Mobile target signal tracking method and system based on deep learning and bayesian confidence weighting

Abstract

The application provides a mobile target signal tracking method and system based on deep learning and Bayesian confidence weighting, and belongs to the field of wireless communication testing.The application comprises the following steps: setting a first transceiving antenna and a second transceiving antenna, wherein the first transceiving antenna is used as a tracked target antenna, and the second transceiving antenna is used as a tracking antenna; obtaining forward transmission scattering parameters between the first transceiving antenna and the second transceiving antenna; performing feature extraction on the forward transmission scattering parameters through a deep learning network; performing real-time evaluation through Bayesian confidence weighting according to the extracted features to generate an observation confidence; controlling the second transceiving antenna according to the observation confidence; and repeating the process of forward transmission scattering parameter acquisition, feature extraction, real-time evaluation and second transceiving antenna control to realize mobile target signal tracking.The application can improve the tracking response speed, signal retention accuracy, testing efficiency and dynamic environment adaptability of a tracking testing device.

Description

Technical Field

[0001] This invention relates to the field of wireless communication testing and automatic tracking control technology, and more specifically, to a method and system for tracking moving target signals based on deep learning and Bayesian confidence weighting. Background Technology

[0002] As fifth-generation mobile communication continues to evolve towards the millimeter-wave band, high-frequency bands such as 28GHz and 39GHz have become important development directions for next-generation wireless communication systems due to their advantages of large bandwidth, high data rate, and low latency. However, millimeter-wave signals have strong directivity, narrow beamwidth, and high free-space loss, which places extremely stringent requirements on the beam pointing accuracy, spatial positioning accuracy, and environmental stability of antenna testing systems.

[0003] In dynamic scene testing, when the target under test undergoes displacement or attitude changes, the receiver needs to adjust the antenna pointing in real time to maintain optimal signal reception quality. Existing millimeter-wave antenna testing systems mostly employ manual adjustment or fixed-track scanning based on mechanical turntables. Manual adjustment requires operators to enter an anechoic chamber to manually adjust the antenna position, which is not only inefficient but also causes disturbances to the electromagnetic environment of the anechoic chamber by the human body and equipment, affecting test reliability. While traditional mechanical turntables can achieve a certain degree of automation, they typically use open-loop control with preset paths, lacking adaptive tracking capabilities based on real-time signal feedback, making it difficult to automatically adjust the receiving antenna pointing according to the real-time positional changes of the moving target. Patent CN118275787B proposes a full-spectrum microwave and millimeter-wave testing system that collects multi-dimensional parameters through waveform information detection and environmental information detection modules, and uses a control module to calculate waveform performance evaluation factors, improving the accuracy of waveform analysis. However, the core function of this system lies in the comprehensive evaluation of multi-parameter signal quality; its testing process focuses on signal-level analysis and judgment, without addressing real-time tracking of changes in the spatial position of the moving target or automated adjustment of the receiving antenna pointing, and lacks the ability to drive the tracking antenna to point at the moving target based on real-time signal strength feedback. Patent CN118971961B proposes an intelligent adaptive microwave and millimeter-wave testing system. Its intelligent control module can automatically select appropriate frequency bands and parameters based on target distance, resolution, environment type, and signal bandwidth, and adjust weighting coefficients through iterative optimization. However, the system's intelligent adaptive capability is mainly reflected in frequency band selection and parameter optimization at the signal generation level. Its focus is on the generation, reception, and analysis of electromagnetic signals, lacking the ability to track the spatial position of moving targets and automatically adjust the pointing of the receiving antenna based on signal strength feedback. This results in significant functional deficiencies in millimeter-wave dynamic testing scenarios requiring continuous signal tracking of moving targets.

[0004] In summary, the aforementioned existing technologies have made some progress in intelligent evaluation and adaptive parameter selection for millimeter-wave testing, but they all share a common defect: a lack of real-time signal tracking capability for moving targets. That is, they cannot automatically adjust the direction of the receiving antenna to maintain maximum signal reception strength based on changes in the position and attitude of the moving target through closed-loop feedback of signal strength. Summary of the Invention

[0005] In view of this, the present invention proposes a moving target signal tracking method and system based on deep learning and Bayesian confidence weighting to solve the problems existing in the prior art.

[0006] To achieve the above objectives, this invention proposes a method and system for tracking moving target signals based on deep learning and Bayesian confidence weighting, comprising: A first transceiver antenna and a second transceiver antenna are configured, wherein the first transceiver antenna serves as the target antenna to be tracked, and the second transceiver antenna serves as the tracking antenna; Obtain the forward transmission scattering parameters between the first and second transceiver antennas; Feature extraction of the forward propagation scattering parameters is performed using a deep learning network; Based on the extracted features, a real-time evaluation is performed using Bayesian confidence weighting to generate observation confidence levels; The second transceiver antenna is controlled based on the observation confidence level. The process of acquiring forward transmission scattering parameters, extracting features, evaluating in real time, and controlling the second transceiver antenna is repeated to achieve moving target signal tracking.

[0007] Optionally, the first transceiver antenna is adjusted in position and horizontal and pitch angles, while the second transceiver antenna is adjusted only in horizontal and pitch angles, in order to perform moving target signal tracking tests.

[0008] Optionally, the deep learning network employs LSTM and Transformer networks.

[0009] Optionally, the extracted features include trend components, fluctuation components, and transient anomaly probability estimates.

[0010] Optionally, a threshold judgment is performed on the observation confidence. Based on the judgment result, the second transceiver antenna is controlled by a gradient ascent search strategy guided by the observation confidence. When the observation confidence passes the verification and the signal strength is stable, the second transceiver antenna performs attitude lock-up and fine-tuning. When the signal strength decreases, the process is restarted and repeated.

[0011] On the other hand, the present invention provides a moving target signal tracking system based on deep learning and Bayesian confidence weighting, corresponding to the above method, comprising: A first transceiver antenna and a second transceiver antenna, wherein the first transceiver antenna is mounted on a first two-dimensional gimbal, and the first two-dimensional gimbal is mounted on a one-dimensional displacement platform; the second transceiver antenna is mounted on a second two-dimensional gimbal, and the second two-dimensional gimbal is fixedly installed. Both the first and second transceiver antennas are equipped with vector network analyzers, wherein the vector network analyzers are connected to a host computer. The host computer is connected to a main control board and a servo control board; the main control board and the one-dimensional displacement platform are used to control the position of the first transceiver antenna; the servo control board is connected to the first two-dimensional cloud platform and the second two-dimensional cloud platform to control the orientation angle of the first transceiver antenna and the second transceiver antenna.

[0012] Optionally, both the first two-dimensional cloud platform and the second two-dimensional cloud platform are equipped with servo motors for rotation in different dimensions. The servo motor drive board is connected to the servo motors in the first two-dimensional cloud platform and the second two-dimensional cloud platform for controlling the output shaft rotation of the servo motors.

[0013] Optionally, the host computer is equipped with a deep learning network and a Bayesian confidence weighting algorithm, and the host computer is wirelessly connected to the main control board and the servo control board.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) Dual-end two-dimensional gimbal architecture, flexible control and strong scene adaptability. The present invention divides the device into a moving target end (one-dimensional displacement + two-dimensional gimbal rotation and pitch) and a fixed tracking end (two-dimensional gimbal rotation and pitch). Both ends use two-dimensional gimbals to achieve angle adjustment. The displacement of the target end and the pose of the two-dimensional gimbal are manually controlled by the operator via Bluetooth through a tablet computer, which can simulate various spatial motion trajectories of the target being tested; the angle of the two-dimensional gimbal at the tracking end is automatically controlled by the host computer's deep learning and Bayesian confidence weighted algorithm, and the tracking and maintenance of the target signal is achieved only through rotation and pitch. There is no mechanical coupling between the two ends, and the test scene configuration is flexible.

[0015] (2) Closed-loop automatic signal strength tracking with fast tracking response and high accuracy. This invention uses the amplitude of the forward transmission scattering parameter (S21) between the two antennas in real time as signal strength feedback by vector network analysis. The two-dimensional gimbal at the tracking end is driven to rotate and pitch in the direction of signal enhancement by a confidence-guided gradient ascent search strategy. This achieves automatic tracking and continuous maintenance of the maximum signal strength of the moving target without the need for manual repeated entry into the anechoic chamber to adjust the antenna pointing. This significantly improves the automation level of millimeter-wave target signal tracking test.

[0016] (3) The deep learning and Bayesian confidence-weighted fusion algorithm provides strong anti-interference capabilities and high tracking confidence. A deep learning network is deployed on the host computer at the tracking end to extract features and identify transient anomalies in the vector network S21 data. An improved Bayesian confidence-weighted algorithm is introduced to dynamically adjust the tracking weights of the horizontal and pitch angles at the tracking end based on the real-time confidence of the data. This effectively solves the problem of identifying non-stationary noise and transient outliers under rapid target maneuvering conditions, overcomes the confidence fusion bottleneck under asynchronous information and dynamic coupling, and achieves high-confidence continuous tracking in complex dynamic environments. This ensures that the angle control accuracy at the tracking end is better than 1 degree, meeting the high-precision requirements for antenna beam alignment in the millimeter-wave band.

[0017] (4) Direct signal acquisition via vector network, simple hardware, and accurate measurement. This invention directly utilizes the existing vector network in the anechoic chamber for signal strength acquisition, eliminating the need for additional RF signal strength detection circuits, thus simplifying the system hardware composition. At the same time, it leverages the accuracy and dynamic range advantages of the vector network as a professional RF measurement instrument to ensure the reliability of the S21 signal strength data.

[0018] (5) Bluetooth wireless remote control plus tablet visualization operation, which is convenient to use. Operators can view all information such as the horizontal angle, pitch angle, real-time curve of vector network S21, and position status of each axis of the target end in real time through a tablet outside the dark room. They can also switch between manual calibration mode and automatic tracking mode at any time, realizing remote unmanned operation in the dark room test environment and reducing the interference of human body entering the dark room on the electromagnetic environment.

[0019] (6) Unified structure of dual-end two-dimensional gimbal, with strong versatility. This invention designs both the moving target end and the tracking end as rotation and pitch mechanisms based on two-dimensional gimbals. The target end is additionally equipped with a one-dimensional displacement platform to form a three-dimensional motion platform (displacement + horizontal rotation + pitch rotation). The tracking end is fixed at the test position and tracks and maintains the target signal only by adjusting the angle of the two-dimensional gimbal. The use of a unified two-dimensional gimbal structure at both ends reduces the complexity of mechanical design and maintenance. The antenna clamps are universal and interchangeable, and can be adapted to antennas of different sizes and types without changing the antenna clamps, thus having good versatility and scalability. Attached Figure Description

[0020] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. In the drawings: Figure 1 This is an overall structural diagram of the moving target signal tracking test device based on deep learning and Bayesian confidence weighting in an embodiment of the present invention. Figure 2 This is a schematic diagram of a two-dimensional gimbal structure shared by the tracking end and the moving target end in an embodiment of the present invention; Figure 3 This is a schematic diagram of the three-degree-of-freedom motion platform structure of the tracking end in an embodiment of the present invention; Figure 4 This is a hardware architecture block diagram of the moving target signal tracking test device in an embodiment of the present invention; Figure 5 This is a block diagram illustrating the principle of the deep learning and Bayesian belief-weighted pursuit control algorithm in this embodiment of the invention. Figure 6 This is a flowchart illustrating the automatic tracking process of moving target signals guided by deep learning and Bayesian confidence in this embodiment of the invention. Figure 7 This is a schematic diagram of the system deployment of the moving target signal tracking test device in a millimeter-wave anechoic chamber test environment according to an embodiment of the present invention. Detailed Implementation

[0021] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0022] This invention addresses the problems of existing technologies by proposing a moving target signal tracking and testing method and apparatus based on deep learning and Bayesian confidence weighting. It organically combines the multi-degree-of-freedom motion capability of the moving target with the rotation and pitch capabilities of the tracking end's two-dimensional gimbal. A deep learning network is used to extract features and identify transient anomalies in the vector network signal (VRF) S21. An improved Bayesian confidence weighting algorithm dynamically adjusts the horizontal rotation and pitch angles of the tracking end, achieving automatic tracking and continuous maintenance of the maximum signal strength of the moving target. This solves the problem of separation between target motion and signal tracking in existing technologies. Through this design, the tracking and testing apparatus of this invention exhibits significant advantages in tracking response speed, signal maintenance accuracy, testing efficiency, and adaptability to dynamic environments, providing an excellent solution for testing millimeter-wave target signals.

[0023] In light of the above, the technical solution of the present invention will be described in detail as follows: This invention provides a moving target signal tracking method and system based on deep learning and Bayesian confidence weighting. Addressing the problems in millimeter-wave band moving target signal tracking tests, such as non-stationary noise and transient anomalies in real-time measurement data from vector network analyzers, rapid changes in signal characteristic distribution introduced by target motion, and the lack of real-time state discrimination and uncertainty quantification capabilities in traditional methods when facing non-stationary dynamic signals, this invention proposes a moving target signal tracking test method and apparatus based on deep learning and Bayesian confidence weighting.

[0024] The present invention provides a moving target signal tracking method based on deep learning and Bayesian confidence weighting, comprising the following: A first transceiver antenna and a second transceiver antenna are set up. The position, horizontal angle and elevation angle of the first transceiver antenna can be adjusted by remote control. The position of the second transceiver antenna is fixed, but the horizontal angle and elevation angle can be adjusted. Collect the forward transmission scattering parameters between the first and second transceiver antennas; Feature extraction of forward propagation scattering parameters is performed using a deep learning network. The extracted features include trend components, fluctuation components, and transient anomaly probability estimates of signal strength. Based on the extracted features, the confidence of the second receiving antenna is evaluated in real time using an improved Bayesian confidence weighting algorithm to generate observation confidence, which includes the observation confidence of the two control channels corresponding to the horizontal and pitch angles. The observation confidence level is judged by a tracking threshold. When it is below the tracking threshold, it is judged as low confidence; otherwise, it is judged as high confidence. When the confidence level is low, the second transceiver antenna is controlled to make a wide range of adjustments. When the confidence level is high, the second transceiver antenna is controlled to perform a fine search. When the signal is stable and the confidence level is verified, the transceiver antenna is finely adjusted according to the confidence level. Tracking will automatically restart when the signal drops.

[0025] Specifically, the step size of large-scale adjustments and fine-tuning differs. Once the signal reaches the expected confidence level, the position is maintained, and the confidence level is checked again to proceed with the next fine-tuning to pursue higher accuracy.

[0026] Specifically, the signal stability and confidence verification process includes: performing time stability verification through sliding variance and exponential smoothing, that is, ensuring that the signal strength can indeed stabilize within the expected range for a period of time after locking; then performing spatial consistency detection through neighborhood gradient to eliminate unreliable isolated spikes; and finally verifying by comprehensively judging whether the phase is continuous and the antenna direction.

[0027] Specifically, the confidence level is dynamically adjusted by fine-tuning parameters through an adaptive parameter regulator. The static gain is dynamically adjusted by PID control to determine the priority between fast and stable signals. The confidence level is updated by Bayesian weights to ensure that the signal strength gradually converges to the expected confidence level range.

[0028] It should be noted that the first transceiver antenna is used as the target antenna to determine the position, horizontal angle, and elevation angle; the second transceiver antenna is used as the antenna passively adjusted by the aforementioned network and algorithm. Specifically, control commands are generated for the second transceiver antenna by combining PID control with a neural network feedforward compensator, and the horizontal and vertical angles of the second transceiver antenna are adjusted by the control commands.

[0029] On the other hand, the moving target signal tracking system, i.e., the moving target signal tracking system testing device provided by the present invention, relates to the following: The moving target signal tracking test device proposed in this invention is a dedicated device designed to solve the problems of dynamic uncertainty in radio frequency measurement and high-confidence pointing tracking under rapid target maneuvering in millimeter-wave band moving target signal tracking.

[0030] The tracking and testing device includes: a target end capable of three-dimensional motion (possessing three degrees of freedom: one-dimensional linear displacement and two-dimensional gimbal rotation and pitch, with a first transceiver antenna mounted on it, allowing operators to manually control the pose of each axis via Bluetooth using a tablet PC software) and a tracking end in a fixed position (possessing two degrees of freedom: two-dimensional gimbal rotation and pitch, with a second transceiver antenna mounted on it, allowing the two-dimensional gimbal angle to be automatically controlled by a deep learning and Bayesian confidence weighted algorithm from the host computer); it also includes related data acquisition and analysis modules: a vector network analyzer, a host computer, a main control board, and a servo control board.

[0031] The forward transmission scattering parameters (S21) between the two antennas are collected in real time by a vector network analyzer as signal strength feedback. The host computer at the tracking end runs a deep learning network to extract features and identify transient anomalies in the S21 data transmitted in real time by the vector network analyzer. An improved Bayesian confidence weighting algorithm is introduced. Based on the real-time confidence of the data, the horizontal rotation angle and pitch angle of the two-dimensional gimbal at the tracking end are dynamically adjusted by the control board so that the antenna at the tracking end always points in the direction of the moving target, thereby realizing automatic tracking and continuous maintenance of the maximum signal strength of the moving target.

[0032] Specifically, its motion target end uses one-dimensional linear displacement as the adjustable distance dimension and two-dimensional gimbal horizontal and pitch rotation as the adjustable attitude dimension. The three-degree-of-freedom coordinated motion can simulate various spatial motion trajectories of the target in the actual scene. Specifically, each axis of the target end is manually controlled by the operator via Bluetooth through a tablet computer software, enabling diverse target motion trajectory settings. The tracking end also adopts a two-dimensional gimbal structure, fixed in the test environment without displacement, with a minimum angular resolution of 1 degree, and has 270-degree horizontal rotation and 180-degree pitch rotation capabilities, covering the physical requirements for full-space angular tracking of moving targets at the mechanical kinematic level.

[0033] Preferably, both 2D gimbals share the same servo drive board. This board simultaneously drives two servos on the moving target gimbal and two servos on the tracking gimbal. However, only the two servos on the tracking gimbal are automatically controlled by a deep learning and Bayesian confidence weighted algorithm from the host computer, while the two servos on the moving target gimbal are manually controlled by the operator via the host computer. The host computer tablet communicates with the STM32 main control board and the servo drive board wirelessly via Bluetooth to issue control commands, achieving spatial separation between the control flow and the RF measurement flow. This avoids electromagnetic interference introduced by the test personnel entering the anechoic chamber, ensuring the purity of the measurement environment.

[0034] Specifically, at the tracking control algorithm level, this invention addresses the dynamic uncertainties caused by the exacerbation of nonlinear factors such as friction and clearance in the mechanical transmission of the two-dimensional gimbal at the tracking end at extreme positions, as well as the non-stationary transient anomalies in measurement data caused by external electromagnetic interference and rapid target maneuvering. The invention innovatively integrates a deep learning network, an improved Bayesian confidence-weighted algorithm, and a PID-neural network composite controller set in the control board within the tracking end's host computer. The deep learning network (using an LSTM time-series network and a Transformer network) extracts features from the real-time S21 time-series data transmitted by the vector network analysis (VNA), outputting the trend component, fluctuation component, and transient anomaly probability estimate of the signal strength. Based on the above feature discrimination results, the improved Bayesian confidence-weighted algorithm performs real-time evaluation and dynamic weighting of the observation confidence scores for the two control channels of the tracking end's horizontal rotation angle and pitch angle. It integrates the basic response output of the PID controller with the feedforward compensation of the neural network to generate the optimal angle control command for the two-dimensional gimbal at the tracking end. Here, the observation confidence score represents the observation reliability weight.

[0035] The process for the improved Bayesian confidence-weighted algorithm is as follows: The algorithm input parameters include: Trend components, symbolized as It comes from a deep learning network (LSTM / Transformer). It is the long-period signal trend extracted from the S21 time series, reflecting the signal intensity changes caused by the target motion.

[0036] Wave component, symbolized as Also derived from deep learning networks. It represents the short-term jitter amplitude of S21, reflecting the intensity of high-frequency noise such as multipath effects and environmental disturbances.

[0037] Transient anomaly probability, symbolized as This is also provided by a deep learning network. It is a probability estimate of whether the current sampling point belongs to non-stationary noise or outliers, with a value ranging from 0 to 1.

[0038] Current S21 amplitude, sign is It originates from the raw data of the vector network and is used to calculate the direction of the signal gradient.

[0039] PID control input, symbol: , from the PID controller, represents the basic control response output.

[0040] Feedforward compensation amount, symbolized as It comes from a neural network feedforward compensator and is used to compensate for transmission nonlinearities (such as friction and backlash).

[0041] The algorithm processing flow includes the following: Step 1: Channel confidence calculation (Bayesian posterior update) After the deep learning network outputs trend components, fluctuation components, and transient anomaly probabilities, the Bayesian model, based on historical priors, calculates the posterior channel confidence for the horizontal and pitch angle channels of the tracking device. and .

[0042] The calculation formula is:

[0043] in, Transient anomaly probability The driving force; the fluctuation component F(t) affects the variance parameter of the likelihood function. The greater the fluctuation, the more dispersed the likelihood function, resulting in lower reliability. It is a historical prior reliable probability, which will be dynamically updated as data accumulates during the tracking process.

[0044] Step 2: Dynamic weighted fusion (control quantity synthesis) Using the channel confidence level obtained in step 1 as weights, the PID basic output and neural network feedforward compensation are fused to obtain the final control input for each channel. Its expression is:

[0045] In the formula, The credibility of the corresponding channel (i.e. or ), This is the balance coefficient (adjustable parameter) between the PID controller and the feedforward. When the confidence level is high, the PID controller plus feedforward provides a normal response, resulting in full control gain; when the confidence level is low, the overall control gain is suppressed, thereby preventing error tracking caused by abnormal data. Note: In the formula... and These correspond to the feedforward compensation quantity and the PID control quantity, respectively, and the specific mapping is determined by the implementation.

[0046] Step 3: Confidence-guided gradient search (search strategy switching) In the confidence-guided gradient ascent search strategy, the channel confidence output by the Bayesian module is used as a dynamic adjustment factor for the search step size and search strategy, driving the tracking gimbal to gradually approach and lock the angle position corresponding to the maximum signal strength along the S21 signal enhancement direction.

[0047] Input data: The following real-time and historical data are also required when performing gradient search and state determination: Current S21 amplitude, sign is The data is collected in real time from the vector network.

[0048] The amplitude of S21 at the previous moment, with the sign as This comes from historical cache.

[0049] Current horizontal angle, symbol: This information is obtained from feedback from the encoder at the tracking end.

[0050] Current pitch angle, symbol: This information is obtained from feedback from the encoder at the tracking end.

[0051] Horizontal angle channel reliability, symbolized by Output by the Bayesian module.

[0052] Pitch angle channel reliability, symbolized by The output is also from the Bayesian module.

[0053] Credibility threshold, symbolized as , is a system preset parameter.

[0054] The specific process includes: In the confidence-guided gradient ascent search strategy, the channel confidence output by the Bayesian module is used as a dynamic adjustment factor for the search step size and search strategy, driving the tracking gimbal to gradually approach and lock the angle position corresponding to the maximum signal strength along the S21 signal enhancement direction.

[0055] Lock status flag, symbol is The destination is a state machine, used to determine whether the system has locked the position of the maximum signal strength.

[0056] First, calculate the aggregate confidence level. Then, based on that value and the threshold The comparison results determine the current search pattern.

[0057] if (i.e., high confidence) then enters a fine-grained search mode: using small step-size gradient ascent to approach and lock onto the signal peak; when approaching the peak, it automatically switches to an even smaller step size for fine alignment. Specifically, based on... and The sign of the difference between the two values ​​determines the search direction, and small steps are used to adjust and thus approach and lock the signal peak.

[0058] previous moment S21 amplitude Conversely, if (For low confidence levels), a coarse scan mode is used: a large step size is used to search for signal enhancement directions in order to quickly eliminate interference from abnormal data. Specifically, based on... and The sign of the difference between them determines the direction of the search signal enhancement.

[0059] Output data: During execution, the following four additional pieces of information will also be output for the next cycle of control and state management: The updated horizontal angle, symbolized as follows: The destination is the PID-neural network fusion controller, which serves as the target angle reference input for the next control cycle.

[0060] The updated pitch angle, symbolized as follows: The destination is the same as above (PID-neural network fusion controller), and it also serves as the target angle reference input for the next cycle.

[0061] Current search pattern marker, symbolized as The destination is the tracking workflow state machine, used to indicate which mode is currently in: coarse scan, fine search, lock hold, or re-tracking.

[0062] Step 4: Dynamic threshold re-tracking (adaptive trigger sensitivity) The trigger threshold for re-tracking is automatically adjusted based on the current confidence level, thereby achieving a balance between lock stability and response speed.

[0063] When the confidence level is high, the trigger threshold is increased to avoid false triggering of re-tracking due to short-term minor fluctuations, which helps to maintain the stability of the lock.

[0064] When the confidence level is low, the trigger threshold is lowered to quickly respond to signal drift and promptly trigger the re-search and tracking process.

[0065] Algorithm output: After processing, the algorithm outputs the following three key pieces of information, which are used to drive the actuator and manage the state machine: Optimal horizontal angle command, symbol: It transmits the command via Bluetooth to the STM32 main control board, and then drives the horizontal servo at the tracking end via the servo driver board, which is a fused horizontal rotation angle control command.

[0066] Optimal pitch angle command, symbol: Its transmission path is the same as the horizontal angle, ultimately driving the pitch servo at the tracking end, which is the fused pitch angle control command.

[0067] Aggregate confidence, symbolized as This value is sent to the tracking workflow state machine to determine which stage of the process the system is currently in: searching, locking, monitoring, or restarting.

[0068] In automatic tracking mode, the system uses the real-time collected S21 signal strength as feedback and drives the two-dimensional gimbal at the tracking end to rotate and pitch in the direction of signal enhancement using a confidence-guided gradient ascent search strategy, gradually approaching and locking the pointing angle corresponding to the maximum signal strength. When the target movement causes the signal strength to decrease, the Bayesian confidence weighting module dynamically triggers a re-search and tracking process based on the confidence level, thus achieving high-confidence continuous tracking and maintenance of moving target signals under non-stationary noise and rapid maneuvering conditions.

[0069] This invention also implements a Bluetooth-based wireless remote control function. Operators can manually control the one-dimensional displacement platform (speed adjustment and forward / reverse rotation) of the moving target end and the horizontal rotation and pitch angles of the target end's two-dimensional gimbal via a tablet computer software. Simultaneously, operators can manually control the horizontal rotation and pitch angles of the tracking end's two-dimensional gimbal for calibration. The software supports two working modes: manual calibration mode, where operators independently control the three axes of the target end and the two axes of the tracking end via the tablet interface, allowing for precise calibration of the initial positional relationship between the target and tracking ends; and automatic tracking mode, where, after one-button start, the tracking end's two-dimensional gimbal automatically rotates and pitches based on the output of a deep learning and Bayesian confidence-weighted algorithm, continuously tracking the moving target and maintaining maximum signal strength. During this time, the target end can be manually controlled or continue moving along a preset trajectory, completing a fully unmanned automated tracking test process. Furthermore, the tracking end's two-dimensional gimbal control algorithm incorporates an end-effector buffer strategy: when the gimbal approaches its rotation limit angle, it actively reduces the movement speed and adjusts the drive waveform to effectively suppress mechanical vibration and overshoot caused by a sudden increase in friction, ensuring the smoothness and safety of the tracking process. The one-dimensional slide rail at the moving target end is equipped with a dual safety mechanism of hardware limit switches and software limit protection, constructing a complete safety constraint system from the mechanical layer to the algorithm layer.

[0070] The above content will be described in detail with reference to the relevant accompanying drawings and specific embodiments: The moving target signal tracking test device proposed in this invention consists of a moving target end, a tracking end, a signal acquisition unit, and an intelligent control unit. The moving target end has three degrees of freedom: one-dimensional linear displacement and two-dimensional gimbal rotation and pitch. It is equipped with a first transceiver antenna, and each axis is manually controlled by the operator via Bluetooth on a tablet. The tracking end is fixed to the test environment, has two degrees of freedom: two-dimensional gimbal rotation and pitch, and is equipped with a second transceiver antenna. The angle is automatically controlled by a deep learning and Bayesian confidence weighting algorithm from a host computer. The signal acquisition unit uses a vector scattering network to obtain the amplitude of the S21 transmission coefficient between the two antennas as intensity feedback. The intelligent control unit uses a tablet computer as the host computer and integrates a deep learning feature extraction network, an improved Bayesian confidence weighting algorithm, and a PID-neural network fusion controller. It communicates with an STM32 main control board via Bluetooth and sends drive signals to both the target end (manual) and the tracking end (algorithm) via the same servo drive board.

[0071] Among them, such as Figure 1 As shown, Figure 1 This is an overall structural diagram of the moving target signal tracking test device based on deep learning and Bayesian confidence weighting of the present invention. In the moving target end, a one-dimensional displacement platform, a horizontal rotation and elevation two-dimensional gimbal, and a first transceiver antenna are set in the target end. The first transceiver antenna is set on the horizontal rotation and elevation two-dimensional gimbal, which is used to adjust the elevation angle of the first transceiver antenna. The horizontal rotation and elevation two-dimensional gimbal is set on the one-dimensional displacement platform, and the antenna is moved one-dimensionally through the one-dimensional displacement platform.

[0072] The tracking device includes a fixed base, a two-dimensional gimbal, and a second transceiver antenna. The fixed base is in a fixed position, the two-dimensional gimbal is mounted on the fixed base, and the second transceiver antenna is mounted on the two-dimensional gimbal.

[0073] The vector network analyzer is connected to two antennas via an RF cable. The vector network analyzer is wirelessly connected to a tablet computer to transmit S12 parameters. The tablet computer is connected to the STM32 main control board via Bluetooth. The STM32 main control board is connected to the same servo drive board. The servo drive board simultaneously drives four servos for the two 2D gimbals at the moving target end and the tracking end. The two servos of the tracking end 2D gimbal are automatically controlled by the host computer using deep learning and Bayesian belief weighted algorithms, while the two servos of the moving target end 2D gimbal are manually controlled by the operator.

[0074] like Figure 2 As shown, Figure 2This is a schematic diagram of a two-dimensional gimbal structure shared by the tracking end and the moving target end of the present invention. By setting a fixed direction for the servo motors, rotation in different dimensions can be achieved. The two-dimensional gimbal includes a vertically positioned servo motor, and a lateral servo motor's mounting bracket is mounted on the output shaft of the vertically positioned servo motor, controlling its horizontal rotation. The gimbal mounting surface is fixed to the output shaft of the laterally positioned servo motor, controlling the pitch angle of the gimbal mounting surface. Horizontal and pitch rotation can be achieved through the cooperation of two servo motors in a single two-dimensional gimbal. Figure 2 The arrangement of the two-dimensional gimbal base, horizontal rotation axis, and pitch rotation axis is shown. The transceiver antenna is mounted on the gimbal mounting surface by a clamp. The electrical connection between the same servo drive board and the four servos at both ends and the feedback potentiometer is also shown.

[0075] like Figure 3 As shown, Figure 3 This is a schematic diagram of the three-degree-of-freedom motion platform structure of the tracking end of the present invention. The second transceiver antenna of the tracking end is fixedly installed, and the first transceiver antenna of the target end is tracked. The first transceiver antenna is set on a two-dimensional gimbal, and the two-dimensional gimbal of the target end is set on a one-dimensional linear displacement slide rail. The one-dimensional linear displacement slide rail is moved in one dimension by a screw drive mechanism on the fixed base. The position of the screw drive mechanism is controlled by a DC geared motor. The output of the DC geared motor is controlled wirelessly or wiredly through the main control board. At the same time, front and rear limit switches are set at the boundary positions of the movement range. The front and rear limit switches are connected to the main control board. When the fixed base contacts or almost contacts the front and rear limit switches, the main control board controls the DC geared motor to stop rotating, and the screw drive mechanism stops moving to perform limit positioning. Figure 3 The installation position of the second transceiver antenna, the arrangement of the one-dimensional linear displacement slide rail (including front and rear limit switches, DC geared motor and lead screw transmission mechanism), the horizontal rotation and pitch two-dimensional gimbal, and the electrical connection between the STM32 main control board and the motor, limit switches, and servo drive board (which drives the target end and tracking end servos simultaneously) are shown.

[0076] like Figure 4 As shown, Figure 4This is a hardware architecture block diagram of the moving target signal tracking test device of the present invention, showing a unified control architecture: the STM32 main control board is responsible for the DC motor drive, limit detection, and Bluetooth communication of the one-dimensional slide rail, and is connected to the same servo drive board, which simultaneously drives the two servos of the two-dimensional gimbal at the moving target end and the two servos of the two-dimensional gimbal at the tracking end; the two servos of the two-dimensional gimbal at the tracking end receive automatic control commands from the host computer using deep learning and Bayesian confidence weighting algorithms, while the two servos of the two-dimensional gimbal at the moving target end receive manual control commands issued by the operator through the host computer. The tablet computer communicates with the STM32 main control board via Bluetooth and also connects to the vector network spectroscopy system via a standard interface to obtain S21 signal strength data.

[0077] like Figure 5 As shown, Figure 5 This is a block diagram illustrating the principle of the deep learning and Bayesian confidence-weighted tracking control algorithm of this invention. Real-time data input from the S21 vector network is processed by a deep learning feature extraction network (parallel dual-path fusion output of LSTM learning network and Transformer network) to extract trends, fluctuations, and anomaly probabilities. A transient anomaly discrimination module identifies non-stationary noise and outliers. An improved Bayesian confidence-weighted algorithm performs real-time evaluation of the observation confidence of the two control channels (horizontal rotation angle and pitch angle) at the tracking end. Then, a PID controller and a neural network feedforward compensator are fused to generate the optimal angle control command for the two-dimensional gimbal at the tracking end. Finally, the command is sent via Bluetooth to the STM32 main control board and the servo drive board. The servo drive board drives the complete closed-loop signal flow of the two servos of the two-dimensional gimbal at the tracking end. This servo drive board is simultaneously connected to the two servos of the two-dimensional gimbal at the moving target end, but its control commands are manually issued by the operator without going through the tracking algorithm.

[0078] like Figure 6 As shown, Figure 6 This is a flowchart illustrating the automatic tracking process of moving target signals guided by deep learning and Bayesian confidence in this invention. The process includes: system initialization, real-time data acquisition using the Vector Network Analysis System (S21), deep learning feature extraction and transient anomaly detection, improved Bayesian confidence assessment, confidence level comparison, branching into confidence-guided coarse scanning or fine search with gradient ascent to approach the signal peak, locking and holding until continuous monitoring is triggered by a dynamic threshold based on Bayesian confidence, and automatically restarting tracking after a signal decline is detected. The flowchart is labeled on the left with seven processing stages: initialization, feature extraction and discrimination, confidence assessment, search, locking, monitoring, and restart.

[0079] like Figure 7 As shown, Figure 7This is a schematic diagram of the system deployment of the moving target signal tracking test device of the present invention in a millimeter-wave anechoic chamber test environment. The device is set in a shell of absorbing material, specifically showing the environmental layout of the absorbing material in the anechoic chamber, the tooling installation position and movement range of the moving target end in the anechoic chamber, the fixed installation position of the tracking end in the anechoic chamber, and the spatial layout relationship of the antennas at both ends.

[0080] In the moving target signal tracking test device of the present invention, the implementation methods of the moving target end and the tracking end include, but are not limited to, the following two configuration schemes: I. Economical Integrated Implementation Method This solution integrates a one-dimensional displacement platform and a two-dimensional gimbal at the target end into a compact three-axis motion platform. Displacement is achieved by a stepper motor driving a synchronous belt slide, while angle adjustment uses digital servos with built-in position feedback. Each axis is manually controlled by the operator. The tracking end uses a fixed-base two-dimensional gimbal, driven by digital servos and automatically controlled by a host computer algorithm. Electronic control uses an ARM core board, which controls the stepper motor driver via PWM and IO, and connects to the same servo driver board to simultaneously drive four servos at both ends (two for receiving algorithm commands at the tracking end and two for receiving manual commands at the target end). Signal strength feedback is achieved by a simple RF power detection module consisting of a broadband detector and operational amplifier, converting the antenna coupling signal at the tracking end into an analog voltage for ADC sampling. The host computer communicates with the core board via USB serial port, deploying a lightweight deep learning network (one-dimensional CNN or three-layer fully connected layer) for feature extraction and anomaly detection. A Bayesian confidence weighting module then dynamically adjusts the tracking weights for the horizontal and pitch angles of the tracking end. The antenna clamp uses a quick-release elastic gripper, which is universal at both ends. The system is mounted on an aluminum alloy base plate, making it suitable for teaching demonstrations and small-scale R&D testing.

[0081] II. High-performance distributed collaborative implementation methods This solution addresses high-precision requirements by upgrading core components. The target-end displacement platform utilizes an AC servo motor paired with a ball screw and optical encoder to achieve micron-level repeatability. The horizontal pitch 2D gimbal uses a direct-drive brushless motor (0.05-degree resolution, equipped with a high-precision encoder), with each axis manually operated by the operator. The tracking-end 2D gimbal also uses a direct-drive brushless motor, fixed to a marble base or air-bearing vibration isolation platform, with its angle automatically controlled by a host computer algorithm. Electronic control employs a distributed real-time network, with each axis driver and main controller connected to a CAN bus (two drivers at the target end receive manual commands, and two at the tracking end receive algorithm commands). Interaction with the host computer is via Ethernet / WiFi, with the underlying closed-loop completed autonomously by each node. Signal acquisition is centered on a vector network gauging system, simultaneously reading the amplitude and phase of the S21 signal. The tracking algorithm is a full-featured version: a multi-layer deep neural network (LSTM / Transformer) extracts trends, fluctuations, and anomaly probabilities from the S21 time-series data; a Bayesian confidence-weighted algorithm evaluates and dynamically weights the horizontal and pitch angles of the tracking device; PID control and neural network feedforward compensation are integrated to generate optimal angle commands; a confidence-guided gradient ascent strategy drives the gimbal to rotate in the direction of signal enhancement; near peak values, it automatically switches to small-step fine-tuning search; and it automatically re-tracks when signal fluctuations exceed thresholds. Limit protection adds torque limiting in addition to hardware switches and software limits, forming triple protection.

[0082] Both solutions can be equipped with a visual coarse positioning module (global shutter camera + ring light source). The host computer quickly obtains the rough pose of the target through image recognition, guides the tracking gimbal to point in the target direction, and then the S21 provides feedback to complete the fine alignment. In addition to manual / automatic control of each axis and S21 curve display, the host computer software supports online fine-tuning of deep learning model parameters (importing pre-trained weights or incremental learning), as well as macro recording, multi-task sequence editing, and target trajectory pre-programming.

[0083] Furthermore, through a dynamic threshold adjustment mechanism, the re-tracking trigger sensitivity can be automatically changed according to the Bayesian confidence level: the threshold is increased at high confidence to avoid false triggering, and the threshold is decreased at low confidence to quickly respond to signal drift. In manual calibration mode, operators can independently control the target displacement and the 2D gimbal angle, as well as the tracking end's 2D gimbal angle; the system can record multiple calibration configurations. This invention integrates deep learning feature discrimination with Bayesian confidence weighting, allowing manual simulation of motion scenarios at the target end and automatic maintenance of maximum signal strength by the tracking algorithm, achieving high-confidence real-time tracking and continuous maintenance of moving target signals, which has significant advantages in scientific research, production, and commercial testing.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A moving target signal tracking method based on deep learning and Bayesian confidence weighting, characterized in that, include: A first transceiver antenna and a second transceiver antenna are configured, wherein the first transceiver antenna serves as the target antenna to be tracked, and the second transceiver antenna serves as the tracking antenna; Obtain the forward transmission scattering parameters between the first and second transceiver antennas; Feature extraction of the forward propagation scattering parameters is performed using a deep learning network; Based on the extracted features, a real-time evaluation is performed using Bayesian confidence weighting to generate observation confidence levels; The second transceiver antenna is controlled based on the observation confidence level. The process of acquiring forward transmission scattering parameters, extracting features, evaluating in real time, and controlling the second transceiver antenna is repeated to achieve moving target signal tracking.

2. The method according to claim 1, characterized in that, The first transceiver antenna is adjusted in position and horizontal and pitch angles, while the second transceiver antenna is adjusted only in horizontal and pitch angles, in order to perform moving target signal tracking tests.

3. The method according to claim 1, characterized in that, The deep learning network uses LSTM and Transformer networks.

4. The method according to claim 1, characterized in that, The extracted features include trend components, fluctuation components, and transient anomaly probability estimates.

5. The method according to claim 1, characterized in that, A threshold judgment is made on the observation confidence. Based on the judgment result, the second transceiver antenna is controlled by a gradient ascent search strategy guided by the observation confidence. When the observation confidence passes the verification and the signal strength is stable, the second transceiver antenna performs attitude lock and fine adjustment. When the signal strength decreases, the process is restarted and repeated.

6. A moving target signal tracking system based on deep learning and Bayesian confidence weighting, corresponding to the method described in any one of claims 1-3, characterized in that, include: A first transceiver antenna and a second transceiver antenna, wherein the first transceiver antenna is disposed on a first two-dimensional gimbal, and the first two-dimensional gimbal is disposed on a one-dimensional displacement platform. The second transceiver antenna is mounted on the second two-dimensional gimbal, which is fixedly installed. Both the first and second transceiver antennas are equipped with vector network analyzers, wherein the vector network analyzers are connected to a host computer. The host computer is connected to a main control board and a servo control board; the main control board and the one-dimensional displacement platform are used to control the position of the first transceiver antenna; the servo control board is connected to the first two-dimensional cloud platform and the second two-dimensional cloud platform to control the orientation angle of the first transceiver antenna and the second transceiver antenna.

7. The system according to claim 6, characterized in that, Both the first two-dimensional cloud platform and the second two-dimensional cloud platform are equipped with servo motors for rotation in different dimensions. The servo motor drive board is connected to the servo motors in the first two-dimensional cloud platform and the second two-dimensional cloud platform for controlling the output shaft rotation of the servo motors.

8. The system according to claim 6, characterized in that, The host computer is equipped with a deep learning network and a Bayesian confidence weighting algorithm, and the host computer is wirelessly connected to the main control board and the servo control board.

Images