6g integrated sensing and communication method with deep intelligent analysis of radio frequency characteristics
By constructing a high-dimensional heterogeneous radio frequency feature tensor and a deep reinforcement learning-based autonomous tuning mechanism, the positioning error and trajectory drift problems of radio frequency sensing methods in complex scenarios are solved, achieving high-precision continuous target tracking and improving robustness and adaptability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING UNIV OF POSTS & TELECOMM
- Filing Date
- 2026-05-29
- Publication Date
- 2026-07-10
Smart Images
Figure CN122372933A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of 6G integrated sensing, intelligent analysis of radio frequency signals, target localization and dynamic tracking technology, and in particular to a 6G integrated sensing target tracking method based on radio frequency front-end fingerprint modeling, heterogeneous feature fusion, deep intelligent localization and reinforcement learning autonomous tuning. Background Technology
[0002] With the development of 6G mobile communication technology, future wireless networks will gradually possess the integrated capabilities of communication, sensing, computing, and intelligence. Integrated sensing technology has become an important direction for achieving target localization, environmental perception, and dynamic tracking. Existing radio frequency (RF) sensing methods typically rely on single or low-dimensional features such as received signal strength, channel state information, angle of arrival, and Doppler shift for target localization and state estimation, achieving certain results in ideal environments. However, in complex scenarios such as non-line-of-sight obstruction, multipath propagation, noise interference, and high-speed target maneuvering, RF signal characteristics are prone to distortion. Traditional model-driven methods struggle to maintain stable accuracy, easily leading to problems such as increased localization errors, trajectory drift, and even tracking interruptions. Simultaneously, while existing data-driven methods possess strong nonlinear feature extraction capabilities, they are heavily reliant on training data, lack cross-scenario generalization ability, and lack effective constraints on target motion patterns and RF propagation mechanisms. Therefore, there is an urgent need for a 6G integrated sensing target tracking method that can fully exploit deep RF features and integrate physical priors and intelligent learning mechanisms to improve the accuracy, robustness, and adaptability of target localization and continuous tracking in complex environments. Summary of the Invention
[0003] The purpose of this invention is to address the shortcomings and deficiencies of the existing technology by proposing a 6G integrated sensing method for deep intelligent analysis of radio frequency features. Through modeling and deep intelligent analysis of radio frequency signal features, this method can efficiently and accurately complete the target tracking task.
[0004] The technical solution adopted by this invention to solve its technical problem is: a target tracking method based on radio frequency feature depth intelligent analysis, the method comprising the following steps:
[0005] Step 1: Establish a ground-based multi-antenna base station and collect the radio frequency characteristics of the communication target;
[0006] Step 2: Establish a non-ideal characteristic model of the RF front end, and model and analyze the RF fingerprint features;
[0007] Step 3: Combine multiple dimensions of radio frequency signal features and radio frequency fingerprint features to construct a high-dimensional heterogeneous radio frequency feature tensor for 6G integrated sensing target tracking;
[0008] Step 4: Establish a task-driven deep intelligent localization model to calculate the target position and motion state;
[0009] Step 5: Construct a deep reinforcement learning-based autonomously tuned prior-data hybrid fusion tracking algorithm, which integrates the target motion physical prior, the filtered tracking model, and the deep data-driven prediction model;
[0010] Step 6: Establish an autonomous optimization mechanism based on deep reinforcement learning to correct and optimize the fusion tracking algorithm to achieve high-precision continuous target tracking.
[0011] Compared with the prior art, the technical solution of this invention has the following advantages:
[0012] 1. This invention introduces non-ideal characteristics of the radio frequency front end, such as I / Q imbalance, power amplifier nonlinearity, and phase noise, into the 6G integrated sensing target tracking process, and constructs a high-dimensional heterogeneous radio frequency feature tensor that integrates spatial, frequency, time, and hardware attributes. This can more fully explore the deep feature information in the radio frequency signal and improve the target localization and state characterization capabilities in complex environments.
[0013] 2. This invention constructs a prior-data hybrid tracking algorithm with autonomous tuning based on deep reinforcement learning. It combines the physical prior of target motion with deep data-driven correction and introduces a deep reinforcement learning mechanism to dynamically adjust the prior-data fusion weights, the intensity of data-driven correction, and the trajectory continuity constraint parameters. When the target motion pattern is obvious, it plays a stable predictive role of the physical prior model. In scenarios with non-line-of-sight occlusion, multipath interference, or strong target maneuvering, it enhances the data-driven correction role, thereby improving the accuracy, robustness, and adaptability of continuous target tracking. Attached Figure Description
[0014] Figure 1 This is a flowchart of the 6G integrated sensing method for deep intelligent analysis of radio frequency features according to the present invention;
[0015] Figure 2 This is a comparison chart of the tracking errors of the tracking method of the present invention with those of other methods;
[0016] Figure 3 This is a comparison chart of the root mean square error of the method of the present invention with other tracking methods. Detailed Implementation
[0017] The invention will now be described in further detail with reference to the accompanying drawings.
[0018] like Figure 1 As shown, this invention proposes a 6G integrated sensing method for deep intelligent analysis of radio frequency features. The specific operation steps of this method are as follows:
[0019] Step 1: Establish a ground-based multi-antenna base station and collect the radio frequency characteristics of the communication target;
[0020] The received signal is represented as:
[0021]
[0022] in, In order to receive signals, The signal transmitted by the base station. For time slots, The imaginary unit, The number of effective transmission paths, Typically represents the line-of-sight path. The amplitude attenuation factor, Indicates the signal along the first The time delay of the propagation path to the receiver satisfies , The physical length of the path. The speed of light; For Doppler frequency shift; For additive noise, it is usually modeled as having a mean of zero and a variance of 0. Complex Gaussian white noise is used to characterize receiver thermal noise and background interference;
[0023] Through the above observations, multi-dimensional radio frequency observation data containing array spatial response, subcarrier frequency response, continuous-time evolution characteristics, and radio frequency hardware attributes can be obtained, providing an input basis for subsequent radio frequency fingerprint extraction, deep intelligent positioning, and dynamic tracking.
[0024] Step 2: Establish a non-ideal characteristic model of the RF front end, and model and analyze the RF fingerprint features;
[0025] Physical modeling is performed to address the non-ideal characteristics of the radio frequency front-end in communication targets, such as I / Q imbalance, power amplifier nonlinearity, and phase noise, as important sources of the target's radio frequency fingerprint features. First, the I / Q imbalance is modeled, assuming an ideal complex baseband signal... There is an amplitude imbalance coefficient. and phase imbalance coefficient When the I / Q imbalance occurs, the signal is represented as:
[0026]
[0027] in, for The conjugate signal, and These are the I / Q imbalance coefficients, defined as follows:
[0028]
[0029]
[0030] in, This indicates the degree of amplitude mismatch between the I-path and the Q-path. This indicates the phase deviation between the I and Q paths, from which I / Q amplitude imbalance, phase imbalance, and image leakage characteristics can be extracted.
[0031] Secondly, the nonlinearity of the power amplifier is modeled. Let the input signal of the power amplifier be... The output signal of the power amplifier is then expressed as:
[0032]
[0033] in, It is a nonlinear order. For the first The higher-order nonlinear coefficients can be used to extract RF fingerprint features such as amplitude compression, AM-AM distortion, AM-PM distortion, and higher-order nonlinear coefficients.
[0034] Next, phase noise is modeled, assuming the power amplifier input signal is... The output signal of the power amplifier is then expressed as:
[0035]
[0036] Phase noise can be modeled as a random walk process:
[0037]
[0038]
[0039] in, For phase perturbation terms, This represents the phase noise variance.
[0040] Based on the above model, construct the target radio frequency fingerprint feature vector:
[0041]
[0042] in, The I / Q amplitude imbalance coefficient is used to characterize the degree of gain mismatch between the I and Q paths. The I / Q phase imbalance coefficient is used to characterize the phase deviation between the I and Q paths. These are the I / Q in-phase signal coefficients, used to characterize the degree to which ideal signal components are retained under I / Q imbalance conditions. The I / Q mirror leakage coefficient is used to characterize the leakage strength of the conjugate mirror signal. These are the first-order and third-order power amplifiers, respectively. First-order nonlinear coefficients, For carrier frequency offset, The error vector amplitude is used to characterize the hardware differences of the target device, the degree of signal distortion, and its dynamic characteristics over time.
[0043] Step 3: Combine multiple dimensions of radio frequency signal features and radio frequency fingerprint features to construct a high-dimensional heterogeneous radio frequency feature tensor for 6G integrated sensing target tracking;
[0044] After obtaining multi-antenna, multi-subcarrier, and multi-slot received signals and RF fingerprint features, the spatial, frequency, temporal, and hardware attribute dimensions are uniformly organized to construct a high-dimensional heterogeneous RF feature tensor for 6G integrated sensing target tracking, defining the first... The channel state information for each time slot is as follows:
[0045]
[0046] in, Indicates the first The receiving antenna is at the first The subcarrier, the first The complex baseband signal received in each time slot Indicates the transmitter is at the 1st The subcarrier, the first Reference signals or pilot signals transmitted in each time slot, Indicates the first The antenna, the first The channel response at each subcarrier is combined with all antenna, subcarrier, time slot, and RF fingerprint attributes to obtain a high-dimensional heterogeneous RF feature tensor:
[0047]
[0048] in, For the number of spatial antennas, The number of frequency-dimensional subcarriers. For the time dimension observation length, For the radio frequency feature dimension, each element in the tensor is represented as:
[0049]
[0050] in, For the first A radio frequency feature mapping function is used to extract corresponding feature quantities from channel response and radio frequency fingerprint;
[0051] To reduce the impact of differences in feature dimensions on deep network training, tensors are normalized:
[0052]
[0053] in, and The first The mean and standard deviation of class features, To prevent constants with zero denominators, the normalized high-dimensional heterogeneous radio frequency characteristic tensor... As input for subsequent deep intelligent localization models.
[0054] Step 4: Establish a task-driven deep intelligent localization model to calculate the target position and motion state;
[0055] Based on the high-dimensional heterogeneous radio frequency feature tensor obtained in step 3, a task-driven deep intelligent localization model is established. This model consists of a multimodal preprocessing module, a feature enhancement module, a deep feature extraction module, and a localization status output module. First, the channel phase is unwrapped, and let the... The subcarrier at the ... The original phase of each time slot is:
[0056]
[0057] Because the phase is confined to Within the range, when the target's motion causes the phase to cross the boundary, a phase jump will occur. Therefore, the unwound phase is defined as:
[0058]
[0059] in, For integer compensation terms, the values satisfy:
[0060]
[0061] To further utilize the physical consistency among multiple subcarriers, a multi-frequency joint ranging objective function is constructed:
[0062]
[0063] Among them, among them, This is the estimated distance to the target. For the first The weights of each subcarrier, For the first Subcarrier frequencies, Given the speed of light, this formula reduces the impact of single subcarrier phase noise and deep fading on ranging results through multi-subcarrier phase consistency constraints.
[0064] Then, the normalized high-dimensional heterogeneous radio frequency feature tensor is input into the deep intelligent localization network. Through an adaptive selection strategy based on dynamic gated routing, the network architecture is dynamically distributed on demand, as follows:
[0065]
[0066] in, For deep feature extraction networks, For network parameters, This is a depth feature vector.
[0067] The localization state output layer outputs the target's position and motion state based on the depth feature vector:
[0068]
[0069]
[0070] in, To locate the solution network, For network parameters, , This is the estimated two-dimensional location of the target. , This is an estimate of the target speed.
[0071] To enhance the model's adaptability in different scenarios, a cross-scenario model decision-making mechanism is further constructed. The environmental state features are defined as follows:
[0072]
[0073] in, For signal-to-noise ratio, For multipath strength index, This is a non-line-of-sight occlusion index. For location reliability, To determine the target maneuver intensity, fusion weights are generated for different localization models or feature branches based on environmental state characteristics:
[0074]
[0075] in, , and These are learnable parameters;
[0076] The final deep intelligent localization output is:
[0077]
[0078] in, The number of candidate models or feature branches. For the first Through the above mechanism, the system can adaptively select the appropriate positioning solution method based on the current channel state, target motion state, and positioning confidence level, using the output of each positioning model or feature branch.
[0079] Step 5: Construct a deep reinforcement learning-based autonomously tuned prior-data hybrid fusion tracking algorithm, which integrates the target motion physical prior, the filtered tracking model, and the deep data-driven prediction model;
[0080] Define the target state vector as:
[0081]
[0082] in, , Let the coordinates be the position coordinates of the target in a two-dimensional plane. , Given the velocity components of the target in two directions, establish the target state transition equation based on the uniform motion model:
[0083]
[0084] in, The predicted state is based on physical priors. For process noise, the state transition matrix Represented as:
[0085]
[0086] in, The time interval between adjacent observation times;
[0087] Simultaneously, a data-driven prediction model is constructed to compensate for state prediction errors caused by complex maneuvers and nonlinear propagation:
[0088]
[0089] in, For data-driven prediction networks, For network parameters, The state correction estimated by the deep model is represented as follows:
[0090]
[0091] The hybrid prediction state is obtained by weighted fusion of the physical prior prediction state and the data-driven prediction state:
[0092]
[0093] in, This is the prior data fusion weight. When the channel environment is stable and the target motion pattern is obvious, increase... Increase the weights of the physical prior model; reduce the weights when the target undergoes strong maneuvering, occlusion, or non-line-of-sight propagation. This enhances the corrective effect of data-driven models;
[0094] Therefore, the target tracking status is directly output:
[0095]
[0096] The target location is:
[0097]
[0098] The target speed is:
[0099]
[0100] The target's direction of motion is:
[0101]
[0102] Through the above steps, a hybrid tracking system combining physical prior prediction and data-driven correction is achieved, which improves tracking stability in complex propagation environments and target maneuvering scenarios while preserving the interpretability of the motion model.
[0103] Step 6: Establish an autonomous optimization mechanism based on deep reinforcement learning to correct and optimize the fusion tracking algorithm to achieve high-precision continuous target tracking;
[0104] First, define the state space of the reinforcement learning agent. At each time step, the agent receives the current environment state, localization state, and tracking state, and constructs a state vector:
[0105]
[0106] in, Given the current signal-to-noise ratio, For multipath strength index, This is a non-line-of-sight occlusion index. For location reliability, For the prior data fusion weight of the previous time step, The intensity coefficient is adjusted based on the data from the previous moment. The deviation between the physical prior predicted state and the data-driven tracking state is defined as:
[0107]
[0108] when A large discrepancy indicates a significant inconsistency between the prior physical model and the data-driven model, which may be caused by target maneuvering, occlusion, multipath mutations, or localization anomalies. The target tracking continuity error is defined as:
[0109]
[0110] This item measures the magnitude of change in the current tracking result relative to the motion model's prediction. When If the value is too large, it indicates that the target state may change abruptly, or that the current tracking results are unstable.
[0111] Define the reinforcement learning action space as:
[0112]
[0113] in, For the prior-data fusion weights in step 5, To correct the intensity coefficient based on data, To track the continuity constraint coefficients;
[0114] The reinforcement learning policy network is represented as follows:
[0115]
[0116] in, For policy networks, For policy network parameters;
[0117] Furthermore, the reinforcement learning output parameters are fed back to step 5 to fine-tune the prior-data hybrid tracking results. The tracking result in step 5 after adding continuity constraints can be expressed as:
[0118]
[0119] The third term is a continuity constraint correction term, used to suppress abrupt output changes that may occur in data-driven models under complex environments. When the positional confidence is high and the target maneuver is significant, the reinforcement learning agent reduces... Increase To enhance data-driven correction; when the channel is stable and the target motion is smooth, the reinforcement learning agent increases... , reduce To enhance the dominant role of the physical prior model; when the tracking results show drastic changes, the reinforcement learning agent increases its power. To enhance trajectory continuity constraints;
[0120] To train the reinforcement learning policy network, the reward function is defined as follows:
[0121]
[0122] in, As a reward weighting coefficient, For positional error, For speed error, To track continuity error, The deviation between the prior branch and the data branch. To determine location confidence, the agent optimizes the policy network parameters with the goal of maximizing long-term cumulative reward:
[0123]
[0124] in, As a discount factor, To track sequence length;
[0125] Through the aforementioned deep reinforcement learning self-tuning mechanism, the system can dynamically adjust the fusion weights and correction intensity in step 5 based on factors such as signal-to-noise ratio, multipath intensity, non-line-of-sight occlusion degree, location reliability, and the deviation between the prior model and the data model. This enables the prior-data hybrid tracking algorithm to possess adaptive tuning capabilities for complex scenarios. Simulation experiments comparing the proposed method with existing classical methods on the same open-source dataset yielded comparative tracking error results. Figure 2 As shown, the root mean square error is compared to... Figure 3 As shown.
[0126] It should be noted that the above description of the embodiments is only for the purpose of helping to understand the method and core idea of this application. For those skilled in the art, several improvements and modifications can be made to this application without departing from the principle of this application, and these improvements and modifications are also within the protection scope of the claims of this application.
Claims
1. A 6G integrated sensing method for deep intelligent analysis of radio frequency characteristics, characterized in that, The method includes the following steps: Step 1: Establish a ground-based multi-antenna base station and collect the radio frequency characteristics of the communication target; Step 2: Establish a non-ideal characteristic model of the RF front end, and model and analyze the RF fingerprint features; Step 3: Combine multiple dimensions of radio frequency signal features and radio frequency fingerprint features to construct a high-dimensional heterogeneous radio frequency feature tensor for 6G integrated sensing target tracking; Step 4: Establish a task-driven deep intelligent localization model to calculate the target position and motion state; Step 5: Construct a deep reinforcement learning-based autonomously tuned prior-data hybrid fusion tracking algorithm, which integrates the target motion physical prior, the filtered tracking model, and the deep data-driven prediction model; Step 6: Establish an autonomous optimization mechanism based on deep reinforcement learning to correct and optimize the fusion tracking algorithm to achieve high-precision continuous target tracking.
2. The 6G integrated sensing method for deep intelligent analysis of radio frequency characteristics according to claim 1, characterized in that, In step 1, the received signal is represented as follows: ； in, In order to receive signals, The signal transmitted by the base station. For time slots, The imaginary unit, The number of effective transmission paths, Typically represents the line-of-sight path. The amplitude attenuation factor, Indicates the signal along the first The time delay of the propagation path to the receiver satisfies , The physical length of the path. The speed of light; For Doppler frequency shift; For additive noise, it is usually modeled as having a mean of zero and a variance of 0. Complex Gaussian white noise is used to characterize receiver thermal noise and background interference; Through the above observations, multi-dimensional radio frequency observation data including array spatial response, subcarrier frequency response, continuous-time evolution characteristics, and radio frequency hardware attributes are obtained, providing an input basis for subsequent radio frequency fingerprint extraction, deep intelligent positioning, and dynamic tracking.
3. The 6G integrated sensing method for deep intelligent analysis of radio frequency characteristics according to claim 1, characterized in that, In step 2, physical modeling is performed on the I / Q imbalance, power amplifier nonlinearity, and phase noise non-ideal characteristics of the radio frequency front-end of the communication target, which serve as an important source of the target radio frequency fingerprint features. First, we model the I / Q imbalance, assuming the ideal complex baseband signal is... There is an amplitude imbalance coefficient. and phase imbalance coefficient When the I / Q imbalance occurs, the signal is represented as: ； in, for The conjugate signal, and These are the I / Q imbalance coefficients, defined as follows: ； ； in, This indicates the degree of amplitude mismatch between the I-path and the Q-path. This indicates the phase deviation between the I and Q paths, from which I / Q amplitude imbalance, phase imbalance, and image leakage characteristics can be extracted. Secondly, the nonlinearity of the power amplifier is modeled; let the input signal of the power amplifier be... The output signal of the power amplifier is then expressed as: ； in, It is a nonlinear order. For the first The higher-order nonlinear coefficients are used to extract RF fingerprint features such as amplitude compression, AM-AM distortion, AM-PM distortion, and higher-order nonlinear coefficients. Next, phase noise is modeled, assuming the power amplifier input signal is... The output signal of the power amplifier is then expressed as: ； Phase noise can be modeled as a random walk process: ； ； in, For phase perturbation terms, This represents the phase noise variance. Based on the above model, construct the target radio frequency fingerprint feature vector: ； in, The I / Q amplitude imbalance coefficient is used to characterize the degree of gain mismatch between the I and Q paths. The I / Q phase imbalance coefficient is used to characterize the phase deviation between the I and Q paths. These are the I / Q in-phase signal coefficients, used to characterize the degree to which ideal signal components are retained under I / Q imbalance conditions. The I / Q mirror leakage coefficient is used to characterize the leakage strength of the conjugate mirror signal. These are the first-order and third-order power amplifiers, respectively. First-order nonlinear coefficients, For carrier frequency offset, The error vector amplitude is used to characterize the hardware differences of the target device, the degree of signal distortion, and its dynamic characteristics over time.
4. The 6G integrated sensing method for deep intelligent analysis of radio frequency characteristics according to claim 1, characterized in that, In step 3, after obtaining the received signals from multiple antennas, multiple subcarriers, and multiple time slots, as well as the RF fingerprint features, the spatial dimension, frequency dimension, time dimension, and hardware attribute dimension are uniformly organized to construct a high-dimensional heterogeneous RF feature tensor for 6G integrated sensing target tracking, defining the first... The channel state information for each time slot is as follows: ； in, Indicates the first The receiving antenna is at the first The subcarrier, the first The complex baseband signal received in each time slot Indicates the transmitter is at the 1st The subcarrier, the first Reference signals or pilot signals transmitted in each time slot, Indicates the first The antenna, the first The channel response at each subcarrier is combined with all antenna, subcarrier, time slot, and RF fingerprint attributes to obtain a high-dimensional heterogeneous RF feature tensor: ； in, For the number of spatial antennas, The number of frequency-dimensional subcarriers. For the time dimension observation length, For the radio frequency feature dimension, each element in the tensor is represented as: ； in, For the first A radio frequency feature mapping function is used to extract corresponding feature quantities from channel response and radio frequency fingerprint; To reduce the impact of differences in feature dimensions on deep network training, tensors are normalized: ； in, and The first The mean and standard deviation of class features, To prevent constants with zero denominators, the normalized high-dimensional heterogeneous radio frequency characteristic tensor... As input for subsequent deep intelligent localization models.
5. The 6G integrated sensing method for deep intelligent analysis of radio frequency characteristics according to claim 1, characterized in that, In step 4, based on the high-dimensional heterogeneous radio frequency feature tensor obtained in step 3, a task-driven deep intelligent positioning model is established. This model consists of a multimodal preprocessing module, a feature enhancement module, a deep feature extraction module, and a positioning status output module. First, the channel phase is unwrapped, and let the... The subcarrier at the ... The original phase of each time slot is: ； Because the phase is confined to Within the range, when the target's motion causes the phase to cross the boundary, a phase jump will occur. Therefore, the unwound phase is defined as: ； in, For integer compensation terms, the values satisfy: ； To further utilize the physical consistency among multiple subcarriers, a multi-frequency joint ranging objective function is constructed: ； Among them, among them, This is the estimated distance to the target. For the first The weights of each subcarrier, For the first Subcarrier frequencies, Given the speed of light, this formula reduces the impact of single subcarrier phase noise and deep fading on ranging results through multi-subcarrier phase consistency constraints. Then, the normalized high-dimensional heterogeneous radio frequency feature tensor is input into the deep intelligent localization network. Through an adaptive selection strategy based on dynamic gated routing, the network architecture is dynamically distributed on demand, as follows: ； in, For deep feature extraction networks, For network parameters, It is a depth feature vector; The localization state output layer outputs the target's position and motion state based on the depth feature vector: ； ； in, To locate the solution network, For network parameters, , This is the estimated two-dimensional location of the target. , This is an estimate of the target speed. To enhance the model's adaptability in different scenarios, a cross-scenario model decision-making mechanism is further constructed; the environmental state features are defined as follows: ； in, For signal-to-noise ratio, For multipath strength index, This is a non-line-of-sight occlusion index. For location reliability, To determine the target maneuver intensity, fusion weights are generated for different localization models or feature branches based on environmental state characteristics: ； in, , and These are learnable parameters; The final deep intelligent localization output is: ； in, The number of candidate models or feature branches. For the first Through the above mechanism, the system can adaptively select the appropriate positioning solution method based on the current channel state, target motion state, and positioning confidence level, using the output of each positioning model or feature branch.
6. The 6G integrated sensing method for deep intelligent analysis of radio frequency characteristics according to claim 1, characterized in that, In step 5, a prior-data hybrid fusion tracking algorithm is constructed. This algorithm fuses the target motion physical prior, the filtered tracking model, and the depth data-driven prediction model to improve the accuracy of continuous target tracking. The target state vector is defined as follows: ； in, , Let the coordinates be the position coordinates of the target in a two-dimensional plane. , Given the velocity components of the target in two directions, establish the target state transition equation based on the uniform motion model: ； in, The predicted state is based on physical priors. For process noise, the state transition matrix Represented as: ； in, The time interval between adjacent observation times; Simultaneously, a data-driven prediction model is constructed to compensate for state prediction errors caused by complex maneuvers and nonlinear propagation: ； in, For data-driven prediction networks, For network parameters, The state correction estimated by the deep model is represented as follows: ； The hybrid prediction state is obtained by weighted fusion of the physical prior prediction state and the data-driven prediction state: ； in, For prior data fusion weights; when the channel environment is stable and the target motion pattern is obvious, increase Increase the weights of the physical prior model; reduce the weights when the target undergoes strong maneuvering, occlusion, or non-line-of-sight propagation. This enhances the corrective effect of data-driven models; Therefore, the target tracking status is directly output: ； The target location is: ； The target speed is: ； The target's direction of motion is: ； Through the above steps, a hybrid tracking system combining physical prior prediction and data-driven correction is achieved, which improves tracking stability in complex propagation environments and target maneuvering scenarios while preserving the interpretability of the motion model.
7. The 6G integrated sensing method for deep intelligent analysis of radio frequency characteristics according to claim 1, characterized in that, In step 6, an autonomous optimization mechanism based on deep reinforcement learning is established. According to the environmental state, location confidence, model bias and tracking continuity, the fusion weights and correction parameters required in step 5 are dynamically output, thereby realizing the closed-loop adaptive optimization of the tracking algorithm. First, define the state space of the reinforcement learning agent; at each time step, the agent receives the current environment state, localization state, and tracking state, and constructs a state vector: ；； in, Given the current signal-to-noise ratio, For multipath strength index, This is a non-line-of-sight occlusion index. For location reliability, For the prior data fusion weight of the previous time step, The intensity coefficient is adjusted based on the data from the previous moment. The deviation between the physical prior predicted state and the data-driven tracking state is defined as: ； when When the value is large, it indicates a significant inconsistency between the physical prior model and the data-driven model; The target tracking continuity error is defined as: ； This item measures the magnitude of change in the current tracking result relative to the motion model's prediction result; when If the value is too large, it indicates that the target state may change abruptly, or that the current tracking results are unstable. Define the reinforcement learning action space as: ； in, For the prior-data fusion weights in step 5, To correct the intensity coefficient based on data, To track the continuity constraint coefficients; The reinforcement learning policy network is represented as follows: ； in, For policy networks, For policy network parameters; Furthermore, the reinforcement learning output parameters are fed back to step 5 to optimize the prior-data hybrid tracking results; the tracking results in step 5 after adding continuity constraints are expressed as follows: ； The third term is a continuity constraint correction term, used to suppress sudden outputs that may occur in data-driven models under complex environments; when the positional confidence is high and the target maneuver is significant, the reinforcement learning agent reduces... Increase To enhance data-driven correction; when the channel is stable and the target motion is smooth, the reinforcement learning agent increases... , reduce To enhance the dominant role of the physical prior model; when the tracking results show drastic changes, the reinforcement learning agent increases its power. To enhance trajectory continuity constraints; To train the reinforcement learning policy network, the reward function is defined as follows: ； in, As a reward weighting coefficient, For positional error, For speed error, To track continuity error, The deviation between the prior branch and the data branch. To determine location confidence, the agent optimizes the policy network parameters with the goal of maximizing long-term cumulative reward: ； in, As a discount factor, To track sequence length.