Large unmanned aerial vehicle multi-modal data transmission method and system based on compressed sensing

Abstract

The application discloses a large unmanned aerial vehicle multi-modal data transmission method and system based on compressed sensing, relates to the field of digital information transmission, and comprises the following steps: based on a data priority label, multi-modal sensor data is divided into safety-critical timing data and high-dimensional auxiliary sensing data; a target measurement matrix is determined by combining an instantaneous channel capacity evaluation value and an importance weight parameter; a sparse basis projection transformation is performed to separate common sparse components and special sparse components; linear projection operations are respectively performed on the common sparse components and the special sparse components based on the target measurement matrix to obtain dimension-reduced feature measurement vectors; the safety-critical timing data is mapped to a low-delay time slot of a sending queue, and the dimension-reduced feature measurement vectors are filled into residual capacity time slots to generate a multi-modal multiplexing data stream, so that data transmission of the multi-modal sensor data is performed. By implementing the application, efficient and reliable downlink transmission of airborne heterogeneous data to a ground station can be realized under limited wireless bandwidth.

Description

Technical Field

[0001] This application relates to the field of digital information transmission, and in particular to a method and system for multimodal data transmission of large unmanned aerial vehicles based on compressed sensing. Background Technology

[0002] With the evolution of the air logistics industry, large cargo drones need to collect various environmental and airframe status information in real time during daily flight operations. In this scenario, the airborne communication system needs to rely on limited wireless channel bandwidth to continuously transmit massive amounts of multimodal sensor data, including flight attitude, photoelectric images, and radar point clouds, to the ground control center in order to maintain remote monitoring of the flight status.

[0003] In related technologies, UAV sensor data transmission schemes rely on independent compressed sensing methods to process massive amounts of data. During airborne data acquisition, this scheme configures independent measurement matrices for sensors of different modalities. After receiving the raw data streams from each sensor, the airborne processor independently performs sparse sampling and linear projection operations on the data according to its matched measurement matrix, generating mutually independent compressed measurement vectors. Subsequently, the airborne communication module queues all the compressed measurement vectors from the sensors and sends them into the wireless communication channel and to the ground station in a fixed periodic order.

[0004] However, when a drone encounters a sudden physical disturbance under complex weather conditions, causing transient changes in the images from multiple sensors and severe attitude fluctuations, the periodic queuing method of independently collecting multimodal data using a fixed matrix will generate a large number of redundant measurement vectors that repeatedly describe the same physical disturbance event. This will instantly cause congestion in the wireless communication channel and severely delay the transmission of critical flight control data. Summary of the Invention

[0005] This application provides a method and system for multimodal data transmission of large UAVs based on compressed sensing, which can realize efficient and reliable downlink of airborne heterogeneous data to ground stations under limited wireless bandwidth.

[0006] In a first aspect, this application provides a method for multimodal data transmission of large unmanned aerial vehicles (UAVs) based on compressed sensing, applied to a data processing system. The method includes: dividing acquired multimodal sensor data into uncompressible safety-critical time-series data and high-dimensional auxiliary sensing data to be compressed, based on preset data priority labels; extracting instantaneous channel capacity assessment values ​​from external uplink signaling and determining a target measurement matrix by combining the importance weight parameters of the high-dimensional auxiliary sensing data; performing sparse basis projection transformation on the high-dimensional auxiliary sensing data to separate common sparse components reflecting cross-modal common physical responses and proprietary sparse components corresponding to each sensor mode; performing linear projection operations on the common sparse components and proprietary sparse components based on the target measurement matrix to obtain a dimensionality-reduced feature measurement vector containing cross-modal feature information; mapping the safety-critical time-series data to low-latency time slots in the transmission queue and filling the remaining capacity time slots in the transmission queue with the dimensionality-reduced feature measurement vector to generate a multimodal multiplexed data stream to be transmitted; and writing the multimodal multiplexed data stream to the network communication port in response to the transmission permission identifier of the external communication link to execute the data transmission of multimodal sensor data.

[0007] In the above embodiments, the data processing system prioritizes the low-latency transmission of security-critical data within a limited bandwidth by using data priority division and channel capacity feedback, in conjunction with cross-modal joint sparse projection, and maximizes the transmission efficiency of auxiliary sensing data. This effectively eliminates redundancy between heterogeneous signals and improves the overall communication throughput.

[0008] In some embodiments of the first aspect, the step of extracting the instantaneous channel capacity assessment value from the external uplink signaling and determining the target measurement matrix by combining the importance weight parameters of the high-dimensional auxiliary sensing data specifically includes: obtaining the total available bandwidth data corresponding to the instantaneous channel capacity assessment value, and obtaining the minimum compression ratio lower limit data of each high-dimensional auxiliary sensing data; performing constraint optimization operations by combining the total available bandwidth data, the importance weight parameters, and the minimum compression ratio lower limit data to obtain the target compression ratio data of each high-dimensional auxiliary sensing data; and generating the corresponding target measurement matrix by combining the target compression ratio data.

[0009] In the above embodiments, the data processing system performs constraint optimization based on available bandwidth and minimum compression ratio limit, accurately allocates the target compression ratio of each mode and generates the corresponding measurement matrix, so that the system has the dynamic adaptability to cope with time-varying fading channels, and ensures that limited resources are tilted to maintain the fidelity of high-importance data when link quality fluctuates.

[0010] In some embodiments of the first aspect, the step of generating a corresponding target measurement matrix by combining the target compression ratio data specifically includes: obtaining a modality type identifier of the high-dimensional auxiliary sensing data; matching the target projection seed data in a preset projection seed library by combining the modality type identifier; performing a matrix expansion operation by combining the target compression ratio data and the target projection seed data to obtain a candidate measurement matrix; calculating the mutual coherence coefficient between the candidate measurement matrix and a preset sparse basis dictionary; and determining the candidate measurement matrix as the target measurement matrix when the mutual coherence coefficient is less than a preset orthogonality threshold.

[0011] In the above embodiments, the data processing system extracts projection seeds by combining modal identifiers and generates candidate matrices by compression ratio expansion. The orthogonality of the matrix with the sparse base dictionary is verified by mutual coherence coefficients, ensuring that the measurement matrix satisfies the constrained equidistant property, which is beneficial for the lossless reconstruction of ground control nodes.

[0012] In conjunction with some embodiments of the first aspect, in some embodiments, the step of performing a sparse basis projection transformation on the high-dimensional auxiliary sensing data to separate the common sparse component reflecting the common physical response across modes and the proprietary sparse component corresponding to each sensor mode specifically includes: mapping the high-dimensional auxiliary sensing data to a preset sparse basis dictionary to obtain an initial sparse coefficient matrix; calculating the cross-correlation matrix between the initial sparse coefficient matrices of different modes; extracting the overlapping feature vectors in the initial sparse coefficient matrix based on the cross-correlation matrix, and determining the overlapping feature vectors as common sparse components; and removing the common sparse components from the initial sparse coefficient matrix to obtain proprietary sparse components.

[0013] In the above embodiments, the data processing system extracts the cross-correlation coefficients between sparse coefficient matrices of different modes to lock the overlapping feature vectors, decouples the fuselage physical response into common components and proprietary components, breaks through the limitation of independent evaluation of single modes, and substantially reduces the total amount of sampling data required for joint reconstruction.

[0014] In conjunction with some embodiments of the first aspect, in some embodiments, the step of calculating the cross-correlation matrix between the initial sparse coefficient matrices of different modes specifically includes: obtaining the time synchronization stamp data of the initial sparse coefficient matrices of different modes; performing a time-domain alignment operation on the initial sparse coefficient matrices of different modes in combination with the time synchronization stamp data to obtain an aligned sparse matrix; extracting the low-frequency envelope feature vectors in the aligned sparse matrix; and calculating the inner product value of the low-frequency envelope feature vectors of different modes to obtain the cross-correlation matrix.

[0015] In the above embodiments, the data processing system uses time synchronization stamps to align sparse matrices and extracts low-frequency envelope feature vectors to calculate inner product values, accurately filtering out high-frequency noise interference unique to each sensor mode, and ensuring the numerical stability and accuracy of cross-modal cross-correlation coefficient calculation.

[0016] In conjunction with some embodiments of the first aspect, in some embodiments, after the step of writing a multimodal multiplexed data stream to a network communication port in response to a transmission permission identifier of an external communication link to perform the data transmission of multimodal sensor data, the method further includes: receiving link quality feedback data sent by a ground control node; obtaining channel fading parameters in the link quality feedback data; and updating the instantaneous channel capacity assessment value in combination with the channel fading parameters.

[0017] In the above embodiments, the data processing system uses the channel fading parameters fed back by the ground control node to update the instantaneous channel capacity assessment value in real time, and constructs a closed-loop monitoring link that accurately reflects the available transmission capacity of the physical layer, thus avoiding network congestion caused by blindly sending data in an open loop.

[0018] In some embodiments of the first aspect, after updating the instantaneous channel capacity assessment value with channel fading parameters, the method further includes: calculating the difference between the updated instantaneous channel capacity assessment value and the total amount of data to be transmitted in the current transmission queue to obtain a capacity gap value; when the capacity gap value is less than zero, extracting the secondary priority label of the dimensionality reduction feature measurement vector in the transmission queue; combining the secondary priority label and the capacity gap value, removing the dimensionality reduction feature measurement vector corresponding to the amount of data from the transmission queue to obtain a truncated multiplexed data stream; and rewriting the truncated multiplexed data stream into the network communication port.

[0019] In the above embodiments, when the capacity gap value is less than zero, the data processing system actively truncates and removes low-priority measurement vectors based on the tags, and establishes a congestion blocking mechanism in extremely harsh communication environments, ensuring that the high-priority feature vectors retained in the current queue can be sent out.

[0020] In a second aspect, embodiments of this application provide a data processing system comprising: one or more processors and a memory; the memory is coupled to the one or more processors and is used to store computer program code, the computer program code including computer instructions, wherein the one or more processors invoke the computer instructions to cause the data processing system to perform the method described in the first aspect and any possible implementation thereof.

[0021] Thirdly, embodiments of this application provide a computer-readable storage medium including instructions that, when executed on a data processing system, cause the data processing system to perform the method described in the first aspect and any possible implementation thereof.

[0022] Fourthly, embodiments of this application provide a computer program product containing instructions that, when the computer program product is run on a data processing system, cause the data processing system to perform the method described in the first aspect and any possible implementation thereof.

[0023] Understandably, the data processing system provided in the second aspect, the computer storage medium provided in the third aspect, and the computer program product provided in the fourth aspect are all used to execute the methods provided in the embodiments of this application. Therefore, the beneficial effects they can achieve can be referred to the beneficial effects in the corresponding methods, and will not be repeated here.

[0024] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0025] 1. By adopting a cross-modal sparse structure derived from common physical excitation, redundant transmission between heterogeneous signals is effectively eliminated. At the same time, a high-priority protection channel with physical isolation level is established at the resource scheduling level. This effectively solves the problems in the existing technology, such as the difficulty of static measurement matrix to adapt to time-varying fading channels and the low bandwidth contention and flight safety hazards caused by the low compression efficiency of heterogeneous sensors. As a result, it realizes the efficient and reliable transmission of massive heterogeneous airborne data to the ground station without relying on the constraints of ultra-wideband hardware links. This ensures that the flight status of large cargo platforms under complex operating conditions can be reconstructed and monitored by the ground control center in real time and with high fidelity.

[0026] 2. By adopting a cross-layer closed-loop adaptive feedback architecture that integrates physical layer channel state awareness and application layer data importance rating, the airborne coding end can finely allocate the compression ratio of each visual or point cloud data according to the actual available network throughput in each transmission cycle. This effectively solves the system bottleneck in existing technologies where the use of equal compression ratios or preset static parameters leads to the complete loss of a large amount of sensing data due to link congestion when multipath fading occurs during UAV maneuvering. This maximizes the transmission potential of the communication system in time-varying and complex electromagnetic environments, and ensures that the data processing system can still perform smooth degradation of each mode according to importance weight when facing a cliff drop in available bandwidth, thus avoiding the phenomenon of global paralysis of the upper-layer flight monitoring link caused by lower-level data congestion.

[0027] 3. By employing a joint sparse analysis mathematical framework for heterogeneous sensor observation data, the vibration frequency band of the fuselage and the visual displacement pixels are correlated and mapped to a unified low-dimensional subspace. This eliminates the redundant coding overhead caused by different hardware modes repeatedly capturing the same physical event. It effectively solves the technical barrier of existing independent compressed sensing schemes limited by single modes failing to explore the joint structure between signals, resulting in a high minimum measurement threshold required for overall reconstruction. This enables deep reorganization and transformation of the original high-dimensional airborne data while maintaining the output accuracy of the ground depth reconstructor. It also significantly reduces the transmission bandwidth requirements of the airborne radio frequency front-end, allowing conventional commercial communication frequency bands to carry the composite transmission of high-definition multispectral video streams and space radar point cloud data. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the topology of the data processing system in an embodiment of this application;

[0029] Figure 2 This is a schematic diagram of a data transmission interaction process of the data processing system in an embodiment of this application;

[0030] Figure 3 This is a flowchart illustrating a method for multimodal data transmission of large unmanned aerial vehicles based on compressed sensing, as described in this application.

[0031] Figure 4 This is a schematic diagram of the physical device structure of a data processing system in an embodiment of this application. Detailed Implementation

[0032] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification of this application, the singular expressions “a,” “an,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to any or all possible combinations including one or more of the listed items.

[0033] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0034] Large cargo drones need to provide real-time feedback of multi-dimensional status parameters to ground control nodes during flight operations. This includes heterogeneous physical information such as six-axis data from the inertial measurement unit, photoelectric and infrared dual-light video images, LiDAR spatial point cloud sequences, and airframe vibration and strain data. Traditional wireless sensor network architectures typically operate in an independent acquisition mode, resulting in massive and highly correlated redundant payload data streams generated when multiple sensors simultaneously capture the same environmental stimuli such as airflow disturbances. Limited by the bandwidth constraints of existing UAV data transmission links such as orthogonal frequency division multiplexing, the continuously increasing aggregated data volume can easily lead to channel overload.

[0035] Based on this, this application mines the joint sparse prior features among heterogeneous modes and integrates the channel quality closed-loop feedback scheduling strategy. At the data generation end, it implements dimensionality reduction and adaptive feature compression of cross-modal common dimensions and deploys a deep unfolding reconstruction network at the receiving end to restore the original physical field. This is to smooth out transient data surges and ensure the deterministic transmission of security-critical information.

[0036] Please see Figure 1 This is a schematic diagram of the topology of the data processing system in this application embodiment; Figure 1 The diagram shows a collaborative topology architecture that includes an airborne terminal 10, a data transmission link, and a ground station 11.

[0037] The airborne terminal 10 contains a hardware cluster responsible for physical signal sensing, which specifically includes an IMU / flight controller 1011, an EO / IR camera 1012, a LiDAR point cloud 1013, and a vibration / stress sensor 1014.

[0038] The aforementioned sensing hardware has been pre-labeled with initial priority tags during the device driver initialization phase: the IMU / flight controller 1011 is labeled P0, the EO / IR camera 1012 and vibration / stress sensor 1014 are labeled P1, and the LiDAR point cloud 1013 is labeled P2. The priority determination module 102 is responsible for reading and verifying these pre-labeled tags, performing modal sparse basis estimation operations, and calculating intermodal correlation matrices. For extended sensor devices without pre-labeled tags, this module dynamically assigns a P1 or P2 level tag based on data characteristics.

[0039] The adaptive joint CS measurement matrix generation module 103 (hereinafter referred to as the measurement strategy module 103) is coupled in series with the priority determination module 102. It is responsible for the compression ratio allocation optimization algorithm based on CSI feedback, cross-modal RIP joint design rules and random seed management to generate or update the measurement matrix set.

[0040] The graded compression sampling and quantization module 104 (hereinafter referred to as the compression processing module 104) receives the allocation instructions of the above measurement matrix and performs differentiated encoding for different grades: for P0 key data, it performs no compression or light encoding logic to ensure that the end-to-end latency is less than 10ms; for P1 auxiliary data, it performs CS compression rate encoding in the range of 30% to 50% to limit the latency to less than 100ms; for P2 low priority data, it performs CS compression rate encoding in the range of 10% to 30% to limit the latency to less than 1s.

[0041] The encoded stream processed by the airborne terminal 10 is sent downlink via a data transmission link covering COFDM, 5G or satellite channels.

[0042] The other side of the data transmission link is connected to the ground station 11, which includes a depth unfolding CS reconstructor 111 (referred to as reconstructor 111) that runs on a ground GPU server. It is responsible for parallel execution of the ISTA-Net++ image / video modal reconstruction process, the ADMM-DAD vibration / IMU temporal modal reconstruction process, and the cross-modal common sparse component fusion layer processing.

[0043] The final reconstructed physical state variables are pushed to the multimodal data fusion + CSI estimation + compression rate feedback controller 112 (hereinafter referred to as controller 112). On the one hand, controller 112 transfers the reconstructed signal to the flight management system for service use, and on the other hand, performs channel quality estimation and sends the uplink CSI feedback back to the airborne terminal 10 along the data transmission link to maintain the adaptive closed-loop operation of the entire system.

[0044] exist Figure 1 In the topology shown, the interaction logic between the modules includes:

[0045] The IMU / flight controller 1011 in the sensing hardware cluster outputs a raw sequence of six-axis acceleration and angular velocity values ​​generated at a fixed sampling period. Its data packet header carries a P0-level priority tag written by the device driver during the initialization phase. The EO / IR camera 1012 outputs a two-dimensional pixel matrix data stream in frames, with each frame carrying a timestamp and a P1-level priority tag. The LiDAR point cloud 1013 outputs a set of three-dimensional coordinate points in scan cycles, with each set of points containing reflection intensity attributes and carrying a P2-level priority tag. The vibration / stress sensor 1014 outputs a multi-channel time-domain vibration waveform sampling value sequence, carrying a P1-level priority tag.

[0046] When the above four types of data arrive at the priority determination module 102 in parallel through the internal bus of the airborne terminal, the priority determination module 102 first reads the priority tag in the header of each data packet and performs a traffic splitting determination. The P0 tag data is directly bypassed to the uncompressed channel of the compression processing module 104 without performing any transformation operation. The data carrying P1 and P2 tags are kept inside the priority determination module 102 and the subsequent sparsity estimation and intermodal correlation matrix calculation operations are performed.

[0047] The data object transmitted from the priority determination module 102 to the measurement strategy module 103 includes two parts: first, the estimated sparsity numerical vectors of each P1 and P2 mode, which are scalar sets used to indicate the number of non-zero coefficients of each mode signal under the corresponding sparse basis; second, the cross-correlation matrix between modes, which is a symmetric square matrix, where the element in the i-th row and j-th column represents the quantized value of the degree of sparse support overlap between the i-th mode and the j-th mode.

[0048] After receiving the above sparsity information, the measurement strategy module 103 combines the instantaneous channel capacity assessment value contained in the CSI feedback data received in reverse through the data transmission link, performs constraint optimization to determine the target compression ratio of each mode, and generates a set of measurement matrices of the corresponding dimension from the projection seed library accordingly, and transmits the set of measurement matrices to the compression processing module 104.

[0049] The input received by the compression processing module 104 includes three items: the P0 level raw data bypassed by the priority determination module 102, the common sparse components and each proprietary sparse component after sparse basis projection separation output by the priority determination module 102, and the measurement matrix set output by the measurement strategy module 103.

[0050] The compression processing module 104 performs only light channel coding on the P0 data and directly encapsulates it into a transmission frame; it performs linear projection operations on the measurement matrix on the common and proprietary sparse components to obtain a dimension-reduced measurement vector, and then performs quantization and channel coding encapsulation. All encapsulated frames are merged into a unified transmission queue and written into the downlink channel of the data transmission link.

[0051] On the ground station 11 side, the input received by the reconstructor 111 is a compressed data packet transmitted via a data link and demodulated by the physical layer. The P0 frame is directly decoded and restored to the original IMU data and sent to the flight management system. The dimensionality reduction measurement vectors in the P1 and P2 frames are sent to the GPU parallel reconstruction network. This network uses the output of the cross-modal common sparse component fusion layer as the prior constraint condition for each modal independent reconstruction branch to complete the signal recovery.

[0052] The controller 112 receives the mode recovery signals output by the reconstructor 111. On the one hand, it performs multi-source data fusion and then hands it over to the flight management system. On the other hand, it uses the pilot symbol sequence in the received data packet to estimate the downlink channel fading state, encapsulates the estimated channel fading parameters into a CSI feedback message, and sends it back to the airborne terminal 10 through the uplink. The feedback message is received by the measurement strategy module 103 to trigger the next round of measurement matrix update.

[0053] Based on the above, the system's operating state can include four modes, and the state transition logic is as follows:

[0054] (1) Cold start mode: Entered after the system is powered on, using conservative compression ratio configuration, the duration is until the first valid CSI feedback message is received (usually 2-5 transmission cycles, about 200-500ms).

[0055] (2) Normal adaptive mode: CSI feedback is received and the channel capacity assessment value is stable at Upon entering the above state, the system will run according to the real-time optimized compression ratio configuration;

[0056] (3) Channel degradation mode: Triggered when the capacity gap value is <0 or the CSI feedback message loss rate is >30% for 3 consecutive cycles, the system locks the compression rate of each mode to the minimum limit. To reduce computational overhead and ensure a minimum level of reliable transmission, compression ratio optimization is paused.

[0057] (4) Emergency Simplified Mode: When the channel capacity falls below the safety threshold (usually) When the channel is at 50% capacity, the system transmits only P0 level data and suspends all P1 / P2 data. Once the channel is restored, it will automatically switch back to normal mode.

[0058] The mode transition determination is executed in the measurement strategy module 103 of each transmission cycle, with a state machine transition delay of <5ms, ensuring the stability of the system when the link quality fluctuates drastically.

[0059] Please see Figure 2 This is a schematic diagram of a multimodal sensor data efficient transmission and interaction process in the data processing system of this application embodiment. Figure 2 This demonstrates the sequential interaction between the airborne terminal 10 and the ground station 11 around a complete transmission cycle. The interaction starts with the sensor data acquisition of the airborne terminal 10 and ends with the CSI feedback message from the ground station 11 being sent back to the airborne terminal 10. The interaction involves the orderly connection of various stages such as data classification, sparsity estimation, compression ratio optimization, matrix generation, linear projection, channel transmission, priority routing, deep reconstruction, and channel state estimation.

[0060] On the airborne end 10 side, after the system powers on and completes hardware self-test and all sensor device ready flags are valid, each sensor starts concurrent acquisition according to the preset raw sampling rate. This process is the execution content of step S201. Its output is a multi-channel parallel data stream continuously output on the airborne internal bus. The protocol header of each data packet carries the priority tag field written by the device driver during the initialization phase.

[0061] When any data packet arrives at the priority determination buffer, step S202 is triggered. The system reads the header tag of the data packet and determines whether it belongs to P0 level flight safety critical data. This determination is a single enumeration value comparison operation, and the computational overhead is negligible. If the determination result is P0, the data packet is bypassed and forwarded to the transmission buffer entry in step S206 to wait for packaging and transmission. This bypass path physically bypasses all subsequent sparse transformation and compression operation modules, ensuring that only one bus scheduling cycle delay occurs between the completion of sensor acquisition and the data entering the RF transmission buffer. If the determination result is not P0, the data packet is routed to the compression queue in step S203 for further processing.

[0062] After entering step S203, which is the sliding window DCT / wavelet sparsity calculation, the system needs to wait for the accumulated amount of data to be compressed in each mode within the current sliding time window to reach the minimum estimated window length requirement before it can start online sparsity estimation. The window length is determined by the period of the lowest effective frequency component in each modal signal. For fuselage vibration signals, it usually corresponds to data segments on the order of tens of milliseconds, while for video frames, it corresponds to a single frame or a set of image blocks of several consecutive frames.

[0063] If the window data is ready, the system selects the corresponding preset sparse basis dictionary based on the physical characteristics of each mode: for image and video modes, a two-dimensional discrete cosine transform (DCT) basis or wavelet basis is selected as the sparse basis dictionary; for vibration and IMU time series signals, a Fourier basis or wavelet packet decomposition basis is selected as the sparse basis dictionary; for LiDAR point clouds, a sparse dictionary based on spatial clustering is selected. The system performs a projection transformation (i.e., orthogonal or overcomplete decomposition) on the corresponding sparse basis dictionary for each mode data within the window, sorts the transformed coefficients in descending order of absolute value, and counts the number of non-zero coefficients exceeding the energy cutoff threshold. This number is the estimated sparsity of the mode at the current time.

[0064] Simultaneously, the system performs position-by-position cross-comparison of the transform coefficient vectors of multiple modes within the same time window to evaluate which sparse domain index positions exhibit synchronous amplitude activation characteristics of different modes, thereby outputting a cross-correlation matrix that quantifies the joint sparsity among the modes. The final output of step S203 is the real-time sparsity scalar set of each mode and the cross-correlation matrix between modes, both of which are written to the on-chip shared register set for subsequent steps to read.

[0065] After obtaining the sparsity information, the system needs to determine a suitable compression ratio based on the real-time channel status. Since wireless channels are constantly changing due to factors such as multipath fading and weather conditions, a fixed compression ratio configuration cannot guarantee transmission reliability when the channel deteriorates, nor can it fully utilize bandwidth resources when the channel conditions are good. Therefore, the system must obtain channel quality information fed back from the ground station before it can perform adaptive measurement matrix generation. This process is implemented through step S204. It should be noted that the sparsity estimation process in step S203 and the CSI feedback reception process in step S204 can be executed asynchronously and in parallel. When the airborne unit has not received a new CSI feedback message in the current cycle, step S204 will use the old channel capacity assessment value cached in the previous cycle, without blocking the calculation process of step S203.

[0066] Step S204 is triggered by the airborne radio frequency receiving channel detecting a complete CSI feedback message from the ground station, which, after CRC verification, sets the arrival flag. This message carries the ground station's assessment of the downlink quality in the previous transmission cycle. The system extracts the instantaneous channel capacity assessment value from this message and converts it into the total available transmission bit rate. Then, combined with the mode sparsity information written to the shared register in step S203, it determines the minimum compression ratio lower limit required to ensure target reconstruction accuracy for each mode. Subsequently, with the goal of maximizing the total information retention under importance weighting, the total available bit rate as an equality constraint, and the minimum compression ratio lower limit for each mode as an inequality constraint, an optimization model is constructed and solved. Finally, a set of parameter vectors containing the target compression ratio values ​​for each P1 and P2 mode is output. This vector is directly passed to the input of the matrix generation module in step S205. It is important to emphasize that if the airborne terminal does not receive a new CSI feedback message in the current cycle, the system will continue to perform compression ratio calculation using the old channel capacity assessment value cached in the previous cycle, ensuring that the data processing pipeline is not blocked due to lack of feedback.

[0067] Step S205 involves deterministically generating the measurement matrix after receiving the target compression ratio parameter vector. The system reads the corresponding pseudo-random number generator seed value from non-volatile memory based on each modality type identifier. After initializing the generator state using the seed, it fills the matrix entries element-wise according to the number of rows determined by the target compression ratio and the number of columns determined by the original signal dimension, and performs row normalization. The memory overhead of this process is only storing the seed value itself, not the complete matrix, making it suitable for resource-constrained onboard computing environments. After generation, the system also needs to perform cross-coherence coefficient verification to confirm that the orthogonality between the candidate matrix and the sparse basis dictionary meets the preset threshold requirement before it can be finalized as the target measurement matrix for the current cycle. At this point, the measurement matrix is ​​ready in the on-chip SRAM, and the hardware multiply-accumulate array in step S206 can directly address its row addresses.

[0068] While the measurement matrix is ​​ready, the sparse basis projection transformation performed on the high-dimensional auxiliary sensing data in step S203 has completed the separation of common sparse components from proprietary sparse components. These sparse components, as the source data to be compressed, together with the measurement matrix generated in step S205, constitute all the input elements required for the compressed sensing linear projection operation. The system then proceeds to the actual dimensionality reduction and compression stage, i.e., step S206.

[0069] Step S206 receives two inputs: first, the common sparse component and the mode-specific sparse component separated after the sparse basis projection transformation in step S203; second, the target measurement matrix ready in step S205. The multiply-accumulate array performs dot-matrix accumulation on each row of the measurement matrix with the vector of the sparse component to be compressed. After all rows are processed, the scalars of the dot products are concatenated to form a dimensionality-reduced measurement sub-vector with a dimension much smaller than the length of the original signal. The common component and each specific component produce their own sets of compressed measurement sub-vectors after this operation. The sum of these sub-vectors constitutes the dimensionality-reduced feature measurement vector for this period, whose data volume has been proportionally reduced relative to the original signal according to the target compression ratio. That is, let the sparse component vector obtained after the sparse basis transformation of a certain mode be... Its length is N, and the measurement matrix generated in step S205 The system performs matrix-vector multiplication when the dimension is M rows and N columns, with M being much smaller than N. , where the output vector The length of M is the dimensionality reduction measurement vector of this mode.

[0070] Step S207 performs unified protocol encapsulation on the dimensionality-reduced feature measurement vector and the bypassed P0 raw data: fixed-point quantization is performed on the measurement vector to convert the floating-point representation into a fixed-bit-width integer; LDPC or Turbo code forward error correction is applied; and a frame header field is inserted at the beginning of each coded block to mark the P0, P1, or P2 level identifier and frame synchronization code, forming a standardized protocol frame sequence. This sequence is arranged into the transmission queue according to priority from high to low.

[0071] Step S208 is triggered when the channel idle flag of the physical layer RF baseband chip is valid. The direct memory access controller moves the protocol frames in the queue to the RF transmission buffer according to the frame priority order. After the baseband chip performs OFDM modulation, digital-to-analog conversion and power amplification, it is radiated to the space channel through the antenna. At this point, the downlink data transmission action of a complete processing cycle of the airborne end is completed.

[0072] On the ground station 11 side, step S211 is initiated when the downlink radio frequency signal power captured by the receiving antenna array exceeds the receiving sensitivity threshold. The radio frequency front end sequentially performs low noise amplification, downconversion, analog-to-digital conversion, OFDM demodulation, channel equalization and channel decoding operations to recover the multi-priority mixed data packet sequence that has been confirmed to be correct by CRC check.

[0073] Step S212 performs priority identification reading and path distribution on the recovered data packets: packets with frame header marked P0 are sent to the instant decoding channel in step S213, and packets marked P1 or P2 are sent to the deep reconstruction channel in step S214. Step S213 performs dequantization on the P0 packets to directly recover the original IMU values ​​or flight control command sequences and write them to the real-time database of the flight control monitoring system. The end-to-end latency of this path is controlled within 10 milliseconds and does not involve any CS reconstruction calculations.

[0074] Step S214 feeds the dimensionality-reduced measurement vectors P1 and P2 into a deep unfolded CS reconstruction network deployed on a ground-based GPU server. For the measurement vectors of image and video modalities, the network uses the ISTA-Net++ architecture to perform multi-layer feedforward inference. Each layer structurally corresponds to one proximal gradient step of a traditional iterative shrinkage thresholding algorithm, but the threshold parameters and transformation operators are learned end-to-end, allowing a fixed number of layers to approximate the reconstruction accuracy of traditional algorithms after dozens of iterations. For the measurement vectors of vibration and IMU temporal modalities, the network uses an ADMM-DAD unfolded architecture. Each layer corresponds to an alternating update of the original and dual variables, where the analysis operator is replaced by a learned dictionary trained on the statistical characteristics of vibration signals instead of a manually designed regularization operator. After each modality independently reconstructs and produces a preliminary recovered signal, it enters the cross-modal fusion reconstruction stage in step S215. In this stage, the independent reconstruction results of the previously obtained common sparse components are injected as priors into the intermediate feature layers of each modality's reconstruction network. Specifically, the common component reconstruction vector is processed through a learnable linear mapping and then element-wise added to the intermediate features of each modality. This forces each modality to output a consistent recovered value at the sparse domain location corresponding to the common physical excitation, thereby utilizing cross-modal mutual information to reduce the residual error of each single-modal reconstruction. Step S215 outputs the final reconstructed signals of each modality after joint enhancement.

[0075] In some embodiments, if the ground station detects that the dimensionality reduction measurement vector of a frame is completely lost due to channel depth fading, or if the residual still exceeds a preset threshold after reconstruction algorithm iterations, the system will trigger a retransmission request mechanism: sending a retransmission request signaling for the frame to the airborne terminal. Upon receiving the request, the airborne terminal will extract the original data of the corresponding frame from its local cache (the airborne terminal maintains a rolling cache of the original data for the most recent N periods), recompress it with a higher compression ratio (20%-30%), and then send it preferentially. If multiple retransmissions still fail, the ground station will use a temporal interpolation algorithm based on adjacent frames to generate an approximate reconstruction result for the frame and mark it as "low confidence" for upper-layer service reference.

[0076] Step S216 formats the reconstructed signal into the standard data interface format required by the flight management system and writes it to the system bus for use by the situation display, trajectory planning and decision support modules.

[0077] Steps S217 and S211 are executed in parallel in time: The system uses the known pilot symbol sequence embedded in the received data packet and the local reference sequence to perform cross-correlation calculations to estimate the frequency-selective fading characteristics and equivalent signal-to-noise ratio of the current downlink channel. Based on this, an updated instantaneous channel capacity assessment value is obtained, and this assessment value, along with the channel fading trend parameters, is encapsulated into a CSI feedback message and sent back to the airborne terminal 10 via the uplink RF link. After the feedback message arrives at the airborne terminal, it will be read and parsed in step S204 of the next processing cycle, so that the compression rate allocation of the airborne terminal can track the changes in the real channel state, completing the full-system adaptive closed-loop control across the airborne terminal and the ground station.

[0078] The entire interaction process is executed cyclically within each transmission cycle. The cycle length is determined by the greater of the sensor’s maximum acquisition frame interval and the round-trip delay of the CSI feedback message, and is typically in the range of tens to hundreds of milliseconds.

[0079] The following is a detailed description of the process of the method provided in this implementation. Please refer to [link / reference]. Figure 3 This is a flowchart illustrating a method for multimodal data transmission of large unmanned aerial vehicles based on compressed sensing, as described in this application.

[0080] S301. Based on the preset data priority labels, the acquired multimodal sensor data is divided into uncompressible safety-critical time-series data and high-dimensional auxiliary sensing data to be compressed.

[0081] Among them, multimodal sensor data represents heterogeneous generic datasets captured by various detectors deployed around the cargo drone; data priority labels represent level identification codes written into the underlying drivers of each device during the system initialization phase to distinguish business importance; safety-critical time-series data represents the set of basic variables directly related to the aircraft's attitude control and positioning navigation; and high-dimensional auxiliary perception data represents environmental survey payload information with a large data volume but tolerating a certain degree of reconstruction error.

[0082] Specifically, the data processing system initiates this step upon receiving concurrent data streams aggregated from various hardware acquisition channels. The system parses the Ethernet or bus protocol header fields of the data packets, extracts the embedded service identity identifier, and looks up the corresponding data priority tag according to the mapping table. When the current data source is determined to be from the flight control board or inertial navigation unit, the system marks it as uncompressible and allocates the corresponding data to the safety-critical time-series data buffer pool. When the current data source is determined to be from a downward-looking camera, infrared thermal imager, or 3D laser scanner, the system marks it as requiring compression and directs it to the high-dimensional auxiliary sensing data buffer pool. This classification and isolation operation ensures that subsequent allocation of computing resources avoids interference with basic safety control commands, guaranteeing the timeliness of core flight data.

[0083] In some embodiments, non-standard extended sensor devices may generate new data frames that are not included in the preset tag dictionary. In response, the system will activate the anomaly classification daemon process, uniformly downgrade unidentified data to the lowest priority category tag to be compressed, and append a pending review flag to the end of the frame before reporting it to the monitoring module, to prevent unknown massive amounts of data from crowding out the dedicated processing cycle of secure time-series data.

[0084] After data classification, the core challenge for the system is how to compress the high-dimensional assisted sensing data. Because the wireless channel between the airborne UAV and the ground station exhibits time-varying characteristics due to factors such as weather conditions, flight attitude, and terrain obstruction, the available bandwidth varies significantly at different times. Using a fixed compression rate configuration will lead to data congestion and packet loss when channel conditions deteriorate, or wasted bandwidth resources when channel conditions are good. Therefore, the system must dynamically adjust the compression parameters for each mode based on the real-time channel status, requiring the acquisition of current channel capacity assessment information, i.e., proceeding to step S302.

[0085] S302. Extract the instantaneous channel capacity assessment value from the external uplink signaling, and combine it with the importance weight parameters of the high-dimensional auxiliary sensing data to determine the target measurement matrix.

[0086] Among them, the external uplink signaling represents the heartbeat protocol packet sent from the ground station to the UAV along the radio frequency control link; the instantaneous channel capacity assessment value represents the upper limit of the current available transmission rate calculated by the ground station based on the received signal-to-noise ratio and bit error rate; the importance weight parameter represents the multiplicative coefficient assigned to different types of auxiliary sensing applications, which determines the bandwidth allocation ratio when the channel degrades; and the target measurement matrix represents a random sampling projection matrix with a specific number of rows and columns and satisfying the constrained equidistant property.

[0087] Specifically, before performing compression operations on the data in the high-dimensional auxiliary sensing buffer pool, the data processing system retrieves environmental reference values. The system retrieves the latest arriving uplink signaling parsing block from the shared register of the RF demodulation module and reads the instantaneous channel capacity assessment value. Simultaneously, the system traverses the queued high-dimensional auxiliary sensing data types in the current buffer pool, extracting the importance weight parameters corresponding to each type. Based on the total available bandwidth represented by the instantaneous channel capacity assessment value, the system calculates the expected transmission bit rate allocated to each mode at this moment, using the importance weight parameters as a skew basis, and converts this transmission bit rate into a specific compression dimension ratio. Finally, the system generates or calls the target measurement matrix with the corresponding row and column dimensions based on the calculated dimension ratio parameters, constructing a mathematical basis for subsequent dimension reduction projection.

[0088] In some embodiments, the data processing system acquires the total available bandwidth data corresponding to the instantaneous channel capacity assessment value, and acquires the minimum compression ratio lower limit data for each high-dimensional auxiliary sensing data; combines the total available bandwidth data, importance weight parameters, and minimum compression ratio lower limit data to perform constraint optimization operations to obtain the target compression ratio data for each high-dimensional auxiliary sensing data; and generates the corresponding target measurement matrix by combining the target compression ratio data.

[0089] Among them, the minimum compression ratio lower limit data represents the lower bound of the minimum effective information retention ratio that does not produce serious semantic distortion at the decompression end, as defined by the specific physical characteristics of the sensor; the target compression ratio data represents the precise dimensionality reduction projection retention ratio coefficient assigned to each sensor after mathematical game compromise.

[0090] Specifically, after identifying the instantaneous channel capacity assessment value, the data processing system initiates a fine-grained bandwidth allocation mechanism. The system extracts the assessment value and converts it into quantified total available bandwidth data. It then retrieves the minimum compression ratio limit data from various sensor-driven configuration items to ensure image readability or point cloud density, thus defining the operational red lines. The system constructs an optimization equation with the objective of maximizing the overall importance-weighted information content. The system uses the total available bandwidth data as an equality constraint and the minimum compression ratio limit data as an inequality truncation boundary constraint. The system inputs importance weight parameters as weighting coefficients for the data rate terms of each modality in the objective function, tilting the optimization solution towards higher-importance modes by allocating more bandwidth. Constraint optimization operations are then performed in multi-dimensional space, outputting mode-specific target compression ratio data that satisfies all boundary conditions. Finally, the system uses these dynamically calculated scaling factors to shape the target measurement matrix of the corresponding size.

[0091] The constraint optimization operation is based on channel state information periodically fed back from the ground station. The ground station periodically (at intervals ΔT) feeds back instantaneous channel capacity estimates to the airborne terminal via uplink signaling. (Unit: Mbps). The airborne terminal executes an adaptive mechanism based on the priority of each modal data.

[0092] The specific data processing logic for this constraint optimization operation is as follows: Suppose there are L types of high-dimensional auxiliary sensing data modes waiting for compressed transmission, and the original signal vector of the l-th type of sensor at time t is... , dimension Each signal exists in its respective sparse basis. The following is represented as:

[0093] ;in It is a sparse coefficient vector.

[0094] The importance weight corresponding to the l-th type of sensor is: (Higher priority) The smaller the value, the more likely it is to prioritize the integrity of high-priority signals. The system denotes the target compression ratio of the l-th type of sensor as . The compression ratio is defined as the length of the measurement vector. Compared with the original signal dimension The ratio, that is: Then the l-th type of sensor has a compression ratio The corresponding data rate during transmission is: ;in This refers to the number of quantization bits for each sample value of this mode; The term represents the redundancy overhead introduced by channel coding. In practical systems, Adaptive adjustment based on current channel quality: when channel quality is good, use a high code rate (e.g., 5 / 6) to improve spectral efficiency; when channel quality is poor, use a low code rate (e.g., 1 / 2) to enhance error correction capability.

[0095] The complete optimization problem constructed by the airborne end based on the priority of each modal data is as follows:

[0096] ;

[0097] Subject to:

[0098] ; ;

[0099] in The minimum allowable compression ratio for Class I sensors (critical safety data) =1.0 means no compression, load auxiliary data =0.1).

[0100] The physical meaning of this optimization objective function is to maximize the importance of high-importance modes (smaller) within a limited bandwidth. (Value) to achieve a higher compression ratio and thus retain more information.

[0101] Since the objective function is linear, the equality constraints are linear, and the inequality constraints are box constraints, this problem belongs to linear programming, and its optimal solution has the following analytical structure:

[0102] First, assign a minimum compression ratio lower limit to each modality. Required minimum bandwidth, calculate the total minimum bandwidth:

[0103] ;

[0104] from The remaining allocable bandwidth is obtained by deducting from the remaining bandwidth:

[0105] ;

[0106] Define the unit bandwidth importance gain density for each mode as:

[0107] ;

[0108] System by Sort the modalities from smallest to largest (because) The smaller the value, the higher the priority. The smaller the value, the higher the priority of that mode in acquiring resources. Without loss of generality, assume that after sorting... In sequence The mode with the highest gain density (minimum η) is assigned until its compression ratio reaches the upper limit of 1. The maximum additional bandwidth increment that the l-th type of sensor can absorb is:

[0109] ;

[0110] The allocation algorithm performs the following iterative process:

[0111] initialization ;

[0112] For the modes sorted by η :

[0113] ;

[0114] ;

[0115] The final output target compression ratio data is the compression ratio of each mode. Value set.

[0116] For example: Suppose there are two modes to be compressed, mode A is an EO / IR video keyframe, and its parameters are... =65536 (corresponding to a 256×256 pixel block). =8 bits =0.7, =0.3; Mode B is a LiDAR non-keyframe point cloud with the following parameters: =32768, =12 bits =0.3, =0.1. If the instantaneous channel capacity estimate fed back by the ground station... =2621440bit / s, then:

[0117] The guaranteed bandwidth is:

[0118] =0.3*65536*8+0.1*32768*12=196608bit / s;

[0119] The remaining allocable bandwidth is:

[0120] =2621440-196608=2424832bit / s;

[0121] Calculate the return density:

[0122] ;

[0123] ;

[0124] Because (that is, Mode B has higher priority, because =0.3< =0.7), it is preferentially allocated to Mode B. The maximum additional bandwidth that Mode B can absorb is:

[0125] =(1-0.1)*32768*12=353894bit / s;

[0126] Since =<2424832> =353894, so =1, update =2424832-353894=2070938bit / s.

[0127] The maximum additional bandwidth that Mode A can absorb is:

[0128] =(1-0.3)*65536*8=367002bit / s;

[0129] Since =<2070938> =367002, so =1.

[0130] If is small resulting in insufficient remaining bandwidth, for example =400000bit / s, then =400000-196608=203392bit / s, since 203392< =353894, then:

[0131] =0.1+0.518=0.618;

[0132] =0.3 (maintain the guaranteed minimum value);

[0133] Through the above process, the target compression ratio data output by the system is the finally determined value set for each mode.

[0134] In addition, when the system is started for the first time or re-establishes a connection after a long-time link interruption, the airborne end adopts a preset conservative initialization strategy: the instantaneous channel capacity evaluation value is initialized to the minimum guaranteed rate of the communication link (usually set to 50%-60% of the nominal link capacity), and the compression ratio of each mode is uniformly initialized to its minimum compression ratio lower limit To ensure the reliability of the first frame of data transmission to the greatest extent possible, the system switches to normal adaptive operating mode after receiving the first CSI feedback message. This cold start protection mechanism ensures that the system can safely initiate the data transmission process even when there is no historical channel information.

[0135] In some embodiments, the data processing system acquires the modality type identifier of the high-dimensional auxiliary sensing data; matches the target projection seed data in a preset projection seed library with the modality type identifier; performs matrix expansion operation with the target compression ratio data and the target projection seed data to obtain a candidate measurement matrix; calculates the mutual coherence coefficient between the candidate measurement matrix and a preset sparse basis dictionary; and determines the candidate measurement matrix as the target measurement matrix when the mutual coherence coefficient is less than a preset orthogonality threshold.

[0136] Among them, the modality type identifier represents the classification key value used to indicate whether the physical source of the input data belongs to visual spectrum, acoustic vibration or laser ranging; the target projection seed data represents the extracted benchmark micro matrix used as the starting point for expansion and derivation; the candidate measurement matrix represents the temporary matrix carrier whose volume has met the standard but has not yet undergone strict isometry verification; the mutual coherence coefficient represents the scalar index that quantitatively evaluates the degree of orthogonality and independence between the random measurement system and the inherent sparse dictionary of the signal; and the preset orthogonality threshold represents the upper limit boundary judgment value for determining whether the temporary matrix will cause coherent ambiguity in subsequent reconstruction.

[0137] Specifically, the data processing system avoids executing high-overhead, purely random generation logic when constructing a measurement matrix. The system extracts the modality type identifier from the current high-dimensional auxiliary sensing data. Based on this identifier, the system accesses a low-memory, pre-defined projection seed library to retrieve the target projection seed data that best matches the current physical features. The system uses the calculated target compression ratio data to determine the required row and column expansion ratio. The system performs matrix expansion operations on the target projection seed data to unfold a candidate measurement matrix of full size in memory. To prevent matrix defects, the system retrieves a pre-defined sparse base dictionary for this modality. The system performs numerous inner product calculations to obtain the maximum cross-coherence coefficient between the candidate measurement matrix row vectors and the dictionary column vectors. The system compares this coefficient with a pre-defined orthogonality threshold; if the coefficient is below this threshold, it indicates that the matrix possesses good decryption and decoupling characteristics, and at this point, it is solidified as the target measurement matrix.

[0138] For scenarios involving sensor hardware expansion or replacement, the system has an online adaptive projection seed generation mechanism:

[0139] (1) When a new type of sensor is detected, the system first determines its data dimensions. The sampling rate and modality type (image / time series / point cloud) are used to retrieve the nearest seed from the preset seed library as the initial template;

[0140] (2) During the ground self-test phase before flight, the airborne terminal collects sample data of the sensor in a static environment, performs rapid sparsity assessment and RIP verification (usually takes 10-30 seconds), and automatically optimizes the seed parameters until the mutual coherence coefficient meets the threshold.

[0141] (3) The generated new seed and its corresponding sparse base dictionary parameters are synchronized to the ground station through an encrypted channel and persistently stored for use in subsequent flight cycles;

[0142] (4) For different flight environments (plateau, ocean, city), the system maintains an environment label index table. During the mission planning stage, the corresponding sparse base dictionary parameter set is preloaded according to the environment label preset by the route, without the need for repeated calibration.

[0143] This mechanism shortens the adaptation cycle of sensor extensions from factory calibration (several days) to on-site self-learning (tens of seconds), which can improve the flexibility of system deployment.

[0144] Specifically, the projection seed library and matrix expansion operations include:

[0145] The projection seed library is a compact data structure pre-generated and stored in the onboard non-volatile memory during the system's factory calibration phase. It stores one projection seed data entry for each modality type. Each entry contains two items: first, a deterministic pseudo-random number generator to initialize the seed value; and second, a seed value initialized by the generator. The seed value is a 64-bit unsigned integer; the second is an enumeration value that identifies the type of random distribution recommended for this modality. For example, labeling image and video modalities as Gaussian distribution types. Vibration time-series modes are labeled as Bernoulli distribution types, and LiDAR point cloud modes are labeled as sparse binary distribution types.

[0146] The execution process of matrix expansion operation is as follows:

[0147] First, based on the target compression ratio data and the original signal dimension The number of rows required to calculate the measurement matrix:

[0148] ;in This indicates the floor function.

[0149] The measurement matrix parameters need to be dynamically adjusted based on the estimated sparsity. The minimum number of measurements required to satisfy the Restricted Isometry Property (RIP) condition is:

[0150] ;

[0151] in denoted as the sparsity of the l-th mode, and c is the scaling factor associated with the RIP constant.

[0152] Using seed value Initialize the internal state of the pseudo-random number generator (commonly including Mersenne Twister or linear congruential generators) according to the distribution type corresponding to that mode. Generate a dimension element by element matrix .

[0153] For Gaussian distributions, generate random matrices that follow a standard normal distribution:

[0154] ;

[0155] For Bernoulli distribution types, the generated values ​​take positive and negative values ​​with equal probability. Random matrix:

[0156] =0.5;

[0157] After generation, perform L2 norm normalization on each row of the matrix to make the Euclidean norm of each row vector equal to 1.

[0158] ;

[0159] Normalized matrix This is the candidate measurement matrix. Combining the joint measurement scheme, each of the L types of signals is subjected to its own measurement matrix. The measurement vector is obtained:

[0160] ;

[0161] Combining airborne linear projection, the operation is as follows:

[0162] ;

[0163] Due to the deterministic nature of pseudo-random number generators, the same seed value... Same dimension parameters The exact same matrix can be reproduced on any platform, so the ground station reconstruction end only needs to receive the seed value and the current compression ratio parameter to regenerate the same measurement matrix locally for reconstruction calculation, without having to transmit the complete matrix content over a wireless link.

[0164] The precise calculation method for the mutual coherence coefficient is as follows:

[0165] Let the candidate measurement matrix be Furthermore, each row has been normalized, and the default sparse base dictionary is... in The number of atoms in the dictionary (for a complete base) For supercomplete dictionaries Calculate matrix multiplication:

[0166] ;

[0167] Mutual coherence coefficient is defined as The maximum absolute value of all elements in the set:

[0168] ;

[0169] in This represents the vector inner product. The smaller the maximum absolute value, the better the orthogonality between the row space of the measurement matrix and the column space of the sparse basis dictionary. According to compressed sensing theory, the smaller the mutual coherence coefficient, the higher the probability that the measurement matrix satisfies the constrained isometric property (RIP), thus ensuring higher reconstruction accuracy.

[0170] Due to real-time sparsity estimation There may be a statistical error of ±20% when the system determines the number of rows in the measurement matrix. When a safety margin factor β (typical value 1.3-1.5) is introduced, the corrected formula for calculating the measurement number is as follows: ;in To estimate sparsity, The margin factor is rounded up. While ensuring that the RIP conditions are met with a high probability, this margin factor only increases the amount of transmitted data by about 30%, which is far lower than the redundancy of independent compression schemes.

[0171] Furthermore, when the ground station reconstructor detects abnormal reconstruction residuals (residuals exceeding twice the threshold for three consecutive frames), it adds an "undersampling warning" flag to the CSI feedback message. Upon responding to this warning, the airborne unit increases the compression rate of the corresponding mode by 10%-15% in the next cycle (i.e., increases the number of measurements), forming a negative feedback loop for reconstruction quality. Experimental data shows that this adaptive error correction mechanism can converge the reconstruction error to a normal level within five cycles.

[0172] The system will calculate the mutual coherence coefficient. Orthogonal threshold For comparison, this threshold is typically set as follows:

[0173] ;

[0174] Where the constant Select based on actual compression ratio and sparsity requirements. If the conditions are met:

[0175] ;

[0176] Then confirm the candidate matrix The orthogonality requirement is met and the result is solidified into a target measurement matrix for use. If the mutual coherence coefficient is greater than or equal to the threshold, the system applies an incremental perturbation to the seed value:

[0177] ;

[0178] Regenerate the candidate matrix and perform the validation again, repeating this process until the validation passes or the maximum number of retries is reached. (Usually set to 10 times). If the maximum number of retries is reached and the system still fails, the system will revert to using the old measurement matrix that has been verified in the previous cycle to ensure continuous and uninterrupted system operation.

[0179] For example: Suppose the original signal dimension of a certain video mode is... =16384 (corresponding to a 128×128 pixel block), target compression ratio Then measure the number of rows in the matrix. =6553. Using the seed value. =12345 Initialize the Gaussian random number generator, generate a standard Gaussian matrix of 6553×16384, and then perform row normalization to obtain the candidate matrix Φ. If the sparse basis used for this mode is a two-dimensional DCT dictionary. (Dimension 16384×16384), calculate the matrix product and then iterate through the matrix to find the maximum absolute value. Assume... =0.021. Set threshold. The seed value was 0.0124. Since 0.021 > 0.0124, the verification failed. The system updated the seed to [a different value]. =12346 is regenerated and validated again until a certain iteration is obtained. <0.0124 Passed the check.

[0180] S303. Perform sparse basis projection transformation on the high-dimensional auxiliary sensing data to separate the common sparse component reflecting the common physical response across modes and the specific sparse component corresponding to each sensor mode.

[0181] Among them, sparse basis projection transformation represents the operation of mapping the data of the original sampling domain to a transformation domain with very few non-zero elements through an orthogonal or overcomplete dictionary matrix; common sparse components represent the underlying structural support vectors that exhibit strong consistency or synchronous mutation in the sparse domain of heterogeneous sensor data; proprietary sparse components represent the independent noise or unique detail feature vectors recorded by each sensor due to its physical measurement mechanism.

[0182] Specifically, the data processing system selects a matching transformation benchmark based on the intrinsic properties of each modality's data. The system performs three-dimensional discrete cosine transform on the high frame rate image sequence and wavelet decomposition on the fuselage multi-point stress time series, transferring the high-dimensional auxiliary sensing data from the dense original observation domain to a sparse representation domain. In the sparse domain, the system performs cross-scanning between modalities to identify high-amplitude activation regions that appear synchronously in the sparse matrices of different modalities within the same time slice. The system extracts and groups these common excitation features, corresponding to fuselage flutter and simultaneous image blurring caused by strong airflow encountered by the aircraft, into common sparse components. These common components are then subtracted from the original sparse representations of each modality, and the remaining residual information is defined as the specific sparse components of each modality. This step removes redundant representations at the physical level.

[0183] In some embodiments, component separation can be achieved in several ways: Optionally, the data processing system concatenates and fuses sparse vectors from multiple modalities; extracts the first principal component vector sequence from the concatenated matrix through principal component analysis; maps this principal component vector back to each modal space as a common sparse component; and performs matrix subtraction to separate the residuals of the proprietary sparse components. Optionally, the data processing system establishes a joint dictionary learning iterative model; adds a joint regularization penalty term to the iterative objective function to measure cross-modal structural similarity; solves for the common sparse support set using the alternating direction multiplier method; and performs orthogonal matching pursuit projection on the original signal based on this support set to separate the common and proprietary components. It is understood that other methods can also be used to achieve the decoupling and separation of cross-modal joint features, which are not limited here.

[0184] In some embodiments, the data processing system maps high-dimensional auxiliary sensing data to a preset sparse base dictionary to obtain an initial sparse coefficient matrix; calculates the cross-correlation matrix between the initial sparse coefficient matrices of different modalities; extracts overlapping feature vectors in the initial sparse coefficient matrix by combining the cross-correlation matrix, and determines the overlapping feature vectors as common sparse components; removes the common sparse components in the initial sparse coefficient matrix to obtain proprietary sparse components.

[0185] Among them, the overlapping feature vector represents the common component expression that exhibits strong matching and co-frequency oscillation characteristics at multiple modality dictionary activation positions.

[0186] Specifically, the data processing system implements deep feature analysis to overcome the performance limitations of independent compressed sensing. The system drives high-dimensional auxiliary sensing data to undergo orthogonal or redundant projection decomposition through a pre-set sparse base dictionary, transcoding the difficult-to-analyze raw observations into initial sparse coefficient matrices. The system captures multiple initial sparse coefficient matrices generated by different physical sensors. The system performs cross-traversal operations on the activation index regions of these matrices to derive their cross-correlation matrices. Based on these cross-correlation matrices, the system locates coordinate point sequences that exhibit high-frequency overlap and positive correlation mapping patterns in each modality. The system extracts and merges the data entities corresponding to these point sequences into overlapping feature vectors reflecting unified aerodynamic interference or body resonance, qualitatively identifying them as common sparse components. The system then controls the initial matrices of each modality to perform subtraction operations or orthogonal spatial projection rejection operations to eliminate these common sparse components, confirming the purified retained results as proprietary sparse components, achieving deep deduplication and isolation of data dimensions.

[0187] After obtaining the cross-correlation matrix, the system performs the determination and extraction of overlapping feature vectors, specifically including:

[0188] Let the original signal vectors of the airborne L-type sensor at time t be respectively The dimensions are respectively Each signal exists in its respective sparse basis. The following is represented as ,in It is a sparse coefficient vector.

[0189] During flight, large cargo drones exhibit physical causal relationships between signals from various modal sensors: airflow disturbances are simultaneously reflected in IMU acceleration, fuselage vibration sensors, and wing strain gauges; load displacements are simultaneously reflected in EO / IR image changes and load stress sensors. This application proposes establishing a joint sparse model of "common sparse support + modal-specific sparse components," namely:

[0190] ;

[0191] in It is a common sparse component shared by all modes (reflecting a common physical excitation source). This is a proprietary sparse component for the type l sensor.

[0192] Let the initial sparse coefficient vectors obtained after the sparse basis projection transformation of the L-type sensor be respectively... , where the sparse coefficient vector of the l-th type of sensor The dimension is the sparse domain dimension D. After time-domain alignment, the sparse coefficients of each mode correspond to the same physical time in the time dimension.

[0193] For each sparse domain index position Extract the absolute values ​​of the coefficients for each mode at that position and construct a correlation measure. Calculate the normalized correlation coefficient between different modes l and k at position j:

[0194] ;

[0195] This coefficient reflects the energy correlation of different modes at this sparse domain location.

[0196] Perform common sparse support determination: If position j satisfies the following condition, then mark it as a common support point: ;

[0197] in The preset threshold for determining common support is typically set between 0.7 and 0.9. This condition requires that the correlation coefficients of all pairwise combinations of modes exceed the threshold, ensuring that the location indeed corresponds to the common response of the same physical excitation source.

[0198] After traversing all D sparse domain positions to complete the determination, the common support index set is obtained:

[0199] ;

[0200] For each public support location Extract the sparse coefficients of each mode at that position and calculate a weighted average to obtain the common sparse component at that position:

[0201] ;

[0202] Among them, the weighting coefficient The signal-to-noise ratio (SNR) of each mode at that location is determined by:

[0203] ;

[0204] ;

[0205] in This represents the noise variance estimate for the l-th type of sensor. Modes with higher signal-to-noise ratios contribute more to the common components.

[0206] For the proprietary sparse components of each mode, perform a stripping operation:

[0207] ;

[0208] For example: Suppose that the vibration mode and the video mode each obtain their lengths after sparse basis transformation. A sparse coefficient vector with a value of 1024. At index position... At points =37,128,512, the correlation coefficients for the two modes are respectively... =0.92, =0.85, =0.78, all exceeding the set judgment threshold. =0.75, then these three positions are marked as the common support set. ={37,128,152}. Assume the position is... At point 37, the signal-to-noise ratio of the vibration mode is 15 dB, i.e., the linearity value. The signal-to-noise ratio of the video modality is 12dB, which is a linear value. The weighting coefficients are:

[0209] 0.666, 0.334;

[0210] The common sparse component at this location is:

[0211] =0.666* +0.334* ;

[0212] The coefficients for the remaining non-common support locations are retained as the proprietary sparse components of each mode.

[0213] In addition, to address situations where there is transitive correlation between multimodalities, the system employs a two-stage strategy when performing common support determination:

[0214] In the first phase, strict global consistency conditions are used. Identify strong public support sets ;

[0215] In the second stage, for those who did not pass the first stage but whose pairwise correlation coefficients exceed the reduction threshold... At a position (typically set to 0.6-0.7), the system constructs a modal association graph and performs connectivity analysis: if this position forms a connected subgraph in the association graph and covers more than 70% of the modal nodes, it is included in the weak common support set. The final common support set is the union of the two: This grading and determination mechanism ensures accuracy while also taking into account the transmission effect of physical excitation between different modes.

[0216] In some embodiments, the data processing system acquires the time synchronization stamp data of the initial sparse coefficient matrices of different modalities; performs a time-domain alignment operation on the initial sparse coefficient matrices of different modalities in combination with the time synchronization stamp data to obtain an aligned sparse matrix; extracts the low-frequency envelope feature vectors in the aligned sparse matrix; and calculates the inner product values ​​of the low-frequency envelope feature vectors of different modalities to obtain the cross-correlation matrix.

[0217] Among them, the aligned sparse matrix represents the adjusted matrix array after phase calibration to ensure that data in the same column are mapped to the same real physical occurrence time; the low-frequency envelope eigenvector represents the smooth transition form expression reflecting the main trend of the physical event after filtering out high-frequency spike jitter; the inner product value represents the sum and product of the projections of the two eigenvectors onto each other in the spatial dimension, which is used to characterize the similarity between the vector angle and the trend of the signal it represents.

[0218] Specifically, the data processing system performs refined front-end preprocessing for cross-correlation calculations to ensure the rigor of physical correlation assessment. The system parses the protocol tails of the initial sparse coefficient matrices for different modes and extracts time synchronization stamp data. The system compares the phase lead or lag differences of each time synchronization stamp. Using a reference mode as an anchor point, the system performs resampling and interpolation correction, and performs time-domain alignment operations on different modes to output an aligned sparse matrix with strictly aligned columns to the time grid. The system applies window moving average filtering or low-pass convolution operations to the aligned sparse matrix to eliminate high-frequency measurement noise glitches unique to each hardware component, retaining low-frequency envelope feature vectors that reflect the core morphology of airflow shear or rigid displacement. The system calls the vector multiplication calculation unit to process the low-frequency envelope feature vectors of different modes to obtain inner product values. The system normalizes the paired inner product values ​​and fills them into the coefficient grid to generate the final cross-correlation matrix for determining correlation relationships.

[0219] In some embodiments, alignment and envelope extraction calculations can be implemented in various ways: Optionally, the data processing system selects the mode with the lowest sampling rate as the clock axis anchor point; performs downsampling on the initial sparse coefficient matrix of the high sampling rate mode based on the time synchronization stamp data to generate a time-domain alignment matrix; applies a discrete Fourier transform to the generated matrix to transfer it to the frequency domain, truncates the high-frequency components, and then performs an inverse transform to generate low-frequency envelope feature vectors; uses the vector dot product summation formula to calculate the spatial inner product value of pairwise envelope vectors and outputs the correlation. Optionally, the data processing system establishes a spline interpolation function surface fitting of the scattered initial sparse coefficient matrix; resamples at the same reference time grid point to output a smooth phase-aligned sparse matrix; directly uses morphological opening and closing operations to extract the connection between troughs and peaks to construct low-frequency envelope feature vectors; calculates the cosine similarity of the feature vectors and scales them to represent the inner product value to construct a cross-correlation matrix. It is understood that other methods can also be used to implement the low-frequency correlation measure of multi-source time series features under phase alignment, which is not limited here.

[0220] After the separation operation, the system has deconstructed the high-dimensional auxiliary sensing data of each modality into two parts: common sparse components and proprietary sparse components. However, these sparse components are still in the sparse domain space, and their dimensions are comparable to the original signal, thus not achieving substantial data volume compression. To achieve the goal of compressed transmission, the system needs to use the target measurement matrix determined in step S302 to perform a dimension reduction projection operation on the separated sparse components, mapping the high-dimensional sparse vectors into low-dimensional measurement vectors, thereby significantly reducing the amount of data that needs to be transmitted through the wireless channel, thus forming the encoding stage in the compressed sensing framework.

[0221] S304. Based on the target measurement matrix, perform linear projection operations on the common sparse components and the proprietary sparse components respectively to obtain a dimension-reduced feature measurement vector containing cross-modal feature information.

[0222] Among them, the linear projection operation represents the matrix-vector multiplication operation that multiplies a high-dimensional column vector by a dimension-reducing matrix to generate a low-dimensional observation result; the dimension-reduced feature measurement vector containing cross-modal feature information represents a non-adaptive linear hash value sequence that is significantly reduced in size after compressed sensing encoding but retains the ability to recover the original signal at the reconstruction end.

[0223] Specifically, the data processing system extracts the separated common sparse components and the proprietary sparse components of each hardware mode. The system divides the target measurement matrix determined in the previous steps into a global projection operator applicable to the common components and a local projection operator applicable to each proprietary component. The global projection operator is used to project the common sparse components, and its dimension is determined based on the joint sparsity of all modes. Since the common components are shared among the modes, they only need to be transmitted once and can be reused during ground station reconstruction, thus significantly reducing the total number of measurements. The local projection operators correspond to the proprietary sparse components of each mode, and their dimensions are determined based on the remaining sparsity of each mode after deducting the common support. Compared with traditional schemes, this separate projection strategy can reduce the total measurement vector length by 30%-50% while maintaining the same reconstruction accuracy. For example, when the number of measurements required for independent compression of the three modes are M1=3000, M2=2500, and M3=2000 (total 7500), and this scheme identifies a common support ratio of 40%, then the number of common measurements... =1200, with the proprietary measurement numbers for each modality being M1'=1800, M2'=1500, and M3'=1200, totaling 5700, saving 24% of transmission bandwidth.

[0224] The system controls a hardware multiply-accumulate array to perform dot-multiplication and accumulation operations on common sparse components and the global projection operator. Simultaneously, it controls other independent computation threads to perform dot-multiplication and accumulation operations on each specific sparse component and its corresponding local projection operator. The system collects the dense low-dimensional numerical streams output by these independent threads and performs concatenation and encapsulation to generate a dimensionality-reduced feature measurement vector representing the overall state of the current monitoring period. At this stage, no complex feature recognition of the data content is required; high-ratio compression can be achieved solely through pure linear dimensionality reduction.

[0225] The ground station receives measurement vectors for each mode. After obtaining the corresponding measurement matrix seed and compression ratio parameters, the measurement matrix for each mode is first regenerated locally. Then, using cross-modal joint sparse priors, the following joint reconstruction optimization problem is constructed:

[0226] ;

[0227] The first term is the data fidelity term, which measures the fitting error between the reconstructed signal projected onto the measurement matrix and the actual measured value; the second term is the L1 norm regularization term for the common sparse components, which ensures the common part is sparsity; the third term is the sum of the L1 norm regularization terms for each mode-specific sparse component, which ensures that the mode-specific part also remains sparse; the fourth term... This is a cross-modal structure regularization term used to encode sparse support correlations between modes.

[0228] The multi-sensor joint sparse recovery theory guarantees that, compared with independent reconstruction of each mode, the joint reconstruction scheme, while maintaining the same reconstruction accuracy, reduces the total number of measurements. This can significantly reduce the amount of data transmitted. This is because of the common sparse support. It can be shared across different modes without repeated transmission, thereby improving the overall compression efficiency.

[0229] This optimization problem can be solved using the Alternating Direction Multiplier Method (ADMM) or the Iterative Threshold Shrinkage Algorithm (ISTA). The specific iterative process includes:

[0230] (1) Fixed common sparse components Update the proprietary sparse components of each modality separately. ;

[0231] (2) Fix the proprietary sparse components of each mode, based on the extracted common support set Update common sparse components ;

[0232] (3) Based on the joint sparse model, synthesize the complete sparse coefficient vector of each mode. ;

[0233] (4) Reconstruct the original signals of each mode by sparse basis inverse transform. .

[0234] The iteration termination condition is that the change in the objective function value between two consecutive iterations is less than a preset threshold or the maximum number of iterations is reached.

[0235] S305. Map safety-critical time-series data to low-latency time slots in the transmission queue, fill the remaining capacity time slots of the transmission queue with the dimensionality-reduced feature measurement vector, and generate a multimodal multiplexed data stream to be transmitted.

[0236] Among them, the sending queue represents the first-in-first-out message queuing buffer between the encoding processing layer and the hardware physical layer; the low-latency slot represents the data carrier segment with the highest scheduling priority and the least queuing waiting time in time division multiplexing or frame structure design; the remaining capacity slot represents the available transmission channel capacity space that is idle in the current sending cycle after meeting the core scheduling requirements; and the multimodal multiplexed data stream represents a composite structure communication sequence that mixes original precise control timing data and compressed feature data.

[0237] Specifically, the data processing system executes data framing logic at the media access control layer. The system reserves a fixed-length secure transmission bandwidth block (i.e., a low-latency time slot) at the beginning of the current time segment and loads the security-critical timing data from the front-end directly into this block without inspection. Subsequently, the system calculates the difference between the total available length of the current transmission queue and the length already occupied by security-critical data. Based on this difference, the system sorts the dimensionality-reduced feature measurement vectors (after linear projection to reduce volume) according to the importance of their sub-modalities and sequentially fills them into the remaining capacity time slots of the transmission queue until they are full or there is no more data to fill. The system adds checksums, frame synchronization codes, and priority scheduling headers to the integrated queue data, ultimately solidifying it into a well-structured multimodal multiplexed data stream.

[0238] After the multimodal multiplexed data stream is generated, the system does not perform physical transmission. Because the wireless communication channel uses a carrier sense multiple access (CSMI) mechanism, multiple devices share the same spectrum resource. The system must wait for the underlying RF baseband chip to complete channel idle detection and grant transmission permission before pushing the data to the physical layer. This handshake mechanism effectively avoids data collisions and ensures the reliability of the communication link.

[0239] In some embodiments, before executing step S306, the system first calculates the difference between the current data volume of the transmission queue and the updated instantaneous channel capacity assessment value. If the difference is negative (i.e., the data volume exceeds the channel capacity), the system executes a congestion control procedure: removing some dimensionality-reduced feature measurement vectors from low to high priority until the queue data volume does not exceed the channel capacity, to ensure that high-priority data can still be reliably transmitted under extremely poor channel conditions.

[0240] S306. In response to the transmission permission identifier of the external communication link, write the multimodal multiplexed data stream to the network communication port to perform the data transmission of multimodal sensor data.

[0241] Among them, the transmission license identifier represents the ready status level signal fed back to the main control core by the underlying RF baseband chip or antenna controller based on the channel idle listening protocol; the network communication port represents the physical medium independent interface register for interconnection between the airborne computing unit and the microwave or millimeter-wave RF transmitting front end.

[0242] Specifically, after framing, the data processing system enters a polling and listening state to wait for physical layer signals. The system detects the level status of the general-purpose input / output pins connected to the RF baseband board. When a high-level pulse indicating a transmit permission is detected, the system locks the valid holding window for that pulse. The system calls the direct memory access controller, bypassing the central processing unit, to quickly move the multimodal multiplexed data stream, which has been constructed in the system's main memory, to the hardware transmit buffer of the network communication port in block mode. After receiving the complete data block, the network communication port follows the underlying frequency domain modulation and digital-to-analog conversion circuit logic to convert the multiplexed data stream frequency and radiate it into atmospheric space, completing the closed-loop data downlink physical execution steps from the airborne end to the ground control node.

[0243] In some embodiments, the data processing system receives link quality feedback data from the ground control node; obtains channel fading parameters from the link quality feedback data; and updates the instantaneous channel capacity assessment value by combining the channel fading parameters.

[0244] Among them, the ground control node represents the integrated receiving control entity configured in the remote command center, which includes a high-power antenna array and an error correction and decoding server; the link quality feedback data represents the communication channel health report file formed by the remote node based on the analysis of the pilot signal of the downlink multimode multiplexed data stream frame header; the channel fading parameter represents the quantitative indicator factor reflecting the degree of amplitude attenuation of radio waves under multipath effects or atmospheric obstruction; and the instantaneous channel capacity assessment value represents the actual limit throughput that the system can predict within the current microsecond-level transmission window to traverse the severe fading channel without loss.

[0245] Specifically, while continuously transmitting data, the data processing system maintains and monitors a low-baud-rate uplink receiving channel dedicated to control information. The system captures and verifies link quality feedback data demodulated from the ground control node via the uplink channel. The system parses specific segments of this feedback data to extract channel fading parameters describing the gradient of carrier signal-to-noise ratio decline and the characteristics of a sharp increase in bit error rate. The system reads the channel capacity limit baseline of the previous cycle from its main memory. Using the extracted channel fading parameters and the built-in Shannon capacity approximation formula, the system performs compensation degradation or recovery gain calculations on the baseline model, ultimately overwriting the old data to update the instantaneous channel capacity assessment value. This process ensures that the queuing pressure of airborne data can immediately adapt to changes in the electromagnetic characteristics of the real physical space.

[0246] In some embodiments, capacity assessment values ​​can be updated in multiple ways: Optionally, the data processing system monitors control messages from the underlying transceiver to obtain link quality feedback data; locates the message parameter segment and reads the modulation and coding strategy degradation indication as a direct representation of channel fading parameters; calls a capacity conversion lookup table to compare the new absolute throughput value corresponding to the degradation indication; and directly replaces the old instantaneous channel capacity assessment value with the new value found to complete the system state machine refresh. Optionally, the data processing system reads the received signal-to-noise ratio absolute floating-point value from the ground feedback message as the main body of the link quality feedback data; calculates the rate of change of the first derivative of the signal-to-noise ratio to determine the transient trend of the channel fading parameters; performs a weighted fusion operation on the original instantaneous channel capacity assessment value and the attenuation Kalman filter iterative prediction value constructed using the rate of change; and outputs the fused and smoothed converged value as the instantaneous channel capacity assessment value for the new period. It is understood that other methods can also be used to implement channel state prediction correction under uplink and downlink asymmetric states, which are not limited here.

[0247] Furthermore, for communication links employing Frequency Division Duplex (FDD), the channel characteristics are non-reciprocal due to the different uplink and downlink operating frequencies. Therefore, when estimating the downlink channel capacity using uplink CSI feedback, the system introduces a frequency mapping correction coefficient. This coefficient is calibrated offline based on uplink / downlink frequency spacing, antenna configuration, and historical statistical data, with a typical value range of 0.7-0.9. The updated formula is corrected as follows: = ·f( ), fading parameters), where f() is the capacity update function. For time division duplex (TDD) links, since uplink and downlink share the same frequency and the time interval is short, the channel reciprocity is good, and it is directly set to =1.0. During the initialization phase, the system automatically identifies the duplex mode and loads the corresponding correction coefficient by reading the communication module configuration register, without requiring manual intervention.

[0248] In some embodiments, the data processing system calculates the difference between the updated instantaneous channel capacity assessment value and the total amount of data to be transmitted in the current transmission queue to obtain a capacity gap value; when the capacity gap value is less than zero, it extracts the secondary priority label of the dimensionality reduction feature measurement vector in the transmission queue; combining the secondary priority label and the capacity gap value, it removes the dimensionality reduction feature measurement vector corresponding to the amount of data from the transmission queue to obtain a truncated multiplexed data stream; and it rewrites the truncated multiplexed data stream into the network communication port.

[0249] Among them, the capacity gap value represents a certain degree of surplus capacity or overload deficit that can characterize the actual carrying capacity of the channel compared with the data queuing pressure.

[0250] Specifically, the data processing system assesses congestion risk within each underlying physical clock tick. The system extracts the newly updated instantaneous channel capacity assessment value and subtracts it from the total amount of data waiting to be transmitted in the current transmission queue. The system then determines the calculated capacity gap. A value greater than or equal to zero indicates sufficient bandwidth, and the system allows data transmission as planned; a value less than zero indicates an overload congestion event. The system performs a reverse scan of the sorted, reduced-dimensional feature measurement vectors in the transmission queue, parsing the secondary priority tag attached to the header of each vector. Following the secondary priority from low to high in reverse order, the system progressively selects the corresponding low-priority vectors based on the overload capacity gap value and performs dequeue destruction until the gap is filled. The system recalculates the total checksum for the remaining data frames after the weight reduction process and encapsulates them into a truncated multiplexed data stream matching the current bandwidth. The system clears the hardware suspend state of the network communication port and rewrites the truncated multiplexed data stream to restore secure downlink operation.

[0251] Furthermore, when the system discards some dimensionality-reduced measurement vectors due to capacity gaps, the ground station reconstructor employs a cascaded reconstruction strategy: it prioritizes reconstructing the corresponding signal using the received complete modal measurement vectors, and propagates the reconstruction results to the missing modes through a cross-modal fusion layer, providing prior constraints for the missing modes. Specifically, if the measurement vector of mode l is discarded, the reconstructor first reconstructs the common sparse components. (This component is highly robust because it is shared by all modes), and then the statistical model of the modality-specific component of historical frames is combined with... Conditional prediction and reconstruction are performed. Simulation results show that this mechanism can still maintain more than 80% reconstruction fidelity of the target mode even when the measurement vector of a single mode is completely missing. In addition, the ground station carries the missing mode identifier in the uplink feedback, and the airborne terminal prioritizes allocating a higher compression rate to that mode in the next cycle to compensate for information loss.

[0252] Meanwhile, to meet the differentiated latency requirements of data with different priorities, the ground station reconstructor can adopt a heterogeneous computing resource allocation strategy, namely:

[0253] (1) P0 level safety-critical timing data bypasses the GPU reconstruction process, is directly decoded on the CPU and immediately sent to the flight management system, with an end-to-end latency of 3-8ms.

[0254] (2) The reconstruction task of P1 level auxiliary data is submitted to the high priority queue of the GPU. Lightweight ISTA-Net (12-layer expansion) is adopted and mixed precision inference (FP16) is enabled. The reconstruction time of a single frame 256×256 image block is compressed to 25-40ms. Combined with parallel processing of multiple frames, the throughput latency can be controlled within 100ms.

[0255] (3) P2-level low-priority data is processed asynchronously in the GPU background queue using standard ISTA-Net++ (24-layer expansion), allowing cumulative batch processing to improve throughput, with a latency tolerance of up to 1 second.

[0256] The GPU scheduler employs preemptive priority management, allowing P1 tasks to interrupt the execution of P2 tasks, ensuring the real-time performance of critical data links. Actual flight tests demonstrate that this hierarchical mechanism achieves a latency compliance rate exceeding 95% for each level in multimodal concurrent scenarios.

[0257] In this embodiment, the data interaction barrier of traditional UAV systems, which uses a single fixed-cycle queuing for transmission of various types of sensors, is broken. The cross-modal sparse structure derived from common physical excitation effectively eliminates redundant transmission between heterogeneous signals. At the same time, a high-priority protection channel with physical isolation level is established at the resource scheduling level. This effectively solves the problems of bandwidth contention and flight safety hazards caused by the difficulty of static measurement matrices to adapt to time-varying fading channels and the low compression efficiency of heterogeneous sensors when compressed independently in the prior art. Thus, it realizes the efficient and reliable transmission of massive heterogeneous multimodal data from airborne to ground stations without relying on ultra-wideband hardware links. This ensures that the flight status of large cargo platforms under complex working conditions can be reconstructed and monitored by the ground control center in real time and with high fidelity.

[0258] The data processing system in this application embodiment is described below from a hardware processing perspective. Please refer to [link / reference]. Figure 4 This is a schematic diagram of the physical device structure of a data processing system in an embodiment of this application.

[0259] It should be noted that, Figure 4 The structure of the data processing system shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0260] like Figure 4As shown, the data processing system includes a CPU 401, which can perform various appropriate actions and processes according to a program stored in ROM 402 or a program loaded into RAM 403 from storage section 408, such as executing the methods described in the above embodiments. RAM 403 also stores various programs and data required for system operation. The CPU 401, ROM 402, and RAM 403 are interconnected via bus 404. I / O interface 405 is also connected to bus 404.

[0261] The following components are connected to I / O interface 405: input section 406 including audio input devices, push-button switches, etc.; output section 407 including a liquid crystal display (LCD) and audio output devices, indicator lights, etc.; storage section 408 including a hard disk, etc.; and communication section 409 including a network interface card such as a LAN (Local Area Network) card, modem, etc. Communication section 409 performs communication processing via a network such as the Internet. Drive 410 is also connected to I / O interface 405 as needed. Removable media 411, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 410 as needed so that computer programs read from them can be installed into storage section 408 as needed.

[0262] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program including a computer program for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 409, and / or installed from removable medium 411. When the computer program is executed by CPU 401, it performs the various functions defined in this application.

[0263] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. Each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those shown in the drawings.

[0264] Specifically, the data processing system of this embodiment includes a processor and a memory. The memory stores a computer program. When the computer program is executed by the processor, it implements the large UAV multimodal data transmission method based on compressed sensing provided in the above embodiment.

[0265] In another aspect, this application also provides a computer-readable storage medium, which may be included in the data processing system described in the above embodiments; or it may exist independently and not assembled into the data processing system. The storage medium carries one or more computer programs, which, when executed by a processor of the data processing system, cause the data processing system to implement the compressed sensing-based multimodal data transmission method for large unmanned aerial vehicles provided in the above embodiments.

[0266] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0267] As used in the above embodiments, depending on the context, the term "when..." can be interpreted as meaning "if...", "after...", "in response to determining...", or "in response to detecting...". Similarly, depending on the context, the phrase "when determining..." or "if (the stated condition or event) is interpreted as meaning "if determining...", "in response to determining...", "when (the stated condition or event) is detected", or "in response to detecting (the stated condition or event)".

Claims

1. A large unmanned aerial vehicle multi-modal data transmission method based on compressed sensing, characterized by, Applied to a data processing system, the method includes: Based on the preset data priority labels, the acquired multimodal sensor data is divided into uncompressible safety-critical time-series data and high-dimensional auxiliary sensing data to be compressed. Extract the instantaneous channel capacity assessment value from the external uplink signaling, and combine it with the importance weight parameters of the high-dimensional auxiliary sensing data to determine the target measurement matrix; A sparse basis projection transformation is performed on the high-dimensional auxiliary sensing data to separate the common sparse component reflecting the common physical response across modes and the specific sparse component corresponding to each sensor mode. Based on the target measurement matrix, linear projection operations are performed on the common sparse component and the proprietary sparse component respectively to obtain a dimension-reduced feature measurement vector containing cross-modal feature information. The safety-critical time-series data is mapped to the low-latency time slots of the transmission queue, and the dimensionality-reduced feature measurement vector is filled into the remaining capacity time slots of the transmission queue to generate a multimodal multiplexed data stream to be transmitted. In response to a transmission permission identifier from an external communication link, the multimodal multiplexed data stream is written to the network communication port to perform data transmission of the multimodal sensor data.

2. The method of claim 1, wherein, The step of extracting the instantaneous channel capacity assessment value from the external uplink signaling and combining it with the importance weight parameters of the high-dimensional auxiliary sensing data to determine the target measurement matrix specifically includes: Obtain the total available bandwidth data corresponding to the instantaneous channel capacity assessment value, and obtain the minimum compression ratio lower limit data for each of the high-dimensional auxiliary sensing data; By combining the total available bandwidth data, the importance weight parameters, and the minimum compression ratio lower limit data, a constraint optimization operation is performed to obtain the target compression ratio data for each of the high-dimensional auxiliary sensing data. The target measurement matrix is ​​generated by combining the target compression ratio data.

3. The method of claim 2, wherein, The step of generating the corresponding target measurement matrix by combining the target compression ratio data specifically includes: Obtain the modality type identifier of the high-dimensional auxiliary sensing data; The target projection seed data is matched in a preset projection seed library by combining the modality type identifier; A matrix expansion operation is performed by combining the target compression ratio data and the target projection seed data to obtain the candidate measurement matrix; Calculate the mutual coherence coefficient between the candidate measurement matrix and the preset sparse base dictionary; When the mutual coherence coefficient is less than a preset orthogonality threshold, the candidate measurement matrix is ​​determined as the target measurement matrix.

4. The method of claim 1, wherein, The step of performing a sparse basis projection transformation on the high-dimensional auxiliary sensing data to separate the common sparse component reflecting the common physical response across modes and the specific sparse component corresponding to each sensor mode specifically includes: The high-dimensional auxiliary sensing data is mapped to a preset sparse base dictionary to obtain an initial sparse coefficient matrix; Calculate the cross-correlation matrix between the initial sparse coefficient matrices of different modes; The overlapping feature vectors in the initial sparse coefficient matrix are extracted by combining the cross-correlation coefficient matrix, and the overlapping feature vectors are determined as the common sparse components; The common sparse components in the initial sparse coefficient matrix are removed to obtain the proprietary sparse components.

5. The method of claim 4, wherein, The step of calculating the cross-correlation matrix between the initial sparse coefficient matrices of different modes specifically includes: Obtain the time synchronization stamp data of the initial sparse coefficient matrix for different modalities; By combining the time synchronization stamp data, a time-domain alignment operation is performed on the initial sparse coefficient matrices of different modalities to obtain an aligned sparse matrix; Extract the low-frequency envelope feature vector from the aligned sparse matrix; The inner product values ​​of the low-frequency envelope eigenvectors of different modes are calculated to obtain the cross-correlation matrix.

6. The method of claim 1, wherein, After the step of writing the multimodal multiplexed data stream to the network communication port in response to a transmission permission identifier from an external communication link to perform the data transmission of the multimodal sensor data, the method further includes: Receive link quality feedback data sent by the ground control node; Obtain the channel fading parameters from the link quality feedback data; The instantaneous channel capacity assessment value is updated based on the channel fading parameters.

7. The method of claim 6, wherein, After the step of updating the instantaneous channel capacity assessment value by incorporating the channel fading parameters, the method further includes: The difference between the updated instantaneous channel capacity assessment value and the total amount of data to be transmitted in the current transmission queue is calculated to obtain the capacity gap value; When the capacity gap value is less than zero, extract the secondary priority label of the dimensionality reduction feature measurement vector in the sending queue; By combining the secondary priority label and the capacity gap value, the dimensionality reduction feature measurement vector corresponding to the data volume is removed from the sending queue to obtain the truncated multiplexed data stream; The truncated multiplexed data stream is rewritten into the network communication port.

8. A data processing system, characterized by The data processing system includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code including computer instructions, and the one or more processors call the computer instructions to cause the data processing system to perform the method as described in any one of claims 1-7.

9. A computer-readable storage medium comprising instructions, characterized in that, When the instructions are executed on the data processing system, the data processing system performs the method as described in any one of claims 1-7.

10. A computer program product, characterised in that, When the computer program product is run on a data processing system, the data processing system performs the method as described in any one of claims 1-7.

Images