A method for adaptively constructing an injury probability model by fusing knowledge transfer

Abstract

The application provides a damage probability model adaptive construction method fusing knowledge migration, relates to the technical field of control and regulation, and through construction of a multi-physical-field-fused characteristic index, introduction of physical information guided knowledge migration and establishment of a prediction-control integrated linkage mechanism, not only the accuracy and reliability of dynamic instability prediction are remarkably improved, but also the prediction result is directly applied to active safety control, a seamless closed loop from intelligent perception to active intervention is realized. Under the premise of not increasing hardware cost, through software algorithm upgrading, the safe flight envelope of the aircraft is effectively widened, and the dynamic robustness of the system in a complex extreme environment is enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of control regulation technology, in particular to a damage probability model adaptive construction method fusing knowledge transfer. BACKGROUND

[0002] With the development of modern equipment systems towards higher performance and more intelligentization, the coupling between internal physical subsystems is increasingly close, and the operating environment is also increasingly complex. Therefore, in order to ensure the safety and reliability of equipment under extreme operating conditions, state monitoring and health management technology is evolving from traditional damage accumulation evaluation based on a single physical quantity to dynamic risk prediction based on multi-physical field information fusion. How to extract precursor features that can accurately represent the dynamic stability of the system from complex and heterogeneous data streams, and build a high-performance prediction model that can run in real time on resource-constrained platforms, has become a frontier challenge and core development trend in this field.

[0003] Although the existing technology has made some progress in the field of structural health monitoring, it still faces the following core challenges in dealing with the dynamic instability problems of complex equipment such as high-performance helicopters:

[0004] Limitations of feature representation: existing monitoring methods usually rely on single-dimensional sensor data such as vibration and strain. However, dynamic instability is the result of highly nonlinear coupling between external aerodynamic excitation and internal structural response. Single-dimensional data is difficult to effectively capture the inherent mechanism of such cross-domain coupling, resulting in extracted features that are not sensitive to instability precursors, and unable to provide sufficient advance warning before the system enters a critical state, with a technical bottleneck of warning lag.

[0005] Limitations of prediction model generalization ability: existing prediction model construction mostly adopts pure data-driven artificial intelligence methods. For example, the technical solution disclosed in Chinese patent CN112229585A, although it uses transfer learning to improve the performance of the model on a specific task, its essence still relies on existing data distribution. For dynamic instability, which occurs on the boundary of the flight envelope where training data is sparse or even missing, pure data-driven models lack the constraints of physical laws, and their prediction behavior in the extrapolation region is often unreliable, posing a risk of catastrophic misjudgment, which limits their application in safety-critical fields.

[0006] The disconnect between prediction and control: Current health monitoring systems mostly stop at generating early warning signals, with an information barrier between their output and the flight control system. This "sensing first, responding later" disconnect means that valuable predictive information fails to be translated into effective mitigation actions in a timely manner, missing the optimal window for intervention in the early stages of instability. Therefore, how to directly and seamlessly transform high-precision predictive capabilities into online active safety control capabilities for aircraft is a direction that needs further improvement in existing technologies.

[0007] The information disclosed in the background section above is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0008] The purpose of this invention is to provide an adaptive construction method for a damage probability model that integrates knowledge transfer, so as to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] An adaptive construction method for a damage probability model incorporating knowledge transfer, comprising the following steps:

[0011] S1: Acquire two types of heterogeneous real-time data streams related to key structural parts of the helicopter. The two types of heterogeneous real-time data streams include: a first data stream characterizing external aerodynamic energy input, and a second data stream characterizing the fidelity of information transmission in the internal control loop.

[0012] S2: Based on the time-delay cross-correlation between the first data stream and the second data stream, dynamically calculate the fusion characteristic index of the efficiency of the key structural parts in converting external aerodynamic energy input into internal information transmission loss.

[0013] S3: Construct a high-precision teacher model, wherein the fusion feature index is used as the core input feature to train the teacher model to learn and characterize the dynamic instability critical precursors of the key structural parts;

[0014] S4: Construct a student model with a lower structural complexity than the teacher model;

[0015] S5: Deploy the student model on the airborne computing unit and use the student model to process flight data, including the fused feature index, in real time to generate a dynamic instability warning signal, and further convert the dynamic instability warning signal into a suppressive bias command for the flight control system.

[0016] Compared with the prior art, the beneficial effect of the present invention is that by constructing a frequency-resolved fusion feature index that can deeply integrate the external aerodynamic energy input and the internal information transmission fidelity, the problem of insufficient information representation capability of a single physical dimension is fundamentally solved.

[0017] To address the challenge of low reliability of prediction models in the extrapolation region, a teacher model based on the Physical Information Neural Network (PINN) was constructed. Prior physical knowledge in the field of aeroelasticity was incorporated as a strong constraint into the model training. Combined with an adaptive knowledge distillation strategy, a student model with both high accuracy and lightweight design was generated while ensuring the consistency of predicted physics.

[0018] Ultimately, this invention breaks through the barriers between prediction and control, establishing a predictive-control integrated linkage mechanism that directly transforms the "predictive insight" of the student model into the "active execution" of the flight control system, achieving a fundamental shift from "passive alarm" to "active suppression". Attached Figure Description

[0019] Figure 1 This is a schematic diagram illustrating the overall application logic of the method of the present invention;

[0020] Figure 2 This is a schematic diagram illustrating the execution logic of steps S1 and S2 of the present invention;

[0021] Figure 3 This is a schematic diagram illustrating the execution logic of steps S3 and S4 of the present invention;

[0022] Figure 4 This is a schematic diagram illustrating the execution logic of step S5 of the present invention. Detailed Implementation

[0023] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0024] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0025] Example 1:

[0026] Please see Figure 1 The present invention provides a technical solution:

[0027] An adaptive construction method for a damage probability model incorporating knowledge transfer includes the following steps:

[0028] S1: Acquire two types of heterogeneous real-time data streams related to key structural parts of the helicopter. The two types of heterogeneous real-time data streams include: a first data stream characterizing external aerodynamic energy input, and a second data stream characterizing the fidelity of information transmission in the internal control loop.

[0029] S2: Based on the time-delay cross-correlation between the first data stream and the second data stream, dynamically calculate the fusion characteristic index of the efficiency of the key structural parts in converting external aerodynamic energy input into internal information transmission loss.

[0030] S3: Construct a high-precision teacher model, wherein the fusion feature index is used as the core input feature to train the teacher model to learn and characterize the dynamic instability critical precursors of the key structural parts;

[0031] S4: Construct a student model with a lower structural complexity than the teacher model;

[0032] S5: Deploy the student model on the airborne computing unit and use the student model to process flight data, including the fused feature index, in real time to generate a dynamic instability warning signal, and further convert the dynamic instability warning signal into a suppressive bias command for the flight control system.

[0033] Further explanation: Acquire a first data stream representing the external aerodynamic energy input and a second data stream representing the internal information transmission fidelity; perform multi-resolution decomposition on the first and second data streams to generate multiple pairs of frequency band sub-sequences; and for each pair of frequency band sub-sequences, independently calculate the fusion feature components, and finally combine them into a frequency-resolved ASII spectrum vector represented by the "fusion feature spectrum vector".

[0034] The first data stream is a time series of boundary layer turbulence dissipation rate, and the second data stream is a time series of flight control command-propeller hub vibration transfer entropy.

[0035] Further explanation: Multi-resolution decomposition is achieved by employing the discrete wavelet transform algorithm;

[0036] The specific steps for calculating the fusion feature components are as follows: performing time-delay cross-correlation analysis on each pair of frequency band sub-sequences and identifying the maximum cross-correlation value within the preset physical time delay range as the fusion feature component of that frequency band sub-sequence;

[0037] Before the time-delay cross-correlation analysis, the frequency band sub-sequence is further subjected to Z-score normalization.

[0038] Before combining them into the final fused feature spectrum vector, the calculated fused feature components are mapped using a hyperbolic tangent function. The following is a detailed implementation description of the above:

[0039] The core innovation of this embodiment lies in performing multi-resolution analysis on the original heterogeneous data stream, independently calculating the fusion characteristics of each in multiple key frequency bands, and ultimately forming a characteristic spectrum that can reveal the coupling characteristics of the system under different excitation frequencies. Specifically, the time series of boundary layer turbulence dissipation rate, which characterizes the external high-frequency aerodynamic excitation, and the time series of flight control command-propeller hub vibration transfer entropy, which characterizes the degradation of the internal overall system response, are acquired in parallel. Before the fusion calculation, discrete wavelet transform is introduced to decompose the above two time series, obtaining frequency band subsequences corresponding to different physical phenomena such as low-frequency maneuvering, mid-frequency structural modes, and high-frequency turbulence. Subsequently, for each pair of corresponding frequency band subsequences, its aerodynamic-structural information impedance (ASII) is calculated independently, thereby generating a frequency-resolved ASII spectrum vector. The "frequency-resolved ASII spectrum vector" enables the subsequent fault prediction model not only to sense the changes in the overall coupling strength of the system, but also to accurately identify which specific frequency excitation and system response caused dangerous nonlinear resonance, thereby improving the sensitivity and diagnostic accuracy of capturing frequency-sensitive dynamic instability events such as aeroelastic flutter.

[0040] The key parameters involved in this embodiment are defined as follows. All parameters are generated in real time by the onboard computing unit according to a preset program.

[0041] The dynamic pressure sequence on the rotor blade surface is denoted as This represents a multi-channel time series, with each channel representing the instantaneous value of dynamic pressure at a specific chord length position on a rotor blade. It is acquired through an array of MEMS (Micro-Electro-Mechanical Systems) piezoresistive pressure sensors embedded at key chordal positions (25%, 50%, and 75% chord length) on the leading edge and upper surface of one or more rotor blades. The sensors are aerospace-grade products with high-frequency response (≥10kHz) and high accuracy (≤0.1%FS). The acquisition system synchronously acquires pressure signals from all channels at a sampling frequency of at least 20kHz to form the original dynamic pressure sequence.

[0042] The flight control collective pitch command sequence is denoted as It represents a single-channel time series, which is the instantaneous value of the control command output by the flight control computer to the main rotor collective pitch actuator; it is obtained by listening to and decoding the airborne ARINC-429 or MIL-STD-1553B data bus.

[0043] The triaxial vibration acceleration sequence of the propeller hub is denoted as... It represents a three-channel time series, which respectively represent the instantaneous values ​​of longitudinal, lateral, and vertical vibration accelerations measured by accelerometers mounted on the main rotor hub; acquired by a triaxial piezoelectric accelerometer rigidly mounted on the non-rotating component of the main rotor hub.

[0044] Boundary layer turbulent dissipation rate, denoted as It represents a single-channel time series, characterizing the rate at which the turbulent kinetic energy of a unit mass of fluid is converted into internal energy due to viscosity within the airflow boundary layer near the blade surface; it is indirectly estimated through a linguistic computational model based on Taylor's frozen turbulence assumption; the specific calculation logic is as follows:

[0045] 1.11) Dynamic pressure sequence from the rotor blade surface The channel with the highest energy was selected, and a high-pass filter was applied to it to remove the quasi-static pressure component, thus obtaining the pressure pulsation sequence. .

[0046] 1.12) Based on the linearized form of Bernoulli's equation, the pressure pulsation sequence is... Convert to velocity pulsation sequence The conversion relationship is that velocity fluctuations are directly proportional to pressure fluctuations and inversely proportional to the product of average airflow density and average flow velocity.

[0047] 1.13) For velocity pulsation sequences Calculate its first-order time difference to obtain the time derivative sequence of velocity fluctuations. It should be noted that in this embodiment, "t" represents the discrete time step index in the time series;

[0048] 1.14) According to Kolmogorov's turbulence theory, the turbulence dissipation rate is proportional to the variance of the velocity spatial gradient. Based on Taylor's frozen turbulence assumption, the spatial gradient is replaced by the time derivative. Finally, through calculation... The variance of the sequence within the sliding time window, multiplied by the calibration coefficient. The boundary layer turbulent dissipation rate was obtained. The estimated value. This calibration coefficient. The determination method is as follows: During the offline experimental calibration phase, measurements are taken simultaneously in a wind tunnel for a specific propeller type using both this method and a high-precision laser Doppler velocimeter. The calibration coefficients that minimize the difference between the results calculated by this method and the direct measurement results from the high-precision laser Doppler velocimeter are determined through least-squares fitting. value.

[0049] Calibration coefficients in this embodiment The variance of the velocity pulsation time derivative sequence within the current window is 1.5. Then the boundary layer turbulent dissipation rate at that moment Value .

[0050] Flight control command - rotor hub vibration transfer entropy, denoted as It represents a single-channel time series, quantized from the flight control collective pitch command sequence. To the propeller hub triaxial vibration acceleration sequence The strength of causal information flow is obtained through a linguistic computational model based on the Kraskov-Stegbauer-Glasberg (KSG) estimator; the specific computational logic is as follows:

[0051] 1.21) Flight control collective pitch command sequence and the triaxial vibration acceleration sequence of the propeller hub Perform time delay embedding to construct a high-dimensional state vector; embedding dimension m1 and time delay. Pre-determined using the pseudo-nearest neighbor method and the average mutual information method;

[0052] 1.22) For each point i in the state space, in the highest-dimensional joint space consisting of all state vectors, use Euclidean distance or Chebyshev distance to find the k-th nearest neighbor point to that point, and record the distance. In this embodiment, k is 4.

[0053] 1.23) In the distance Within the defined hypercube, count the number of neighboring points that fall into each of the reduced-dimensional edge spaces;

[0054] 1.24) Substitute the neighbor counts obtained in step 1.23) into a mathematical expression containing the digamma function for summation and averaging. This expression originates from the derivation of the relationship between mutual information and neighborhood statistics in information theory, and its output is the flight control command-propeller hub vibration transfer entropy. The estimated value.

[0055] If, in this embodiment, the calculation results show that, given the current and past flight control commands, the uncertainty in predicting the rotor hub vibration state at the next moment can be reduced from the original 5 bits to 3.5 bits, then the flight control command-rotor hub vibration transfer entropy... The value is 1.5 bits per second;

[0056] Frequency-resolved ASII spectral vector, denoted as It represents a multidimensional vector, where each component represents the aerodynamic-structural information impedance (ASII) value calculated within a specific frequency band. The value range of each component in the "frequency-resolved ASII spectrum vector" is normalized using a hyperbolic tangent function and limited to the range (-1, 1).

[0057] For the above content, perform the following complete calculation process:

[0058] 1.31) In each calculation cycle, obtain the latest segment of the dynamic pressure sequence on the rotor blade surface. Flight control collective pitch command sequence and the triaxial vibration acceleration sequence of the propeller hub Ensure that the timestamps of the data stream are aligned.

[0059] 1.321) Based on the dynamic pressure sequence of rotor blade surface Calculate the boundary layer turbulent dissipation rate Time series.

[0060] 1.322) Based on flight control collective pitch command sequence and the triaxial vibration acceleration sequence of the propeller hub Calculate the entropy of the flight control command-propeller hub vibration transfer. Time series.

[0061] 1.331) The discrete wavelet transform algorithm is adopted, using the Daubechies-4 wavelet basis, to analyze the boundary layer turbulent dissipation rate. The time series is decomposed into three levels to obtain sub-band detail coefficient sequences for low frequency, mid frequency, and high frequency: , , .

[0062] 1.332) Using the same wavelet transform settings, the transfer entropy of flight control commands and rotor hub vibration is calculated. The time series is decomposed to obtain the corresponding sub-band detail coefficient sequence: , , .

[0063] 1.341) for low-frequency subsequence pairs The fusion algorithm is executed to calculate the low-frequency components. .

[0064] 1.342) for mid-frequency subsequence pairs The fusion algorithm is executed to calculate the mid-frequency band components. .

[0065] 1.343) for high-frequency subsequence pairs The fusion algorithm is executed to calculate the high-frequency band components. .

[0066] 1.35) The final output combines the three components to form a frequency-resolved ASII spectrum vector. And use it as the core input feature of the fault prediction model.

[0067] This embodiment employs time-delay cross-correlation analysis within each frequency band; it can directly quantify the linear causal coupling strength between excitation and response. The innovation of this embodiment lies in applying this classical method to specific physical frequency bands after wavelet decomposition, enabling the fusion analysis to focus on specific physical phenomena such as structural modal resonance, thereby achieving diagnostic accuracy. The following are the detailed steps for calculating the ASII component of any frequency band:

[0068] Calculate the reciprocal of the transfer entropy subsequence to obtain the information loss subsequence. To avoid division by zero, the entropy transferred between flight control commands and rotor hub vibration is reduced. Set a positive lower limit that is not less than the machine's precision.

[0069] Within the computation window, Z-score normalization is performed on the dissipation rate subsequence and the information loss subsequence to obtain the standardized sequence. and .

[0070] At the preset maximum physical delay time lag Within the range, calculate the normalized sequences for all discrete time delays. The following are the cross-relationship values; the maximum physical delay in this embodiment. The range is 0 to 100 milliseconds;

[0071] Identify the calculated cross-correlation function in The maximum value within the time delay range; this maximum value is the strongest linear coupling strength between external excitation and internal response degradation within this frequency band.

[0072] The obtained maximum cross-correlation value is mapped using the hyperbolic tangent function, strictly limiting the output value to the (-1, 1) interval, and the strongly correlated signal is nonlinearly amplified. This mapping result is the ASII component value for this frequency band. The following is a detailed explanation of the implementation:

[0073] The core output of this embodiment is a frequency-resolved ASII spectrum vector. The value range of each component, after being mapped by the hyperbolic tangent function, is strictly limited to the interval (-1, 1). When the ASII component value of any frequency band subsequence approaches 1, it indicates a strong positive correlation between external aerodynamic energy input and information transmission loss in the internal control loop within that specific frequency band. This reveals a dangerous physical state: the structural system exhibits efficient energy-damage conversion characteristics at this frequency. That is, minute external excitations can be efficiently "amplified" by the structure and converted into a significant decrease in system control fidelity. This state is a highly credible precursor to dynamic instability events such as aeroelastic flutter.

[0074] When the ASII component value of any frequency band subsequence approaches 0, it indicates that there is no significant linear correlation between external aerodynamic energy input and internal information transmission loss within that specific frequency band. This represents a healthy or robust system state. The structure can effectively dissipate or withstand external energy impacts without significantly affecting its stability as an information transmission channel. This suggests that the system is far from any dangerous resonance or coupling regions within that frequency band.

[0075] The final output frequency-resolved ASII spectrum vector The values ​​of each component are mainly affected by the dynamic characteristics of the time series of two upstream input parameters: boundary layer turbulent dissipation rate. And flight control commands - rotor hub vibration transfer entropy .

[0076] When other conditions remain unchanged, the boundary layer turbulent dissipation rate An increase in the average amplitude or pulsation intensity of a time series is positively correlated with the ASII component values ​​of the corresponding frequency band.

[0077] The core of calculating the ASII component is the cross-correlation function, which is essentially the expectation of the product of two signals. When the boundary layer turbulent dissipation rate... When a signal is enhanced, if there is a potential positive correlation between it and the information-loss signal, the expected value of their product will also increase. This represents a higher "boundary layer turbulence dissipation rate" corresponding to stronger external aerodynamic excitation. "This is a necessary condition for stimulating and exposing the potential weaknesses of the structure."

[0078] When other conditions remain unchanged, the entropy of the flight control command-propeller hub vibration transfer is... The decrease in the average value of the time series is positively correlated with the ASII component value of the corresponding frequency band, i.e., the entropy of the flight control command-propeller hub vibration transfer. It is negatively correlated with ASII; when the flight control command-propeller hub vibration transfer entropy A decrease in efficiency indicates a reduction in the information transmission efficiency from control commands to structural responses, leading to a deterioration in system health; its reciprocal, representing increased information loss, further contributes to this. This is based on the boundary layer turbulent dissipation rate. The same cross-correlation logic applies, but increased information loss will lead to an increase in the ASII component value. This design accurately maps to real-world logic: a system with impaired internal information transmission pathways will deteriorate more severely than a healthy system when faced with external shocks, thus exhibiting a higher ASII value.

[0079] To verify the significant improvement in diagnostic accuracy and sensitivity of the proposed "frequency-resolved ASII spectrum vector" compared to the traditional "single broadband ASII index," the following six sets of control experiments were designed. The experiments simulated the key parameter responses of helicopters under different health states (healthy, early damage with a 10% decrease in mid-frequency modal stiffness, and advanced damage affecting both mid- and high-frequency modes) and different flight environments (calm airflow and strong turbulence). The aim was to demonstrate that this embodiment can accurately locate the fault frequency band and provide early warning with a higher signal-to-noise ratio; details are shown in Table 1 below.

[0080] Table 1: Frequency-Resolved ASII Spectral Vector Study Table

[0081]

[0082] Broadband ASII, used for comparison with existing technical specifications, is calculated by directly calculating the original boundary layer turbulent dissipation rate without discrete wavelet transform decomposition. And flight control commands - rotor hub vibration transfer entropy Time-delay cross-correlation analysis was performed on the time series to obtain a single ASII value that represents the average coupling strength across the entire frequency band.

[0083] In this embodiment, the maximum spectrum ASII is taken as the frequency-resolved ASII spectrum vector. The maximum absolute value of all components (low frequency, medium frequency, high frequency) is used as a quantitative representation of the most dangerous state of the system at present.

[0084] Comparing the data from Experiment 2 (healthy - strong turbulence) and Experiment 4 (early damage - strong turbulence), under approximately the same strong turbulence input (average turbulent dissipation rate ≈ 15.0 m² / s³), the broadband ASII index increased from 0.18 to 0.45, an increase of 150%, calculated as (0.45 - 0.18) / 0.18 × 100%, providing only a vague "potential risk" indication. However, the frequency-resolved ASII spectral vector of this embodiment reveals a completely different physical picture: the mid-frequency ASII component surged from 0.15 to 0.85, an increase of 467%, calculated as (0.85 - 0.15) / 0.15 × 100% ≈ 467%, directly triggering a "critical warning." Meanwhile, the low-frequency and high-frequency components remained almost unchanged. This data quantitatively demonstrates the core value of this embodiment: through multi-resolution decomposition, the weak fault signal (mid-frequency modal stiffness decrease) hidden in broadband noise is successfully decoupled and amplified. Compared to existing technologies, the diagnostic sensitivity of this embodiment is improved by 211%, calculated as (467%-150%) / 150%×100%≈211%. This ability to accurately locate fault bands is something that broadband analysis methods completely lack.

[0085] Comparing Experiment 4 (early damage) and Experiment 6 (advanced damage), as the damage expanded from a single mid-frequency mode to mid-high frequency modes, the broadband ASII increased from 0.45 to 0.72. Although the value increased, it did not provide any additional information about the change in damage mode. In contrast, the spectral vector of this embodiment clearly depicts the evolution path of the fault: the mid-frequency ASII component slightly increased from 0.85 to 0.91, while the high-frequency ASII component significantly jumped from 0.14 to 0.65. This comparative data demonstrates the significant advancement in fault mode diagnosis and evolution tracking achieved by this embodiment. The diagnostic system not only determines "how dangerous" the fault is through the "maximum spectral ASII," but also reveals "where the danger originates" through the distribution pattern of the spectral vector.

[0086] Observe experiments 1, 2, and 3. Under healthy conditions, even when facing strong turbulence (Experiment 2) or early damage in calm airflow (Experiment 3), all ASII components in this embodiment remain at extremely low levels (≤0.15). This demonstrates the robustness of the algorithm design in this embodiment, effectively avoiding false alarms caused by the deterioration of a single condition (strong turbulence only or damage only). As a measure of coupling strength, the ASII index only produces a significant response when both "external excitation" and "internal defects" are present simultaneously. This perfectly matches the real physical triggering mechanism of dynamic instability, proving the feasibility and high reliability of the overall technical solution in this embodiment.

[0087] Further explanation: A teacher model based on a physical information neural network is constructed and trained under the guidance of a composite loss function that includes a data-driven loss term and a physical constraint loss term; and, based on a knowledge distillation strategy, the reasoning logic contained in the teacher model is transferred to a student model with lower structural complexity.

[0088] The input features of the teacher model include frequency-resolved ASII spectrum vectors;

[0089] The physical constraint loss term is obtained by calculating the residual of the predicted output of the teacher model under the preset aeroelastic motion equation.

[0090] Further explanation: The structure of the student model is selected from gradient boosting decision trees;

[0091] The knowledge distillation strategy is adaptive knowledge distillation, wherein the imitation loss weights of the student model for different intermediate layer features of the teacher model are dynamically determined by the gradient contribution of the corresponding intermediate layer activation values ​​of the teacher model to the final prediction result.

[0092] The following is a detailed explanation of the implementation of the above content:

[0093] The core technical feature of this embodiment is: constructing a high-precision teacher model based on a Physical Information Neural Network (PINN) architecture. This model not only uses frequency-resolved ASII spectrum vectors as core input features to learn data-driven instability precursors, but also embeds the aeroelastic motion equation as a differentiable physical constraint into the loss function of the neural network. This forces the model to make predictions that conform to physical laws even in the critical region where data is sparse, thereby generating high-fidelity reasoning logic that contains prior physical knowledge.

[0094] Building upon this foundation, this embodiment constructs a lightweight student model employing an adaptive knowledge distillation strategy guided by feature importance. Specifically, it utilizes a pre-trained teacher model to reason on a large number of samples, extracting the gradient contribution of each neuron's activation value to the final prediction result, thereby quantifying the importance of different "knowledge points" in the teacher model's reasoning logic. Subsequently, during the distillation process, the weights of the loss function are dynamically adjusted, enabling the student model to prioritize and more accurately mimic the reasoning paths and intermediate layer feature expressions identified as "highly important" in the teacher model. Finally, this optimized student model undergoes INT8 quantization, generating a final prediction model that retains the core physical insights of the teacher model while possessing extremely low computational resource consumption and can be directly deployed on onboard computing units. This solves the technical challenge of treating "knowledge" equally in conventional knowledge distillation, which can lead to the dilution of core physical logic during the transfer process, achieving a significant simplification of the model while maintaining almost no loss in core prediction accuracy.

[0095] The key parameters involved in this embodiment are defined as follows. All parameters are generated by the onboard computing unit or offline training server according to a preset program.

[0096] Obtain the frequency-resolved ASII spectrum vector from the calculation results of step S2. ;

[0097] The flight state parameter vector, whose parameter symbols are: , is a multidimensional real number vector, whose physical meaning is: a set of key parameters describing the current macroscopic flight state of the helicopter; specifically obtained by direct measurement or calculation through the airborne atmospheric data computer and inertial navigation system, including standardized values ​​such as flight Mach number, tip Mach number, flight altitude, collective pitch, cyclic pitch, angle of attack, sideslip angle, etc., which are the results of direct measurement values ​​after standardization processing.

[0098] The predicted value of the structural modal damping ratio, with parameter sign as follows: , is a scalar real number, representing the equivalent viscous damping ratio of the i-th critical structural mode predicted by the neural network model; this value is a core indicator for measuring the dynamic stability of the system, and when it approaches zero or becomes negative, it indicates the occurrence of dynamic instability; in this embodiment, "critical structural mode" refers to the inherent fundamental vibration form that plays a decisive role in the dynamic instability phenomenon of the rotor system; the damping ratio of these modes is a direct indicator for measuring the stability of the system. In embodiments of the present invention, the "critical structural mode" specifically, but not limited to, includes the following two types:

[0099] First-order flapping mode: This is the most dominant and lowest-frequency out-of-plane bending vibration of the rotor blade. Its motion is similar to the entire blade oscillating rigidly and holistically around a virtual flapping hinge. This mode directly determines the dynamic response characteristics of the rotor disk.

[0100] The first-order torsional mode is the torsional vibration of a rotor blade about its spanwise pitch axis from the blade root to the blade tip. The motion of this mode directly changes the angle of attack of the blade profile, thus drastically affecting the unsteady aerodynamic forces acting on the blade.

[0101] Data-driven loss value, with parameter symbol as follows: , is a non-negative real number that quantifies the difference between the structural modal damping ratio predicted by the neural network and high-fidelity simulation or experimental data; it is calculated during the model training phase; the calculation of this value originates from the mean squared error in statistics. The calculation steps are as follows: calculate the difference between the predicted structural modal damping ratio of a single sample and its corresponding true label value; then, square this difference; finally, take the arithmetic mean of the squared differences of all samples within a batch to obtain the data-driven loss value for that batch.

[0102] Physical constraint residuals, with parameter symbols as follows: , is a non-negative real vector, and its physical meaning is: the unbalance quantity at both ends of the known aeroelastic motion equation after substituting the output of the neural network into the equation; it is calculated through a differentiable physical model during the model training phase; the calculation of this value originates from the second-order linear ordinary differential equation system in the field of fluid-structure interaction dynamics; the calculation steps are as follows: construct a differential equation describing the flapping-torsional coupled motion of the rotor blades, which includes structural mass, damping, stiffness matrix terms, and aerodynamic terms described by unsteady aerodynamic theory (Theodorsen). Next, by back-calculating the predicted structural modal damping ratio output by the neural network and combining it with other flight parameters, the required dynamic coefficients and aerodynamic forces are obtained. Then, these coefficients are substituted into the motion equation to calculate the resultant force term; the L2 norm of this resultant force term is the physical constraint residual value.

[0103] The physical constraint loss value, whose parameter symbol is: , is a non-negative real number, and its physical meaning is: a quantitative penalty for the physical constraint residual value, used to guide network parameters to optimize in a direction that conforms to physical laws during model training. The calculation of this value originates from norm theory in mathematics. The calculation steps are: to take the arithmetic mean of the L2 norm of the physical constraint residual values ​​of all samples within a batch, and obtain the physical constraint loss value for that batch.

[0104] The physical constraint weighting coefficient has the following parameter sign: , is a pre-defined non-negative real-valued hyperparameter, whose physical meaning is: in the total loss function of the teacher model, it is a modulating factor used to balance the relative importance of data-driven loss values ​​and physical constraint loss values. It is determined through offline experimental calibration. Specifically, the calibration method involves using a grid search approach, and training multiple teacher models using cross-validation within a pre-defined set of candidate values ​​(0.01, 0.1, 1, 10, 100). Finally, the physical constraint weight coefficient value corresponding to the model that best combines prediction accuracy and conformity to physical laws on an independent validation dataset is selected.

[0105] The activation gradient norm of the intermediate layer in the teacher model, with parameter sign as follows: , is a non-negative real number, and its physical meaning is: the comprehensive influence or contribution of the activation values ​​of all neurons in the j-th intermediate layer of the quantified teacher model to the final output structural modal damping ratio prediction value; it is calculated by the automatic differentiation engine built into the deep learning framework during the data preparation stage of knowledge distillation; the calculation of this value is functionally equivalent to applying the chain rule in calculus to perform backpropagation algorithm. The specific calculation steps are as follows:

[0106] Given input samples, perform a complete forward propagation through the teacher model to obtain the final predicted structural damping ratio, and cache the output activation value tensor of the j-th intermediate layer.

[0107] Define a scalar loss function whose value is the predicted modulus damping ratio of the structure itself.

[0108] Initiate the automatic differentiation procedure, requesting the calculation of the gradient of the scalar loss function with respect to the activation tensor of the j-th intermediate layer. The procedure will automatically perform backpropagation, calculating the partial derivatives of the loss function with respect to the activation values ​​of each neuron in that layer, forming a gradient tensor with the same dimensions as the activation tensor of that layer;

[0109] Calculate the square root of the sum of squares of all elements in the gradient tensor, which is the L2 norm of the gradient tensor. The result of this L2 norm calculation is the activation gradient norm of the intermediate layer of the teacher model for that sample in that layer.

[0110] In this embodiment, if the intermediate layer has 10 neurons and the calculated gradient tensor is [0.1, -0.2, ..., 0.5], then the L2 norm of this gradient tensor is: .

[0111] The execution flow for the above calculation results is defined as follows:

[0112] 2.11) Constructing a high-precision teacher model: A training dataset containing multiple flight conditions, where each data point is represented by a frequency-resolved ASII spectrum vector. Flight state parameter vector It consists of the corresponding actual structural modal damping ratio labels.

[0113] 2.12) The parameterized aeroelastic equations of motion used to define physical constraints;

[0114] 2.2) Constructing a Physical Information Neural Network (PINN). This network structure is a multilayer perceptron (MLP) with four hidden layers, each containing 256 neurons, and the activation function is GELU. The input layer receives the concatenated frequency-resolved ASII spectrum vector and flight state parameter vector, and the output layer outputs the predicted structural modal damping ratio. .

[0115] 2.3) Input a batch of training data into PINN, perform forward computation, and obtain a batch of predicted structural modal damping ratios. 2.4) Based on the predicted values ​​and the true labels, calculate the data-driven loss value. 2.5) Substitute the predicted structural modal damping ratio and relevant flight parameters into the aeroelastic motion equations to calculate the physical constraint residuals. And further obtain the physical constraint loss value. 2.6) Based on the preset physical constraint weight coefficients The two loss values ​​are then weighted and summed to obtain the total loss function value. The calculation logic for this total loss function is: the total loss equals the data-driven loss value plus the product of the physical constraint weight coefficient and the physical constraint loss value. 2.7) Using the AdamW (Adam-with-Weigh-Decay) optimizer, the gradient is calculated using the backpropagation algorithm based on the total loss function value, and all network weights and biases of PINN are updated. 2.8) Steps 2.3) to 2.7) are repeated until the model's performance on the validation set converges or the preset number of training epochs is reached. Final output: A trained, high-precision teacher model.

[0116] The process for adaptive knowledge distillation and student model construction is as follows:

[0117] 2.81) Data input: The trained teacher model; and a large-scale, unlabeled “distilled” dataset covering a wide flight envelope, with each data point containing only a frequency-resolved ASII spectrum vector and a flight state parameter vector.

[0118] 2.82) Input the distillation dataset into the teacher model, perform forward propagation, and record the logits (the original output values ​​of the last layer) of its final output as soft labels and activation values ​​of all intermediate layers.

[0119] 2.83) For each sample, perform backpropagation, calculate and record the norm of the activation gradient of each intermediate layer of the teacher model. This numerical value represents the "importance score" of the knowledge at that level.

[0120] 2.84) A student model was constructed, implemented using Light-GBM with gradient boosting decision trees. Its hyperparameters, such as the number and depth of trees, were set to reduce the total number of parameters and prediction computation by at least one order of magnitude compared to the teacher model.

[0121] 2.85) Define an adaptive distillation loss function that incorporates multiple losses. This loss function consists of two parts: the first part is the loss between the student model prediction and the teacher model soft label, characterized by the "KL divergence"; the second part is the imitation loss between the intermediate layer feature representation of the student model and the corresponding intermediate layer activation value of the teacher model, characterized by the "mean squared error".

[0122] When calculating the imitation loss in the second part, the imitation loss term for each layer is weighted. The weight of the imitation loss is proportional to the norm of the activation gradient of the intermediate layer of the corresponding teacher model. It is directly proportional. That is, the larger the gradient norm of a layer, the greater its imitation loss weight, forcing the student model to prioritize learning the most important knowledge in the teacher model.

[0123] 2.86) Using this adaptive distillation loss function, the student model is trained; the final output is a deployable student model that retains core physical insights.

[0124] In the teacher model construction phase of this embodiment, the core multi-parameter fusion is reflected in the construction of the loss function, specifically employing a weighted fusion method based on physical constraints. Specifically, deep fusion is performed during the gradient update process of model training.

[0125] The physical constraints section introduces deterministic physical prior knowledge from the field of aerospace engineering, providing a robust theoretical anchor for the model's learning process. This ensures that the model will not produce predictions that seriously violate the fundamental laws of physics under any circumstances, especially in extrapolation regions where training data is scarce. This constraint plays a crucial role in regularization.

[0126] The fusion mechanism in this embodiment lies in the fact that, through backpropagation, the gradient generated by physical constraints can directly affect the weights of neurons in the network that process frequency-resolved ASII spectral vectors. When learning how to interpret ASII indices, the representation network spontaneously favors feature representations at the gradient level that are both interpretable to the data and compatible with aeroelasticity theory. This fusion, achieved at the gradient flow level, naturally endows the teacher model with the physical interpretability of the input features.

[0127] The physical constraint weight coefficient in the teacher model of this embodiment It was determined through an offline experimental calibration method using grid search and cross-validation, ensuring an optimal balance between data-driven approaches and physical constraints;

[0128] The imitation loss weights in the student model distillation process are dynamically adaptive values. They are determined as follows:

[0129] For each sample in the distillation dataset, compute the activation gradient norm of each of the N intermediate layers of the teacher model. .

[0130] The global average gradient norm of each layer is obtained by averaging the gradient norms of all samples across their respective layers. .

[0131] The global average gradient norm vector is then Soft-Max normalized. That is, the imitation loss weights of the j-th intermediate layer are... The calculation method is as follows: The natural exponent is the sum of the natural exponents of the average gradient norm of all layers.

[0132] During distillation training, the imitation loss of the student model for the j-th intermediate layer feature of the teacher model will be multiplied by the imitation loss weight. Then it is included in the total loss function. This dynamic weighting mechanism based on gradient importance ensures that the optimization resources of the distillation process can be accurately and preferentially allocated to the transfer of "core knowledge" that contributes the most to the final prediction result, achieving efficient and high-fidelity knowledge transfer. The following is a detailed implementation description of the above content:

[0133] When the predicted value of structural modal damping ratio A larger value indicates a more stable structural mode; it also represents a stronger ability of the system to suppress vibrations and a greater distance from the dynamic instability boundary. This embodiment sets the "structural modal damping ratio prediction value" as follows: The stability threshold for "" is 0.02;

[0134] When the predicted value of structural modal damping ratio The smaller the value, the easier it is for the structural mode to reach a critically stable state; the greater the probability of a precursor to dynamic instability, the lower the system's margin against external disturbances; dynamic instability includes aeroelastic flutter.

[0135] When the structural modulus is set to the predicted damping ratio When the value becomes negative, it indicates that the structural mode has entered an unstable state, and the vibration will diverge over time.

[0136] The final output is the predicted structural modal damping ratio. It is mainly affected by the nonlinearity of the following key input parameters, which is solidified by the trained teacher-student model system.

[0137] When any component value in the frequency-resolved ASII spectral vector increases from 0 towards the absolute value of 1 or -1, the predicted value of the structural modal damping ratio... There is a strong negative correlation. The core physical meaning of the ASII index is the conversion efficiency of external energy to internal damage; when the absolute value of ASII in a frequency band increases, it indicates that the energy coupling of the system in that frequency band is enhanced, the structural integrity is reduced, and thus its ability to dissipate vibrational energy is reduced.

[0138] During training, physical constraint residuals The existence of this constraint imposes constraints on the parameter updates of the teacher model, reducing the deviation between the model's predictions and aeroelastic theory. Even in regions lacking real instability data, larger physical constraint residuals generate larger gradients, forcing the model to adjust its weights to output structural modal damping ratio predictions that better conform to the physical equations. This ensures that when faced with unseen hazardous conditions characterized by high ASII values, the predicted damping ratio of the model's behavior rapidly decreases, enabling it to follow known physical laws.

[0139] To verify the significant advancements of the proposed "prediction model based on physical information and adaptive distillation" compared to existing technologies, the following comparative experiment was designed. The experiment compared the prediction performance of three models under different flight conditions: 1) Comparison Method 1: Pure Data-Driven NN (same structure as the teacher model but using only data to drive the loss value). 1) The trained neural network); 2) The teacher model (PINN) of this invention; 3) The student model of this invention. The experiments aim to demonstrate the comprehensive advantages of this invention in terms of extrapolation prediction accuracy, model lightweighting, and knowledge fidelity. All data are derived from a test set generated by a high-fidelity aeroelastic coupling analysis platform, which has no overlap with the model training set; see Table 2 below for details:

[0140] Table 2: Predictive Performance Study under Different Flight Conditions

[0141]

[0142] Maximum Spectrum ASII: Take the frequency-resolved ASII spectrum vector The maximum absolute value of all components is used as the comprehensive input indicator for the current working condition risk level.

[0143] The formula for calculating prediction error is: |predicted value - true value| / true value × 100%, which is used to quantify prediction accuracy.

[0144] Model parameter count / inference time: Model storage usage and single forward inference time measured on the same embedded hardware platform, used to evaluate the lightweight nature of the model;

[0145] Comparing the data from experiments B1 and B2 (extrapolation domain, i.e., critical high-risk conditions not encountered by the model during training), the predictions of Method 1 (pure data-driven NN) completely failed, with prediction errors ranging from 162.5% to 1700%. The 1700% error is calculated as |0.018-0.001| / 0.001×100%, incorrectly providing an overly optimistic damping ratio far exceeding the true value. Conversely, the teacher model (PINN) of this invention, due to the physical constraints of the aeroelasticity equations, still follows physical laws in its prediction behavior even in the data-sparse region, with prediction errors controlled within 12.5% ​​to 100%; the absolute error is only |0.002-0.001| / 0.001×100%, accurately capturing the critical trend of the damping ratio rapidly decreasing to near zero. This set of data demonstrates that by introducing a physical information neural network, the core pain point of prediction failure in the extrapolation region of traditional data-driven models is solved. The introduction of physical constraints enables the model to perform reliable inference even in the absence of data.

[0146] The performance of the teacher model and the student model of this invention were compared in Experiments B1 and B2. The student model's parameter count (0.9MB) and inference time (1.8ms) were reduced by 94.3% and 93.2% respectively compared to the teacher model (15.8MB, 26.5ms). However, in the most critical instability condition (Experiment B2), although the student model's prediction error (200%) was higher than the teacher model's (100%), its predicted value (0.003) was highly consistent with the actual value (0.001) and the teacher model's predicted value (0.002) in both magnitude and trend, completely preserving the core information that "the system is tending towards instability." This demonstrates the effectiveness of the "adaptive knowledge distillation based on gradient contribution" strategy. It ensures that during the extreme compression of the model, the core inference logic about the critical state, guided by physical information in the teacher model, is preferentially and faithfully transferred. Ultimately, a lightweight model that can run in real-time on resource-constrained airborne hardware without sacrificing critical fault prediction capabilities was obtained.

[0147] Observing experiments A1 and A2 (within the training domain), all models exhibited high prediction accuracy, demonstrating the feasibility of the base model. However, considering the resource evaluation data from experiment C1, only the student model of this invention met the stringent requirements for airborne real-time deployment in terms of model size (<1MB) and inference speed (<2ms). The entire experimental data forms a complete chain of evidence: the PINN teacher model addresses the issues of prediction accuracy and reliability, while adaptive knowledge distillation and quantization resolve the feasibility of model deployment, ultimately providing an end-to-end dynamic instability prediction solution that combines high accuracy and high efficiency.

[0148] Further explanation: Using a student model, a sequence of predicted values ​​for future dynamic stability is generated based on real-time input data; a predictive risk gradient value for quantifying the rate of risk evolution is calculated based on the predicted value sequence; and an inhibitory bias instruction for adjusting the original manipulation command is generated based on the predictive risk gradient value.

[0149] The real-time input data of the student model includes frequency-resolved ASII spectrum vectors;

[0150] The predictive risk gradient value is obtained by performing linear regression fitting on the predicted value sequence and taking its slope.

[0151] Further explanation: The suppression bias instruction is generated through a preset nonlinear mapping module, which is used to nonlinearly map the predictive risk gradient value into a control bias;

[0152] The step of generating the inhibitory bias instruction is performed only when the predictive risk gradient value is lower than a preset negative activation threshold.

[0153] The implementation method of the "preset nonlinear mapping module" is selected from a two-dimensional lookup table;

[0154] The implementation method of the "preset nonlinear mapping module" is selected from radial basis function neural networks;

[0155] The suppression bias command is applied to the collective pitch control channel and the cyclic pitch control channel of the main rotor.

[0156] The following is a detailed explanation of the implementation of the above content:

[0157] The core technical feature of this embodiment lies in achieving a closed loop from intelligent perception to proactive safety control through an integrated prediction-control linkage mechanism. Utilizing a student model deployed on the airborne computing unit, high-frequency inference is performed on the real-time flight data stream, including frequency-resolved ASII spectral vectors, to generate a sequence of predicted structural modal damping ratios within a short future time window. This sequence is then subjected to time-series gradient calculation to obtain a predictive risk gradient value. This gradient value directly quantifies the evolution rate of dynamic instability risk.

[0158] The predictive risk gradient value is fed into a nonlinear mapping module, which, according to a preset mapping rule based on the aircraft dynamics model, nonlinearly converts the risk gradient into suppressive bias commands for the main rotor collective pitch and cyclic pitch control channels. These bias commands are superimposed on the pilot's or autopilot's original control commands, enabling proactive and precise adjustments to the rotor aerodynamic load distribution without significantly affecting normal flight performance. This suppresses nascent unstable aeroelastic modes predicted by ASII index anomalies. The predictive model's "output" is directly translated into the flight control system's "actions," forming a seamless closed loop from risk quantification and trend prediction to proactive suppression. This significantly improves the aircraft's dynamic stability boundary in complex flight environments without increasing hardware costs.

[0159] The key parameters involved in this embodiment are defined as follows. All parameters are generated in real time by the onboard computing unit according to a preset program.

[0160] The "predicted value sequence" is characterized as a sequence of predicted values ​​for structural modal damping ratios, with parameter symbols as follows: , is a real-number time series containing N elements; its physical meaning is: a continuous prediction of the structural modal damping ratio for the next N time steps by a student model; this sequence is obtained by including frequency-resolved ASII spectrum vectors. and flight state parameter vector The sliding time window data is input into the student model and obtained by cyclic prediction or multi-step output prediction; the example of the "structural modal damping ratio prediction value sequence" in this embodiment is [0.015, 0.013, 0.011, 0.009, 0.007];

[0161] Predictive risk gradient value, the parameter symbol is: , is the average rate of change of the predicted structural modal damping ratio sequence within the prediction time window, i.e., the evolution rate of dynamic instability risk; its range is real numbers, negative values ​​indicate a decrease in damping ratio as risk increases, and positive values ​​indicate a decrease in risk; the calculation of the predictive risk gradient value originates from the finite difference method or linear regression in numerical analysis; the predicted structural modal damping ratio sequence and its corresponding time point sequence are subjected to least squares linear fitting to obtain the best fitting line; then, the slope of this fitting line is the predictive risk gradient value.

[0162] With a time step of 10 milliseconds, the slope obtained through linear fitting is... Then the predictive risk gradient value It is -0.2.

[0163] The suppressive bias instruction vector, whose parameter notation is: , is a two-dimensional real vector. Its physical meaning is: a small adjustment that needs to be superimposed on the original control commands of the flight control system to actively suppress dynamic instability; it is calculated through a nonlinear mapping module; the calculation of this suppressive bias command vector originates from the gain scheduling concept in control theory. The calculation steps are: constructing a two-dimensional look-up table or a small radial basis function (RBF) neural network as the nonlinear mapping module; the input of this module is the current flight state parameter vector. This is used to determine the current flight envelope area and the predictive risk gradient value. The output consists of two components of the suppression bias command vector, corresponding to the total offset channel bias, respectively. and periodic variable pitch channel offset The specific values ​​of the two-dimensional lookup table or RBF neural network were determined through offline experimental calibration.

[0164] The offline experimental calibration method is defined as follows: In a high-fidelity aeroelastic coupling simulation environment for aircraft, virtual control biases of different magnitudes and directions are systematically injected for multiple typical flight state points, and their effect on suppressing the downward trend of modal damping ratio induced by high ASII values ​​is observed; through genetic algorithms or particle swarm optimization, the method that can maximize the cancellation of predictive risk gradient values ​​at each state point with minimal control input is found. The negative trend is used to maximize the optimal combination of control biases that improve the modal damping ratio; these optimal solutions are then filled into a two-dimensional lookup table or used to train an RBF network.

[0165] This embodiment is currently... The flight state is high Mach number forward flight, and the nonlinear mapping module outputs a lookup table. This indicates that a total pitch reduction of -0.5% and a periodic pitch adjustment of +0.2% need to be applied.

[0166] The following is a complete description of the execution flow for the above content:

[0167] 3.1) The input is a frequency-resolved ASII spectrum vector. and flight state parameter vector ;

[0168] Student models deployed on the airborne computing unit; and raw collective pitch and cyclic pitch control commands generated by the pilot or autopilot;

[0169] 3.2) Feed the real-time input data into the student model to generate a sequence of predicted structural modal damping ratios within a short future time window. .

[0170] 3.3) The generated sequence of predicted structural modal damping ratios Perform a linear fit, calculate its slope, and obtain the predictive risk gradient value. .

[0171] 3.4) Check the predictive risk gradient value Is it less than the preset negative activation threshold? This embodiment sets a negative activation threshold. for .

[0172] Execution path 1: If the predictive risk gradient value Less than the preset negative activation threshold If this indicates that the risk is increasing at a non-negligible rate, then the "suppress instruction generation" step is executed.

[0173] Execution path 2: If the predictive risk gradient value Not less than the preset negative activation threshold If the system state is stable or improving, skip the "suppress instruction generation" step and directly output the zero bias instruction.

[0174] For "suppression command generation": combine the current flight state parameter vector and the predicted risk gradient value. The input is fed into a preset nonlinear mapping module to calculate the suppression bias command vector. .

[0175] 3.5) The calculated suppression bias command vector The original control commands are vector-added to generate the final control commands, which are then actively suppressed and adjusted, and sent to the main rotor actuators.

[0176] 3.6) The final output is a finely tuned rotor control command designed to actively suppress dynamic instability.

[0177] The core fusion mechanism in this embodiment is based on the feedforward control law superposition method of predictive risk gradient. This method deeply integrates the "prediction" and "control" aspects, and the rationale for its selection is as follows:

[0178] Traditional feedback control relies on existing errors to generate control action, resulting in a time lag. This method uses predictive risk gradient values. Using predictions of future risk trends to generate control instructions is feedforward control, which can suppress instability before it fully manifests.

[0179] By employing lookup tables or RBF networks with nonlinear mapping modules, the magnitude and direction of control actions can be adaptively adjusted according to the current flight status and the rate of risk deterioration. Stronger suppression commands are generated in high-risk areas of the flight envelope, or when the risk gradient is extremely high, while intervention is minimal in safe areas.

[0180] In this embodiment, the mapping relationship within the nonlinear mapping module is determined through the aforementioned offline simulation optimization. The objective function of this optimization process is formally described as: maximizing the improvement in the predicted structural modal damping ratio while minimizing the norm of the control bias command. This is a multi-objective optimization problem, the solution of which inherently balances control effectiveness and control cost, ensuring the accuracy, efficiency, and low intrusion of the suppression effect.

[0181] The following is a detailed explanation of the implementation of the above content:

[0182] When the suppressive bias instruction vector The closer the vector approaches zero, the more stable the characterization system's judgment of the current flight state; in this state, the more easily the active suppression function of this embodiment can remain silent, resulting in less intervention for the pilot or autopilot. When the vector is zero, there is no intervention;

[0183] When the suppressive bias instruction vector As the norm increases from zero, it indicates that the system predicts the risk of dynamic instability at a faster rate; the larger the norm, the greater the adjustment applied to rotor control, and the stronger the suppression effect. The direction of the vector determines the specific combination of control biases to most effectively counteract specific unstable modes under the current flight state.

[0184] The predictive risk gradient value exhibits a non-linear positive correlation with the norm of the inhibitory bias instruction vector. Specifically, this correlation only holds if the predictive risk gradient value... Below the negative activation threshold This relationship is only activated at that time. As the predictive risk gradient value... As the absolute value of the risk gradient increases, the norm of the inhibitory bias instruction vector also increases nonlinearly; the predictive risk gradient value It directly quantifies the "acceleration" of risk. Small negative values ​​indicate slow deterioration, while large negative values ​​indicate rapid deterioration. Intervention is only necessary when the risk shows a clear and rapid deterioration trend, and the intensity of intervention should match the urgency of the deterioration. This threshold activation mechanism, based on the "rate of change" rather than the "instantaneous value," reduces the possibility of false alarms and unnecessary interventions caused by small fluctuations near the critical stability region, accurately reflecting the advanced control concept of "on-demand and appropriate."

[0185] The directions of the flight state parameter vector and the suppression bias command vector exhibit a complex nonlinear mapping relationship;

[0186] Same predictive risk gradient value The underlying aeroelastic mechanisms and most effective suppression methods differ depending on the flight state, such as high-speed forward flight or hovering. At high speeds, collective pitch needs to be reduced to lower propeller disk load, while during hovering, cyclic pitch adjustment is necessary to mitigate tip vortex interference. The flight state parameter vector, as a key input to the nonlinear mapping module, transmits the current flight background to the system. The mapping rules, fixed within the module and obtained through offline optimization, ensure that for each specific flight state, the generated suppression command combination is determined based on the mode most likely to cause instability in that state. This demonstrates the high degree of adaptability of the control strategy to the flight environment.

[0187] To verify the synergistic gain effect of the "predictive-control integrated" linkage mechanism, the following comparative experiment was designed. The experiment simulated the helicopter's response when encountering a "dynamically instable" upwash in a high-fidelity aeroelastic coupling analysis environment. Two scenarios were compared: 1) the baseline flight control system (controlled only by the pilot / autopilot, without the intervention of the scheme in this embodiment); 2) the active suppression system of this embodiment (suppression commands of this embodiment are superimposed on the baseline flight control). The experiment aims to demonstrate the outstanding substantive features and significant progress of this embodiment in suppressing instability trends and improving safety margins; see Table 3 below for details:

[0188] Table 3: Research on Integrated Prediction and Control

[0189]

[0190] Minimum instantaneous modal damping ratio: The minimum value of the actual modal damping ratio reached within the entire time window after the disturbance occurs. This value is used to assess the dynamic stability of the system; the smaller the value, the more dangerous it is, and a value less than zero indicates instability.

[0191] Minimum safety margin: Equivalent to the minimum instantaneous modal damping ratio, used to intuitively assess the stability boundary of the system.

[0192] Compare the data from Experiment A5 and Experiment A6. During stable cruise (Experiment A5), the predictive risk gradient value is -0.01, which is higher than the set negative activation threshold. The negative activation threshold in this example is [value missing]. Therefore, the inhibitory bias command is zero, and the system does not intervene. In the initial stage of the disturbance (Experiment A6), the minimum instantaneous modal damping ratio (0.031) is within the safe range, but the predictive risk gradient value has already decreased. This indicates that the system predicted the risk was accelerating. In this embodiment, the system was activated and generated a small suppression command in advance. This demonstrates the forward-looking nature of this embodiment's intervention based on "trend" rather than "threshold". Compared to traditional passive systems that wait for the risk value to cross a fixed threshold before issuing an alarm, this embodiment can provide precise and low-intrusion intervention in the early stages of instability, even before any macroscopic vibration indicators show abnormalities.

[0193] Comparing the performance of Experiment B5 (baseline) and Experiment B6 (this embodiment) during the mid-stage of the disturbance, under the same external disturbance, i.e., the same ASII and risk gradient, the minimum safety margin of the baseline flight control system had dropped to a dangerous 0.004. In contrast, the system in this embodiment successfully maintained the minimum safety margin at 0.012 by applying a collective pitch of -0.85% and a cyclic pitch offset of +0.32%. The improvement in safety margin compared to the baseline is calculated as: (Safety margin in this embodiment - Baseline safety margin) / Baseline safety margin × 100%. In Experiment B6, the improvement was (0.012 - 0.004) / 0.004 × 100% = 200.00%. This 200% improvement in safety margin demonstrates that this embodiment generates a synergistic gain effect by linking the "sensing" capability of the predictive model with the "execution" capability of the flight control system. The predictive model provides the control system with a prediction of the instability state, while the control system provides a direct physical path to realize the prediction, ultimately pulling the helicopter back from the brink of instability.

[0194] Comparing the final results of Experiment C5 (baseline) and Experiment C6 (this embodiment), under continuous perturbation, the baseline system eventually failed to suppress vibration divergence, with the minimum safety margin dropping to -0.002, indicating the occurrence of dynamic instability. In contrast, the system in this embodiment, due to effective suppression in the early stages, successfully curbed further deterioration of the risk, ultimately stabilizing the system at a critical safety state with a margin of 0.006, thus avoiding instability. This data ultimately proves that this embodiment not only improves performance indicators but also fundamentally changes the system's safety boundary, effectively preventing catastrophic dynamic instability accidents caused by external perturbations.

[0195] Based on the statistical distribution characteristics of predictive risk gradient values ​​under numerous simulation conditions, and combined with expert experience in flight control law design, the following three-level intervention intensity classification standard was developed. This standard aims to ensure the timeliness, effectiveness, and smoothness of intervention, and has been optimized in an independent validation dataset, as detailed in Table 4 below:

[0196] Table 4: Criteria for Differentiating the Intensity of Three-Tier Interventions

[0197]

[0198] This invention achieves a leap from high-precision instability precursor characterization to active flight safety control, generating significant synergistic gains that cannot be achieved by a single technology, specifically manifested in the following four aspects:

[0199] 1. In existing technologies, purely data-driven prediction models often produce catastrophic errors when facing dangerous conditions near the instability boundary that are not covered by training data, due to a lack of guidance from physical laws. This invention fundamentally solves this problem by constructing a teacher model based on a Physical Information Neural Network (PINN) architecture in step S4, incorporating the aeroelastic equations of motion as a strong constraint into model training. This ensures that even in data-sparse critical regions, the model's prediction behavior still follows the first principles of physics, greatly improving the model's generalization ability and prediction accuracy in the extrapolation region, and ensuring the absolute reliability of instability precursor detection.

[0200] 2. Due to its complexity, the high-precision PINN teacher model cannot be directly deployed on resource-constrained onboard computing units. The adaptive knowledge distillation strategy employed in step S4 of this invention is not a simple result imitation, but rather a dynamic adjustment of the imitation loss weights of the student model by analyzing the gradient contribution of the activation values ​​of the intermediate layers of the teacher model to the final prediction. This ensures that the core reasoning logic about critical states, guided by physical information, in the teacher model is preferentially and faithfully transferred. Ultimately, while achieving a model compression rate of over 90%, the "wisdom" of the teacher model is preserved to the maximum extent, generating a student model that combines high precision with extreme lightweight design.

[0201] 3. Traditional safety systems often respond based on state "thresholds," which are subject to time delays and prone to false alarms. In step S5 of this invention, a student model is innovatively used to generate a sequence of predicted values ​​for future stability and to calculate its predictive risk gradient. This gradient quantifies the rate of risk evolution, shifting the trigger mechanism for system intervention from "whether the current state is dangerous" to "whether the future trend is deteriorating." This forward-looking intervention, driven by "trend gradient," enables precise and low-intrusive early suppression before any macroscopic vibration indicators show anomalies.

[0202] 4. This invention directly transforms the high-fidelity prediction capability generated in step S4 into the physical execution capability of the flight control system through the prediction-control integrated linkage mechanism in step S5. The predictive risk gradient value is converted into a suppressive bias command for the main rotor in real time and adaptively through a nonlinear mapping module. This constitutes a seamless information-physical closed loop from "ASII index perception" to "student model prediction" and then to "flight control command adjustment". This closed loop directly transforms the "insight" of the prediction model into the "execution capability" of the flight control system, transforming the entire system from a passive "warning device" into an active "stabilizer". The ultimate effect is that, without increasing any hardware costs, the safe flight envelope of the aircraft is essentially broadened through a pure software algorithm upgrade.

[0203] It should be noted that all calculation formulas in this application employ regression analysis, including but not limited to machine learning algorithms, to deeply analyze the collected parameters and identify their natural trends and interrelationships. Specialized software, such as Python's Scikit-learn library or the R language, is used to automatically generate mathematical models that match the data. Then, cross-validation and other methods are used to objectively evaluate the model performance, and continuous feedback and optimization are combined to ensure that the created formulas truly reflect the inherent laws of the data, thereby guaranteeing their effectiveness and accuracy. In all calculation formulas in this application, the parameters in each formula undergo dimensionless processing within a consistent range to ensure that different physical quantities are compared on the same scale; dimensionless processing techniques include, but are not limited to, min-max-normalization and Z-score standardization.

[0204] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An adaptive construction method for a damage probability model integrating knowledge transfer, characterized in that, The specific steps include: S1: Acquire two types of heterogeneous real-time data streams related to key structural parts of the helicopter. The two types of heterogeneous real-time data streams include: a first data stream representing external aerodynamic energy input, and a second data stream representing the fidelity of information transmission in the internal control loop; perform multi-resolution decomposition on the first and second data streams to generate multiple pairs of frequency band sub-sequences; and for each pair of frequency band sub-sequences, independently calculate the fused feature components, and finally combine them into a frequency-resolved ASII spectrum vector represented by the "fused feature spectrum vector"; multi-resolution decomposition is achieved by using a discrete wavelet transform algorithm. The first data stream is a time series of boundary layer turbulence dissipation rate, and the second data stream is a time series of flight control command-propeller hub vibration transfer entropy. The specific steps for calculating the fusion feature components are as follows: performing time-delay cross-correlation analysis on each pair of frequency band sub-sequences and identifying the maximum cross-correlation value within the preset physical time delay range as the fusion feature component of that frequency band sub-sequence; Before the time-delay cross-correlation analysis, the frequency band sub-sequence is further subjected to Z-score normalization. Before combining them into the final fused feature spectrum vector, the calculated fused feature components are mapped using a hyperbolic tangent function. When the ASII component value of any frequency band subsequence is closer to 1, it indicates that the positive correlation between external aerodynamic energy input and information transmission loss of internal control loop is stronger within that frequency band subsequence. When the ASII component value of any frequency band subsequence is closer to 0, it indicates that the linear correlation between external aerodynamic energy input and internal information transmission loss is weaker within that frequency band subsequence. S2: Based on the time-delay cross-correlation between the first data stream and the second data stream, dynamically calculate the fusion characteristic index of the efficiency of the key structural parts in converting external aerodynamic energy input into internal information transmission loss. S3: Construct a high-precision teacher model, wherein the fusion feature index is used as the core input feature to train the teacher model to learn and characterize the dynamic instability critical precursors of the key structural parts; A teacher model based on a physical information neural network is constructed and trained under the guidance of a composite loss function that includes a data-driven loss term and a physical constraint loss term; and, based on a knowledge distillation strategy, the reasoning logic contained in the teacher model is transferred to a student model with lower structural complexity. The input features of the teacher model include frequency-resolved ASII spectrum vectors; The physical constraint loss term is obtained by calculating the residual of the predicted output of the teacher model under the preset aeroelastic motion equation; S4: Construct a student model with a lower structural complexity than the teacher model; the structure of the student model is selected from a gradient boosting decision tree. The knowledge distillation strategy is adaptive knowledge distillation, wherein the imitation loss weights of the student model for different intermediate layer features of the teacher model are dynamically determined by the gradient contribution of the corresponding intermediate layer activation values ​​of the teacher model to the final prediction result. S5: Deploy the student model on the airborne computing unit and use the student model to process flight data, including the fused feature index, in real time to generate a dynamic instability warning signal, and further convert the dynamic instability warning signal into an inhibitory bias command for the flight control system. Using a student model, a sequence of predicted values ​​for future dynamic stability is generated based on real-time input data; a predictive risk gradient value for quantifying the rate of risk evolution is calculated based on the predicted value sequence; and an inhibitory bias instruction for adjusting the original manipulation command is generated based on the predictive risk gradient value. The real-time input data of the student model includes frequency-resolved ASII spectrum vectors; The predictive risk gradient value is obtained by performing linear regression fitting on the predicted value sequence and taking its slope.

2. The adaptive construction method for a damage probability model integrating knowledge transfer according to claim 1, characterized in that: The suppression bias instruction is generated by a preset nonlinear mapping module, which is used to nonlinearly map the predictive risk gradient value into a control bias. The step of generating the inhibitory bias instruction is performed only when the predictive risk gradient value is lower than a preset negative activation threshold. The implementation method of the "preset nonlinear mapping module" is selected from a two-dimensional lookup table; The suppression bias command is applied to the collective pitch control channel and the cyclic pitch control channel of the main rotor.

3. The adaptive construction method for a damage probability model integrating knowledge transfer according to claim 2, characterized in that: When the inhibitory bias command vector is closer to zero, it indicates that the system's judgment of the current flight state is more stable. As the norm of the suppressive bias command vector increases from zero, it indicates that the rate at which the system predicts the risk of dynamic instability increases; the larger the norm, the greater the adjustment applied to rotor control, and the stronger the suppressive effect.

Images