ComplexMama-2-based urban multipath environment arrival angle stability estimation method

Abstract

The invention provides an urban multipath environment arrival angle stable estimation method based on ComplexMama-2, and the method comprises the steps: carrying out the channel gain and phase unification processing of multi-channel antenna array complex baseband data, carrying out the framing according to a time window, and carrying out the arrangement according to frequency sub-bands, and obtaining a broadband array observation sequence; a broadband Capon direction-finding spectrum is calculated, a main peak is extracted, a candidate angle of arrival is screened, candidate angle complex response is obtained according to array guide vector matching projection, and the cross-band phase alignment degree and the adjacent frame continuity degree are calculated to form multipath distinguishing features; inputting the candidate angle feature sequence into a ComplexMamba-2 complex state space network to generate a direct path tracking state, and outputting direct path credibility and a state prediction angle; target candidate angles are screened according to the credibility, time alignment is carried out under the angular velocity constraint to obtain an arrival angle continuous result, a pointing control and sector search instruction is generated according to a credibility threshold value and is issued to a photoelectric turntable, a directional antenna or an interception carrier for execution, and thus the continuity and pointing stability of the arrival angle result under the multipath condition are improved.

Description

Technical Field

[0001] This invention relates to the field of array signal processing and pointing control technology, and in particular to a method for stabilizing the angle of arrival in urban multipath environments based on ComplexMamba-2. Background Technology

[0002] In urban applications such as low-altitude protection and vehicle-mounted reconnaissance and strike, array receivers can acquire complex baseband data of UAV control / image transmission signals through multi-channel antennas, and use the array direction finding output to provide pointing information for optoelectronic turntables, directional antennas, or intercept vehicles. Because streets, buildings, glass curtain walls, and metal components in factories easily generate reflections and obstructions, the propagation link often exhibits non-line-of-sight and strong multipath superposition. At the same time, the receiver may simultaneously have direct paths and multiple reflection paths, making the direction finding output susceptible to environmental changes and resulting in unstable indications.

[0003] Existing angle-of-arrival (AOA) estimation methods are typically based on array steering vector and spatial spectrum estimation, such as Capon / MVDR, MUSIC / ESPRIT, and their broadband extensions. In engineering implementations, complex baseband data is often framed, sub-bands are divided in the frequency domain, and the sub-band covariance matrix is ​​estimated. A broadband direction-finding spectrum is obtained through cross-band power synthesis. Peak detection is then performed on the direction-finding spectrum, and the main peak or several peaks are selected as candidate directions of arrival. Response amplitude and phase information can be obtained by combining thresholding, angle grid search, and steering vector projection. To obtain continuous output, the system can also smooth and track the frame-by-frame angle results using moving averages, Kalman filtering, or threshold-based correlation strategies, and implement a sector search process after lock-out on the pointing control side.

[0004] The aforementioned methods are prone to the situation where the "spectral main peak corresponds to the reflection path" under strong multipath and non-line-of-sight conditions, causing the angle to switch between the direct path and the reflection path frame by frame. When candidate peaks are rearranged or the peak width is broadened in adjacent frames, it is difficult to stably maintain continuous tracking of the same path based on single-peak selection or simple threshold correlation. On the other hand, broadband synthesis often mainly utilizes power information and does not make sufficient use of cross-band phase consistency and the continuous characteristics of candidate angles in adjacent frames, making it difficult to provide a stable and reliable direction reference for pointing control when the environment changes rapidly.

[0005] Therefore, a stable estimation method for the angle of arrival in urban multipath environments that can overcome the shortcomings of the existing technology is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a stable angle of arrival estimation method based on ComplexMamba-2 for urban multipath environments. The core technical problem to be solved by this application is: under strong multipath / non-line-of-sight propagation conditions in urban areas, how to stably identify and continuously track the angle of arrival output that can represent the true direction of arrival from broadband array observations, so as to avoid misdirection caused by jumps between the direct path and the reflection path when driving the turntable, directional antenna or interceptor vehicle to point.

[0007] The method for stabilizing the angle of arrival in urban multipath environments based on ComplexMamba-2 according to embodiments of the present invention includes:

[0008] S1. Receive complex baseband data from a multi-channel antenna array, divide the complex baseband data into frames according to time windows and organize them according to frequency sub-bands to obtain a broadband array observation sequence;

[0009] S2. Calculate the broadband Capon direction finding spectrum for each frame based on the broadband array observation sequence to obtain the main peak of the direction finding spectrum. Filter the top K candidate angles of arrival based on the angular interval between the main peaks and the width of the main peak. Perform matching projection on the candidate angles of arrival according to the array steering vector to obtain the complex response of the candidate angles. Calculate the cross-band phase alignment degree based on the complex response of the candidate angles and calculate the duration of adjacent frames to form multipath discrimination features. Summarize the candidate angles of arrival and multipath discrimination features frame by frame to obtain the candidate angle feature sequence.

[0010] S3. Input the candidate angle feature sequence into the ComplexMamba-2 complex state space network. The ComplexMamba-2 complex state space network includes a complex input encoding layer, a complex state recursion layer, and an angle decoding layer. The complex input encoding layer performs complex encoding on the candidate angle complex response and multipath discrimination features. The complex state recursion layer generates the direct path tracking state based on the complex encoding. Under the constraint of the direct path tracking state, the cross-band phase alignment degree and the duration of adjacent frames are fused to calculate the direct path confidence. The angle decoding layer decodes the direct path tracking state to obtain the state prediction angle.

[0011] S4. Filter the candidate angles of the candidate angle feature sequence according to the direct path confidence, and perform time alignment of the candidate angles of the target under the constraint of the upper limit of angular velocity to obtain the continuous result of the arrival angle and the confidence of the target.

[0012] S5. Generate a pointing control command based on the target credibility and credibility threshold. When the target credibility is greater than or equal to the credibility threshold, the pointing control command points to the continuous result of the arrival angle. When the target credibility is less than the credibility threshold, the pointing control command points to the state prediction angle and generates a search sector command.

[0013] S6. Send pointing control commands and sector search commands to the optoelectronic turntable, directional antenna, or interceptor vehicle to drive the optoelectronic turntable, directional antenna, or interceptor vehicle to perform pointing and sector search.

[0014] Optionally, S1 is as follows:

[0015] The complex baseband data of the multi-channel antenna array is used as input. Gain and phase consistency processing is performed according to the array channel identifier. The consistency result is checked by channel amplitude threshold and phase deviation threshold to maintain the channel amplitude-phase relationship.

[0016] Complex baseband data is divided into frames according to a preset time window length. The sampling of each channel is aligned with the start and end points of the time window. A preset overlap ratio is used to control the overlap area of ​​adjacent time windows, forming a frame sequence organized by time windows.

[0017] The data of each frame is organized into frequency sub-bands according to the preset frequency sub-band boundaries. A unified frequency sub-band boundary table and sub-band order are used to bind the complex value sequence of each frequency sub-band with the channel index, forming a sub-band set organized by frequency sub-band.

[0018] The results of framing and frequency sub-banding are concatenated according to the dual indexes of time frame and frequency sub-band to obtain the broadband array observation sequence. The broadband array observation sequence contains multi-channel complex values ​​of each frame and frequency sub-band and retains the channel amplitude-phase relationship.

[0019] Optionally, S2 is as follows:

[0020] The broadband array observation sequence is input in time frames. The broadband Capon direction finding spectrum is calculated based on the multi-channel complex values ​​of each frame. The main peak of the direction finding spectrum is extracted using the peak significance threshold, and the main peak angle and main peak width are used as the main peak attributes.

[0021] Based on the angular interval and peak width between the main peaks of the direction finding spectrum, the first K candidate angles of arrival are selected using the angular interval threshold and the peak width threshold, and the selection results are divided into a set of candidate angles of arrival by frame.

[0022] For each candidate angle of arrival, a matching projection is performed according to the array steering vector to map the multi-channel complex values ​​of the broadband array observation sequence at the angle to the candidate angle complex response, and organized into vector form according to frequency sub-bands;

[0023] The cross-band phase alignment degree is calculated based on the phase difference of the candidate angle complex response in each frequency sub-band, and the phase consistency is checked by using the phase alignment threshold to obtain the cross-band phase alignment degree of each candidate angle of arrival.

[0024] The persistence of adjacent frames is calculated based on the consistency of candidate angle of arrival in adjacent time frames. A persistence threshold is used to identify consecutive occurrences, thereby obtaining the persistence of adjacent frames for each candidate angle of arrival.

[0025] The cross-band phase alignment degree and the persistence of adjacent frames are combined to form a multipath discrimination feature, and then concatenated with the candidate angle complex response one by one according to the candidate angle of arrival to obtain the structured candidate angle of arrival feature.

[0026] The candidate angle of arrival (AOA) features are aggregated frame by frame to generate a candidate angle feature sequence. The order of the candidate AOA and the frequency subband index remain unchanged, so that the candidate angle feature sequence can be used in the complex input coding layer of the ComplexMamba-2 complex state space network.

[0027] Optionally, the cross-band phase alignment degree is calculated by a phase alignment function, which is as follows:

[0028] ;

[0029] in, Candidate arrival angle The degree of cross-band phase alignment, For the number of frequency sub-bands, For frequency sub-band index, The candidate angular complex response vector In frequency subband Complex values ​​at that location, for The complex conjugate, for The range, Adjacent frequency sub-bands and The weights are adopted. Obtain, complete After calculation, the phase alignment threshold is used. right The verification will be lower than of Marked as a low-alignment candidate.

[0030] Optionally, S3 specifically refers to:

[0031] The candidate angle feature sequence is input into the ComplexMamba-2 complex state space network according to time frames. The complex input coding layer, complex state recursion layer and angle decoding layer are connected in sequence. The indices of K candidate arrival angles and frequency subband indices of each frame are passed to the complex input coding layer with each frame.

[0032] In the complex input coding layer, the candidate angle complex response corresponding to each candidate angle of arrival is used as a complex input to perform complex linear mapping to obtain the candidate angle complex embedding. The multipath discrimination feature and the candidate angle of arrival are used as the input of the gated branch to generate gate coefficients and channel-level modulation is performed on the candidate angle complex embedding. The output is a single-frame complex code bound to the frequency sub-band index.

[0033] In the complex state recursive layer, frame-by-frame complex encoding is received and the direct path tracking state is maintained. The direct path tracking state is updated by channel, and the gated complex encoding is injected into the state update, so that the direct path tracking state evolves continuously over time and serves as a constraint source for subsequent confidence calculation.

[0034] In the confidence output header of the angle decoding layer, the matching score is calculated based on the direct path tracking status and the complex embedding of the candidate angles in each frame. The channel-weighted matching rule is used to maintain the correspondence between the matching score and the candidate angle of arrival index.

[0035] Under the constraints of direct path tracking, the matching score is fused with the cross-band phase alignment degree and the persistence of adjacent frames to generate K direct path confidence scores for each frame, while retaining the influence paths of the cross-band phase alignment degree and the persistence of adjacent frames.

[0036] In the angle output head of the angle decoding layer, a state prediction angle is generated based on the direct path tracking status, so that the state prediction angle expresses the time extrapolation of the arrival angle and is bound to the corresponding time frame index;

[0037] The direct path confidence and state prediction angle are used as the output of the ComplexMamba-2 complex state space network, maintaining consistency with the frame and candidate arrival angle index of the candidate angle feature sequence, and providing input for subsequent filtering and pointing control branches.

[0038] Optionally, the reliability of the direct path is calculated using a reliable feedback function, which is specifically:

[0039] ;

[0040] in, For time frames Internal candidate index The corresponding direct path reliability, For time frame indexing, As a candidate index, The number of candidate angles of arrival per frame. To match scores, For cross-band phase alignment degree, The duration of adjacent frames, To match score weighting coefficients, This is a weighting coefficient for cross-band phase alignment. The weighting coefficients represent the duration of adjacent frames. The interaction item weight coefficient, and For the summation index, It is a frame-normalized stability coefficient and is a preset positive real constant.

[0041] Optionally, S4 specifically refers to:

[0042] The candidate angle feature sequence is read by time frame, and a binding is established between the direct path confidence of each frame and the candidate angle of arrival index to form the confidence binding result of the candidate angle of arrival set by frame, while maintaining the correspondence with the complex response of the candidate angle.

[0043] For each frame's credibility binding result, the maximum direct path credibility criterion is used to filter the target candidate angles. The filtered target candidate angles are then spliced ​​together in time frame order to form a target candidate angle sequence, and the corresponding direct path credibility and the frequency sub-band index of the candidate angle of arrival are retained.

[0044] Under the constraint of the upper limit of angular velocity, the target candidate angle sequence is time-aligned. The upper limit of angle difference is determined according to the upper limit of angular velocity constraint and the length of time window. The upper limit of angle difference criterion is used to limit the angle difference between adjacent time frames. During the time alignment process, the candidate arrival angle index and the frequency sub-band index are kept consistent to obtain the aligned target candidate angle sequence.

[0045] The aligned target candidate angle sequence is used as the continuous result of the angle of arrival, and the corresponding direct path confidence is summarized into the target confidence by time frame, keeping it consistent with the frame index of the candidate angle feature sequence.

[0046] Optional, S5 specifically includes:

[0047] Using the target confidence level and confidence level threshold as input, threshold judgment is performed on each frame to establish a branch selection relationship between the continuous results of the angle of arrival and the state prediction angle, keeping it consistent with the candidate angle of arrival index and binding it with the time frame index;

[0048] For frames where the threshold is determined to be greater than or equal to the confidence threshold, a pointing control instruction is generated, the pointing angle of the pointing control instruction is set to the continuous result of the angle of arrival, and the target confidence is bound to the time frame index.

[0049] In frames where the threshold is determined to be less than the confidence threshold, a pointing control instruction is generated and the pointing angle is set as the state prediction angle. At the same time, a search sector instruction is generated, using the state prediction angle as the center angle, and the search sector boundary is determined using a preset sector angle range. This is then bound to the time frame index, and the candidate arrival angle index and frequency sub-band index are passed along with the instruction.

[0050] The pointing control instructions and search sector instructions generated in each time frame are summarized in chronological order, keeping them consistent with the frame index of the candidate corner feature sequence, and the instruction update interval is set using the time window length.

[0051] Optionally, step S6 specifically includes:

[0052] The pointing control command and the search sector command are organized in time frame order and bound to the time frame index. The receiving target is determined for the photoelectric turntable, directional antenna or intercept vehicle, and the time frame index is kept consistent with the time frame index in the pointing control command and the search sector command.

[0053] Send a pointing control command to the receiving object, use the pointing angle in the pointing control command as the execution angle, and set the execution clock according to the time window length to maintain a consistent transmission relationship with the candidate angle of arrival index and frequency sub-band index;

[0054] After the receiving object completes the pointing, a sector search is initiated according to the sector search instruction. The center angle and sector angle range in the sector search instruction are used as the search boundary, and the execution sequence of the sector search is advanced according to the time frame index.

[0055] When the angle of arrival is continuously updated or the state prediction angle is updated, the pointing and sector search are performed cyclically according to the time frame index, so that the pointing control command and the sector search command are continuously effective on the photoelectric turntable, directional antenna or intercept vehicle.

[0056] The beneficial effects of this invention are:

[0057] (1) This proposal proposes an improved method for estimating the angle of arrival in urban multipath environments. While retaining the extraction of the main peak from array observations using the broadband Capon direction finding spectrum, the "candidate angle" is changed from a single main peak output to a set of the top K candidates constrained by the angle interval between main peaks and the width of the main peak. Furthermore, array steering vector matching projection is used for each candidate angle to obtain the complex response of the candidate angle across frequency subbands. This design enables subsequent discrimination to no longer rely solely on the power level of the direction finding spectrum, but can utilize the cross-band phase evolution information of the candidate angle complex response and the consistency information of adjacent frames to construct multipath discrimination features of cross-band phase alignment degree and adjacent frame persistence. This mechanism reduces the probability of the reflection path being directly selected due to instantaneous power dominance, providing an input basis for stably locking the path representing the true direction of arrival in the case of strong multipath and candidate peak rearrangement.

[0058] (2) This proposal proposes a novel complex state-space direct path tracking and credibility generation mechanism based on ComplexMamba-2. The candidate angle complex response is linearly mapped in complex form to obtain complex embeddings. Multipath discrimination features and candidate angles are used to generate gating coefficients. Channel-level modulation and intra-frame weighted convergence are applied to the complex embeddings to explicitly suppress the interference of low phase alignment and low persistence candidates on the temporal state during the coding stage. Subsequently, the direct path tracking state is maintained in the complex state recursion layer, so that the state evolves continuously between frames and serves as a source of constraints. In the angle decoding stage, the matching score is calculated with the candidate embeddings based on the state. The matching score is fused with the cross-band phase alignment degree, the persistence of adjacent frames and their interaction terms through a credibility feedback function to output the direct path credibility for each candidate. At the same time, the state prediction angle is obtained from the state decoding. Compared to schemes that only perform post-processing smoothing on angle results or rely solely on threshold correlation, this proposal incorporates "complex phase consistency, inter-frame persistence, and temporal state constraints" into the same generation path, enabling the credibility to have an expression that can be used for control decisions, and providing a predictive basis consistent with the tracking state when the credibility is insufficient.

[0059] (3) This proposal presents a method for stable estimation of the angle of arrival from complex baseband observations to a pointing control closed loop. First, the array amplitude and phase structure is maintained by channel gain and phase consistency. Then, a broadband array observation sequence is formed by dividing the data into frames and frequency sub-bands according to time windows. The feature sequence is organized with fixed K candidates and fixed sub-band indices to ensure consistent transmission of candidate indices and frequency sub-band indices in each processing stage, facilitating frame-by-frame backtracking and re-examination at the execution end. At the output end, target candidate angles are selected for each frame based on the direct path confidence level, and time alignment is performed under the constraint of the upper limit of angular velocity to limit unreasonable angle jumps. Furthermore, the pointing branch is controlled by a confidence threshold. At high confidence level, continuous angle of arrival results are output, and at low confidence level, the result is converted to a state prediction angle and a search sector command centered on the prediction angle is generated simultaneously. This closed-loop strategy enables photoelectric turntables, directional antennas, or interception vehicles to avoid mispointing caused by angle jumps when multipath dominance, loss of lock, or obstruction occurs, and maintains continuous handling capability for the direction of target arrival through sector search. Attached Figure Description

[0060] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0061] Figure 1 The flowchart shows a method for stabilizing the angle of arrival in urban multipath environments based on ComplexMamba-2 proposed in this invention.

[0062] Figure 2This is a flowchart of the broadband array observation sequence construction for a stable estimation method of angle of arrival in urban multipath environments based on ComplexMamba-2 proposed in this invention;

[0063] Figure 3 This is a flowchart illustrating the generation of candidate angle feature sequences for a method for stabilizing the angle of arrival in urban multipath environments based on ComplexMamba-2, as proposed in this invention.

[0064] Figure 4 The flowchart of ComplexMamba-2 complex state-space network processing for a method for stabilizing the angle of arrival in urban multipath environments based on ComplexMamba-2 proposed in this invention is shown below.

[0065] Figure 5 This is a flowchart of the target candidate angle selection and time alignment process for a stable estimation method of angle of arrival in urban multipath environments based on ComplexMamba-2 proposed in this invention.

[0066] Figure 6 This is a flowchart illustrating the generation of pointing control commands and search sector commands for a method for stabilizing the angle of arrival in urban multipath environments based on ComplexMamba-2, as proposed in this invention.

[0067] Figure 7 This is a flowchart illustrating the instruction issuance and execution closed-loop process of a method for stabilizing the angle of arrival in urban multipath environments based on ComplexMamba-2 proposed in this invention.

[0068] Figure 8 This is a schematic diagram of a non-line-of-sight direction finding and pointing method for a stable estimation method of angle of arrival in urban multipath environments based on ComplexMamba-2 proposed in this invention.

[0069] Figure 9 This is a spatial diagram of multiple candidate angles of arrival at the same time for a multipath environment angle of arrival stability estimation method based on ComplexMamba-2 proposed in this invention. Detailed Implementation

[0070] In Example 1, reference Figures 1 to 9 A method for stabilizing the angle of arrival in urban multipath environments based on ComplexMamba-2 includes:

[0071] S1. Receive complex baseband data from a multi-channel antenna array, divide the complex baseband data into frames according to time windows and organize them according to frequency sub-bands to obtain a broadband array observation sequence;

[0072] S2. Calculate the broadband Capon direction finding spectrum for each frame based on the broadband array observation sequence to obtain the main peak of the direction finding spectrum. Filter the top K candidate angles of arrival based on the angular interval between the main peaks and the width of the main peak. Perform matching projection on the candidate angles of arrival according to the array steering vector to obtain the complex response of the candidate angles. Calculate the cross-band phase alignment degree based on the complex response of the candidate angles and calculate the duration of adjacent frames to form multipath discrimination features. Summarize the candidate angles of arrival and multipath discrimination features frame by frame to obtain the candidate angle feature sequence.

[0073] S3. Input the candidate angle feature sequence into the ComplexMamba-2 complex state space network. The ComplexMamba-2 complex state space network includes a complex input encoding layer, a complex state recursion layer, and an angle decoding layer. The complex input encoding layer performs complex encoding on the candidate angle complex response and multipath discrimination features. The complex state recursion layer generates the direct path tracking state based on the complex encoding. Under the constraint of the direct path tracking state, the cross-band phase alignment degree and the duration of adjacent frames are fused to calculate the direct path confidence. The angle decoding layer decodes the direct path tracking state to obtain the state prediction angle.

[0074] S4. Filter the candidate angles of the candidate angle feature sequence according to the direct path confidence, and perform time alignment of the candidate angles of the target under the constraint of the upper limit of angular velocity to obtain the continuous result of the arrival angle and the confidence of the target.

[0075] S5. Generate a pointing control command based on the target credibility and credibility threshold. When the target credibility is greater than or equal to the credibility threshold, the pointing control command points to the continuous result of the arrival angle. When the target credibility is less than the credibility threshold, the pointing control command points to the state prediction angle and generates a search sector command.

[0076] S6. Send pointing control commands and sector search commands to the optoelectronic turntable, directional antenna, or interceptor vehicle to drive the optoelectronic turntable, directional antenna, or interceptor vehicle to perform pointing and sector search.

[0077] In this embodiment, step S1 specifically includes:

[0078] In this embodiment, the receiver of the multi-channel antenna array outputs complex baseband data of the multi-channel antenna array, assuming the number of array channels is . The channel is marked as Discrete sampling time is Each channel at the sampling time Corresponding to a complex sample value To ensure the stability of the array statistical structure in the subsequent broadband Capon direction finding spectrum calculation, and to ensure that the array phase structure is not destroyed by the channel link error when matching and projecting according to the array steering vector, gain and phase consistency processing is performed on each channel before entering the framing and frequency subband sorting.

[0079] Establish channel gain compensation coefficients for each channel based on the array channel identifier. With channel phase compensation coefficient And select a reference channel identifier. As a standardization reference, a segment of length is selected. The verification data segment consists of continuous... The sampling points consist of 100 sampling points, and the same sampling interval is used for all channels. For each channel... Calculation and benchmark channels Complex proportionality coefficient: This refers to the coefficient within the verification data segment. With reference channel The conjugates of the values ​​are multiplied point by point and averaged to obtain the phase difference and relative gain between channels. This average result is then normalized according to the average power of the reference channel's verification data segment to obtain the characteristic channel. relative channel The complex proportionality coefficient will be used to compensate for the channel gain. Set the channel phase compensation coefficient to the reciprocal of the amplitude of the complex scaling factor. Set it to the opposite of the phase of the complex proportional coefficient, and... and Together After obtaining the standardized complex baseband data, the channel amplitude threshold is then applied. Phase deviation threshold Verify the standardization results: Calculate the amplitude ratio deviation and phase difference deviation between each channel and the reference channel after standardization within the verification data segment. The amplitude ratio deviation should not exceed [the specified value]. And the phase difference deviation does not exceed The consistency result is confirmed to be valid, thus maintaining the channel amplitude phase relationship for subsequent array steering vector matching projection;

[0080] The unified complex baseband data is processed according to a preset time window length. Frames are divided and a preset overlap ratio is used. The sliding step size of adjacent frames is determined. The sliding step size is the product of the time window length and one minus the overlap ratio. For each frame, a unified start and end point of the time window is used as the frame boundary. The samples of all channels within the same frame boundary are truncated into multi-channel complex data blocks of that frame, thus forming a frame sequence organized by time window. In order to avoid the array phase structure drift caused by the intra-frame sampling misalignment between multiple channels, the sampling index of each channel is aligned with the same frame boundary to ensure that the data blocks of each channel in the same frame are synchronized in time.

[0081] For each frame, the multi-channel complex data block is processed according to the preset frequency sub-band boundary table. Perform frequency subband organization and frequency subband boundary table. The frequency range and sub-band order of each frequency sub-band are given, and the number of frequency sub-bands is given. Subband index is Perform a frequency domain transformation on each channel of each frame to obtain the complex frequency sequence of that frame, and then sort the frequency points of the complex frequency sequence according to... Mapping to the corresponding sub-band index indexes for the same sub-band The complex values ​​in the frequency domain within the sub-band are generated using a deterministic convergence rule. This convergence rule involves averaging the complex values ​​of all frequency points within the sub-band, thus forming a complex sub-band value for each channel in each frame. Each complex subband value retains phase information within its subband, which is used to form candidate angle complex responses and support cross-band phase alignment calculations.

[0082] The framing results and frequency subband organization results are indexed by time frame. With frequency subband index Double-index concatenation is performed to obtain the broadband array observation sequence. Each of them For length is The multi-channel complex value vector contains the complex subband values ​​of each channel in the frequency subband of the time frame, and retains the channel amplitude-phase relationship. The broadband array observation sequence is used as a unified input for subsequent steps to calculate the broadband Capon direction finding spectrum frame by frame and extract the main peak of the direction finding spectrum. It is also used to obtain the candidate angle complex response by matching and projecting it according to the array steering vector after the candidate angle of arrival is determined. This ensures that the candidate angle complex response contains the stable array phase structure and the phase evolution information between the frequency subbands, which meets the pre-data requirements of the ComplexMamba-2 complex state space network for the candidate angle feature sequence input structure.

[0083] In this embodiment, step S2 specifically includes:

[0084] Step S2 uses a broadband array observation sequence As input, where For time frame indexing, This is the frequency sub-band index, and the number of frequency sub-bands is... each For length is The multi-channel complex-valued vectors, in order to support stable estimation of the broadband Capon direction-finding spectrum, are used in each time frame. The corresponding complex baseband data will be processed according to the preset number of snapshots. Divided into Each intra-frame snapshot is processed, and frequency sub-band adjustment is performed on each intra-frame snapshot in the same manner as in step S13, to obtain the intra-frame snapshot set. ,in For intra-frame snapshot index and The range of values ​​is to ,Will The array covariance estimate is used to construct the covariance estimate for each frequency subband to avoid the covariance becoming irreversible due to a single vector;

[0085] For each frame Frequency subband Calculate the array covariance matrix The calculation method is to The multi-channel complex value vectors of each intra-frame snapshot are externally accumulated, summed, and averaged, with diagonal loading coefficients applied. right Diagonal loading is performed to ensure reversibility, followed by angle-searching the grid. The broadband Capon direction finding spectrum is calculated angle by angle, where This is a set of angles with a preset angle step size, where each element in the set is a candidate angle. For each angle With each frequency sub-band Calculate the array steering vector ,Will and The Capon power estimate for the subband is obtained by performing a quadratic operation on the inverse matrix. Then, the Capon power estimates for all frequency subbands are summed by subband index and normalized to obtain the broadband Capon direction-finding spectrum of the frame. ,exist The above uses the peak significance threshold. Extracting the main peak of the direction-finding spectrum: Perform local maxima detection on the angle grid, calculate the spectral difference of each local maximum relative to its left and right neighbors as significance, and filter out peaks with significance less than 1. The peak value was recorded, and the angle of the main peak was preserved. With the width of the main peak ,in The main peak index is used to distinguish frames within the same time period. Within different main peaks, the width of the main peak is determined by searching the spectral values ​​around that peak and decreasing them to the peak value multiplied by a preset decrease threshold. The angle difference at corresponding positions is obtained, where The relative decrease percentage threshold;

[0086] According to the same frame The main peak angles of all internal directional spectra With the width of the main peak To filter candidate angles of arrival, first sort the main peaks by significance from highest to lowest, then iterate through them and apply angle interval constraints: for the current main peak, calculate its angle difference with the set of selected main peaks; if the angle difference is less than the angle interval threshold... The peak is discarded to avoid generating too many candidates near a single wide peak, and then the peak width constraint is applied: the peak width is greater than the peak width threshold. Discarding the main peak at that time suppresses unstable peaks formed by multipath broadening. Selecting peaks sequentially according to the above constraints continues until the previous peak is obtained. candidate arrival angles ,in Candidate index and The range of values ​​is to When the number of effective main peaks is insufficient When using a preset fill angle The candidate arrival angles are supplemented and the corresponding candidates are marked as filler candidates to maintain the fixed input structure of the subsequent candidate angle feature sequences;

[0087] For each candidate angle of arrival Matching projection is performed according to the array steering vector to obtain the candidate angular complex response; specifically, for each frequency sub-band... Calculate the array steering vector and compared it with the broadband array observation sequence of that frame. The complex response components with angle matching are obtained by performing a complex inner product. To avoid the influence of the steering vector energy difference between different angles on the response amplitude, the complex inner product result is normalized according to the steering vector energy. The complex response components of all frequency sub-bands are then indexed by frequency sub-band. By concatenating the sequences sequentially, we obtain the candidate angle complex response vector. and its first The complex values ​​of each frequency sub-band are denoted as ,in Include Each complex value retains the phase evolution information between frequency subbands;

[0088] Based on candidate angle complex response vector Calculate cross-band phase alignment The accumulation is performed using normalized complex correlation of adjacent frequency sub-bands. During accumulation, higher-energy adjacent sub-band pairs are assigned higher weights, which are obtained by multiplying the complex response amplitudes of adjacent sub-bands. This determines the cross-band phase alignment. Calculated using the phase alignment function:

[0089] ;

[0090] in, Candidate arrival angle The degree of cross-band phase alignment, For the number of frequency sub-bands, For frequency sub-band index, The candidate angular complex response vector In frequency subband Complex values ​​at that location, for The complex conjugate, for The range, Adjacent frequency sub-bands and The weights are adopted. Obtain, complete After calculation, the phase alignment threshold is used. right The verification will be lower than of Marked as low-aligned candidates for subsequent multipath discrimination feature construction;

[0091] The persistence of adjacent frames is calculated based on the consistency of candidate angle of arrival indices in adjacent time frames. To handle the rearrangement of candidate sets between adjacent frames, an angle-related threshold is used. Associate candidates: For candidate angles of arrival in the current frame The search angle difference in the candidate angle of arrival set of the previous frame does not exceed [a certain value]. The nearest neighbor candidate is identified and a relationship is established. If a relationship is established, the persistence count is incremented by one; otherwise, the persistence count is reset to zero. The persistence count is maintained for a preset persistence window length. The interval is truncated and normalized to obtain the duration of adjacent frames, with values ​​ranging from zero to one. Using a persistence threshold Identifying consecutive occurrences will be lower than of Marked as a low-persistence candidate;

[0092] Cross-band phase alignment Duration of adjacent frames Merging to form multipath distinguishing features ,in Composed of two real numbers in a fixed order, corresponding to the cross-band phase alignment degree and the duration of adjacent frames respectively, the candidate angular complex response vector is... Multipath Differentiation Features With candidate angle of arrival By candidate index By piecing them together one by one, structured candidate angle of arrival features are obtained. , the same frame Inside The structured candidate angle of arrival features are summarized in candidate index order to obtain the candidate angle feature sequence. and indexed by time frame across all time frames. Concatenation to form candidate angular feature sequences Candidate angular feature sequence Keeping the candidate angle of arrival order and frequency sub-band index unchanged, the candidate angle feature sequence It can directly input the complex input coding layer of the ComplexMamba-2 complex state space network and maintain a structural correspondence that is consistent with the multipath discrimination features of the subsequent gated branches.

[0093] In this embodiment, step S3 specifically includes:

[0094] Step S3 uses candidate angle feature sequences As input, Indexed by time frame Organization, time frame number Each frame contains Structured candidate angle of arrival features ,in For candidate indexes, each From the candidate angular complex response vector Multipath Differentiation Features With candidate angle of arrival constitute, Indexed by frequency subband Organization and containing Complex values ​​of a frequency sub-band Based on cross-band phase alignment Duration of adjacent frames Constructed in a fixed order, the network will be implemented as follows: It is formed by splitting into real and imaginary channels. One real number channel, and with , , Concatenate them into the encoded input of the complex input encoding layer;

[0095] Candidate angle feature sequence The data is input into the ComplexMamba-2 complex state space network in time frame order. The ComplexMamba-2 complex state space network consists of a complex input coding layer, a complex state recursion layer, and an angle decoding layer. For each frame, candidate indices are... With frequency subband index The frame is passed to the complex input coding layer, so that the complex input coding layer remains... The frequency subband order remains unchanged, and is maintained With candidate index The correspondence;

[0096] The complex input coding layer employs a candidate angle complex response coding branch concatenated with a multipath discrimination feature gating branch. The candidate angle complex response coding branch... Perform a complex linear mapping to obtain the complex embedding of candidate angles. ,in Include Multiple channels, For the complex embedding dimension, a complex linear mapping multiplies and adds the real and imaginary channels respectively, and then cross-combines them to generate the output real and output imaginary channels, so that the array phase structure corresponding to the angle of arrival is in Maintaining multipath differentiation features and gating branches and As a real number input, it is generated via a real number fully connected neuron. Each gate coefficient And use a preset activation function to adjust the gating coefficient Limit the gating coefficient to zero to one. right Perform channel-level modulation to obtain gated candidate angle complex embeddings The modulation method is to The real and imaginary channels are scaled synchronously to obtain executable single-frame complex encoding. For each candidate index based on Calculate candidate gating scores Candidate Gating Score Adopting The gating coefficients are obtained by taking the arithmetic mean of all gating coefficients within the same time frame. Normalization is performed by using... Divide by Obtain candidate convergence weights ,in For candidate summation index, To preset positive real stability coefficients, finally each By candidate convergence weight Perform complex weighted summation to obtain the complex code of a single frame. and make With time frame index Bind, while retaining each Used for subsequent matching score calculation;

[0097] Complex state recursion layer receives frame-by-frame complex encoding And maintain the direct path tracking status. ,in Include A complex number of channels and initial state The complex state recursive layer employs either all-zero complex vectors or learnable initial vectors, and utilizes ComplexMamba-2 complex state space unit pairs. Perform a channel-by-channel state update, which includes a state preservation component and an input injection component: The state-preserving components are obtained through complex linear transformation. The input injection component is obtained through complex linear transformation, and then the two are combined according to channels to form... The input injection component is determined by the gated... The convergence of resources enables right and The common representation is time-continuous information, which is passed to the angle decoding layer as a constraint source for subsequent direct path credibility calculation;

[0098] The confidence output head of the angle decoding layer is in direct path tracking state. Complex embedding of gated candidate angles in each frame Calculate matching score The matching score calculation uses a channel-weighted matching rule: for and Channel-by-channel complex multiplication and accumulation is performed, where each channel is multiplied by the complex conjugate of the state channel value and the candidate angle complex embedding channel value, and then multiplied over all channels. The channels are summed to obtain a complex matching quantity. The magnitude of the complex matching quantity is then taken to obtain the channel matching quantity. The channel matching quantity is then distributed according to the channel weight vector. Weighted summation ,in Let be a learnable real weight vector and be... Each complex channel corresponds to a certain number of channels, thus maintaining... With candidate index correspond;

[0099] In step S35, in the direct path tracking state Matching scores under constraints Phase alignment with cross-band Duration of adjacent frames The reliability of the direct path is obtained by fusion. , integration adopts Perform intra-frame normalization and introduce and Interaction items are used to maintain the explicit participation of multipath distinguishing features in the credibility generation path, direct path credibility. Calculated using a reliable feedback function:

[0100] ;

[0101] in, For time frames Internal candidate index The corresponding direct path reliability, For time frame indexing, As a candidate index, The number of candidate angles of arrival per frame. To match scores, For cross-band phase alignment degree, The duration of adjacent frames, To match score weighting coefficients, This is a weighting coefficient for cross-band phase alignment. The weighting coefficients represent the duration of adjacent frames. The interaction item weight coefficient, and For the summation index, It is an intra-frame normalized stability coefficient and a preset positive real constant;

[0102] The angle output head of the angle decoding layer tracks the state based on the direct path. Generate state prediction angle State prediction angle Real angle and indexed by time frame Binding, angle output head will The real and imaginary channels are concatenated to form a real state vector, which is then input into a fully connected neuron and output from it. and to Apply a preset angle range constraint to satisfy the angle representation domain;

[0103] Determine the reliability of the direct path for each frame. With state prediction angle As the output of the ComplexMamba-2 complex state-space network, the output is kept consistent with the candidate angular feature sequence. Time frame index Consistency, and maintaining the credibility of the direct route. With candidate index This consistency ensures that the candidate angles are selected based on the direct path confidence level, and that the pointing control instructions and search sector instructions are generated under the confidence level threshold branch.

[0104] In this embodiment, step S4 specifically includes:

[0105] Step S4 uses the aforementioned candidate angle feature sequence The output of the ComplexMamba-2 complex state-space network is used as the input to perform target candidate angle filtering and time alignment, assuming the time frame index is... and The range of values ​​is to Candidate index is and The range of values ​​is to The frequency subband index is and The range of values ​​is to Each frame The corresponding candidate angle of arrival is Each frame The corresponding direct route reliability is Each frame The corresponding candidate angular complex response vector is and Indexed by frequency subband organize;

[0106] Candidate angle feature sequence Indexed by time frame Read and extract all candidate angles of arrival within the same time frame. With candidate angular complex response vector And read the reliability of all direct paths within the same time frame. For each candidate index Establish a ternary binding relationship and include candidate indexes. With candidate angle of arrival Reliability of direct route Bind while preserving the candidate angle complex response vector With candidate index The correspondence remains unchanged and is maintained during the binding process. Internal frequency sub-band index The order remains unchanged, allowing subsequently selected candidate indices to directly locate the corresponding candidate angular complex response vector and frequency sub-band index; for each time frame The credibility binding results are filtered for target candidate angles using the maximum direct path credibility criterion. Specifically, within the same time frame... Inner traversal of all candidate indices And compare , choose to Candidate index for finding the maximum value As the target candidate index for this time frame, the target candidate index Corresponding candidate angle of arrival The target candidate angle for this time frame is determined, and the target candidate angles for all time frames are indexed by time frame. Sequential splicing to form a target candidate angle sequence At the same time, the reliability of the direct path for each frame is determined. With candidate angular complex response vector Follow Synchronous retention, making Frequency subband index Maintain traceability;

[0107] The target candidate angle sequence under the constraint of upper limit of angular velocity Time alignment is performed, and the upper limit constraint for angular velocity is set to the upper limit of angular velocity. Characterization, The time window length is set as the preset upper limit of the angle change rate. Characterization, The duration of the time window corresponding to adjacent time frames is determined by the time window length and the sampling period, based on the upper limit of angular velocity. Duration of the time window Determine the upper limit of the angle difference , the upper limit of angle difference The maximum allowable angle change between adjacent time frames, and by using and The aligned target candidate angle sequence is obtained by multiplication and denoted as follows: and will Initialize to For each time frame from Increment to Execution angle difference upper limit criterion: calculation and The angle difference is taken as the absolute value, and the absolute value does not exceed [the value of the angle difference]. season equal The absolute value exceeds At that time, according to The sign determines the angle update direction, and in the angle update direction, the sign is used to determine the angle update direction. Set as Plus The limiting step, thus and The angle difference satisfies the upper limit criterion for angle difference, and the target candidate index is maintained in each frame during time alignment. Unchanged, and maintained Internal frequency sub-band index The order remains unchanged, ensuring that the aligned angle values ​​maintain a consistent binding relationship with the corresponding candidate index and frequency sub-band index;

[0108] Aligned target candidate angle sequence The results are determined as continuous angles of arrival, and the direct path confidence level for each frame is calculated. Indexed by time frame Summarize to form target credibility ,in With angle of arrival continuous results in time frame index To maintain consistency, the arrival angle continuity results and target confidence are both consistent with the candidate angle feature sequence. The frame indexes are consistent.

[0109] In this embodiment, step S5 specifically includes:

[0110] Step S5 uses the aforementioned arrival angle continuous results Target credibility With state prediction angle This serves as input to generate control instructions and sector search instructions. indexed by time frame and The range of values ​​is to , For time frames The credibility of the target For time frames The angle of arrival is a continuous result. For time frames State prediction angle, For time frames The corresponding target candidate index, For the number of frequency sub-bands, For frequency subband index and The range of values ​​is to Time window length The time window length used during frame division, with a sampling period of [value missing]. The duration of the time window is determined by the length of the time window. With sampling period The confidence threshold is obtained by multiplication and denoted as . The preset sector angle range is denoted as ;

[0111] Target credibility With credibility threshold As input, for each time frame Perform threshold judgment and establish branch selection relationship for each Generate branch selection marker ,when When Set as a continuous result branch for the angle of arrival, when When Set as a state prediction corner branch, and mark the branch selection flag. With time frame index Bind and set the target candidate index Follow Records are maintained to ensure that branch selection results remain traceable and consistent across the candidate angle of arrival index dimension;

[0112] To satisfy Time frames for continuous result branches of the angle of arrival Generate pointer control instructions ,in For time frames The pointing control command will The pointing angle is set to and the credibility of the target Time frame index With target candidate index Write This is used by the subsequent execution end for frame-by-frame scheduling and candidate index backtracking, indexing the frequency sub-bands. The range of values to As an index table field This transmission ensures that the execution end has a consistent index basis when it needs to read the frequency sub-band organization order of the candidate angle complex response;

[0113] To satisfy Time frames for predicting angle branches of the state Generate pointer control instructions Simultaneously generate search sector instructions ,in For time frames The search sector command will The pointing angle is set to and the credibility of the target Time frame index With target candidate index Write ,Will The center angle is set to Set the search sector boundary to and and index the time frame Write To form a frame-by-frame executable search window, indexing target candidates With frequency subband index The range of values to Follow This ensures that the candidate angle of arrival index and the frequency subband index are organized in a consistent manner when the execution end of the search sector re-examines the candidate angle of arrival within the sector.

[0114] Pointing control commands generated in each time frame With the search sector command Indexed by time frame The ascending order is summarized into an instruction sequence, while maintaining the relationship between the instruction sequence and the candidate angle feature sequence. The frame indexes are consistent, and the time window duration is used to set the instruction update interval so that adjacent time frames are consistent. and The corresponding instructions are updated and issued according to the duration of the time window, thereby aligning the pointing control instructions and the search sector instructions in the time dimension with the frame structure of the candidate corner feature sequence.

[0115] In this embodiment, step S6 specifically includes:

[0116] Step S6 refers to the aforementioned control command. With the search sector command As input, closed-loop scheduling of pointer and sector search is completed at the execution end. indexed by time frame and The range of values ​​is to , For time frames The pointing control command contains a pointing angle field and a time frame index field. For time frames The sector search command includes a center corner field, a sector boundary field, and a time frame index field. For time frames The target candidate index, For the number of frequency sub-bands, For frequency subband index and The range of values ​​is to Time window length The number of sampling points contained in each frame, with a sampling period of . Execution rhythm Execute the beat for the time length corresponding to each frame. By lengthening the time window With sampling period The result is obtained by multiplication, and the receiving object is denoted as... ,in For equipment identification of optoelectronic turntables, directional antennas, or interception vehicles, the execution angle is denoted as... ,in For time frames The target pointer angle for writing to the receiving object, the sector center pointer angle is denoted as . The left boundary of the sector is denoted as The right boundary of the sector is denoted as The preset scan step size is denoted as ,in This represents the angular interval between adjacent scan angles during sector search.

[0117] Point to control command With the search sector command Indexed by time frame The instructions are organized in ascending order into an instruction queue, and the time frame of each instruction is indexed. Write to the queue header field to form a sequentially readable time frame binding relationship, and determine the receiving target for the photoelectric turntable, directional antenna, or interceptor vehicle. The receiving object This is a preset device type identifier used to select the corresponding control interface and angle execution channel at the execution end, and to receive the target. Time frame index of the instruction queue Binding storage enables the receiving object Indexing the same time frame Read and Maintain a consistent time frame index;

[0118] To the receiving object Send pointing control command And trigger the execution of the pointed-to command. The pointer angle field in the code is used as the execution angle. In the receiving object Side general Write to the position controller input register and start the position closed loop, so that the receiving object... Rotate from the current angle to According to the execution rhythm Set execution timing: Index each time frame. In time length Within the control cycle, the issuance and execution confirmation of the control command are completed, and the time frame index is set. As an execution confirmation return field, the target candidate index With frequency subband index Follow The data is transmitted together, allowing the executing end to record the "time frame index". —Target candidate index —Frequency Subband Index The correspondence between "" is used to subsequently backtrack the candidate angle complex response and frequency subband organization order using the same indexing rules;

[0119] In the receiving object After the pointer is completed, follow the sector search instructions. Initiate sector search and read The center angle field and sector boundary field use the center angle as the sector center pointing angle. The sector boundary is used as the left sector boundary. And the right boundary of the sector and with to As the search boundary, according to the preset scan step size Generate scan angle sequence: from Start button Increment to Obtain the discrete scan angle set, and drive the receiving object in the order of the set. The execution proceeds point by point, performing a preset dwell time for each discrete scan angle and outputting the corresponding execution confirmation, which carries a time frame index. Thus indexed by time frame The execution sequence of sector search is advanced, and the target candidate indexes are... With frequency subband index Follow The execution records of sector searches are transmitted and written to the search log to maintain an organization consistent with that of the candidate angle of arrival index and frequency subband index.

[0120] Indexed by time frame during continuous result updates of angle of arrival or state prediction angle updates. The execution loop performs pointer and sector searches, with the execution end performing each execution cycle. Incrementing time frame index at the end And read the next frame and When the corresponding time frame index of If it exists, first execute step S62 to complete the pointing control, then execute step S63 to complete the sector search, when the corresponding time frame index... of If it does not exist, only step S62 is executed to complete the pointing control and enter the next time frame index, by indexing by time frame. Cyclic scheduling makes the pointer point to the control command. With the search sector command It remains continuously active on optoelectronic turntables, directional antennas, or intercept vehicles, and maintains consistency with time frame indexes and target candidate indexes. and frequency subband index The consistent transitive relationship.

[0121] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for stabilizing the angle of arrival in urban multipath environments based on ComplexMamba-2, characterized in that, include: S1. Receive complex baseband data from a multi-channel antenna array, divide the complex baseband data into frames according to time windows and organize them according to frequency sub-bands to obtain a broadband array observation sequence; S2. Calculate the broadband Capon direction finding spectrum for each frame based on the broadband array observation sequence to obtain the main peak of the direction finding spectrum. Filter the top K candidate angles of arrival based on the angular interval between the main peaks and the width of the main peak. Perform matching projection on the candidate angles of arrival according to the array steering vector to obtain the complex response of the candidate angles. Calculate the cross-band phase alignment degree based on the complex response of the candidate angles and calculate the duration of adjacent frames to form multipath discrimination features. Summarize the candidate angles of arrival and multipath discrimination features frame by frame to obtain the candidate angle feature sequence. S3. Input the candidate angle feature sequence into the ComplexMamba-2 complex state space network. The ComplexMamba-2 complex state space network includes a complex input encoding layer, a complex state recursion layer, and an angle decoding layer. The complex input encoding layer performs complex encoding on the candidate angle complex response and multipath discrimination features. The complex state recursion layer generates the direct path tracking state based on the complex encoding. Under the constraint of the direct path tracking state, the cross-band phase alignment degree and the duration of adjacent frames are fused to calculate the direct path confidence. The angle decoding layer decodes the direct path tracking state to obtain the state prediction angle. S4. Filter the target candidate angles based on the candidate angle feature sequence according to the direct path confidence, and perform time alignment of the target candidate angles under the constraint of the upper limit of angular velocity to obtain the continuous result of the arrival angle and the target confidence. S5. Generate a pointing control command based on the target credibility and credibility threshold. When the target credibility is greater than or equal to the credibility threshold, the pointing control command points to the continuous result of the arrival angle. When the target credibility is less than the credibility threshold, the pointing control command points to the state prediction angle and generates a search sector command. S6. Send pointing control commands and sector search commands to the optoelectronic turntable, directional antenna, or interceptor vehicle to drive the optoelectronic turntable, directional antenna, or interceptor vehicle to perform pointing and sector search.

2. The method for stable estimation of angle of arrival in urban multipath environments based on ComplexMamba-2 according to claim 1, characterized in that, S1 specifically refers to: The complex baseband data of the multi-channel antenna array is used as input. Gain and phase consistency processing is performed according to the array channel identifier. The consistency result is checked by channel amplitude threshold and phase deviation threshold to maintain the channel amplitude-phase relationship. Complex baseband data is divided into frames according to a preset time window length. The sampling of each channel is aligned with the start and end points of the time window. A preset overlap ratio is used to control the overlap area of ​​adjacent time windows, forming a frame sequence organized by time windows. The data of each frame is organized into frequency sub-bands according to the preset frequency sub-band boundaries. A unified frequency sub-band boundary table and sub-band order are used to bind the complex value sequence of each frequency sub-band with the channel index, forming a sub-band set organized by frequency sub-band. The results of framing and frequency sub-banding are concatenated according to the dual indexes of time frame and frequency sub-band to obtain the broadband array observation sequence. The broadband array observation sequence contains multi-channel complex values ​​of each frame and frequency sub-band and retains the channel amplitude-phase relationship.

3. The method for stable estimation of angle of arrival in urban multipath environments based on ComplexMamba-2 according to claim 1, characterized in that, S2 specifically refers to: The broadband array observation sequence is input in time frames. The broadband Capon direction finding spectrum is calculated based on the multi-channel complex values ​​of each frame. The main peak of the direction finding spectrum is extracted using the peak significance threshold, and the main peak angle and main peak width are used as the main peak attributes. Based on the angular interval and peak width between the main peaks of the direction finding spectrum, the first K candidate angles of arrival are selected using the angular interval threshold and the peak width threshold, and the selection results are divided into a set of candidate angles of arrival by frame. For each candidate angle of arrival, a matching projection is performed according to the array steering vector to map the multi-channel complex values ​​of the broadband array observation sequence at the angle to the candidate angle complex response, and organized into vector form according to frequency sub-bands; The cross-band phase alignment degree is calculated based on the phase difference of the candidate angle complex response in each frequency sub-band, and the phase consistency is checked by using the phase alignment threshold to obtain the cross-band phase alignment degree of each candidate angle of arrival. The persistence of adjacent frames is calculated based on the consistency of candidate angle of arrival in adjacent time frames. A persistence threshold is used to identify consecutive occurrences, thereby obtaining the persistence of adjacent frames for each candidate angle of arrival. The cross-band phase alignment degree and the persistence of adjacent frames are combined to form a multipath discrimination feature, and then concatenated with the candidate angle complex response one by one according to the candidate angle of arrival to obtain the structured candidate angle of arrival feature. The candidate angle of arrival (AOA) features are aggregated frame by frame to generate a candidate angle feature sequence. The order of the candidate AOA and the frequency subband index remain unchanged, so that the candidate angle feature sequence can be used in the complex input coding layer of the ComplexMamba-2 complex state space network.

4. The method for stable estimation of angle of arrival in urban multipath environments based on ComplexMamba-2 according to claim 3, characterized in that, The cross-band phase alignment degree is calculated by the phase alignment function, which is as follows: ； in, Candidate arrival angle The degree of cross-band phase alignment, For the number of frequency sub-bands, For frequency sub-band index, The candidate angular complex response vector In frequency subband Complex values ​​at that location, for The complex conjugate, for The range, Adjacent frequency sub-bands and The weights are adopted. Obtain, complete After calculation, the phase alignment threshold is used. right The verification will be lower than of Marked as a low-alignment candidate.

5. The method for stable estimation of angle of arrival in urban multipath environments based on ComplexMamba-2 according to claim 1, characterized in that, S3 specifically refers to: The candidate angle feature sequence is input into the ComplexMamba-2 complex state space network according to time frames. The complex input coding layer, complex state recursion layer and angle decoding layer are connected in sequence. The indices of K candidate arrival angles and frequency subband indices of each frame are passed to the complex input coding layer with each frame. In the complex input coding layer, the candidate angle complex response corresponding to each candidate angle of arrival is used as a complex input to perform complex linear mapping to obtain the candidate angle complex embedding. The multipath discrimination feature and the candidate angle of arrival are used as the input of the gated branch to generate gate coefficients and channel-level modulation is performed on the candidate angle complex embedding. The output is a single-frame complex code bound to the frequency sub-band index. In the complex state recursive layer, frame-by-frame complex encoding is received and the direct path tracking state is maintained. The direct path tracking state is updated by channel, and the gated complex encoding is injected into the state update, so that the direct path tracking state evolves continuously over time and serves as a constraint source for subsequent confidence calculation. In the confidence output header of the angle decoding layer, the matching score is calculated based on the direct path tracking status and the complex embedding of the candidate angles in each frame. The channel-weighted matching rule is used to maintain the correspondence between the matching score and the candidate angle of arrival index. Under the constraints of direct path tracking, the matching score is fused with the cross-band phase alignment degree and the persistence of adjacent frames to generate K direct path confidence scores for each frame, while retaining the influence paths of the cross-band phase alignment degree and the persistence of adjacent frames. In the angle output head of the angle decoding layer, a state prediction angle is generated based on the direct path tracking status, so that the state prediction angle expresses the time extrapolation of the arrival angle and is bound to the corresponding time frame index; The direct path confidence and state prediction angle are used as the output of the ComplexMamba-2 complex state space network, maintaining consistency with the frame and candidate arrival angle index of the candidate angle feature sequence, and providing input for subsequent filtering and pointing control branches.

6. The method for stable estimation of angle of arrival in urban multipath environments based on ComplexMamba-2 according to claim 5, characterized in that, The reliability of the direct path is calculated by the reliable feedback function, which is as follows: ； in, For time frames Internal candidate index The corresponding direct path reliability, For time frame indexing, As a candidate index, The number of candidate angles of arrival per frame. To match scores, For cross-band phase alignment degree, The duration of adjacent frames, To match score weighting coefficients, This is a weighting coefficient for cross-band phase alignment. The weighting coefficients represent the duration of adjacent frames. The interaction item weight coefficient, and For the summation index, It is a frame-normalized stability coefficient and is a preset positive real constant.

7. The method for stable estimation of angle of arrival in urban multipath environments based on ComplexMamba-2 according to claim 1, characterized in that, S4 specifically refers to: The candidate angle feature sequence is read by time frame, and a binding is established between the direct path confidence of each frame and the candidate angle of arrival index to form the confidence binding result of the candidate angle of arrival set by frame, while maintaining the correspondence with the complex response of the candidate angle. For each frame's credibility binding result, the maximum direct path credibility criterion is used to filter the target candidate angles. The filtered target candidate angles are then spliced ​​together in time frame order to form a target candidate angle sequence, and the corresponding direct path credibility and the frequency sub-band index of the candidate angle of arrival are retained. Under the constraint of the upper limit of angular velocity, the target candidate angle sequence is time-aligned. The upper limit of angle difference is determined according to the upper limit of angular velocity constraint and the length of time window. The upper limit of angle difference criterion is used to limit the angle difference between adjacent time frames. During the time alignment process, the candidate arrival angle index and the frequency sub-band index are kept consistent to obtain the aligned target candidate angle sequence. The aligned target candidate angle sequence is used as the continuous result of the angle of arrival, and the corresponding direct path confidence is summarized into the target confidence by time frame, keeping it consistent with the frame index of the candidate angle feature sequence.

8. The method for stable estimation of angle of arrival in urban multipath environments based on ComplexMamba-2 according to claim 1, characterized in that, S5 specifically refers to: Using the target confidence level and confidence level threshold as input, threshold judgment is performed on each frame to establish a branch selection relationship between the continuous results of the angle of arrival and the state prediction angle, keeping it consistent with the candidate angle of arrival index and binding it with the time frame index; For frames where the threshold is determined to be greater than or equal to the confidence threshold, a pointing control instruction is generated, the pointing angle of the pointing control instruction is set to the continuous result of the angle of arrival, and the target confidence is bound to the time frame index. In frames where the threshold is determined to be less than the confidence threshold, a pointing control instruction is generated and the pointing angle is set as the state prediction angle. At the same time, a search sector instruction is generated, using the state prediction angle as the center angle, and the search sector boundary is determined using a preset sector angle range. This is then bound to the time frame index, and the candidate arrival angle index and frequency sub-band index are passed along with the instruction. The pointing control instructions and search sector instructions generated in each time frame are summarized in chronological order, keeping them consistent with the frame index of the candidate corner feature sequence, and the instruction update interval is set using the time window length.

9. The method for stable estimation of angle of arrival in urban multipath environments based on ComplexMamba-2 according to claim 1, characterized in that, Step S6 is as follows: The pointing control command and the search sector command are organized in time frame order and bound to the time frame index. The receiving target is determined for the photoelectric turntable, directional antenna or intercept vehicle, and the time frame index is kept consistent with the time frame index in the pointing control command and the search sector command. Send a pointing control command to the receiving object, use the pointing angle in the pointing control command as the execution angle, and set the execution clock according to the time window length to maintain a consistent transmission relationship with the candidate angle of arrival index and frequency sub-band index; After the receiving object completes the pointing, a sector search is initiated according to the sector search instruction. The center angle and sector angle range in the sector search instruction are used as the search boundary, and the execution sequence of the sector search is advanced according to the time frame index. When the angle of arrival is continuously updated or the state prediction angle is updated, the pointing and sector search are performed cyclically according to the time frame index, so that the pointing control command and the sector search command are continuously effective on the photoelectric turntable, directional antenna or intercept vehicle.

Images