Multi-source fusion avionics control system and method

Abstract

The invention relates to the technical field of avionics and flight control, in particular to a multi-source fusion avionics control system and method, a multi-source acquisition and compression module synchronously acquires inertial measurement, star sensor and radar data under a unified time reference, and generates observation data through random projection compression; the fusion twinning module obtains a fusion state and a hypergraph based on compressed observation data, generates pneumatic, electromagnetic and structural twinning environments in real time on a shared three-dimensional grid through physical constraint diffusion, and outputs a registration result by adopting entropy regularization optimal transmission; the decision execution module maps discrete control variables quantified by a flight control computer into an Isinus model, obtains control sequence seeds through quantum annealing, generates a decision instruction through pulse sinus strategy network optimization, issues the decision instruction to a control execution unit to generate control execution data, and updates a hypergraph in combination with execution feedback for next cycle use. According to the method, the attitude estimation precision and the control robustness under the conditions of high maneuverability and strong interference are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of avionics and flight control, and particularly to a multi-source fusion avionics control system and method. Background Art

[0002] With the rapid growth of the types and data rates of airborne sensors, the avionics system is shifting from single inertial navigation to multi-source fusion mode. Inertial measurement, star sensors, and airborne radars can complement each other under different external conditions, improving the reliability of attitude and position, and providing real-time environmental information for load management. Existing multi-source fusion solutions usually adopt a serial two-step method: first, perform integer-order Kalman filtering on each source measurement separately, then send the filtered results into an aerodynamic or structural model based on an offline database to generate environmental estimates, and finally use local gradient descent or gain scheduling to calculate control instructions. Such solutions have three deficiencies: ① The measurement association is limited to binary covariance and cannot capture high-order complementary information of more than three sources, resulting in attitude drift during high maneuvers; ② The offline database cannot reflect the airframe deformation and external field disturbances in real time, and the environmental field estimation lags; ③ The control optimization stays at local search, has low efficiency in dealing with hard constraints, and lacks the write-back of execution errors, making it difficult for the long-term deviation of the model to converge. Summary of the Invention

[0003] In view of the above-mentioned many problems existing in the prior art, the present invention provides a multi-source fusion avionics control system and method. First, the present invention randomly projects and compresses inertial, star sensor, and radar measurements at the edge side; the fusion twin module uses fractional-order filtering of the Euclidean motion group and ternary mutual information hypergraph to obtain the fusion state and correlation, then generates the aerodynamic, electromagnetic, and structural fields of the shared grid through physical constraint diffusion at one time, and aligns them with radar observations using entropy-regularized optimal transport; the decision execution module maps discrete actions to the Ising model, obtains a globally feasible sequence through quantum annealing, and the pulsed symplectic structure decision network continuously refines and writes back the execution error to the hypergraph.

[0004] A multi-source fusion avionics control system includes: A multi-source acquisition and compression module, configured to synchronously acquire multi-source sensor data including at least inertial measurement data, star sensor data, and radar data under a unified time reference, and perform compression processing on the multi-source sensor data to generate compressed observation data; A fusion twin module, configured to generate fusion state data and hypergraph data based on the compressed observation data, generate twin environment data and twin environment registration results according to the fusion state data and the hypergraph data, and output the twin environment data and the twin environment registration results; A decision execution module, configured to generate decision instruction data based on the twin environment data, the twin environment registration result, and the discrete control variables quantified by a flight control computer, and send the decision instruction data to an independently provided control execution unit to generate control execution data; the decision execution module is further configured to update the hypergraph data based on the control execution data and the execution feedback data for the next cycle call of the fusion twin module.

[0005] Preferably, the multi-source acquisition and compression module synchronously acquires multi-source sensor data including inertial measurement data, star sensor data, synthetic aperture radar data, short-wave infrared optoelectronic array data, global navigation satellite system receiver data, and meteorological sounding data link data, and performs a linear mapping on the multi-source sensor data through a random projection matrix after completing time reference alignment to generate compressed observation data.

[0006] Preferably, before performing a linear mapping on the multi-source sensor data, the multi-source acquisition and compression module whitens the multi-source sensor data using covariance matrices corresponding to various types of acquired data.

[0007] Preferably, the fusion twin module obtains the ternary joint mutual information by calculating the difference between the Shannon entropy and the joint entropy of three groups of node observations, and determines the hyperedge weights based on the ternary joint mutual information to construct hypergraph data.

[0008] Preferably, the fusion twin module processes the attitude state vector using a fractional-order Kalman filtering method on the Euclidean motion group manifold, and uses the filtering result as the fusion state data.

[0009] Preferably, the fusion twin module generates aerodynamic field data, electromagnetic scattering field data, and structural stress field data on a three-dimensional finite element grid using the fusion state data and the hypergraph data, and the three types of field data share the node coordinates and element topologies of the same spatial grid.

[0010] Preferably, the fusion twin module uses an entropy-regularized optimal transport algorithm to register the electromagnetic scattering field data and the radar observation density obtained from the compressed observation data to generate a twin environment registration result.

[0011] Preferably, the decision execution module constructs the discrete control variables into an Ising model, and solves the Ising model on a quantum annealing processor to obtain a control sequence seed.

[0012] Preferably, the decision execution module inputs the control sequence seed, the twin environment data, and the twin environment registration result into a pulse symplectic structure policy network, optimizes the control sequence seed to generate decision instruction data, and after obtaining the control execution data, writes the control execution data and the corresponding execution feedback data into the hypergraph data for the next cycle call of the fusion twin module.

[0013] A multi-source fusion avionics control method, which is applied to the multi-source fusion avionics control system, includes the following steps: S1. Synchronously collect multi-source sensor data under a unified time reference. The multi-source sensor data at least includes inertial measurement data, star sensor data, and radar data. Perform compression processing on the multi-source sensor data to generate compressed observation data; S2. Generate fusion state data and hypergraph data based on the compressed observation data; S3. Generate twin environment data and twin environment registration results according to the fusion state data and the hypergraph data, and output the twin environment data and the twin environment registration results; S4. Generate decision instruction data based on discrete control variables, the twin environment data, and the twin environment registration results, and send the decision instruction data to a control execution unit to generate control execution data; S5. Write the control execution data and execution feedback data into the hypergraph data for the next cycle to execute step S2.

[0014] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows: By introducing a ternary mutual information hypergraph to perform high-order coupling mapping on inertial measurement, star sensor, and radar measurements, the complete mining of complementary information between multi-source observations is realized, and attitude drift under high-maneuver conditions is avoided.

[0015] By using a diffusion network with physical constraints to generate an aerodynamic field, an electromagnetic field, and a structural stress field in real time on a shared grid, the online reflection of airframe deformation and external disturbances is realized, and the lag problem of the offline database is eliminated.

[0016] By mapping discrete control variables to an Ising model and globally searching on a quantum annealing processor, and then continuously refining in combination with a pulse symplectic structure strategy network, the solution of control instructions that simultaneously satisfies hard constraints and global optimality is realized, and the defect of instability prone to local search is solved.

[0017] By writing control execution data and execution feedback data back to the hypergraph in the decision execution module, closed-loop self-evolution is realized, enabling the model deviation to converge quickly. Description of the Drawings

[0018] Figure 1 It is an interaction diagram of the system of the present invention; Figure 2 It is an entropy-regularized optimal transport registration flowchart in the present invention; Figure 3 It is a quantum annealing-pulse symplectic strategy optimization flowchart in the present invention; Figure 4 It is a schematic flowchart of the method of the present invention. Detailed implementation manners

[0019] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present disclosure. In the following detailed description, for the sake of explanation, numerous specific details are set forth in order to provide a comprehensive understanding of the embodiments of the present disclosure.

[0020] As Figure 1 shown, a multi-source fusion avionics control system includes: A multi-source acquisition and compression module, configured to synchronously acquire multi-source sensor data including at least inertial measurement data, star sensor data, and radar data under a unified time reference, and perform compression processing on the multi-source sensor data to generate compressed observation data; The multi-source acquisition and compression module undertakes the responsibility of input unification for the multi-source fusion avionics control system. First, the internal time reference unit of the module is driven by a temperature-compensated rubidium atomic clock, and the pulse-per-second signal is distributed to the inertial measurement component, the star sensor component, and the radar component, so that the three types of sensors generate raw observation data at the same absolute moment. When each component outputs, a hardware timestamp field is written simultaneously. The timestamp belongs to the category of "first-time synchronization data", and subsequent links complete cross-source alignment based on this. The time synchronization accuracy depends on the short-term stability characteristics of the master clock, which is less than 10 picoseconds in this system, and can ensure that the track errors from different data sources in subsequent filtering accumulate within the millimeter level.

[0021] After data synchronization, the module starts the dimensionality reduction and compression process. Since the 6D vectors of the inertial measurement component, the star catalog of the star sensor image, and the radar echo grid data have significant differences in the spatial dimension, directly splicing is likely to lead to the curse of dimensionality. The module uses a random projection matrix to complete one-time compression, mapping the high-dimensional multi-source observations into low-dimensional compressed observation data. The core calculation expression is as follows:

[0022] where represents the original multi-source observation vector spliced according to the timestamp, is a random projection matrix with independently valued elements, the number of rows is , and the number of columns is , is the compressed observation data. The row-column ratio depends on the bandwidth budget, and as long as the Johnson-Lindenstrauss lemma is satisfied, the Euclidean distance can be approximately maintained unchanged. To reduce the uncertainty of random mapping, the module uses a true random number generator based on quantum tunneling noise to generate when the system is powered on, and then is solidified in the on-chip read-only memory to ensure repeatability.

[0023] Whitening processing is a necessary step before dimensionality reduction. The module utilizes the noise covariance matrix obtained through offline calibration to calculate its inverse square root matrix and performs a linear transformation on the vector before whitening:

[0024] where is the observation vector with covariance unit normalization. Performing random projection on the whitened vector can avoid the situation where a large-scale dimension of a certain type of sensor dominates all the principal components, thereby enhancing the numerical stability of the compressed observation data for subsequent state estimation.

[0025] To reduce the end-side latency, the module's hardware implementation adopts a system-on-chip design. The inertial measurement unit outputs at a rate of 1000 Hz, the star sensor outputs at a rate of 10 Hz, and the radar unit outputs at a rate of 40 Hz. Inside the module, an event-driven buffer mechanism is used: when any sensor generates a new frame, the corresponding buffer is written to and a synchronization check is triggered. If all three buffers are marked as "unprocessed", the module immediately reads three frames and splices them to form the original multi-source observation vector. This design not only avoids high-sampling-rate sensors from waiting but also ensures the consistency of all sensor frame numbers. The size of the spliced vector is approximately one-tenth of the original size, and the single-frame processing latency does not exceed 200 microseconds.

[0026] The multi-source acquisition and compression module encapsulates the output compressed observation data according to a custom protocol: the frame header contains a timestamp, a frame count, and an integrity check code, and the frame body stores the whitening coefficient index, the compressed observation data vector, and additional sensor health information in field order. The data is sent to the fusion twin module through an avionics full-duplex Ethernet bus. Since the random projection operation approximately preserves the inner product of the vector, the fusion twin module can directly perform attitude state filtering in the compressed space without restoring the high-dimensional original data. This strategy saves more than 70% of the link bandwidth and reduces the storage space pressure.

[0027] Actual flight examples show that in the Mach 0.8 cruise state, the module processes and sends approximately 1300 frames of compressed observation data per second, and the average power consumption of the whole machine is less than 5 watts. Compared with the uncompressed scheme, the total data volume is reduced by approximately 80%, and the increase in the root mean square error of the attitude of the fused state data is less than 0.1 degree, meeting the requirements of high-maneuver avionics training. In summary, the multi-source acquisition and compression module provides a hardware-implementable, low-computation-overhead, and tightly data-coupled observation front-end solution through high-precision synchronization, covariance whitening, and random projection compression, providing a unified, low-latency, and information-complete input for the subsequent fusion twin module.

[0028] Preferably, the multi-source sensor data synchronously collected by the multi-source acquisition and compression module includes inertial measurement data, star sensor data, synthetic aperture radar data, short-wave infrared optoelectronic array data, global navigation satellite system receiver data, and meteorological sounding data link data. After completing the time reference alignment, a linear mapping is performed on the multi-source sensor data through a random projection matrix to generate compressed observation data.

[0029] The multi-source acquisition and compression module of the multi-source fusion avionics control system undertakes the responsibility of unifying the data entry, ensuring both spatio-temporal consistency and controlling the bus bandwidth. The module consists of four parts: a time reference unit, a sensor synchronization unit, a whitening unit, and a random mapping unit. The time reference unit uses a temperature-compensated rubidium atomic clock to broadcast the absolute time to the inertial measurement component, star sensor component, synthetic aperture radar component, short-wave infrared optoelectronic array component, global navigation satellite system receiver component, and meteorological sounding data link component through a pulse-per-second signal. Each component samples immediately when receiving the trigger edge and writes a timestamp field to the local hardware register. The introduction of a unified time reference enables observations from different sources to be aligned at the nanosecond level, and the subsequent state estimation process can still maintain an angular displacement error of less than 1 arcminute in a high-maneuver scenario.

[0030] The sensor synchronization unit polls the output buffers of the 6 types of sensors according to the timestamps. When it detects that all 6 frames of data at the same moment have arrived, it immediately triggers a splicing operation to generate an original multi-source observation vector. . Here The dimension is affected by the number of star points of the star sensor and the size of the radar echo grid, and can reach tens of thousands of dimensions. If directly transmitted to the fusion twin module, it will cause a sharp increase in the Ethernet load. To reduce the dimension and maintain the information content, the module first applies covariance equalization to in the whitening unit. The noise covariance matrix obtained by offline calibration is denoted as , and the inverse square root matrix is obtained using the Sherman-Morrison formula. After linear transformation, a whitened vector is obtained. The variances of the elements of

[0031] are constrained to the same order of magnitude, making the influence of different sensors on the transformation result more uniform in the random mapping stage. . The matrix has rows and columns, where the element takes values of +1 or -1 and satisfies that the row vectors are pairwise independent. Each element obeys the Rademacher distribution with a mean of 0 and a variance of 1, and only A non-zero value is used to form a sparse structure, which not only ensures the approximate orthogonality of random projection but also reduces the amount of multiplication operations. Represents the sparse trace random projection matrix The number of non-zero elements retained in each row. The matrix is generated and solidified once by a quantum tunneling noise random source after the system is powered on. The core mapping calculation formula is:

[0032] In the formula, Represents the compressed observation data vector, Represents the dimension of the original multi-source observation vector, Represents the compressed dimension, satisfying , Represents the random projection matrix, Represents the observation vector after covariance whitening.

[0033] This mapping follows the Johnson-Lindenstrauss lemma and approximately preserves the inner product of any two frame vectors in the probabilistic sense, thus ensuring that the random projection does not damage the measurement innovation component used for Kalman filtering. By experimentally tuning When is approximately 10% of the original dimension, the bus bandwidth can be compressed to less than 20% of the original scheme. In the flight simulation experiment, the root mean square error increment of the system attitude is less than 0.05 degrees. The vector after random mapping is written into the Ethernet frame body in a fixed-point 16-bit format and sent to the fusion twin module in real time; the frame header includes a timestamp, a frame count, and a cyclic redundancy check field.

[0034] To verify the operability of the compression strategy, an actual installation test was carried out on a certain type of UAV platform. When the UAV is cruising at an altitude of 9000 meters at 0.8 Mach, the inertial measurement component outputs at a rate of 1000 Hz, the star sensor outputs at a rate of 10 Hz, the radar component outputs at a rate of 40 Hz, and the output rates of the remaining sensors are between 5 Hz and 50 Hz. The module hardware uses a system-on-chip solution, in which the multiplication of the whitening matrix is completed by a parallel multiplier-accumulator array, and the single-frame processing delay is 120 microseconds; the random projection is completed by a systolic array pipeline, with a delay of 80 microseconds. The whole machine generates about 1300 frames of compressed observation data per second, and the actual link load does not exceed 1.2 gigabits per second, which is 80% lower than the uncompressed transmission scheme. At the same time, the fusion twin module completes fractional-order filtering and hypergraph update only by receiving compressed observation data, and the mean square error of the output heading angle increases by only 0.03 degrees compared with the full-dimensional input, meeting the high dynamic load requirements of avionics training.

[0035] In another set of maneuvering tests, the drone entered a high angle of attack maneuver with an 8g turn. The inertial sensor saturation clock was about 80 milliseconds, and the random mapping remained effective. The back-end filter converged within 1 second after restoring the linear range. This embodiment shows that even in extreme attitude changes and temporary failure of some sensors, random mapping can still provide sufficient measurement information and compensate for missing nodes through hypergraph redundant mutual information, thereby maintaining the system's ability to continuously estimate the heading.

[0036] Preferably, before performing linear mapping on the multi-source sensor data, the multi-source acquisition and compression module performs whitening processing on the multi-source sensor data using a covariance matrix corresponding to each type of acquired data.

[0037] The multi-source acquisition and compression module needs to explain the whitening process before performing random projection linear mapping. The so-called whitening process in this system refers to the use of the measured noise covariance matrix of various types of acquisition data to perform linear transformation on the observation vectors spliced ​​at the same time, so that different sensor channels enter the subsequent random projection under the condition that the mean is zero and the covariance matrix is ​​converted into a unit matrix. After whitening, the variance of each dimension of data is consistent and the cross-correlation is zero, thereby avoiding the problem of high-energy channels dominating the mapping direction during random projection and improving the numerical stability of compressed observation data in Kalman filtering.

[0038] First, during the UAV ground test and wind tunnel calibration phase, long-time noise data is collected within the static and dynamic envelopes for each type of sensor, and the covariance matrix is ​​estimated offline. First, during the UAV ground test and wind tunnel calibration phase, long-time noise data is collected for each type of sensor within the static and dynamic envelopes, and the covariance matrix is ​​estimated offline. Represents a column vector composed of inertial measurement data, star sensor catalog feature vectors, synthetic aperture radar echo feature vectors, etc. spliced ​​by timestamp, with a dimension of .make Represents the covariance matrix of the vector on the calibration set, with dimension .Will Doing eigendecomposition yields:

[0039] in is the eigenvector matrix, is a diagonal matrix of eigenvalues. Define the whitening matrix:

[0040] Then the observation vector after whitening is:

[0041] in The covariance matrix is the identity matrix, and each component is uncorrelated.

[0042] The whitening matrix is solidified into the read-only memory of the on-chip system during the manufacturing process. During flight, after the time synchronization unit detects that the data of six types of sensors are complete, the whitening operation is started in the on-chip system pipeline. The pipeline uses a 32-way parallel multiply-accumulate array to ensure that a 4×4 block matrix-vector multiplication can be completed within a 40-nanosecond clock cycle; the entire dimensional multiplication is divided into several blocks for parallel processing. To avoid the influence of temperature drift on the eigen-decomposition result, the system recalibrates and issues a new whitening matrix after each maintenance calibration.

[0043] After whitening, the variances of the vector elements converge, and the random projection matrix is no longer dominated by a single high-energy channel in the subsequent linear mapping. System tests show that the ratio of the maximum channel variance to the minimum channel variance of the vector after random projection without whitening can reach 80; after whitening, this ratio is maintained within 1.3. Corresponding to the fractional Kalman filter of the fusion twin module, the root mean square error of the attitude in the simulation is reduced by about 15%. In addition, the whitening operation weakens the influence of large-energy radar spurs on the random projection result, reduces the dynamic range requirement of the subsequent multiplier, and compresses the power consumption of the on-chip system by about 10%.

[0044] In an embodiment, during a 30-minute continuous high-maneuver flight test, the aircraft body successively experiences level flight at 0.8 Mach, 8g roll, and a 90-degree angle of attack maneuver. The system continuously outputs whitened vectors and compresses them to 1280 dimensions through random mapping. The convergence time of the fusion twin module always remains within 200 milliseconds; compared with the non-whitened version, the convergence time in the extreme maneuver section increases to 300 milliseconds. After post-event evaluation, the average value of the overall heading angle error of the whitened version is 0.12 degrees, and that of the non-whitened version is 0.19 degrees. This result shows that the whitening operation makes an obvious contribution to the state estimation accuracy when the dynamic envelope expands.

[0045] As Figure 2 shown, the fusion twin module is used to generate fusion state data and hypergraph data based on the compressed observation data, generate twin environment data and twin environment registration results according to the fusion state data and the hypergraph data, and output the twin environment data and the twin environment registration results; The fusion twin module undertakes the core computing function of the multi-source fusion avionics control system. Its workflow can be divided into four stages: state fusion, hypergraph construction, environment twin generation, and environment registration. First, after receiving the compressed observation data, the module needs to filter the attitude state on the Euclidean motion group manifold. The specific method is to denote the state vector containing the attitude quaternion, three-dimensional position, three-dimensional velocity, and inertial measurement bias as and denote the measurement vector from the compressed observation data as 。The filter aims to minimize the cost function:

[0046] where is the measurement mapping, is the state propagation operator, and are the measurement and process covariances respectively. The fused state data is obtained through a fractional-order prediction-correction recursion. This method can weaken the instantaneous mismatch between measurement and process noise in the high-speed maneuvering section, and the convergence speed is about 25% higher than that of the integer-order Kalman filter.

[0047] After obtaining the fused state data, the module constructs a multi-source correlation hypergraph structure through ternary joint mutual information. Let , , represent the whitened vectors of the current frames of the , , -th class sensors respectively. The corresponding hyperedge weights are calculated by:

[0048] where represents the Shannon entropy. After the Laplacian decomposition of the hypergraph, the hypergraph data and hypergraph spectrum coefficients are output, which encode the coupling relationship of multi-source observations in the frequency domain and provide high-frequency priors for subsequent scene twins.

[0049] In the environmental twin generation stage, a physical constraint diffusion model is adopted. On the one hand, the fused state data and hypergraph spectrum coefficients are concatenated into a conditional tensor; on the other hand, a three-dimensional structured grid is used as the analytical domain. The neural network injects the incompressible flow equation and the static elastic equation as constraints during the reverse sampling process and outputs the aerodynamic field, electromagnetic scattering field and structural stress field, and the three fields constitute the twin environmental data. Compared with the traditional per-module analysis scheme, this method couples three physical domains at one time, and the generation speed is increased by about ten times.

[0050] In the twin environment registration stage, the deviation problem between the model and the real observation is solved. Let the predicted electromagnetic scattering density be , and the radar-measured density be . The module solves it through entropy-regularized optimal transport:

[0051] to obtain the transport matrix and the registration potential field, and outputs the twin environment registration result. After registration, the matching error between the aerodynamic field and the measured wing pressure decreases from 15% to 6% in the wind direction mutation scenario, and the radar scattering intensity error decreases from 22% to 8%.

[0052] To verify the operability of the module, flight tests were conducted on a certain type of tactical UAV platform. Two scenarios were set for the tests: one was level flight at 0.8 Mach, and the other was a high-maneuver dive turn at 8g. In the level flight scenario, the fused twin module took an average of 8 milliseconds to complete one frame of twin generation and registration; in the high-maneuver scenario, the time increased to 11 milliseconds. The real-time attitude estimation error was always less than 0.15 degrees. After the twin environment data and registration results were used by the decision execution module to generate instructions, the trajectory holding error of the whole aircraft was within 0.3 meters, meeting the requirements for precise ground navigation.

[0053] After system integration, the data interfaces of each module adopted a unified hierarchical key-value format. Each frame of output of the fused twin module included three parts: one was the fusion status data, the second was the hypergraph data, and the third was the twin environment data and twin registration results encoded in octal grids. Such an organization method could directly parse the target information and environmental constraints on the side of the decision execution module, avoiding secondary format conversion. The actual installation test showed that compared with the traditional modular independent computing scheme, the fused twin module scheme reduced the power consumption by about 35% with equivalent hardware configuration, the bandwidth requirement decreased by about 70%, and significantly improved the robustness of the system under extreme envelopes.

[0054] All in all, the fused twin module transforms discrete, multi-source, and heterogeneous sensing information into a unified and continuously updatable environment and self-state description through four-level processes of state fusion, hypergraph association, physical coupled twins, and entropy-regularized registration, which not only enhances the response speed of the decision-making layer to sudden working conditions but also provides a high-fidelity prior for subsequent cycles, realizing the closed-loop self-evolution of the sensing-decision link of the avionics control system.

[0055] Preferably, the fused twin module obtains the ternary joint mutual information by calculating the difference between the Shannon entropy and the joint entropy of three groups of node observations, and determines the hyperedge weights based on the ternary joint mutual information to construct the hypergraph data.

[0056] The fused twin module needs to understand the correlation patterns among the inertial measurement vector, the star point features of the star sensor, and the radar scattering intensity within milliseconds so that the downstream physical field generation can receive both spatial and frequency-domain priors simultaneously. Therefore, the present invention introduces a ternary joint mutual information hypergraph. The traditional mutual information measures the dependence relationship between two variables and cannot reveal the new information generated after the intervention of the "third source" in a multi-source scenario. The present invention regards the inertial measurement node, the star sensor node, and the radar node as a triple, and quantifies the interaction of the three sources through the ternary joint mutual information. For each timestamp , the whitened inertial measurement vector is denoted as , the star map encoding vector of the star sensor is denoted as , and the radar echo grid vector is denoted as . Using the Shannon entropy to describe the amount of information, it is defined as:

[0057] where is the probability of the -th discretized value of the inertial measurement vector. Similarly, and can be obtained. The joint probability density of the three is calculated to obtain the joint entropy:

[0058] The three-way joint mutual information is defined as:

[0059] where measures the new or redundant information when the three sources appear simultaneously. If is positive, it means that the effective information provided by the three-source combination is greater than the sum of their individual independent information, and there is complementarity between the sensors; if is negative, it means there is redundancy. The module maps to the hyperedge weights:

[0060] After the exponential mapping, the weights are always positive and increase monotonically with the mutual information, facilitating the use of multiplication and accumulation rather than sign judgment in hardware.

[0061] The estimation methods of entropy and joint entropy directly affect the weight accuracy. "Entropy" refers to the Shannon entropy of each single-source observation vector (i.e., the marginal entropy), while "joint entropy" refers to the Shannon entropy of the joint distribution of the three-source observation vectors. The module discretizes the vector value range using the fixed-width histogram binning method and combines the logarithmic quantization of the radar echo amplitude in segments to reduce the information bias caused by the long-tailed distribution. To improve real-time performance, the entropy calculation is performed on a programmable logic array: the histogram counting, logarithm lookup table, and accumulator are pipelined, and the triple sum operation is completed in 250 nanoseconds. The entropy value is written into the on-chip cache in single-precision floating-point format and then enters the mutual information calculation pipeline. The mutual information result is written into the hypergraph structure buffer together with the corresponding node index.

[0062] When constructing the hypergraph data, the node set includes inertial measurement nodes, star sensor nodes, and radar nodes, and the hyperedge set consists of all triples. The relationship between hyperedges and nodes is stored in the form of an incidence tensor. Then, the incidence tensor is mapped to a weighted Laplacian matrix for the next spectral analysis. The eigenvector matrix obtained from the spectral decomposition is used both for high-frequency noise suppression in fractional-order filtering and as the conditional encoding vector for the physical field generation network. For example, in the high angle-of-attack state, there is a strong coupling between radar scattering and attitude, the mutual information weight increases with the change of attitude, and the high-frequency components output by the spectral decomposition automatically amplify the weight of the radar channel in the twin generation network, realizing data-driven dynamic channel scheduling.

[0063] To verify the effectiveness of the ternary mutual information hypergraph, the present invention constructs two sets of comparison systems: one uses traditional binary mutual information for mapping, and the other uses the ternary mutual information proposed by the present invention for mapping. Using the same batch of flight data as input, the mean absolute error between the output twin environment data and the aerodynamic pressure field actually measured by the aircraft is detected. The results show that the traditional binary mutual information scheme has an 18% error in the sharp roll condition, and the error of the present invention's scheme is 11%; in the sudden headwind condition, the traditional scheme has an error of 24%, and the error of the present invention's scheme is 13%. The improvement is mainly attributed to the fact that the ternary mutual information can capture the synchronous changes between the high-frequency jitter of inertial measurement and the radar stray noise, thereby strengthening the prior expression of the flow field disturbance in the mapping stage.

[0064] In terms of engineering implementation, to ensure a 50 Hz cycle speed, the module adopts a sliding window strategy: updating the histogram count for the most recent 16 frames of data, and re-estimating the entropy and mutual information as the window slides. The count update is completed by a circular buffer, writing new frame data while discarding the oldest frame data; the differential update of the entropy value and mutual information does not require a complete recalculation, greatly reducing the use of FPGA resources. One sliding update takes 80 microseconds, meeting the real-time constraint.

[0065] In the embodiment, the system is tested for 6 hours of continuous operation on the flight control test bench. The platform adds ±5-degree random attitude disturbances to the IMU excitation, adds 3% Gaussian noise to the star sensor images, and adds 6 dB pulse interference to the radar amplitude. The dynamic distribution of the ternary mutual information weights can expose the data mutation source in real time: when the radar amplitude mutates, the weights of the radar-related terms in the matrix drop rapidly, and the spectral matrix rotation result projects the radar nodes into the high-frequency subspace, and the filter automatically reduces its influence. The peak value of the system attitude estimation error is only 0.4 degrees; using the binary mutual information scheme, a 1.1-degree spike appears. This shows that the scheme of the present invention has stronger robustness.

[0066] Preferably, the fusion twin module uses a fractional-order Kalman filtering method to process the attitude and pose state vector on the Euclidean motion group manifold, and uses the filtering result as the fusion state data.

[0067] The core task of the fusion twin module is to simultaneously estimate the attitude, position, and velocity under continuous high-dynamic conditions. Traditional attitude filtering mostly relies on the integer-order Kalman algorithm, which assumes that the process noise and measurement noise remain stable within each sampling interval. However, in scenarios such as sharp maneuvers, radar interference, or star map truncation, the measurement covariance shows obvious historical dependence. To solve this problem, the present invention constructs the state estimation model on the Euclidean motion group manifold SE3 and introduces a fractional-order Kalman filter to capture the cross-frame correlated noise through a non-integer-order memory kernel.

[0068] The Euclidean motion group manifold SE3 is composed of the rotation part SO3 and the translation vector combination, which can fully represent the rigid body attitude and position.

[0069] Let the state vector be composed of the quaternion and the position , and then append the velocity and the inertial measurement bias to be extended to . In the prediction stage, the quaternion is updated through the exponential map, and the position and velocity are propagated by inertial measurement integration. To introduce historical memory, the filter adds a fractional-order prediction-correction operator of order , and its core recurrence is:

[0070]

[0071] where is the sampling period, is the Lie algebra element of the previous period, is the measurement vector, is the measurement mapping matrix. The introduction of the fractional power makes the increment show power-law memory, and the corresponding noise covariance gradually decays in time according to the long-tailed kernel. Taking between 0 and 1 can degenerate to the classical Kalman in the stationary section and retain historical information in the non-stationary section.

[0072] In terms of implementation, the module uses the Lie group exponential-logarithmic mapping look-up table to avoid taking the power of the rotation matrix in real time. The prediction formula is implemented with fixed-point arithmetic in the digital signal processor. The fractional exponentiation of the correction formula is approximated by a piecewise power series, and the order is truncated at the 5th order to control the delay. The entire filter has a running delay of about 300 microseconds per frame, meeting the system's 50 Hz closed-loop requirement. Since the Euclidean motion group manifold naturally encapsulates the rotation and translation coupling, there is no need for an external alignment transformation between the attitude and the position, and the fused state data can be directly output for subsequent physical twin generation.

[0073] The effect verification was carried out through two rounds of flight tests. In the first round on a straight flight path, the fractional-order filtering and integer-order filtering performed similarly, and the root mean square error of the attitude was less than 0.05 degrees. In the second round, maneuvers of 8g dive and 90-degree angle of attack were adopted, resulting in instantaneous distortion of radar measurements and a large number of missing star points in the star sensor. At this time, the output of the integer-order filtering had an angle drift exceeding 0.8 degrees, while the error of the fractional-order filtering was controlled within 0.23 degrees. The reason is that the power-law memory assigns higher weights to historical low-noise measurements, effectively suppressing the covariance inflation of abnormal frames. The generated fused state data maintains the continuity of pose parameters in the twin environment solution, avoiding fluid mesh remapping, thereby shortening the twin generation time by approximately 25%.

[0074] The embodiments show that in the scenario of crossing the thunderstorm area, the radar input is affected by Doppler leakage, and the inertial measurement is coupled with vibration noise. The fractional-order filtering uses a 4-second window in the past to perform memory compensation for velocity drift, reducing the velocity error from 3 meters per second to 1 meter per second. Combined with the supergraph frequency-domain correction, the estimation error of the aerodynamic load is reduced from 18% to 9%, providing a more reliable environmental prior for the decision-making execution module.

[0075] Preferably, the fused twin module generates aerodynamic field data, electromagnetic scattering field data, and structural stress field data on a three-dimensional finite element grid using the fused state data and supergraph data. The three types of field data share the node coordinates and element topologies of the same spatial grid.

[0076] The fused twin module undertakes the responsibility of instantaneously synthesizing three-dimensional physical fields in the multi-source fusion avionics control system. The inputs of the module include the fused state data and the supergraph data reflecting the correlation of multi-source observations. The fused state data provides the attitude, position, velocity, and zero-bias estimation of inertial measurement of the aircraft on the manifold of the Euclidean motion group; the supergraph data encodes the coupling relationship of each sensor measurement in the frequency domain in the form of a high-dimensional feature vector. The module first constructs a three-dimensional finite element grid within the envelope of the airframe. The grid node coordinates and element topologies are generated offline during the mission planning phase, and their resolution is determined based on the minimum structural feature length and the radar wavelength. The grid space remains fixed, enabling different physical fields to share the same coordinate basis and interpolation function, avoiding cross-domain interpolation errors.

[0077] The physical field generation uses a conditional diffusion network. The conditional vector is composed of the fused state data and the supergraph spectrum coefficients, denoted as . The initial tensor of the grid is a Gaussian noise field, and the network gradually evolves to the target distribution during the reverse process. A conservation constraint kernel function is added during the evolution process:

[0078] where is the velocity field, and are the electric field and the magnetic field respectively, is the stress tensor, is the magnetic permeability of the medium. Physical losses are added to the end-to-end minimization of the overall loss function of the model to ensure that the aerodynamic field data, electromagnetic scattering field data, and structural stress field data satisfy the conservation laws.

[0079] After generation, the three field data are written into the shared grid node array in tensor form: velocity, pressure, and viscous stress are written into the aerodynamic tensor channel; the composite scattering coefficient is written into the electromagnetic tensor channel; the redundant principal stress is written into the structural tensor channel. Utilizing the topological consistency brought by the shared grid, the decision execution module can directly access multi-domain information at the node level without performing re-interpolation between the three fields.

[0080] The adaptability to scene mutations is achieved through the registration strategy. The module reconstructs the radar observation density based on the compressed observation data and extracts the predicted density from the electromagnetic scattering field data . The two are used to calculate the transport matrix through entropy-regularized optimal transport , and based on this, the diffusion network noise schedule is updated to achieve online correction. This feedback is triggered every twenty milliseconds and can quickly approximate the true scattering distribution when there is a flicker interference in the radar signal.

[0081] Example 1: The aircraft flies level at an altitude of 9000 meters and a Mach number of 0.8, and the module generates field data with a cycle period of 5 milliseconds. Compared with the traditional method of looking up tables in the aerodynamic database and using the radar scattering algebraic model, the aerodynamic pressure error is reduced from 12% to 4%, the scattering cross-section error is reduced from 18% to 8%, and the difference between the structural stress and the multi-body dynamics model remains within 6%.

[0082] Example 2: Execute an 8g rapid roll maneuver, the instantaneous angle of attack of the airframe reaches 60 degrees, and the radar azimuth SNR drops. The module captures the synchronous change of the high-frequency lateral acceleration of the inertial measurement and the radar amplitude through the hyper-spectrum coefficient and automatically increases the radar node weight. With the optimal transport registration, the stable field estimation is restored 45 milliseconds after the maneuver, while without the registration strategy, it takes 110 milliseconds to converge.

[0083] The shared grid strategy saves video memory bandwidth in hardware implementation. The three physical tensors share the node array, saving approximately 40% of the memory compared to independent grids; at the same time, the same node index allows the decision execution module to obtain the three-field vectors in one bus access, achieving zero-copy fusion. This design reduces the overall machine communication delay by approximately 30%, which is particularly crucial for high-maneuver closed loops.

[0084] From a system perspective, the integrated twin module realizes the homomorphic coupling of the aerodynamic, electromagnetic, and structural domains by mapping the integrated state data to a three-dimensional finite element grid, adjusting the cross-source information weights with the help of hypergraph data, and generating a coupled physical field using a conservation-constrained diffusion model; and enhances the adaptability to external mutations through fast correction with entropy-regularized optimal transport. The output twin environment data and twin environment registration results provide a more realistic prior of the external environment and internal structural loads of the aircraft body for the decision-making and execution module, enabling the decision-making and execution module to maintain trajectory control accuracy under extreme working conditions while reducing the risk of aircraft body overload, demonstrating the overall technical advantages of the present invention in integrated avionics control.

[0085] Preferably, the integrated twin module uses the entropy-regularized optimal transport algorithm to register the electromagnetic scattering field data and the radar observation density obtained from the compressed observation data to generate the twin environment registration result.

[0086] The electromagnetic scattering field data output by the integrated twin module comes from a physically constrained diffusion network. Based on the three-dimensional finite element grid, this data gives the complex scattering intensity for each grid cell. At the same time, the multi-source acquisition and compression module records the radar echo amplitude at the same moment in the radar channel. After amplitude calibration and azimuth matching, it is discretized into a grid-based radar observation density. Although the two share the coordinate system, the former is the result generated by the model, and the latter is the measured result, with differences such as amplitude scaling, local offset, and Gaussian speckle noise. If the electromagnetic scattering field data is directly used to drive the decision-making, it will cause distortion in radar threat assessment. Therefore, the present invention introduces entropy-regularized optimal transport registration in the integrated twin module to align the model density with the observation density, output the twin environment registration result, which is used to correct the physical field and update the network parameters.

[0087] Basis of entropy-regularized optimal transport, let be the discrete density vector of the electromagnetic scattering field data on the grid, be the radar observation density vector. The number of elements in the two vectors is the same, and after normalization, they satisfy . Define the cost matrix , where is the Euclidean distance from grid cell to cell multiplied by the normalized scattering attenuation coefficient. The entropy-regularized optimal transport problem is written as:

[0088] where, is the transport matrix, satisfying that the row margin and column margin are equal to , respectively; is the regularization coefficient; is the relative entropy. In the formula Minimize the transportation cost. The entropy regularization term avoids the solution being too sparse and is conducive to high-speed iteration.

[0089] Algorithm implementation: The module uses the Sinkhorn iteration to perform this optimization. Specific steps: 1. Initialize the scaling vectors , .

[0090] 2. Calculate the kernel matrix .

[0091] 3. Repeat in the graphics processing unit:

[0092] where the symbol represents element-wise division.

[0093] 4. Output the transportation matrix when the row margin and column margin errors are both below the threshold:

[0094] Select taking into account both convergence and resolution. By scanning the empirical table, the system uses at an altitude of 6000 meters and a wavelength of 3 cm and can converge within 32 iterations, meeting the 20-millisecond cycle budget. After obtaining the registration potential and updating the transportation matrix of the network, the module calculates the registration potential:

[0095] Take as the additional conditional tensor and input it into the self-attention layer of the diffusion network, replacing the original residual noise schedule, to guide the subsequent sampling process. Experiments show that a single guidance can significantly narrow the gap between the model and the observation. This registration potential is also output to the decision execution module as the registration result of the twin environment and serves as an important indicator in interference detection and signal camouflage evaluation.

[0096] During registration in the level flight scenario, the scattering cross-section error is reduced from 9% when separately output by the generation network to 3%. In the radar deception interference test, the enemy sends an instantaneous gain inversion pulse to introduce false echoes, resulting in extrema appearing in the local grid. The entropy-regularized optimal transport can smooth the extrema to adjacent cells through kernel matrix diffusion. After network update, the sensitivity to abnormal points is reduced, avoiding the body attitude jitter caused by excessive suppression of the control law. During continuous ten-minute interference tests, the pitch angle amplitude of the system is maintained within 0.4 degrees, while the unregistered scheme has peaks above 2 degrees.

[0097] Example: Multiple-segment maneuvering flight tests on a certain type of high-speed UAV platform: Phase 1: Level flight at Mach 0.6. The registration convergence time is 5 milliseconds, and the scattering cross-section comparison error is 2.5%.

[0098] In the second stage, an 8g roll is performed, and the radar signal-to-noise ratio drops sharply. The registration convergence time is 11 milliseconds, and the contrast error of the scattering cross-section is 6%.

[0099] In the third stage, radar pitch perturbation and scintillation interference are added. The registration convergence time is 16 milliseconds, the root mean square error of the attitude increases by 0.08 degrees, and it decreases by 0.3 degrees compared with the unregistered scheme.

[0100] The results show that the entropy-regularized optimal transport of the present invention can quickly and stably correct the twin model under different flight conditions and external interference conditions.

[0101] The hardware resources adopt a 32-by-32 block parallel Sinkhorn kernel, and the floating-point multiply-accumulate unit occupies 17% of the graphics processing unit's computing resources. The video memory consumption per frame of data slice for registration is about 1.5 megabytes, and it can be transferred without copy between the graphics processing unit and the central processing unit.

[0102] As Figure 3 shown, the decision execution module is used to generate decision instruction data according to the twin environment data, the twin environment registration result, and the discrete control variables quantified by the flight control computer, and send the decision instruction data to an independently set control execution unit to generate control execution data; the decision execution module is also used to update the hypergraph data according to the control execution data and the execution feedback data for the next cycle call of the fusion twin module.

[0103] The decision execution module is located downstream of the information chain of the multi-source fusion avionics control system. Its function is to quickly convert the discrete control variables, the twin environment data, and the twin environment registration result into control execution data, and feed back the execution closed-loop information to the hypergraph data, providing the latest maneuver-environment coupling prior for the next round of the fusion twin module. The module as a whole consists of a global search unit, a local optimization unit, a constraint evaluation unit, and an event write-back unit. Among them, the global search unit is responsible for finding the seed of the feasible control sequence in the large-scale combination space; the local optimization unit refines the control sequence in the continuous time domain; the constraint evaluation unit checks the safety boundary in real time; the event write-back unit writes the execution feedback into the hypergraph data to realize the self-evolution of the model.

[0104] The global search unit, the discrete control variables are represented by binary quantities such as rudder surface deflection symbols and thrust adjustment gears. Denote the control sequence of length as the vector . To avoid exhaustive search in the exponential combination space, the present invention uses a quantum annealer to solve the Ising model:

[0105] where is the spin variable, taking the value +1 indicates that the control quantity is activated, and -1 indicates that it is closed; the matrix Stored series and parallel coupling relationships, vectors Store single-spin energy bias. Each coupling coefficient is dynamically calculated from the twin environment data and the registration result. For example, when the aerodynamic load is too high, the positive coupling coefficient of the same vector of adjacent control surfaces is increased to suppress excessive deflection. After a microsecond-level annealing, the quantum annealer outputs the spin configuration with the lowest energy, and the control sequence seed is obtained through transcoding.

[0106] Local optimization unit. The quantum annealer gives a discrete approximate solution, and it still needs to satisfy the aerodynamic and structural differential constraints in the continuous time domain. The local optimization unit uses a pulsed symplectic structure strategy network to perform energy-conserving iteration on the control sequence seed. Define the Hamiltonian:

[0107] is the generalized coordinate, including attitude and position; is the conjugate momentum; is the mass-inertia matrix; the potential energy is given by the twin environment data and the registration result, reflecting the instantaneous aerodynamic and electromagnetic coupling loads. The pulsed symplectic strategy network uses the velocity-type symplectic-Verlet format to iteratively update , making a small correction to the control sequence while ensuring the numerical conservation of the Hamiltonian. The gradient during the iteration is obtained by automatic differentiation, and the network weights are transmitted in a pulsed manner on the on-chip neural chip, with a single-step delay less than 0.2 milliseconds.

[0108] Constraint evaluation unit. The system models the safety boundary as inequality constraints, such as overload and angle of attack . The constraint evaluation unit calculates the penalty function at the end of the pulsed symplectic iteration:

[0109] where the symbol represents the positive part. The Lagrange multipliers and are adaptively updated online. When the penalty exceeds the threshold, the unit triggers constraint projection, truncating the excessive control surface deflection or reducing the thrust command amplitude. The constraint projection keeps the system within the controllable envelope and writes the correction information to the feedback port of the local optimization unit for rapid convergence.

[0110] Event write-back unit. The control instruction data is sent to the control execution unit through the avionics data bus to generate control execution data. After execution, the flight control computer transmits the attitude, load, and self-diagnostic information back every 20 milliseconds, which is called the execution feedback data. The event write-back unit calculates the difference between the control execution data and the execution feedback data. If the difference norm exceeds the preset threshold, it is considered that there is a dynamic model deviation or external interference, and the hypergraph data needs to be updated. The write-back rule is divided into two parts: If the difference is mainly concentrated in the radar-related control surfaces, increase the hyperedge weight between the radar node and the inertial node to enhance the sensitivity of subsequent filtering to radar anomalies; if the difference mainly comes from the pitch-thrust channel, increase the self-loop weight of the inertial node to emphasize the credibility of inertial measurements.

[0111] The updated hypergraph data is read by the fusion twin module in the next cycle, enabling the physical field generation to have an adaptive ability to execution-observation errors.

[0112] In an embodiment, in a continuous 90-second figure-eight maneuver flight test, the system is closed-loop at a rate of 50 Hz. The average search time of quantum annealing is 0.4 ms, the time-consuming of pulse symplectic optimization is 0.9 ms, the time-consuming of constraint evaluation is 0.1 ms, and the overall decision-making delay is 1.4 ms. Compared with the benchmark algorithm (integer-order Kalman-gradient descent optimization), the attitude error is reduced by 40%, and the peak-to-peak value of the control surface is reduced by 18%, indicating that distributed search and symplectic conservation optimization can reduce the actuator load while maintaining trajectory accuracy.

[0113] In another set of anti-interference experiments, external radar glint interference is injected while disturbing the center of pressure. The decision execution module uses the registration result to timely increase the weight of the electromagnetic-inertial coupling term, enabling the control law to suppress unnecessary large control surface responses. The test records show that the maximum roll rate suppression of the system is 60% higher than that of the benchmark algorithm, indicating that the combination of entropy-regularized registration and hypergraph write-back improves the robustness to environmental noise.

[0114] For hardware resources, the quantum annealing interface is directly connected to the annealing controller through PCIe; the pulse symplectic policy network runs on the Loihi-II neural chip, occupying 3 clusters of cores; the constraint evaluation and event write-back logic are deployed on the field-programmable gate array side, with a resource occupancy of about 12%. The entire module consumes 4.8 watts under full-load extreme maneuvers.

[0115] In summary, the decision execution module forms a fast adaptive closed-loop of decision-execution-feedback through global search of quantum annealing, local optimization of symplectic conservation, online constraint projection, and hypergraph event write-back; this closed-loop can not only handle large discrete combination spaces but also meet continuous physical constraints. It shows high trajectory accuracy and actuator life advantages in extreme maneuver, radar interference, and payload mutation scenarios, which is one of the key innovations of the multi-source fusion avionics control system of the present invention.

[0116] Preferably, the decision execution module constructs the discrete control variables into an Ising model and solves the Ising model on a quantum annealing processor to obtain the control sequence seed.

[0117] The decision execution module undertakes the global discrete decision search task. The present invention adopts the quantum annealing technology, maps the discrete control variables in the time domain to the Ising model, obtains the control sequence seed through the quantum annealing processor, and then hands it over to the subsequent pulse symplectic policy network for continuous domain refinement. The discrete control variables refer to the binary encoding of actions such as the opening and closing of rudder surface deflection and the increase and decrease of thrust gears within a fixed prediction time domain. For example, for the next 5 control steps, the left and right aileron deflection directions of each step are represented by binary symbols respectively, and a bit string with a length of 10 is formed in total. In order to utilize the parallel search ability of quantum annealing, the bit string is further transcribed into a spin vector , where the element values are ±1, and the additive cost function is converted into the Ising energy form:

[0118] is the spin vector; is the symmetric coupling matrix, and its element represents the influence on the task index when the actions of the th bit and the th bit are activated or cancelled simultaneously; is the linear bias vector, and its element represents the cost or benefit of activating the th bit action alone. The coupling coefficient is generated by comprehensively scoring the aerodynamic load distribution, electromagnetic cross-section change and structural stress margin provided by the twin environment data: if the cooperation of the two-bit actions can reduce the total load, the corresponding is negative; if the two-bit actions conflict, is positive. The bias vector is obtained by linearly combining the trajectory error and the fuel consumption index.

[0119] In terms of hardware implementation, the decision execution module transmits the matrix and the vector to the quantum annealing controller through the PCIe bus. The internal physical qubit layout of the controller is limited, and it needs to go through a mapping process to embed the complete graph into the sparse quantum coupling graph. This step uses a heuristic mapping algorithm to automatically allocate the chain strength to avoid energy degradation caused by the breakage of the spin chain. The annealing process is divided into three stages: initialization, linear annealing and readout. In the initialization stage, an equal-amplitude transverse field is loaded into each qubit; in the linear annealing stage, the transverse field is gradually reduced while the Ising Hamiltonian is amplified; in the readout stage, the spin directions of all qubits are measured to obtain the sample . The present invention takes 10 independent annealing samples and selects the one with the minimum energy as the control sequence seed. The single annealing takes about 0.25 milliseconds, and the total search time is on the order of 0.5 milliseconds, which is much lower than the more than ten milliseconds of traditional simulated annealing for the same-scale problem.

[0120] To ensure the feasibility of the output sequence, the decision execution module performs a quick legality correction after sampling. The correction rules are provided by the constraint evaluation unit: for example, two consecutive reverse deflections of the ailerons will generate excessive instantaneous torque and need to insert a neutral position; another example is that the number of uses of the maximum thrust gear is restricted by the upper limit of the oil temperature. If the limit is exceeded, the gear with the minimum energy gain will be spin-flipped. After correction, the spin vector is re-encoded into a binary control string and handed over to the local optimization unit.

[0121] Example: In a high-maneuver task with a turning radius of 200 meters and a load of 8g, the decision execution module quantifies a 5-step control sequence seed every 50 milliseconds, that is, a total of 10 bits for the roll direction, pitch increment, and thrust increase and decrease. The average energy of the seeds output by quantum annealing is 28% lower than that of the heuristic gradient climbing method. After impulse symplectic optimization, the trajectory error is controlled within 0.3 meters, while the heuristic method is 0.45 meters. The peak deflection of the control surface is reduced by 12%, and the peak thrust is reduced by 9%, proving that quantum annealing provides a more optimal starting point for subsequent local optimization in the global search stage.

[0122] The benefits of the module are reflected in two aspects: First, quantum annealing can find a discrete sequence close to the global optimum in the case of multi-objective coupling, reducing the probability of the local optimization falling into a suboptimal solution; second, the hardware annealing time grows sublinearly with the problem scale, leaving room for the expansion of a longer prediction time domain and more control channels. Combined with the event write-back mechanism, when the execution feedback shows that the body response is inconsistent with the model prediction, the hypergraph edge weights are adjusted in real time, and the subsequent and will be updated immediately, making quantum annealing tend to generate new sequences that can quickly converge the error in the next cycle.

[0123] Preferably, the decision execution module inputs the control sequence seed, twin environment data, and twin environment registration result into the impulse symplectic structure policy network, optimizes the control sequence seed to generate decision instruction data, and writes the control execution data and the corresponding execution feedback data into the hypergraph data after obtaining the control execution data for the next cycle call of the fusion twin module.

[0124] After receiving the control sequence seed output by quantum annealing, the decision execution module integrates the sequence with the twin environment data and the twin environment registration result into three types of input tensors of the policy network. The twin environment data provides the spatial distribution of aerodynamic load, electromagnetic scattering, and structural stress on a three-dimensional finite element grid; the twin environment registration result indicates the local difference between the model and the measurement in the form of a registration potential; the control sequence seed gives an initial guess of the discrete action. In order to simultaneously meet the requirements of fast convergence and strict physical constraints, the present invention proposes an impulse symplectic structure policy network. This network explicitly embeds the flight dynamics in the form of a Hamiltonian, takes the neural impulse pair as conjugate variables, and realizes policy optimization through symplectic conservation numerical integration.

[0125] The Hamiltonian of the network core is written as:

[0126] where is the generalized coordinate tensor, including the roll angle, pitch angle, yaw angle, and three-axis positions; is the conjugate momentum tensor, corresponding to the membrane potential in the pulsed neural implementation; is the mass-inertia matrix, a constant parameter; is the potential energy function, explicitly dependent on the control sequence seed tensor and the registration potential . Setting the time step to , the network is updated once using the velocity-type symplectic-Verlet scheme:

[0127]

[0128]

[0129] The gradient is calculated by the automatic differentiation engine, whose input includes the local pressure gradient and scattering gradient of the twin environment data, enabling the control instructions to balance aerodynamic efficiency and electromagnetic observability. The neural pulses are stored in fixed-point sixteen-bit format, and the equivalent value is adjusted through gated synaptic weights to ensure the conservation of physical mass. Each symplectic integration cycle iterates three steps, and the number of iterations is automatically terminated when the residual converges to the threshold, usually not exceeding five times.

[0130] For mapping to hardware, the network is deployed on a neural chip with a pulsed timing mechanism. The pulse width is fixed at twenty nanoseconds, and a single-step symplectic operation is completed in twenty pulse cycles; the complete optimization outputs decision instruction data within one millisecond. The decision instruction data is written into the control execution unit through the avionics data bus, and the control execution unit decodes it into analog quantities such as aileron angles and thrust gears within two milliseconds to form control execution data.

[0131] In the closed-loop feedback stage, the flight control computer composes the execution feedback data from the attitude and acceleration obtained by real-time sampling of the sensors, and attaches the execution time tag. The decision execution module calculates the difference between the control execution data and the execution feedback data to obtain the error vector . If exceeds the dynamic threshold, the module triggers event backwriting, mapping the error to the hypergraph nodes: When the error accumulates in the roll angle channel, indicating insufficient response of the left and right ailerons, the module increases the self-loop weight of the aileron node; when the error accumulates in the pitch acceleration and the radar scattering intensity rises abnormally at the same time, the module increases the mutual information weight between the radar node and the pitch node.

[0132] The hypergraph data is updated in-situ without additional replication. Subsequently, the fusion twin module reads the new weights in the next cycle, corrects the priority of three-dimensional field generation, and realizes the self-evolution of perception-decision-making.

[0133] Embodiment: In a low-altitude sharp turn mission, the length of the control sequence seed is ten. After the pulse symplectic network is optimized, the decision instruction reduces the trajectory radius of the aircraft body to 98% of the original plan, while reducing the peak value of the control surface by twelve degrees. A total of Figure 3 times of hypergraph data is written back within two seconds of the closed loop. The mutual information weights between the radar and inertia change dynamically by twenty percentage points, suppressing the radar scattering peak in real time during the subsequent maneuvering section and reducing the enemy detection probability. Compared with the non-symplectic conservation scheme, the course error of the system is reduced by 35%, and the pitch overshoot is reduced by 20%.

[0134] Power consumption and resources: The neural chip occupies 40% of the core, with a peak power consumption of two watts; the event write-back logic occupies 16% of the look-up table in the programmable gate array, with a power consumption of 0.3 watts. Overall, it meets the airborne power margin.

[0135] By fusing the Hamiltonian, pulsed neurons, and hypergraph back-coupling, the decision execution module can not only output high-quality control instructions that meet hard constraints but also adjust the perception map in real time according to the execution error, reflecting the innovative effect of the present invention in the data-physical double closed-loop optimization.

[0136] As Figure 4 shown, a multi-source fusion avionics control method is applied to the multi-source fusion avionics control system, including the following steps: S1. Synchronously collect multi-source sensor data under a unified time reference. The multi-source sensor data at least includes inertial measurement data, star sensor data, and radar data, and perform compression processing on the multi-source sensor data to generate compressed observation data; S2. Generate fusion state data and hypergraph data based on the compressed observation data; S3. Generate twin environment data and twin environment registration results according to the fusion state data and the hypergraph data, and output the twin environment data and the twin environment registration results; S4. Generate decision instruction data based on discrete control variables, the twin environment data, and the twin environment registration results, and send the decision instruction data to the control execution unit to generate control execution data; S5. Write the control execution data and execution feedback data into the hypergraph data for the next cycle to execute step S2.

[0137] S1 acquires multi-source data such as inertial measurement, star sensor, and radar simultaneously under a unified time base, and performs dimensionality reduction and compression in parallel to form compressed observation data in a unified format. S2 uses the compressed observation data for attitude-position fusion estimation, and simultaneously constructs a hypergraph reflecting multi-source correlation to obtain fused state data and hypergraph data. S3 relies on the fused state and hypergraph data to generate aerodynamic, electromagnetic, and structural fields in real time on a three-dimensional grid, performs entropy-regularized registration by comparing the observation density, and outputs the corrected twin environment data and registration results. S4 combines discrete control variables, twin environment data, and registration results, generates decision instructions through global-local joint optimization, and sends them to the control execution unit to generate actual control quantities. S5 collects control execution data and flight feedback, writes the difference between the two into the hypergraph data, drives the next round of state fusion and environment reconstruction, and realizes the self-evolving closed loop of perception-decision-making.

[0138] The above are only embodiments of the present application and are not used to limit the present application. For those skilled in the art, various changes and modifications can be made to the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall be included within the scope of the claims of the present application.

Claims

1. A multi-source fusion avionics control system, characterized in that, Comprising: A multi-source acquisition and compression module, configured to synchronously acquire multi-source sensor data including at least inertial measurement data, star sensor data, and radar data under a unified time reference, and perform compression processing on the multi-source sensor data to generate compressed observation data; A fusion twin module, configured to generate fusion state data and hypergraph data based on the compressed observation data, generate twin environment data and a twin environment registration result according to the fusion state data and the hypergraph data, and output the twin environment data and the twin environment registration result; A decision execution module, configured to generate decision instruction data according to the twin environment data, the twin environment registration result, and discrete control variables quantified by a flight control computer, and send the decision instruction data to an independently provided control execution unit to generate control execution data; the decision execution module is further configured to update the hypergraph data according to the control execution data and execution feedback data for the next cycle call of the fusion twin module.

2. The system according to claim 1, wherein The multi-source sensor data synchronously acquired by the multi-source acquisition and compression module includes inertial measurement data, star sensor data, synthetic aperture radar data, short-wave infrared optoelectronic array data, global navigation satellite system receiver data, and meteorological sounding data link data, and after completing time reference alignment, performs linear mapping on the multi-source sensor data through a random projection matrix to generate compressed observation data.

3. The system according to claim 2, characterized in that, Before performing linear mapping on the multi-source sensor data, the multi-source acquisition and compression module whitens the multi-source sensor data using covariance matrices corresponding to various types of acquired data.

4. The system according to claim 1, wherein The fusion twin module obtains ternary joint mutual information by calculating the difference between the Shannon entropy and the joint entropy of three groups of node observations, and determines hyperedge weights based on the ternary joint mutual information to construct hypergraph data.

5. The system according to claim 4, characterized in that, The fusion twin module processes the attitude state vector using a fractional-order Kalman filtering method on the Euclidean motion group manifold, and uses the filtering result as the fusion state data.

6. The system according to claim 1, wherein The fusion twin module generates aerodynamic field data, electromagnetic scattering field data, and structural stress field data on a three-dimensional finite element grid using the fusion state data and the hypergraph data, and the three types of field data share the node coordinates and element topologies of the same spatial grid.

7. The system according to claim 6, characterized in that, The fusion twin module uses an entropy-regularized optimal transport algorithm to register the electromagnetic scattering field data and the radar observation density obtained from the compressed observation data to generate a twin environment registration result.

8. The system according to claim 1, wherein The decision execution module constructs the discrete control variables into an Ising model, and solves the Ising model on a quantum annealing processor to obtain a control sequence seed.

9. The system according to claim 8, wherein The decision execution module inputs the control sequence seed, the twin environment data, and the twin environment registration result into a pulse symplectic structure policy network, optimizes the control sequence seed to generate decision instruction data, and after obtaining the control execution data, writes the control execution data and the corresponding execution feedback data into the hypergraph data for the next cycle call of the fusion twin module.

10. A multi-source fusion avionics control method, applied to the multi-source fusion avionics control system according to any one of claims 1-9, characterized in that, Including the following steps: S1. Synchronously collect multi-source sensor data under a unified time reference. The multi-source sensor data at least includes inertial measurement data, star sensor data, and radar data. Perform compression processing on the multi-source sensor data to generate compressed observation data; S2. Generate fusion state data and hypergraph data based on the compressed observation data; S3. Generate twin environment data and twin environment registration results according to the fusion state data and the hypergraph data, and output the twin environment data and the twin environment registration results; S4. Generate decision instruction data based on discrete control variables, the twin environment data, and the twin environment registration results, and send the decision instruction data to a control execution unit to generate control execution data; S5. Write the control execution data and execution feedback data into the hypergraph data for the next cycle to execute step S2.

Images