 1: Obtain the orientation of the obstacle relative to the millimeter-wave antenna array;

[0080] The difference in propagation path length from a single antenna element to the obstacle is determined based on the obstacle's location in space.

[0081] The phase offset value is determined based on the difference in propagation path length, and the phase offset value is sent to the corresponding antenna element. The phase offset value is used to ensure that the reflected wave cancels out the main beam.

[0082] Step 2: Divide the preset frequency range into m sub-bands at equal intervals, where m is the preset number. Send a short pulse signal at the center frequency of each sub-band. The receiving end measures the quality of the received short pulse signal and calculates the signal-to-noise ratio.

[0083] Obtain the sub-frequency band Z1 corresponding to the highest signal-to-noise ratio, and smoothly tune the local oscillator to the sub-frequency band Z1;

[0084] It should be noted that when the interference probability output by the prediction model exceeds a preset threshold, the DSP operation unit of the RF front-end is immediately woken up via a soft interrupt or DMA trigger to schedule the interference cancellation algorithm. In step one, the high-precision positioning subsystem integrated in the array control module (which can be based on MIMO direction finding or millimeter-wave radar delay estimation algorithm) is first invoked to obtain the azimuth and elevation angles of the obstacle relative to the array, and this angle information is converted into a spatial vector to determine the precise coordinates of the obstacle in three-dimensional space. Subsequently, the DSP operation unit calculates the propagation path length difference Δd from each antenna element to the obstacle using pre-stored array correction parameters based on the array element geometry and obstacle coordinates. This calculation process is completed in parallel internally through vectorized instructions (SIMD), and all Δd values ​​can be output within microseconds. Next, the phase offset value θ = 2π*Δd / λ is calculated based on the relationship between Δd and the operating wavelength λ, and θ is mapped to a digitally programmable phase control register, which is then transmitted via high-speed SPI or I / O. 2 The C-bus sends signals to the digital phase shifters of each channel to ensure that the reflected waves from obstacles cancel out the phase of the main beam, achieving destructive interference. After phase adjustment, step two is executed: the preset communication frequency band is divided into m sub-bands at equal intervals, and the local oscillator output is rapidly switched at the center frequency of each sub-band via a digital synthesizer (DDS) to transmit short pulse signals with durations ranging from hundreds of nanoseconds to several microseconds. Simultaneously, the receiving DSP starts high-speed ADC sampling, performs windowing and FFT transformation on the received echo signals, calculates the signal power and noise power of each sub-band, and obtains the corresponding signal-to-noise ratio (SNR). The entire sub-band scanning process is implemented through a pipelined parallel design, that is, the pulse transmission and reception of the next band are started simultaneously with the completion of the measurement of one sub-band, ensuring that the SNR evaluation of all m sub-bands is completed within milliseconds. Subsequently, the DSP selects the sub-band Z1 with the highest SNR from the calculation results and performs fine-tuning of the local oscillator frequency using a smooth digital phase-locked loop (PLL) algorithm to suppress instantaneous phase jitter caused by frequency jumps, ultimately stabilizing the local oscillator at the center of the Z1 band. The entire interference cancellation process is executed in parallel within the DSP using a pipeline and interrupt priority, which not only completes environmental perception, phase correction, and frequency band switching within milliseconds, but also evaluates the cancellation effect and performs closed-loop optimization through real-time monitoring feedback loop, thereby significantly improving the system's communication stability and anti-interference capability in complex environments.

[0085] It should be noted that after the DSP processing unit performs the preset interference cancellation steps to complete the interference cancellation, the following steps are also included:

[0086] Interference probabilities greater than the aforementioned interference probability threshold are marked as abnormal probabilities. Within a preset monitoring period, the proportion of the abnormal probabilities to the total interference probabilities is calculated. If the proportion is greater than a preset proportion threshold, an early warning message is sent to alert the user.

[0087] It should be noted that the ratio threshold can be 0.5.

[0088] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A signal transmission method for an antenna system based on DSP, characterized in that, Includes the following steps: A parameter acquisition network for a millimeter-wave antenna array is established. The parameter acquisition network is used to acquire target factors in real time, which are factors that affect the interference of the millimeter-wave antenna array. Based on the target factors, a target vector is generated, and the target vector is input into a pre-trained interference prediction model to output the interference probability within a future time period t, where t is a preset prediction step size. When the interference probability is greater than a preset interference probability threshold, the DSP processing unit integrated in the radio frequency front end is activated, and the DSP processing unit executes preset interference elimination steps to complete the interference elimination.

2. The signal transmission method for a DSP-based antenna system according to claim 1, characterized in that, The process of generating a target vector based on the aforementioned target factors specifically includes: The 360° horizontal scanning range is divided into several intervals, with each interval corresponding to an angle of 30°. The average reflection intensity of the laser echo point within each interval is calculated, where A1 represents the average reflection intensity corresponding to the first interval. The proportion of the number of laser echo points in a single interval to the total number of laser echo points is recorded as the reference proportion, and B1 represents the reference proportion corresponding to the first interval. Acquire the temperature change rate C1, humidity fluctuation amplitude C2, and the three-dimensional acceleration of the obstacle C3; Generate the first vector (A1, A2, ..., A 12 B1, B2, ..., B 12 (C1, C2, C3); The values ​​of the first vector in each dimension are standardized to eliminate dimensions, thus obtaining the target vector.

3. The signal transmission method for a DSP-based antenna system according to claim 2, characterized in that, The process of pre-training the interferometric prediction model includes: A database is established, which stores training samples labeled with interference probabilities. Each training sample is obtained by stacking the target vectors of four adjacent time points in chronological order. An interferometric prediction model is constructed based on a spatiotemporal convolutional neural network. The interferometric prediction model is then trained and validated based on the database to obtain a pre-trained interferometric prediction model.

4. The signal transmission method for a DSP-based antenna system according to claim 2, characterized in that, The interference elimination step specifically includes: Step 1: Obtain the orientation of the obstacle relative to the millimeter-wave antenna array; The difference in propagation path length from a single antenna element to the obstacle is determined based on the obstacle's location in space. The phase offset value is determined based on the difference in propagation path length, and the phase offset value is sent to the corresponding antenna element. The phase offset value is used to ensure that the reflected wave cancels out the main beam. Step 2: Divide the preset frequency range into m sub-bands at equal intervals, where m is the preset number. Send a short pulse signal at the center frequency of each sub-band. The receiving end measures the quality of the received short pulse signal and calculates the signal-to-noise ratio. Obtain the sub-frequency band Z1 corresponding to the highest signal-to-noise ratio, and smoothly tune the local oscillator to the sub-frequency band Z1.

5. The signal transmission method for a DSP-based antenna system according to claim 3, characterized in that, The training samples in the database are based on manually labeled interference probabilities.

6. The signal transmission method for a DSP-based antenna system according to claim 3, characterized in that, The interference prediction model is trained using the backpropagation algorithm and validated using five-fold cross-validation.

7. The signal transmission method for a DSP-based antenna system according to claim 3, characterized in that, The process of training and validating the interferometric prediction model based on the database also includes the following steps: The performance metrics of the validated interferometric prediction model are obtained, including mean square error, Brier score, and ECE. The performance score of the interferometric prediction model is calculated based on the superior-inferior solution distance method and the performance metrics. When the performance score is less than a preset performance score threshold, the hyperparameters of the interferometric prediction model are optimized, and the model is trained and validated again.

8. The signal transmission method for a DSP-based antenna system according to claim 6, characterized in that, Methods for optimizing the hyperparameters of the interferometric prediction model include Bayesian optimization, grid search, and random search.

9. The signal transmission method for a DSP-based antenna system according to claim 1, characterized in that, The target vector is input into a pre-trained interferometry prediction model, and the output of the interferometry probability within the future time period t specifically includes: The target vectors at four consecutive time points are stacked according to the channel-space-time dimension to obtain a three-dimensional tensor; The three-dimensional tensor is sequentially passed through H spatiotemporal convolution units, where H is the preset number of spatiotemporal convolution units. Each spatiotemporal convolution unit performs spatial convolution and temporal convolution operations. A downsampling layer is set after any 0.3H spatiotemporal convolutional units to expand the network's ability to perceive spatiotemporal information over a wider range by progressively reducing the temporal and spatial resolution. Residual connections are introduced between pre-defined deep convolutional units to alleviate gradient vanishing, accelerate convergence, and fuse shallow details with deep abstract features. Global pooling is performed on the final spatiotemporal feature map to obtain a compact feature vector, which is then mapped through one or more fully connected layers and activation functions to predict the probability of interference within a future time period t.

10. The signal transmission method for a DSP-based antenna system according to claim 1, characterized in that, After the DSP arithmetic unit executes the preset interference cancellation steps to complete the interference cancellation, the following steps are also included: Interference probabilities greater than the aforementioned interference probability threshold are marked as abnormal probabilities. Within a preset monitoring period, the proportion of the abnormal probabilities to the total interference probabilities is calculated. If the proportion is greater than a preset proportion threshold, an early warning message is sent to alert the user.

Images