A method for radio frequency output based on communication garbled characters in racing drones

Abstract

This invention discloses a radio frequency (RF) output method based on communication scrambling in a racing drone, comprising the following steps: S1, generating a communication scrambling bitstream; S2, introducing a self-normalized martingale envelope adaptive contraction algorithm based on martingale theory, and performing probabilistic upper bound contraction screening; S3, inputting the structured scrambling bitstream sequence into an improved TIME-LLM model, introducing a scrambling state conditional prefix routing mechanism, and outputting a structured distortion discrimination feature sequence; S4, introducing a posterior coupled update algorithm with confidence-gated Kalman gain for gated modulation; S5, mapping to a set of RF output control parameters; S6, updating the RF output control parameters; S7, if the convergence condition is met, performing progressive backoff control to restore the preset RF output state. This invention achieves adaptive adjustment and stable backoff of RF output control parameters, effectively improving the stability and security of RF output control under abnormal communication conditions.

Description

Technical Field

[0001] This invention relates to the field of wireless communication and radio frequency control technology, and in particular to a radio frequency output method based on communication scrambling in a racing drone. Background Technology

[0002] With the rapid development of unmanned systems, miniaturized flight platforms, and wireless communication applications in complex electromagnetic environments, racing drones are increasingly used in low-altitude reconnaissance, emergency response, racing flight, and combat scenarios. The stability of their wireless communication links and the ability to control radio frequency output have become crucial factors affecting system reliability. Existing racing drone wireless communication systems typically rely on statistical indicators such as signal-to-noise ratio, bit error rate, and packet loss rate to evaluate link quality and adjust transmit power, operating frequency band, or frequency hopping strategy accordingly. However, under complex interference or high-speed motion conditions, the communication receiver often experiences cyclic redundancy check failures and abnormal decoding terminations. In these situations, the large number of abnormal bits generated by the physical receiver layer cannot be effectively utilized by existing link evaluation mechanisms, leading to lag in radio frequency output control or insufficient adjustment basis.

[0003] Meanwhile, existing technologies often use a single time scale or fixed threshold to determine communication anomalies, making it difficult to characterize the evolution of abnormal states over time. This makes them susceptible to transient noise and sudden interference, leading to frequent frequency switching, drastic power fluctuations, and further instability in RF output. On the other hand, traditional filtering or simple statistical methods have limited robustness under non-stationary and strongly disturbed conditions, making it difficult to achieve precise characterization and smooth estimation of abnormal states. Furthermore, existing RF parameter control strategies typically lack clear variable budget constraints and convergence mechanisms, making it difficult to smoothly roll back RF output parameters after communication recovery, resulting in risks of parameter oscillations and control discontinuities.

[0004] Therefore, how to provide a radio frequency output method based on the communication garbled characters of a racing drone is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a radio frequency (RF) output method based on communication scrambling in a racing drone. This invention extracts the original received bit sequence for a corresponding time period under conditions of cyclic redundancy check failure or abnormal decoding termination, forming a communication scrambling bitstream. A multi-time-scale sliding window is applied to the scrambling bitstream to construct a window-level delay statistical sequence. A structured scrambling segment sequence is obtained by combining this with a self-normalized martingale envelope adaptive shrinking algorithm. This structured scrambling segment sequence is then input into an improved TIME-LLM model containing a scrambling reprogramming module, a prefix hint module, a frozen inference module, and a discriminative decoding module. The model outputs a structured distortion discriminative feature sequence, which is then generated using a posterior coupled update algorithm incorporating confidence-gated Kalman gain in a Kalman variational autoencoder. Furthermore, the scrambling state feature sequence is mapped to center frequency control parameters, transmit power control parameters, and transmit timing control parameters. A variable budget constraint and convergence condition determination are applied to the RF output control parameters to complete the update and gradual backoff control of the RF output control parameters. This invention can effectively extract the state information contained in communication scrambled codes, and realize the stable adjustment and reliable recovery of radio frequency output control parameters under abnormal communication conditions. It has the advantages of high radio frequency output stability, good control continuity and low risk of parameter oscillation.

[0006] A radio frequency output method based on communication scrambling in a racing drone according to an embodiment of the present invention includes the following steps:

[0007] S1. During the wireless communication process of the racing drone, if the receiver fails the cyclic redundancy check or the decoding terminates abnormally, the original received bit sequence within the corresponding time period is intercepted at the physical receiving layer to form a communication scrambled bit stream.

[0008] S2. Apply a multi-time-scale sliding window to the communication scrambled bitstream to construct a window-level delay statistics sequence. Based on martingale theory, introduce an adaptive shrinking algorithm with a self-normalized martingale envelope to perform probability upper bound shrinking screening on the delay statistics sequence to obtain a structured scrambled bitstream sequence.

[0009] S3. Input the structured scrambled code fragment sequence into the improved TIME-LLM model. The improved TIME-LLM model includes a scrambled code reprogramming module, a prefix hinting module, a frozen inference module, and a discriminative decoding module. In the prefix hinting module, a scrambled code state conditional prefix routing mechanism is introduced. The prefix is ​​selected according to the structured scrambled code fragment sequence and gated injection is performed to output a structured distortion discriminative feature sequence.

[0010] S4. Input the structural distortion discrimination feature sequence into the Kalman variational autoencoder, and introduce a posterior coupling update algorithm of confidence-gated Kalman gain into the Kalman variational autoencoder to perform gated modulation on the Kalman gain and output the scrambled state feature sequence.

[0011] S5. Map the scrambled state feature sequence to a set of radio frequency output control parameters;

[0012] S6. Apply a variable budget constraint to the RF output control parameters. If the scrambled state feature sequence changes, update the RF output control parameters according to the variable budget constraint.

[0013] S7. If the scrambled state feature sequence meets the convergence condition within a preset continuous time period, then the RF output control parameters are gradually rolled back to restore the preset RF output state.

[0014] Optionally, S1 specifically includes:

[0015] During the wireless communication process of the racing drone, the received data stream output by the receiver at the physical receiving layer is acquired;

[0016] Perform cyclic redundancy check processing on the received data stream and perform decoding processing;

[0017] If the cyclic redundancy check fails or the decoding terminates abnormally, the start and end timestamps corresponding to the failure of the cyclic redundancy check or the abnormal termination of the decoding are determined.

[0018] Based on the start and end timestamps, the original received bit sequence within the corresponding time period is extracted at the physical receiving layer;

[0019] The intercepted original received bit sequence is spliced ​​together in chronological order to form a scrambled communication bit stream.

[0020] Optionally, the step of applying a multi-time-scale sliding window to the garbled communication bitstream to construct a window-level delay statistical sequence specifically involves:

[0021] Associate a timestamp with each original received bit in the garbled bit stream of communication; set multiple window lengths for different time scales and corresponding sliding step sizes;

[0022] Under each time scale condition, the communication garbled code stream is divided into sliding segments according to the timestamp to obtain multiple window segments, each window segment corresponding to the start timestamp and the end timestamp.

[0023] For each window segment, calculate the window delay statistic, which is equal to the difference between the end timestamp and the start timestamp;

[0024] The window delay statistics corresponding to multiple window segments are arranged in order of timestamp to form a window-level delay statistics sequence.

[0025] Optionally, the self-normalized martingale envelope adaptive shrinkage algorithm based on martingale theory is introduced to perform probabilistic upper bound shrinkage filtering on the time delay statistical sequence to obtain a structured scrambled fragment sequence, specifically:

[0026] For each time scale in the window-level latency statistics sequence, construct the cumulative arrival sequence and the cumulative service sequence;

[0027] The queue backlog sequence is calculated based on the cumulative arrival sequence and the cumulative service sequence. The queue backlog sequence is the difference sequence between the cumulative arrival sequence and the cumulative service sequence.

[0028] An upper delay sequence is constructed based on the queue backlog sequence. The upper delay sequence is determined by the queue backlog sequence and the corresponding window delay statistics.

[0029] The queue backlog sequence is self-normalized to obtain a self-normalized statistical sequence, which is determined by the dispersion of the queue backlog sequence and the samples at the corresponding time scale.

[0030] An upper bound sequence of martingale envelope probability is constructed based on a self-normalized statistical sequence, and an adaptive shrinkage process is applied to the upper bound sequence of martingale envelope probability.

[0031] The window segments are filtered based on the upper bound of the shrinking probability and the preset filtering conditions. The window segments that meet the filtering conditions are retained and spliced ​​together in the order of timestamps to form a structured scrambled segment sequence.

[0032] Optionally, the improved TIME-LLM model includes a garbled text reprogramming module, a prefix hinting module, a frozen inference module, and a discriminative decoding module, specifically:

[0033] The scrambling reprogramming module groups the original received bits into a numerical sequence according to the timestamp order in the structured scrambling fragment sequence. The numerical sequence is then subjected to fixed-length processing, which includes padding the numerical sequence with zeros if the length is insufficient and truncating the numerical sequence if the length is excessive. The fixed-length numerical sequence is then mapped to a reprogramming embedding sequence, which is obtained by multiplying the numerical sequence with the reprogramming mapping matrix.

[0034] The prefix suggestion module constructs a prefix candidate set based on the reprogrammed embedding sequence. The prefix candidate set contains multiple prefix embedding sequences. It calculates the fragment state vector corresponding to the reprogrammed embedding sequence. The fragment state vector is obtained by concatenating the mean vector and variance vector of the reprogrammed embedding sequence. The prefix suggestion module introduces a garbled state conditional prefix routing mechanism. It generates routing weights based on the similarity between the fragment state vector and the prefix embedding sequence. It performs a weighted summation of multiple prefix embedding sequences based on the routing weights to obtain a gated prefix embedding sequence. The gated prefix embedding sequence and the reprogrammed embedding sequence are concatenated in the input order to obtain a combined embedding sequence.

[0035] The frozen inference module combines an embedded sequence input parameter that remains unchanged in a sequence inference network. The sequence inference network includes multiple self-attention computation layers and feedforward computation layers. In the self-attention computation layer, query vectors, key vectors, and value vectors are generated based on the combined embedded sequence. Attention weights are calculated based on the correlation between the query vector and the key vector, and the value vector is weighted and summed. The feedforward computation layer performs a nonlinear transformation on the self-attention computation results and outputs a hidden representation sequence.

[0036] The discriminant decoding module calculates the pooling vector of the final hidden representation sequence in the time dimension. The pooling vector is the average value vector of the final hidden representation sequence in the time dimension. The pooling vector is input into the linear discriminant layer to output the structural distortion discriminant feature sequence. The structural distortion discriminant feature sequence is obtained by arranging the output vector of the linear discriminant layer in the order of timestamps.

[0037] Optionally, the posterior coupled update algorithm for introducing confidence-gated Kalman gain into the Kalman variational autoencoder is specifically as follows:

[0038] The structural distortion discrimination feature sequence is input into the encoding network to obtain the latent variable mean vector and latent variable variance vector, and the latent variable sequence is generated based on the latent variable mean vector and latent variable variance vector;

[0039] A state prediction sequence is constructed based on the latent variable sequence. The state prediction sequence is obtained by transforming the state estimation sequence of the previous time step through the state transition matrix.

[0040] An observation residual sequence is constructed based on the state prediction sequence and the structural distortion discrimination feature sequence. The observation residual sequence is obtained by the difference between the structural distortion discrimination feature sequence and the state prediction sequence under the action of the observation matrix.

[0041] The prediction covariance matrix is ​​constructed based on the state prediction sequence and the observation matrix. The prediction covariance matrix is ​​obtained by transforming the state covariance matrix of the previous time step through the state transition matrix and superimposing the process noise covariance matrix.

[0042] Confidence gating coefficients are constructed, which are determined by the segment state vector corresponding to the structural distortion discriminant feature sequence and the time dimension statistics of the structural distortion discriminant feature sequence.

[0043] The gated observation noise covariance matrix is ​​obtained by scaling the observation noise covariance matrix according to the confidence gating coefficient.

[0044] The Kalman gain matrix is ​​calculated based on the prediction covariance matrix, the observation matrix, and the gated observation noise covariance matrix. The Kalman gain matrix is ​​equal to the product of the prediction covariance matrix and the transpose of the observation matrix, multiplied by the inverse of the sum of the products of the observation matrix, the prediction covariance matrix, and the transpose of the observation matrix, and the gated observation noise covariance matrix.

[0045] The state prediction sequence is updated based on the Kalman gain matrix and the observation residual sequence to obtain the state estimation sequence. The state estimation sequence is equal to the state prediction sequence plus the product of the Kalman gain matrix and the observation residual sequence.

[0046] The updated covariance matrix is ​​obtained by updating the state covariance matrix based on the Kalman gain matrix. The updated covariance matrix is ​​equal to the product of the observation matrix and the Kalman gain matrix minus the identity matrix, and then multiplied by the predicted covariance matrix.

[0047] The state estimation sequence is input into the decoding network, which outputs a garbled state feature sequence.

[0048] Optionally, S5 specifically includes:

[0049] Arrange the garbled state feature sequence according to timestamp order to form a feature vector sequence, and then aggregate the feature vector sequence by time dimension to obtain an aggregated feature vector;

[0050] The center frequency control parameters are obtained by multiplying the aggregated feature vector with the pre-stored center frequency mapping matrix and superimposing the center frequency bias vector; the transmit power control parameters are obtained by multiplying the aggregated feature vector with the power mapping matrix and superimposing the power bias vector; and the transmit timing control parameters are obtained by multiplying the aggregated feature vector with the timing mapping matrix and superimposing the timing bias vector. The transmit timing control parameters include the transmit start time, transmit duration, and transmit interval duration.

[0051] Amplitude constraints are applied to the center frequency control parameters, transmit power control parameters, and transmit timing control parameters. The amplitude constraints include limiting the center frequency control parameters to a preset center frequency range, limiting the transmit power control parameters to a preset power range, and limiting the transmit duration and transmit interval duration to a preset timing range, thus forming a set of radio frequency output control parameters.

[0052] Optionally, S6 specifically includes:

[0053] Obtain the scrambled state feature sequence corresponding to adjacent timestamps and perform difference to obtain the feature change amount; calculate the change amount of radio frequency output control parameters based on the feature change amount, which includes the change amount of center frequency control parameters, the change amount of transmit power control parameters and the change amount of transmit timing control parameters.

[0054] Set the center frequency change budget, transmit power change budget, and transmit timing change budget. The center frequency change budget, transmit power change budget, and transmit timing change budget correspond to the maximum allowable change range per unit time.

[0055] The changes in center frequency control parameters, transmit power control parameters, and transmit timing control parameters are compared and limited with the center frequency change budget, transmit power change budget, and transmit timing change budget, respectively, to obtain the limited changes in radio frequency output control parameters.

[0056] The RF output control parameters are updated based on the limited changes in the RF output control parameters to obtain the updated RF output control parameters.

[0057] Optionally, the convergence condition is specifically satisfied as follows:

[0058] Within a preset continuous time period, a garbled state feature sequence is obtained in order of timestamps; the garbled state feature sequence represents a set of garbled state feature vectors arranged in order of timestamps.

[0059] The feature vectors of the scrambled state corresponding to adjacent timestamps are differentially calculated to obtain the sequence of feature changes;

[0060] Perform time-dimensional statistics on the feature change sequence to obtain feature change statistics;

[0061] The feature change statistics are compared with the preset convergence threshold; when the corresponding feature change statistics within the preset continuous time period are all less than or equal to the preset convergence threshold, the garbled state feature sequence is determined to meet the convergence condition.

[0062] Optionally, S7 specifically includes:

[0063] If the garbled state feature sequence meets the convergence condition, the parameter difference between the RF output control parameters and the preset RF output state is obtained, and the parameter difference is reduced step by step according to the timestamp order. The RF output control parameters are then updated based on the parameter difference after the step reduction.

[0064] When the RF output control parameters corresponding to multiple consecutive timestamps are consistent with the preset RF output state, the gradual rollback control of the RF output control parameters is completed.

[0065] The beneficial effects of this invention are:

[0066] (1) By directly intercepting the original received bit sequence of the physical receiving layer during the wireless communication process of the racing machine under the condition of failure of cyclic redundancy check or abnormal termination of decoding, a communication scrambled bit stream is constructed. Combined with multi-time scale sliding window and window-level delay statistical modeling, the timing information under abnormal communication state can be preserved and quantified, thereby improving the integrity and usability of the communication abnormal state representation.

[0067] (2) The window-level time delay statistics sequence is subjected to probability upper bound shrinkage screening by self-normalized martingale envelope adaptive shrinkage algorithm to form a structured scrambled code fragment sequence, which effectively suppresses the influence of random noise and occasional jitter on time delay statistics and enhances the stability and distinguishability of scrambled code fragments in the time dimension;

[0068] (3) Input the structured scrambled code fragment sequence into the improved TIME-LLM model, and combine it with the confidence-gated Kalman gain posterior coupled update algorithm in the Kalman variational autoencoder to generate a continuous and smooth scrambled code state feature sequence, thereby improving the accuracy and robustness of the characterization of communication state evolution.

[0069] (4) By mapping the scrambled state feature sequence to a set of radio frequency output control parameters and introducing a variable budget constraint and convergence condition judgment mechanism, the center frequency control parameters, transmit power control parameters and transmit timing control parameters are subject to restricted updates and gradual back-off control, thereby reducing the risk of sudden changes in radio frequency output parameters and achieving smooth adjustment and reliable recovery of radio frequency output state. Attached Figure Description

[0070] 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:

[0071] Figure 1 This is a flowchart of a radio frequency output method based on communication garbled characters in a racing drone, as proposed in this invention.

[0072] Figure 2 This is a schematic diagram of the structure of the improved TIME-LLM model proposed in this invention;

[0073] Figure 3 This is a data flow diagram of a radio frequency output method based on communication garbled characters in a racing drone, as proposed in this invention. Detailed Implementation

[0074] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0075] refer to Figures 1-3 A method for radio frequency output based on communication garbled characters in a racing drone includes the following steps:

[0076] S1. During the wireless communication process of the racing drone, if the receiver fails the cyclic redundancy check or the decoding terminates abnormally, the original received bit sequence within the corresponding time period is intercepted at the physical receiving layer to form a communication scrambled bit stream.

[0077] S2. Apply a multi-time-scale sliding window to the communication scrambled bitstream to construct a window-level time delay statistical sequence. Based on martingale theory, introduce an adaptive shrinking algorithm with a self-normalized martingale envelope to perform probability upper bound shrinking and filtering on the time delay statistical sequence to obtain a structured scrambled bitstream sequence.

[0078] S3. Input the structured scrambled code fragment sequence into the improved TIME-LLM model. The improved TIME-LLM model includes a scrambled code reprogramming module, a prefix hinting module, a frozen inference module, and a discriminative decoding module. In the prefix hinting module, a scrambled code state conditional prefix routing mechanism is introduced. The prefix is ​​selected according to the structured scrambled code fragment sequence and gated injection is performed to output a structured distortion discriminative feature sequence.

[0079] S4. Input the structural distortion discrimination feature sequence into the Kalman variational autoencoder, and introduce a posterior coupling update algorithm of confidence-gated Kalman gain into the Kalman variational autoencoder to perform gated modulation on the Kalman gain and output the scrambled state feature sequence.

[0080] S5. Map the scrambled state feature sequence to a set of radio frequency output control parameters; the radio frequency output control parameters include center frequency control parameters, transmit power control parameters, and transmit timing control parameters;

[0081] S6. Apply a variable budget constraint to the RF output control parameters. If the garbled state characteristic sequence changes, update the RF output control parameters according to the variable budget constraint.

[0082] S7. If the garbled state feature sequence meets the convergence condition within a preset continuous time period, then the RF output control parameters are gradually rolled back to restore the preset RF output state.

[0083] In this embodiment, S1 specifically refers to:

[0084] During the wireless communication process of the racing drone, the received data stream output by the physical receiving layer is obtained. The received data stream contains received bits arranged in chronological order and corresponding sampling numbers. The sampling number and sampling frequency satisfy the condition that the timestamp is equal to the sampling number divided by the sampling frequency.

[0085] The received data stream is frame delimited, and the start and end positions of the frames are determined according to the preset frame structure to obtain a sequence of received frames arranged in chronological order.

[0086] Cyclic redundancy check (CRC) processing is performed on the received frame sequence. The CRC processing includes a generator polynomial and a redundancy check length. The degree of the generator polynomial is equal to the redundancy check length. The information bits in the received frame correspond to the information polynomial. The information polynomial is multiplied by x raised to the power of the redundancy check length and then moduloed by the generator polynomial to obtain the remainder polynomial. The remainder polynomial corresponds to the redundancy check bits. The redundancy check bits are compared bit by bit with the redundancy check bits carried in the received frame. If any bit is inconsistent, the CRC is determined to have failed.

[0087] The received frame sequence is decoded. The decoding process includes error correction decoding of the encoded bits in the received frame and outputting the decoded bit sequence. If the decoding path does not converge, the verification constraint is not met, or the decoding output is an empty sequence during the decoding process, the decoding is determined to be abnormal and terminated.

[0088] If the cyclic redundancy check fails or the decoding terminates abnormally, determine the start and end timestamps corresponding to the failure of the cyclic redundancy check or the abnormal termination of the decoding. The start timestamp is the start timestamp of the first received frame that fails the cyclic redundancy check or terminates abnormally, and the end timestamp is the end timestamp of the received frames that fail the cyclic redundancy check or terminates abnormally consecutively.

[0089] Based on the start and end timestamps, the physical receiving layer extracts the original received bit sequence within the corresponding time period. The original received bit sequence takes all received bits whose timestamps are greater than or equal to the start timestamp and less than or equal to the end timestamp, and arranges them in ascending order of timestamp.

[0090] The original received bit sequence is spliced ​​together in chronological order to form a communication scrambled bit stream. The communication scrambled bit stream is defined as a continuous sequence of original received bits within the time period corresponding to a failed cyclic redundancy check or an abnormal termination of decoding. The splicing method is to concatenate the original received bit sequences of multiple consecutive failed intervals in ascending order of timestamps. If there is an overlap of timestamps between adjacent intervals, the overlapping bits of the earlier interval are retained and the overlapping bits of the later interval are discarded, resulting in a communication scrambled bit stream corresponding to a monotonically increasing timestamp.

[0091] In this embodiment, a multi-time-scale sliding window is applied to the garbled communication bitstream to construct a window-level delay statistics sequence, specifically as follows:

[0092] A timestamp is associated with each original received bit in the communication scrambled bit stream. The timestamp is determined by the sampling sequence number and sampling frequency of the received bit in the physical receiver layer. The timestamp value is equal to the sampling sequence number divided by the sampling frequency.

[0093] Set multiple window lengths and corresponding sliding step sizes for different time scales. The window length represents the cumulative range of differences between adjacent timestamps, and the sliding step size represents the time interval between the start timestamps of adjacent windows.

[0094] Under each time scale, the communication scrambled bitstream is segmented by sliding based on the timestamp. The sliding segmentation method is to continuously extract multiple window segments on the time axis with the window length as the coverage range and the sliding step size as the displacement. Each window segment contains the original received bits with the timestamp located between the window start timestamp and the window end timestamp.

[0095] Record the start and end timestamps for each window segment. The start timestamp is the minimum timestamp in the window segment, and the end timestamp is the maximum timestamp in the window segment.

[0096] For each window segment, calculate the window delay statistic, which is equal to the end timestamp minus the start timestamp.

[0097] The window delay statistics corresponding to multiple window segments are arranged in ascending order of timestamps to form a window-level delay statistics sequence.

[0098] In this embodiment, a self-normalized martingale envelope adaptive shrinkage algorithm is introduced based on martingale theory to perform probabilistic upper bound shrinkage filtering on the time delay statistical sequence, resulting in a structured scrambled code fragment sequence, specifically:

[0099] For window segments corresponding to different time scales in the window-level delay statistics sequence, a cumulative arrival sequence and a cumulative service sequence are constructed. The cumulative arrival sequence is obtained by accumulating the original number of received bits in the window segment in chronological order, and the cumulative service sequence is obtained by accumulating the number of received bits in the window segment that can be decoded in chronological order.

[0100] The queue backlog sequence is calculated based on the cumulative arrival sequence and the cumulative service sequence. The value of each moment in the queue backlog sequence is equal to the value of the corresponding moment in the cumulative arrival sequence minus the value of the corresponding moment in the cumulative service sequence.

[0101] A delay upper bound sequence is constructed based on the queue backlog sequence. The value of each item in the delay upper bound sequence is equal to the corresponding item in the queue backlog sequence divided by the ratio of the number of received bits in the window segment to the window delay statistic.

[0102] The queue backlog sequence is self-normalized by dividing the corresponding item of the queue backlog sequence by the sample dispersion of the queue backlog sequence in all window segments at the same time scale. The sample dispersion is obtained by summing the squares of the differences between the queue backlog sequence values ​​and the mean, dividing by the number of window segments, and then taking the square root. After self-normalization, a self-normalized statistical sequence is formed.

[0103] A martingale envelope probability upper bound sequence is constructed based on the self-normalized statistical sequence. The martingale envelope probability upper bound sequence represents the probability constraint relationship that the self-normalized statistical sequence does not exceed a preset limit under a given confidence level.

[0104] An adaptive shrinkage process is applied to the upper bound sequence of martingale envelope probability. This process scales the corresponding values ​​of the upper bound sequence of martingale envelope probability by introducing a shrinkage factor related to the sample dispersion and window delay statistics, so that the upper bound of probability is dynamically adjusted as statistical stability changes.

[0105] The window segments are filtered based on the upper bound of the probability after shrinkage and the preset filtering conditions. The preset filtering conditions include that the upper bound of the probability does not exceed the preset probability threshold and the corresponding upper bound of the delay does not exceed the preset delay threshold.

[0106] Window segments that meet the filtering criteria are spliced ​​together in ascending order of timestamps. The splicing method is to directly connect the original received bits with consecutive or adjacent timestamps in adjacent window segments to form a continuous sequence of original received bits, thus obtaining a structured scrambled segment sequence.

[0107] In this embodiment, the self-normalized martingale envelope adaptive shrinkage algorithm is introduced based on martingale theory because the communication scrambled bitstream exhibits strong non-stationarity, burstiness, and long-tail distribution characteristics in the time dimension. The window-level delay statistical sequence shows significant variance drift and abnormal peak accumulation at different time scales. Traditional screening methods based on fixed thresholds or empirical statistical assumptions are difficult to provide a consistent and effective upper bound constraint on delay fluctuations in a probabilistic sense, and are prone to false screening or missed screening under high noise or sudden interference conditions. By constructing cumulative arrival sequences, cumulative service sequences, and queue backlog sequences on the window-level delay statistical sequence, the martingale theory framework is introduced to model the delay evolution process. Furthermore, self-normalization is performed on the queued backlog sequence, so that the probability upper bound expression no longer depends on the fixed scale parameter but adapts to the sample dispersion, thereby obtaining a martingale envelope probability upper bound with consistent constraint ability for different time delay scales. On this basis, an adaptive shrinkage process is applied to the martingale envelope probability upper bound, so that the probability upper bound gradually approaches the actual statistical behavior while maintaining theoretical validity, realizing stable screening of structurally abnormal time delay segments, providing support conditions with probabilistic interpretability and temporal consistency for the construction of structured scrambled segment sequences, and improving the input reliability and overall robustness of the subsequent improved TIME-LLM model and Kalman variational autoencoder processing stage.

[0108] In this embodiment, the improved TIME-LLM model includes a garbled text reprogramming module, a prefix hinting module, a frozen inference module, and a discriminative decoding module, specifically:

[0109] The scrambled code reprogramming module groups the original received bits into a numerical sequence according to the timestamp order in the structured scrambled code fragment sequence. The numerical sequence is then subjected to fixed-length processing, which includes padding the numerical sequence with zeros if the length is insufficient and truncating the numerical sequence if the length is excessive. The fixed-length numerical sequence is then mapped to a reprogramming embedding sequence, which is obtained by multiplying the numerical sequence with the reprogramming mapping matrix.

[0110] The prefix suggestion module constructs a prefix candidate set based on the reprogrammed embedding sequence. The prefix candidate set contains multiple prefix embedding sequences. It calculates the fragment state vector corresponding to the reprogrammed embedding sequence. The fragment state vector is obtained by concatenating the mean vector and variance vector of the reprogrammed embedding sequence. The prefix suggestion module introduces a garbled state conditional prefix routing mechanism. It generates routing weights based on the similarity between the fragment state vector and the prefix embedding sequence. It performs a weighted summation of multiple prefix embedding sequences based on the routing weights to obtain a gated prefix embedding sequence. The gated prefix embedding sequence and the reprogrammed embedding sequence are concatenated in the input order to obtain a combined embedding sequence.

[0111] The frozen inference module combines the embedded sequence input parameters of the sequence inference network without updating them. The sequence inference network contains multiple self-attention computation layers and feedforward computation layers. In the self-attention computation layer, query vectors, key vectors and value vectors are generated based on the combined embedded sequence. Attention weights are calculated by the correlation between the query vector and the key vector and the value vector is weighted and summed. The feedforward computation layer performs nonlinear transformation on the self-attention computation results and outputs the hidden representation sequence.

[0112] The discriminant decoding module calculates the pooling vector of the final hidden representation sequence in the time dimension. The pooling vector is the average value vector of the final hidden representation sequence in the time dimension. The pooling vector is input into the linear discriminant layer to output the structural distortion discriminant feature sequence. The structural distortion discriminant feature sequence is obtained by arranging the output vector of the linear discriminant layer in the order of timestamps.

[0113] In this embodiment, the improved TIME-LLM model is based on the TIME-LLM model. The reason for this improvement is that the TIME-LLM model possesses the ability to freeze the inference structure and represent long-term dependencies in time series modeling, enabling stable inference on non-stationary sequences while keeping parameters unchanged. This makes it suitable for feature extraction requirements in communication scrambled bitstreams under conditions of abrupt changes in statistical characteristics and sparse samples. The improvement involves introducing a scrambled bit reprogramming module on the input side, mapping the original received bits in the structured scrambled bitstream sequence to a fixed-length numerical sequence, and converting it into a reprogrammed embedding sequence through a reprogramming mapping matrix. This aligns the bit-level information of the physical receiving layer with the embedding space of the sequence inference network. A prefix hinting module is introduced before inference, and a prefix candidate set is constructed based on the reprogrammed bitstream. The segment state vector obtained from the embedded sequence is used to generate routing weights based on the similarity between the embedded sequence and the prefix embedded sequence. Weighted summation is then used to form a gated prefix embedded sequence, which is concatenated with the reprogrammed embedded sequence in the input order. This enables the sequence inference network to possess conditional inference capabilities under different scrambled states. The frozen inference module keeps the sequence inference network parameters unchanged. Hidden representation sequences are extracted through multiple self-attention and feedforward computation layers, reducing the impact of small-sample scrambled segments on model stability. On the output side, the discriminative decoding module performs temporal pooling and linear discrimination on the hidden representation sequences, forming a structured distortion discriminative feature sequence arranged in timestamp order. This allows the model output to directly serve the subsequent Kalman variational autoencoder's modeling of scrambled state feature sequences and the generation of the RF output control parameter set.

[0114] In this embodiment, a posterior coupled update algorithm for confidence-gated Kalman gain is introduced into the Kalman variational autoencoder, specifically as follows:

[0115] The structural distortion discrimination feature sequence is input into the encoding network to obtain the latent variable mean vector and latent variable variance vector, and the latent variable sequence is generated based on the latent variable mean vector and latent variable variance vector;

[0116] A state prediction sequence is constructed based on the latent variable sequence. The state prediction sequence is obtained by transforming the state estimation sequence of the previous time step through the state transition matrix.

[0117] An observation residual sequence is constructed based on the state prediction sequence and the structural distortion discrimination feature sequence. The observation residual sequence is obtained by the difference between the structural distortion discrimination feature sequence and the state prediction sequence under the action of the observation matrix.

[0118] The prediction covariance matrix is ​​constructed based on the state prediction sequence and the observation matrix. The prediction covariance matrix is ​​obtained by transforming the state covariance matrix of the previous time step through the state transition matrix and superimposing the process noise covariance matrix.

[0119] Confidence gating coefficients are constructed, which are determined by the segment state vectors corresponding to the structural distortion discriminant feature sequences and the time dimension statistics of the structural distortion discriminant feature sequences. The observation noise covariance matrix is ​​scaled according to the confidence gating coefficients to obtain the gated observation noise covariance matrix.

[0120] The Kalman gain matrix is ​​calculated based on the prediction covariance matrix, the observation matrix, and the gated observation noise covariance matrix. The Kalman gain matrix is ​​equal to the product of the prediction covariance matrix and the transpose of the observation matrix, multiplied by the inverse of the sum of the products of the observation matrix, the prediction covariance matrix, and the transpose of the observation matrix, and the gated observation noise covariance matrix.

[0121] The state prediction sequence is updated based on the Kalman gain matrix and the observation residual sequence to obtain the state estimation sequence. The state estimation sequence is equal to the state prediction sequence plus the product of the Kalman gain matrix and the observation residual sequence.

[0122] The updated covariance matrix is ​​obtained by updating the state covariance matrix based on the Kalman gain matrix. The updated covariance matrix is ​​equal to the product of the identity matrix, the observation matrix, and the Kalman gain matrix, and then multiplied by the predicted covariance matrix. The state estimation sequence is input into the decoding network and the output is a scrambled state feature sequence.

[0123] In this embodiment, a posterior coupled update algorithm with confidence-gated Kalman gain is introduced into the Kalman variational autoencoder. This is because the structural distortion discriminant feature sequence exhibits uneven noise intensity and significant fluctuations in statistical stability with changes in segments over time. Traditional Kalman gain under fixed observation noise covariance conditions is difficult to achieve adaptive balance between different scrambled states, which can easily lead to over-response of instantaneous abnormal features or insufficient suppression of effective distortion information in the state prediction sequence. By coupling the segment state vector corresponding to the structural distortion discriminant feature sequence with the time-dimensional statistics to construct confidence-gated coefficients, the observation noise covariance matrix is ​​scaled and modulated, so that the Kalman gain matrix is ​​dynamically adjusted with the confidence of the scrambled state during inference. This forms a posteriorly consistent coupled update relationship in the processes of latent variable sequence generation, state prediction sequence update, and state covariance matrix recursion, thereby enhancing the performance of the scrambled state feature sequence in terms of temporal continuity, resistance to sudden distortion interference, and state estimation stability.

[0124] In this embodiment, S5 specifically refers to:

[0125] The garbled state feature sequence is arranged in timestamp order to form a feature vector sequence. The feature vector sequence contains multiple feature vectors, and the dimension of the feature vectors is consistent with the output dimension of the Kalman variational autoencoder. The feature vector sequence is aggregated in the time dimension to obtain an aggregated feature vector, which is equal to the dimensional average of the feature vector sequence in the time dimension.

[0126] The center frequency control parameter is obtained by multiplying the aggregated feature vector with the center frequency mapping matrix and superimposing the center frequency bias vector. The value of the center frequency control parameter is equal to the product of the center frequency mapping matrix and the aggregated feature vector plus the center frequency bias vector. Let the dimension of the aggregated feature vector obtained by time aggregation of the scrambled state feature sequence be D, denoted as z, where z is a D-row, 1-column vector; the center frequency mapping matrix is ​​denoted as Wf, where Wf is a 1-row, D-column matrix; and the center frequency bias vector is denoted as bf, where bf is a 1-row, 1-column vector.

[0127] The transmit power control parameters are obtained by multiplying the aggregated feature vector by the power mapping matrix and superimposing the power bias vector. The value of the transmit power control parameters is equal to the product of the power mapping matrix and the aggregated feature vector plus the power bias vector. The power mapping matrix is ​​used to map the overall strength and stability of the scrambled state features to the radio frequency transmit power control parameters, which are linear control mapping parameters. The aggregated feature vector has a dimension of D, denoted as z, where z is a D-row, 1-column vector. The power mapping matrix is ​​denoted as Wp, where Wp is a 1-row, D-column matrix. The power bias vector is denoted as bp, where bp is a 1-row, 1-column vector.

[0128] The transmission timing control parameters are obtained by multiplying the aggregated feature vector with the timing mapping matrix and superimposing it with the timing bias vector. These parameters include the transmission start time, transmission duration, and transmission interval duration. The values ​​of the transmission start time, transmission duration, and transmission interval duration are respectively equal to the product of the timing mapping matrix and the aggregated feature vector plus the corresponding component in the timing bias vector. The timing mapping matrix is ​​used to map the scrambled state features to the control parameters of the radio frequency transmission in the time dimension, including the transmission start time, transmission duration, and transmission interval duration. The aggregated feature vector z is D rows and 1 column; the timing mapping matrix is ​​denoted as Wt, a 3-row, D-column matrix; the timing bias vector is denoted as bt, a 3-row, 1-column vector; and the output vector t is a 3-row, 1-column vector. The first row of t corresponds to the transmission start time; the second row corresponds to the transmission duration; and the third row corresponds to the transmission interval duration.

[0129] Amplitude constraints are applied to the center frequency control parameters, transmit power control parameters, and transmit timing control parameters. The amplitude constraints include limiting the center frequency control parameters to a preset center frequency range, limiting the transmit power control parameters to a preset power range, and limiting the transmit duration and transmit interval duration to a preset timing range, thus forming a set of radio frequency output control parameters.

[0130] In this embodiment, S6 specifically refers to:

[0131] Obtain the scrambled state feature sequence corresponding to adjacent timestamps and perform difference to obtain the feature change amount; calculate the change amount of radio frequency output control parameters based on the feature change amount, which includes the change amount of center frequency control parameters, the change amount of transmit power control parameters and the change amount of transmit timing control parameters.

[0132] Set the center frequency change budget, transmit power change budget, and transmit timing change budget. The center frequency change budget, transmit power change budget, and transmit timing change budget correspond to the maximum allowable change range per unit time.

[0133] The change in the center frequency control parameter is compared with the center frequency change budget. If the change in the center frequency control parameter exceeds the center frequency change budget, the change in the center frequency control parameter is limited to the center frequency change budget.

[0134] The change in the transmit power control parameter is compared with the transmit power change budget. If the change in the transmit power control parameter exceeds the transmit power change budget, the change in the transmit power control parameter is limited to the transmit power change budget.

[0135] The change in launch duration is compared with the budgeted change in launch duration. If the change in launch duration exceeds the budgeted change in launch duration, the change in launch duration is limited to the budgeted change in launch duration.

[0136] The change in the launch interval duration is compared with the budgeted change in the launch interval duration. If the change in the launch interval duration exceeds the budgeted change in the launch interval duration, the change in the launch interval duration is limited to the budgeted change in the launch interval duration.

[0137] The RF output control parameters are updated based on the changes in the defined RF output control parameters. The updated value of the center frequency control parameter is equal to the original center frequency control parameter plus the change in the center frequency control parameter. The updated value of the transmit power control parameter is equal to the original transmit power control parameter plus the change in the transmit power control parameter. The updated value of the transmit duration is equal to the original transmit duration plus the change in the transmit duration. The updated value of the transmit interval duration is equal to the original transmit interval duration plus the change in the transmit interval duration. The updated RF output control parameters are thus obtained.

[0138] In this embodiment, the convergence condition is specifically:

[0139] Within a preset continuous time period, the scrambled state feature sequence is obtained in the order of timestamps. The scrambled state feature sequence consists of multiple scrambled state feature vectors arranged in the order of timestamps. The scrambled state feature vector is the state representation vector output by the Kalman variational autoencoder at the corresponding timestamp.

[0140] Vector difference operations are performed on the scrambled state feature vectors with adjacent timestamps. Vector difference operation is to subtract the components of the same dimension element by element to obtain the feature change sequence. Each element in the feature change sequence corresponds to the magnitude of the state change under adjacent timestamps. Statistics are performed on the feature change sequence in the time dimension. The time dimension statistics include calculating the Euclidean norm of each feature change in the feature change sequence and arranging it in the order of timestamps. The Euclidean norm is equal to the square root of the sum of squares of each component in the feature change, resulting in a feature change statistics sequence.

[0141] The sequence of feature change statistics within a preset continuous time period is compared to determine whether the feature change statistics corresponding to each timestamp are less than or equal to a preset convergence threshold. The preset convergence threshold is the upper limit of the maximum state change amplitude allowed by the system. When all values ​​in the corresponding feature change statistics sequence within the preset continuous time period are less than or equal to the preset convergence threshold, the garbled state feature sequence is determined to meet the convergence condition.

[0142] In this embodiment, S7 specifically refers to:

[0143] After determining that the scrambled state feature sequence meets the convergence condition, the parameter difference between the RF output control parameter corresponding to the current timestamp and the preset RF output state is obtained. The parameter difference consists of the difference between the center frequency control parameter, the difference between the transmit power control parameter, and the difference between the transmit timing control parameter. The parameter difference is the component-wise difference between the current parameter value and the preset parameter value. The parameter difference is reduced step by step according to the timestamp order. The step-by-step reduction method is to multiply the parameter difference by a backoff ratio coefficient less than 1 at each timestamp. The backoff ratio coefficient is a preset constant used to control the parameter backoff speed. The RF output control parameter is updated based on the parameter difference after step-by-step reduction. The updated RF output control parameter is equal to the sum of the preset RF output state parameter and the parameter difference after step-by-step reduction. The above parameter difference reduction and update process is repeated at multiple consecutive timestamps. When the parameter difference between the RF output control parameter corresponding to multiple consecutive timestamps and the preset RF output state is zero, it is determined that the RF output control parameter is consistent with the preset RF output state, and the gradual backoff control of the RF output control parameter is completed.

[0144] Example 1: To verify the feasibility of this invention in practice, it was applied to a scenario involving an integrated wireless link for image transmission and remote control of a power line inspection drone formation. The work area was located at the boundary between the city edge and the factory area, with 5.8GHz co-frequency video transmission, 2.4GHz remote control interference, and intermittent electromagnetic pulse noise. The drones flew back and forth at altitudes of 20m to 80m above the ground, with a maximum link distance of 1.8km. Existing links experience cyclic redundancy check failures or abnormal decoding terminations when interference is superimposed, resulting in screen distortion and control delay jitter at the receiving end. When the transmitting end outputs at a fixed center frequency and fixed transmission power, retransmissions and delay accumulation caused by short-term scrambling further amplify the risk of loss of control. This invention addresses the issue of unusable but still observable original received bit sequences at the physical receiving layer. It generates a scrambled bit stream and reconstructs a stable set of radio frequency output control parameters, reducing link jitter and the probability of disconnection.

[0145] During on-site deployment, the receiver outputs the received data stream at the physical receiving layer and simultaneously outputs the cyclic redundancy check (CRC) processing result and decoding status. If a CRC failure occurs or decoding terminates abnormally, the original received bit sequence is extracted according to the start and end timestamps and concatenated in chronological order to form a garbled communication stream. This garbled stream then enters a multi-timescale sliding window processing phase, with window lengths of 20ms, 50ms, and 100ms, and sliding step sizes of 5ms, 10ms, and 20ms, forming a window-level delay statistical sequence. The window-level delay statistics sequence is subjected to probability upper bound shrinkage screening. Based on the cumulative arrival sequence A(t) and the cumulative service sequence S(t), the queue backlog sequence B(t) = A(t) - S(t) is obtained. Based on the queue backlog sequence and the window delay statistics, the delay upper bound sequence D(t) is obtained. The queue backlog sequence is self-normalized to obtain Z(t) = B(t) / sqrt(V(t) + ε), where V(t) is the sample dispersion at the corresponding time scale. A martingale envelope probability upper bound sequence P(t) is constructed. Adaptive shrinkage is applied to P(t), and window segments are retained according to the screening conditions. The segments are then concatenated in timestamp order to obtain a structured scrambled fragment sequence. The structured scrambled fragment sequence is input into the improved TIME-LLM model, which outputs a structured distortion discrimination feature sequence. This sequence is then input into a Kalman variational autoencoder and subjected to gating modulation to obtain a scrambled state feature sequence. The scrambled state feature sequence is arranged and aggregated according to timestamp order to obtain an aggregated feature vector. This vector is then mapped to center frequency control parameters, transmit power control parameters, and transmit timing control parameters via a center frequency mapping matrix, power mapping matrix, timing mapping matrix, and bias vector. Amplitude constraints are then applied to form a set of RF output control parameters. This set of RF output control parameters is updated under varying budget constraints: center frequency variation budget is 0.6 MHz / s, transmit power variation budget is 1.5 dB / s, transmit duration variation budget is 2 ms / s, and transmit interval duration variation budget is 2 ms / s. The convergence condition is determined using a preset continuous time period of 2 seconds. The feature changes of adjacent timestamps are calculated, and the Euclidean norm sequence is obtained. If the Euclidean norm is less than or equal to 0.08 for 2 consecutive seconds, the convergence condition is met. If met, a gradual backoff control is executed, progressively reducing the RF output control parameter set to a preset RF output state step-by-step according to timestamp order. Backoff is completed consistently for 8 consecutive timestamps.

[0146] Table 1 Comparison of Communication Garbled Code Stream Filtering and Convergence Performance

[0147]

[0148] Table 1 shows that the proportions of cyclic redundancy check failures and decoding abnormal terminations are similar under both methods, indicating that abnormal triggering is unavoidable under the same interference intensity. This invention obtains a structured scrambled code fragment sequence through window-level delay statistics and probability upper bound shrinkage screening, enabling distorted fragments to enter a stable modeling link. The retention rate of structured scrambled code fragments is between 47.8% and 62.5%, supporting the formation of a stable trend between subsequent structured distortion discrimination feature sequences and scrambled code state feature sequences. The median convergence time decreased from 3.8s to 9.4s to 1.2s to 2.6s, and the convergence failure rate decreased to 12.0% under strong interference, demonstrating that a controllable state can still be quickly achieved even under scrambled code conditions.

[0149] Table 2 Comparison of RF output control parameter stability and link recovery effect

[0150]

[0151] Table 2 illustrates the improved stability of the RF output control parameter set under varying budget constraints and gradual backoff control constraints. The peak-to-peak values ​​of the center frequency and transmit power control parameters are significantly reduced, and the jitter in the start time and duration of the transmit timing control parameters decreases, alleviating the delay accumulation observed at the receiver and improving decoding window matching. The 2-second window link recovery success rate increases to 86.0%, and the end-to-end delay at the 95th percentile decreases to 138ms. This indicates that the RF output control parameter set can converge quickly and suppress excessive jumps before meeting the convergence conditions. After meeting the convergence conditions, it returns to the preset RF output state through gradual backoff control, reducing the risk of long-term drift and improving the continuity and controllability of flight missions.

[0152] 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 radio frequency output method based on communication scrambling in a racing drone, characterized in that, Includes the following steps: S1. During the wireless communication process of the racing drone, if the receiver fails the cyclic redundancy check or the decoding terminates abnormally, the original received bit sequence within the corresponding time period is intercepted at the physical receiving layer to form a communication scrambled bit stream. S2. Apply a multi-time-scale sliding window to the communication scrambled bitstream to construct a window-level delay statistics sequence. Based on martingale theory, introduce an adaptive shrinking algorithm with a self-normalized martingale envelope to perform probability upper bound shrinking screening on the delay statistics sequence to obtain a structured scrambled bitstream sequence. S3. Input the structured scrambled code fragment sequence into the improved TIME-LLM model. The improved TIME-LLM model includes a scrambled code reprogramming module, a prefix hinting module, a frozen inference module, and a discriminative decoding module. In the prefix hinting module, a scrambled code state conditional prefix routing mechanism is introduced. The prefix is ​​selected according to the structured scrambled code fragment sequence and gated injection is performed to output a structured distortion discriminative feature sequence. S4. Input the structural distortion discrimination feature sequence into the Kalman variational autoencoder, and introduce a posterior coupling update algorithm of confidence-gated Kalman gain into the Kalman variational autoencoder to perform gated modulation on the Kalman gain and output the scrambled state feature sequence. S5. Map the scrambled state feature sequence to a set of radio frequency output control parameters; S6. Apply a variable budget constraint to the RF output control parameters. If the scrambled state feature sequence changes, update the RF output control parameters according to the variable budget constraint. S7. If the scrambled state feature sequence meets the convergence condition within a preset continuous time period, then the RF output control parameters are gradually rolled back to restore the preset RF output state.

2. The radio frequency output method based on communication scrambling in a racing drone according to claim 1, characterized in that, Specifically, S1 is: During the wireless communication process of the racing drone, the received data stream output by the receiver at the physical receiving layer is acquired; Perform cyclic redundancy check processing on the received data stream and perform decoding processing; If the cyclic redundancy check fails or the decoding terminates abnormally, the start and end timestamps corresponding to the failure of the cyclic redundancy check or the abnormal termination of the decoding are determined. Based on the start and end timestamps, the original received bit sequence within the corresponding time period is extracted at the physical receiving layer; The intercepted original received bit sequence is spliced ​​together in chronological order to form a scrambled communication bit stream.

3. The radio frequency output method based on communication scrambling in a racing drone according to claim 1, characterized in that, The process of applying a multi-time-scale sliding window to the garbled communication bitstream to construct a window-level delay statistics sequence is as follows: Associate a timestamp with each original received bit in the garbled bit stream of communication; set multiple window lengths for different time scales and corresponding sliding step sizes; Under each time scale condition, the communication garbled code stream is divided into sliding segments according to the timestamp to obtain multiple window segments, each window segment corresponding to the start timestamp and the end timestamp. For each window segment, calculate the window delay statistic, which is equal to the difference between the end timestamp and the start timestamp; The window delay statistics corresponding to multiple window segments are arranged in order of timestamp to form a window-level delay statistics sequence.

4. The radio frequency output method based on communication scrambling in a racing drone according to claim 1, characterized in that, The self-normalized martingale envelope adaptive shrinkage algorithm, based on martingale theory, is used to perform probabilistic upper bound shrinkage filtering on the time delay statistical sequence to obtain a structured scrambled fragment sequence. Specifically: For each time scale in the window-level latency statistics sequence, construct the cumulative arrival sequence and the cumulative service sequence; The queue backlog sequence is calculated based on the cumulative arrival sequence and the cumulative service sequence. The queue backlog sequence is the difference sequence between the cumulative arrival sequence and the cumulative service sequence. An upper delay sequence is constructed based on the queue backlog sequence. The upper delay sequence is determined by the queue backlog sequence and the corresponding window delay statistics. The queue backlog sequence is self-normalized to obtain a self-normalized statistical sequence, which is determined by the dispersion of the queue backlog sequence and the samples at the corresponding time scale. An upper bound sequence of martingale envelope probability is constructed based on a self-normalized statistical sequence, and an adaptive shrinkage process is applied to the upper bound sequence of martingale envelope probability. The window segments are filtered based on the upper bound of the shrinking probability and the preset filtering conditions. The window segments that meet the filtering conditions are retained and spliced ​​together in the order of timestamps to form a structured scrambled segment sequence.

5. The radio frequency output method based on communication scrambling in a racing drone according to claim 1, characterized in that, The improved TIME-LLM model includes a garbled text reprogramming module, a prefix hinting module, a frozen inference module, and a discriminative decoding module, specifically: The scrambling reprogramming module groups the original received bits into a numerical sequence according to the timestamp order in the structured scrambling fragment sequence. The numerical sequence is then subjected to fixed-length processing, which includes padding the numerical sequence with zeros if the length is insufficient and truncating the numerical sequence if the length is excessive. The fixed-length numerical sequence is then mapped to a reprogramming embedding sequence, which is obtained by multiplying the numerical sequence with the reprogramming mapping matrix. The prefix suggestion module constructs a prefix candidate set based on the reprogrammed embedding sequence. The prefix candidate set contains multiple prefix embedding sequences. It calculates the fragment state vector corresponding to the reprogrammed embedding sequence. The fragment state vector is obtained by concatenating the mean vector and variance vector of the reprogrammed embedding sequence. The prefix suggestion module introduces a garbled state conditional prefix routing mechanism. It generates routing weights based on the similarity between the fragment state vector and the prefix embedding sequence. It performs a weighted summation of multiple prefix embedding sequences based on the routing weights to obtain a gated prefix embedding sequence. The gated prefix embedding sequence and the reprogrammed embedding sequence are concatenated in the input order to obtain a combined embedding sequence. The frozen inference module combines an embedded sequence input parameter that remains unchanged in a sequence inference network. The sequence inference network includes multiple self-attention computation layers and feedforward computation layers. In the self-attention computation layer, query vectors, key vectors, and value vectors are generated based on the combined embedded sequence. Attention weights are calculated based on the correlation between the query vector and the key vector, and the value vector is weighted and summed. The feedforward computation layer performs a nonlinear transformation on the self-attention computation results and outputs a hidden representation sequence. The discriminant decoding module calculates the pooling vector of the final hidden representation sequence in the time dimension. The pooling vector is the average value vector of the final hidden representation sequence in the time dimension. The pooling vector is input into the linear discriminant layer to output the structural distortion discriminant feature sequence. The structural distortion discriminant feature sequence is obtained by arranging the output vector of the linear discriminant layer in the order of timestamps.

6. The radio frequency output method based on communication scrambling in a racing drone according to claim 1, characterized in that, The posterior coupled update algorithm for introducing confidence-gated Kalman gain into the Kalman variational autoencoder is as follows: The structural distortion discrimination feature sequence is input into the encoding network to obtain the latent variable mean vector and latent variable variance vector, and the latent variable sequence is generated based on the latent variable mean vector and latent variable variance vector; A state prediction sequence is constructed based on the latent variable sequence. The state prediction sequence is obtained by transforming the state estimation sequence of the previous time step through the state transition matrix. An observation residual sequence is constructed based on the state prediction sequence and the structural distortion discrimination feature sequence. The observation residual sequence is obtained by the difference between the structural distortion discrimination feature sequence and the state prediction sequence under the action of the observation matrix. The prediction covariance matrix is ​​constructed based on the state prediction sequence and the observation matrix. The prediction covariance matrix is ​​obtained by transforming the state covariance matrix of the previous time step through the state transition matrix and superimposing the process noise covariance matrix. Confidence gating coefficients are constructed, which are determined by the segment state vector corresponding to the structural distortion discriminant feature sequence and the time dimension statistics of the structural distortion discriminant feature sequence. The gated observation noise covariance matrix is ​​obtained by scaling the observation noise covariance matrix according to the confidence gating coefficient. The Kalman gain matrix is ​​calculated based on the prediction covariance matrix, the observation matrix, and the gated observation noise covariance matrix. The Kalman gain matrix is ​​equal to the product of the prediction covariance matrix and the transpose of the observation matrix, multiplied by the inverse of the sum of the products of the observation matrix, the prediction covariance matrix, and the transpose of the observation matrix, and the gated observation noise covariance matrix. The state prediction sequence is updated based on the Kalman gain matrix and the observation residual sequence to obtain the state estimation sequence. The state estimation sequence is equal to the state prediction sequence plus the product of the Kalman gain matrix and the observation residual sequence. The updated covariance matrix is ​​obtained by updating the state covariance matrix based on the Kalman gain matrix. The updated covariance matrix is ​​equal to the product of the observation matrix and the Kalman gain matrix minus the identity matrix, and then multiplied by the predicted covariance matrix. The state estimation sequence is input into the decoding network, which outputs a garbled state feature sequence.

7. The radio frequency output method based on communication scrambling in a racing drone according to claim 1, characterized in that, Specifically, S5 is: Arrange the garbled state feature sequence according to timestamp order to form a feature vector sequence, and then aggregate the feature vector sequence by time dimension to obtain an aggregated feature vector; The center frequency control parameters are obtained by multiplying the aggregated feature vector with the pre-stored center frequency mapping matrix and superimposing the center frequency bias vector; the transmit power control parameters are obtained by multiplying the aggregated feature vector with the power mapping matrix and superimposing the power bias vector; and the transmit timing control parameters are obtained by multiplying the aggregated feature vector with the timing mapping matrix and superimposing the timing bias vector. The transmit timing control parameters include the transmit start time, transmit duration, and transmit interval duration. Amplitude constraints are applied to the center frequency control parameters, transmit power control parameters, and transmit timing control parameters. The amplitude constraints include limiting the center frequency control parameters to a preset center frequency range, limiting the transmit power control parameters to a preset power range, and limiting the transmit duration and transmit interval duration to a preset timing range, thus forming a set of radio frequency output control parameters.

8. The radio frequency output method based on communication scrambling in a racing drone according to claim 1, characterized in that, Specifically, S6 is: Obtain the garbled state feature sequence corresponding to adjacent timestamps and perform difference to obtain the feature change amount; The changes in radio frequency output control parameters are calculated based on the characteristic changes. These changes include the changes in center frequency control parameters, transmit power control parameters, and transmit timing control parameters. Set the center frequency change budget, transmit power change budget, and transmit timing change budget. The center frequency change budget, transmit power change budget, and transmit timing change budget correspond to the maximum allowable change range per unit time. The changes in center frequency control parameters, transmit power control parameters, and transmit timing control parameters are compared and limited with the center frequency change budget, transmit power change budget, and transmit timing change budget, respectively, to obtain the limited changes in radio frequency output control parameters. The RF output control parameters are updated based on the limited changes in the RF output control parameters to obtain the updated RF output control parameters.

9. The radio frequency output method based on communication scrambling in a racing drone according to claim 1, characterized in that, The specific convergence condition is as follows: Within a preset continuous time period, a garbled state feature sequence is obtained in order of timestamps; the garbled state feature sequence represents a set of garbled state feature vectors arranged in order of timestamps. The feature vectors of the scrambled state corresponding to adjacent timestamps are differentially calculated to obtain the sequence of feature changes; Perform time-dimensional statistics on the feature change sequence to obtain feature change statistics; The feature change statistics are compared with the preset convergence threshold; when the corresponding feature change statistics within the preset continuous time period are all less than or equal to the preset convergence threshold, the garbled state feature sequence is determined to meet the convergence condition.

10. The radio frequency output method based on communication scrambling in a racing drone according to claim 1, characterized in that, Specifically, S7 is: If the garbled state feature sequence meets the convergence condition, the parameter difference between the RF output control parameters and the preset RF output state is obtained, and the parameter difference is reduced step by step according to the timestamp order. The RF output control parameters are then updated based on the parameter difference after the step reduction. When the RF output control parameters corresponding to multiple consecutive timestamps are consistent with the preset RF output state, the gradual rollback control of the RF output control parameters is completed.

Images