A UAV Flight Control Method and System Based on Neuron Activation Maps and Gradient Sensitivity

Abstract

This invention provides a UAV flight control method and system based on neuron activation maps and gradient sensitivity, belonging to the interdisciplinary fields of interpretability and flight control verification. The method includes: S1, evaluating the consistency of UAV flight control logic based on gradient sensitivity analysis and unsupervised clustering; S2, constructing a data-driven activation map of hidden layer neurons for standard flight actions; S3, performing multi-level UAV flight control based on dual reliability assessment and adaptive thresholds. This invention deeply monitors the deviation of UAV flight status by quantifying online activation maps and offline standard flight action activation maps; it automatically mines flight decision patterns from historical experience data using unsupervised clustering technology, effectively reducing redundant false alarms and false intercepts under complex boundary situations, and effectively improving the task execution efficiency of the decision algorithm while maintaining the UAV flight status.

Description

Technical Field

[0001] This invention relates to the field of interpretability and flight control verification technology, specifically to a UAV flight control method and system based on neuron activation maps and gradient sensitivity. Background Technology

[0002] With the widespread application of deep learning and reinforcement learning technologies in aviation decision-making, flight control decision models have gradually become the core of flight control. However, the inherent black-box nature of these models and their sensitivity to input disturbances pose serious challenges to the execution of flight missions.

[0003] Existing technologies face several significant bottlenecks in ensuring the operational stability of unmanned aerial vehicles (UAVs): First, there is a lack of monitoring for internal cognitive anomalies. Current monitoring methods primarily focus on physical violations at the input-output level. When a model encounters an unseen state, such as cognitive illusions caused by out-of-distribution (OOD) inputs, its output actions may not yet violate regulations, but its internal reasoning logic has already deviated from the normal pattern. Traditional runtime traffic management (RTA) can only respond passively the instant after a physical violation occurs, making it difficult to identify model logic collapse in advance. Second, rule design is redundant and conservative. Traditional flight boundaries and control thresholds mostly rely on manual design based on expert experience, making it difficult to accurately cover high-dimensional, nonlinear flight state spaces. To ensure timely flight control, extremely conservative thresholds are often set, which significantly limits the performance of flight control decision models. Third, interpretability is lacking. Existing interpretable XAI methods, such as the Shapley additive interpretation algorithm (SHAP) or the locally interpretable model-independent algorithm (LIME), typically require thousands of sampling perturbations near the decision point, with computation times reaching the second level, which is insufficient to meet the millisecond-level response requirements of aviation control systems. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention aims to provide a UAV flight control method and system based on neuron activation maps and gradient sensitivity. By quantifying the Hamming distance between online activation maps and offline standard flight maneuver activation maps, the deviation of the UAV's flight state from deep inference is monitored. The complex black-box model interpretability verification is transformed into lightweight single-step backward gradient differentiation and binary map bit operations. A data-driven adaptive control mechanism that balances baseline and maneuverability performance is constructed. Unsupervised clustering technology is used to automatically mine flight decision patterns from historical experience data, and the optimal control threshold is dynamically learned based on statistical features. This effectively reduces the system's redundant false alarms and false interceptions under complex boundary situations, and effectively improves the task execution efficiency of the decision algorithm while maintaining the UAV's flight state.

[0005] Specifically, on the one hand, the present invention provides a UAV flight control method based on neuron activation maps and gradient sensitivity, which includes the following steps: S1: Set the UAV flight state vector and flight maneuver response parameters Quantitative sensitivity of UAV flight state characteristic parameters Using Gaussian mixture models to mine the flight control logic of UAVs; selecting decision logic parameters for flight states based on the sensitivity of flight state feature parameters. The key feature set corresponding to the flight action mode is obtained. Determine the logical consistency parameters of the drone. Assess the consistency of the drone flight control logic; S2: The model's internal flight actions are represented by monitoring semantically information-dense hidden layers. Flight state data samples are grouped according to flight action modalities, and the neurons in each hidden layer are identified. Activation frequency in flight action mode Extract the activation vectors of neurons in a specified hidden layer within the UAV flight control decision model, and generate a standard flight action activation map through binarization; calculate the Hamming distance using the standard flight action activation map corresponding to the flight action mode constructed by the flight control logic evaluation, and then generate reliability evaluation parameters for the degree of flight action deviation. Construct a data-driven activation map of hidden layer neurons for standard flight actions; S3: Determine the UAV flight control logic consistency parameters obtained in step S1 Corresponding optimal isolation threshold The reliability assessment parameters of the degree of deviation of flight maneuvers obtained in step S2 Corresponding optimal isolation threshold Double cross-validation is performed on the UAV flight control decisions; based on the optimal adaptive control threshold obtained through data-driven learning from the working characteristic curve, the consistency parameters of the UAV flight control logic are adjusted accordingly. Reliability assessment parameters for deviation from flight maneuvers Independent judgment is made, a multi-level dynamic control strategy is implemented, and control commands are dynamically adjusted to achieve multi-level drone flight control.

[0006] Preferably, step S1 specifically includes: S11: Set the UAV flight state vector Specifically, this includes: pitch angle Angle of attack Height deviation Pitch rate q, roll angle Roll rate m and heading angle and flight maneuver response parameters Quantitative sensitivity of UAV flight state characteristic parameters ; S12: Yes Sensitivity vector of feature parameters for each flight state sample Establish a Gaussian mixture model and use the Gaussian mixture model (GMM) to explore the flight control logic of the UAV. S13: Selecting decision logic parameters for flight status based on the sensitivity of flight status characteristic parameters. The key feature set corresponding to the flight action mode is obtained. Determine the logical consistency parameters of the drone. .

[0007] Preferably, the sensitivity of the UAV flight state characteristic parameters in step S11 is... for: ; in, For the first Flight status characteristic parameters of a drone Sensitivity; The sign for partial derivatives; For the first Values ​​of characteristic parameters of the flight status of an unmanned aerial vehicle; The optimal action response parameters output by the UAV flight control system; This is the index of the flight state characteristic parameters of the first UAV; This represents the total number of characteristic parameters of the UAV's flight status. This is a sensitivity constant to avoid situations where the denominator is 0; For the first The covariance matrix of each UAV flight action mode; For the first Values ​​of characteristic parameters of the flight status of an unmanned aerial vehicle; This is the index of the flight status characteristic parameters of the second UAV.

[0008] Preferably, the UAV logical consistency parameters in step S13 for: ; in, For consistency parameters of UAV flight control logic; For auxiliary parameters of flight control logic; For the first Sensitivity to individual UAV flight status characteristic parameters; This is a set of key characteristic parameters for flight control. For the first A set of key characteristic parameters for flight control; These are the decision logic parameters for flight status.

[0009] Preferably, step S2 specifically includes: S21: Set up multi-level UAV flight control monitoring points, select the flight state feature representation layer in the UAV flight control decision model, and use the threshold function to generate the binary activation map of the UAV flight control monitoring points; S22: Group the flight status data samples according to flight maneuver modes to form groups corresponding to specific flight maneuver modes. Flight status dataset ; Identify each hidden layer neuron Activation frequency in flight action mode ; S23: Generate a standard flight maneuver activation map based on flight control logic evaluation, and calculate the reliability evaluation parameters for the degree of flight maneuver deviation. .

[0010] Preferably, step S21 specifically includes: ; in, For the current flight status data sample in the th The first monitoring layer The binary activation state of a neuron; For the first Monitoring layer The output vector of the flight state feature representation layer for each neuron; For indicator functions; For the first Monitoring layer Number of neurons; For flight control monitoring layer index; This is an index of neurons within the flight control monitoring layer.

[0011] Preferably, the standard flight maneuver activation map in step S23 is specifically as follows: ; in, For flight action modes Lower monitoring layer The Middle The standard activation binary state of each neuron; This is a conditional function; For the first Monitoring layer The activation frequency of each neuron; These are the activation parameters for standard flight maneuvers.

[0012] Preferably, the reliability assessment parameters for the degree of deviation of flight maneuvers in step S23 are... for: ; ; in, Reliability assessment parameters for the degree of deviation from flight maneuvers; Hamming distance for the flight state feature representation layer; Hamming distance for the flight maneuver decision-making layer; The weighting coefficients for the flight state feature representation layer; Weighting coefficients for flight action decision-making; The result is the Hamming distance calculation. This is a function to count the number of activation states of 1 in a binary vector; This represents the activation feature vector of the currently flying sample in the monitoring layer. The standard flight motion activation map vector for the corresponding motion mode, which is fixed in the offline phase; This represents the total number of neurons involved in the calculation within the flight control monitoring layer. This is the neuron activation state vector.

[0013] Preferably, step S3 specifically includes: S31: Obtain UAV flight control logic consistency parameters Corresponding optimal isolation threshold Reliability assessment parameters for deviation from flight maneuvers Corresponding optimal isolation threshold Based on the flight state sample set Determine the consistency parameters of UAV flight control logic Reliability assessment parameters for deviation from flight maneuvers The corresponding execution threshold; optimize the dual-channel adaptive control threshold based on data-driven approaches; S32: Consistency parameters of UAV flight control logic at the fusion input end Reliability assessment parameters for deviation from flight maneuvers Based on dual verification, three levels of monitoring are implemented for dynamic traffic management and control.

[0014] On the other hand, the present invention provides a UAV flight control system based on a UAV flight control method using neuron activation maps and gradient sensitivity, which includes: a data-level processing gradient calculation module, a hidden layer neuron state extraction bit operation module, and a traffic decision module. The data-level processing gradient calculation module connects the multi-channel signal synchronization and noise reduction unit in the UAV flight control process to the inertial measurement unit, positioning module, and airspeed sensor; performs timestamp alignment and cache management on the input data streams of each frequency, and removes abnormal high-frequency noise pulses through a low-pass filtering algorithm to ensure that the data transmitted to the main control processor has high fidelity. The hidden layer neuron state extraction bitwise operation module samples and rapidly compares the internal operating states of the deep learning network structure, directly extracting the data array of the intermediate operation layer and performing dimensionality reduction encoding. The hidden layer neuron state extraction bitwise operation module includes: a network intermediate layer data extraction unit, an array dimensionality reduction discretization encoding unit, and a bit logic operation difference measurement unit. The network intermediate layer data extraction unit reads and copies the floating-point numerical array output by the specific layer, representing the internal operation characteristics while processing the current data stream. The array dimensionality reduction discretization encoding unit receives the extracted floating-point operational array and uses a preset truncation function to forcibly convert continuous variables into discrete state values ​​composed of zeros and ones. The bit logic operation difference measurement unit calls the currently generated binary encoding sequence to quantify the numerical index representing the degree of difference between the current internal operation path and the standard path. The traffic decision module is used for scheduling UAV control permissions and executing signals; specifically, it includes: an offline feature statistics dual-boundary calibration unit, a dual-channel parameter synchronization interval determination unit, and a hierarchical response bus scheduling unit; the offline feature statistics dual-boundary calibration unit statistically analyzes the distribution of evaluation parameters for normal flight state samples to improve the pass rate of control commands under normal flight conditions; the dual-channel parameter synchronization interval determination unit determines the risk level of the current actuation command issued by the main control computer; and the hierarchical response bus scheduling unit directly controls the signal transmission channel of the aircraft's underlying servo mechanism based on the interval determination result.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention moves from passive physical limits to proactive monitoring and protection based on internal cognitive logic, breaking through the limitation of traditional UAV flight control systems that can only rely on physical parameters to exceed limits for post-event control; by quantifying the Hamming distance between the online activation map and the offline standard flight action activation map, it deeply monitors the deviation of the UAV flight state in the deep inference of the model; this mechanism can identify early signs of abnormal feature mapping inside the model before the physical flight state of the UAV deteriorates substantially, providing sufficient time margin for takeover and degraded control.

[0016] (2) This invention overcomes the problem of large computational overhead in airborne environments for complex interpretable algorithms; it transforms the interpretability verification of complex black-box models into lightweight single-step backward gradient differentiation and binary graph bit operation; this architecture design effectively avoids the large-scale sampling iteration required by traditional proxy models, controls the consumption of computing resources and processing delay, and can strictly meet the performance constraints of the high-frequency control cycle of airborne flight control computers.

[0017] (3) This invention constructs a data-driven adaptive control mechanism that takes into account both bottom line and maneuverability; it uses unsupervised clustering technology to automatically mine flight decision patterns from historical experience data and dynamically learns the optimal control threshold based on statistical features, overcoming the subjective limitations and overly conservative problems of traditional preset rules; this mechanism effectively reduces the redundant false alarm and false interception rate of the system under complex boundary situations while maintaining the consistency of decision logic, and effectively improves the task execution efficiency of the decision algorithm while maintaining the flight status of the UAV. Attached Figure Description

[0018] Figure 1 This is a control block diagram of the UAV flight control method based on neuron activation maps and gradient sensitivity of the present invention; Figure 2 Flowchart for offline processing decision logic consistency assessment; Figure 3 Flowchart for online operation decision logic consistency assessment; Figure 4 Flowchart for constructing a standard flight action activation atlas library for hidden layer neurons; Figure 5 Flowchart for adaptive control threshold learning; Figure 6 A flowchart for dual reliability assessment and three-level traffic management RTA dynamic control; Figure 7 A heatmap showing the importance distribution of decision pattern features mined based on Gaussian mixture models during the offline phase; Figure 8 A statistical diagram showing the distribution of Hamming distances between the flight state feature representation layer and the flight action decision layer in the hidden layer neuron activation map; Figure 9 A two-dimensional joint distribution of overall reliability assessment parameters and logical consistency scores, and a regional division map of multi-level decision-making areas for traffic management RTA; Figure 10 This is a time series comparison of the early warning capability of neuronal activation and the evolution of physical parameters under multiple violation scenarios of the present invention; Figure 11 This is a diagram comparing the differences in logical consistency across different control dimensions in embodiments of the present invention. Figure 12 This is a joint distribution diagram of the consistency parameters of the UAV flight control logic in an embodiment of the present invention; Figure 13 This is a time series graph showing the reliability assessment parameters of the consistency parameters of human-machine flight control logic and the degree of deviation of flight actions under pitch violation scenarios in an embodiment of the present invention. Figure 14 This is a time series graph of the reliability evaluation parameters of the consistency parameters of human-machine flight control logic and the degree of deviation of flight actions under the angle of attack violation scenario of this invention. Figure 15 This is a diagram showing the evolution of physical parameters in multiple violation scenarios according to an embodiment of the present invention; Figure 16 This is a distribution diagram of traffic management decision-making during runtime, based on a real-world test of an embodiment of the present invention. Detailed Implementation

[0019] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings.

[0020] This invention proposes a UAV flight control method based on neuron activation maps and gradient sensitivity. Taking a simulation environment equipped with a UAV flight controller based on deep reinforcement learning (DQN) as an example, it verifies and manages UAV flight control under high-dimensional complex conditions. Figure 1 As shown, the consistency of UAV flight control logic is evaluated based on gradient sensitivity analysis and unsupervised clustering; a data-driven activation map of hidden layer neurons for standard flight actions is constructed; and multi-level UAV flight control is performed based on dual reliability evaluation and adaptive thresholds. Specifically, the steps include: Step S1: Evaluate the consistency of UAV flight control logic based on gradient sensitivity analysis and unsupervised clustering. This step overcomes the limitation of traditional UAV flight control rules heavily relying on prior knowledge design, and proposes a data-driven method for extracting and evaluating UAV flight control logic online. Step S1 reflects the improvement in gradient sensitivity through the physical quantity-gradient sensitivity coupling formula and the modal posterior decision formula.

[0021] Step S11: Quantify the sensitivity of UAV flight state feature parameters; let the UAV flight state vector be... Includes: pitch angle Angle of attack Height deviation Pitch rate q, roll angle Roll rate m and heading angle , etc., the optimal motion response parameters output by the UAV flight control system The flight state characteristic parameters are calculated and the output action response parameters are obtained through a single backpropagation. Gradient sensitivity and sensitivity to UAV flight state feature parameters; quantification of UAV flight state feature parameter sensitivity. for: ; in, For the first Flight status characteristic parameters of a drone Sensitivity; The sign for partial derivatives; For the first Values ​​of characteristic parameters of the flight status of an unmanned aerial vehicle; The optimal action response parameters output by the UAV flight control system include discrete action selection results, action score values ​​corresponding to each candidate action, and selection probabilities corresponding to each candidate action. This is the index of the flight state characteristic parameters of the first UAV; This represents the total number of characteristic parameters of the UAV's flight status; in this example, the value is 7. This is a sensitivity constant to avoid situations where the denominator is 0; For the first The covariance matrix of each UAV flight action mode; For the first Values ​​of characteristic parameters of the flight status of an unmanned aerial vehicle; This is the index of the flight status characteristic parameters of the second UAV.

[0022] The above formula incorporates the magnitude of physical quantities into the gradient to couple physics and model, thereby characterizing the influence weight of each physical quantity on the UAV flight control decision under the current flight state.

[0023] Step S12: Use Gaussian Mixture Model (GMM) to mine the UAV flight control logic; during the offline data mining phase, record the... Sensitivity vector of feature parameters for each flight state sample ,in An index for offline flight status samples. This represents the total number of samples. Based on this, [the following is used]... Sensitivity vector of feature parameters for each flight state sample The Gaussian mixture model is established as follows: ; in, The probability density function is the sensitivity of the flight state characteristic parameters. This is the sensitivity vector of the feature parameters for the flight state samples; This is the complete set of optimization parameters for the Gaussian mixture model. ; For the first The mixed weights of each drone flight action mode satisfy the following conditions. ; It is a natural exponential function; For the first The mean vector of the UAV flight maneuver mode represents the typical sensitivity distribution of characteristic parameters of each flight state under the UAV flight maneuver mode, such as pitch angle and roll angle. For the first The covariance matrix of each UAV flight action mode describes the fluctuation of the sensitivity of the flight state feature parameters and the correlation between the features under the flight of the UAV. It is the transpose symbol; Number the flight maneuver modes of the UAV; This represents the total number of flight maneuver modes of the drone.

[0024] The parameters are solved iteratively using the expectation-maximization algorithm. Thus, the first Mean vector of UAV flight action modes For the mean vector Sort the sensitivity of each feature from largest to smallest to obtain ,in, Then the key feature set of mode k is the set of the first m feature indices after sorting, that is, we have ;in, For the first A set of key characteristic parameters for flight control; For the first Key characteristic parameters of flight control.

[0025] The above steps automatically extract the dominant feature set for each UAV flight maneuver mode, i.e., the set of key flight control feature parameters. For example, the key feature set for obtaining pitch control should include attitude and altitude.

[0026] Step S13: Determine the consistency parameters of the UAV flight control logic; during the online operation phase, extract the sensitivity of the current flight state feature parameters. The decision logic parameter for the current flight state is selected by maximizing the posterior probability of the mode. ; in, These are the decision logic parameters for flight status; To select the largest candidate index among all candidate action modalities; For the first Hybrid weights for each drone flight action mode; Sensitivity to current flight state characteristic parameters; For the first The mean vector of the UAV flight maneuver modes; For the first The covariance matrix of each UAV flight action mode; For the first The mean vector of the UAV flight maneuver modes; For the first The covariance matrix of each UAV flight action mode; This is the matrix transpose symbol.

[0027] Therefore, the key feature set corresponding to the flight action mode was selected. Then the current drone logical consistency parameters for: ; in, For the consistency parameters of UAV flight control logic, in the embodiment The higher this parameter is, the more the current decision-making basis conforms to the physical logic learned by massive data; conversely, the lower the parameter is, the more the decision may be driven by environmental noise or irrelevant secondary features. These are auxiliary parameters for flight control logic. To prevent division by zero, the values ​​in this embodiment are [values ​​to be specified]. ; For the first Sensitivity to individual UAV flight status characteristic parameters; The key feature parameter set for flight control was extracted unsupervised from massive amounts of normal flight state data using a Gaussian mixture model (GMM). For the first A set of key characteristic parameters for flight control.

[0028] like Figure 2 The flowchart for offline decision-making logic consistency evaluation illustrates the complete data flow and processing logic from massive historical flight control experience data to online logical consistency verification: It demonstrates how, during the offline processing stage, key flight control feature parameter sets are automatically mined for each flight maneuver mode through feature-sensitive quantification and unsupervised clustering. For example... Figure 3 The flowchart shown is for the online operation decision logic consistency evaluation. It shows that during the online operation phase, the gradient of the Gaussian Mixture Model (GMM) is tracked to calculate the sensitivity of each feature. Combined with the matching of the key flight control feature parameter set and weighted summation, the online logic consistency score reflecting whether the UAV flight control decision logic conforms to the physical logic is finally calculated.

[0029] This invention first collects 6000 normal flight state samples of a UAV in a simulation environment. The state dimensions cover seven physical quantities: pitch angle (Pitch), angle of attack (AoA), altitude deviation (AltD), pitch rate (Q), roll angle (Roll), roll rate (P), and heading angle (Heading). An unsupervised clustering method using a Gaussian mixture model (GMM) is employed to analyze the normalized gradient sensitivity vectors of the normal flight state samples, automatically identifying nine natural flight control modes. The dominant features and success rates of each flight control mode are shown in Table 1.

[0030] Table 1. Characteristics and Success Rate Statistics of Typical Decision-Making Patterns Based on Unsupervised Mining like Figure 7As shown, flight control modes 1 and 3, which include the UAV's flight altitude (ALT) and heading (HY), have extremely high success rates (>98%), while flight control mode 2, which lacks altitude features, has the lowest success rate. Based on this, key flight control feature parameter sets for each mode are extracted for online calculation of UAV flight control logic consistency parameters. .

[0031] Step S2: Construct a data-driven standard flight action activation map of hidden layer neurons. Addressing the issues of flight action matching in hidden layer neuron activation maps and the lack of visibility into the internal states of deep learning black-box models, this invention proposes characterizing the model's internal flight actions by monitoring semantically information-dense hidden layers. Activation vectors of neurons in specified hidden layers within the UAV flight control decision model are extracted and binarized to generate a standard flight action activation map. Hamming distance is calculated between this map and the standard flight action activation map corresponding to the flight action modality constructed through flight control logic evaluation in the offline phase. This generates a reliability assessment parameter reflecting the degree of deviation of the UAV flight control decision model's internal flight actions. The specific process for constructing the data-driven standard flight action activation map of hidden layer neurons in this embodiment of the invention is as follows: In the offline stage, normal flight state data samples under each flight action modality are collected; the output of the specified hidden layer is extracted and a binary standard flight action activation map is generated; and the activation frequency of each neuron is calculated. And a standard flight action activation map of flight action modes is generated through a logic evaluation of greater than 0.5. Step S2 reflects the improvement of neuron activation maps through the hidden layer activation mapping formula and the Hamming deviation quantization formula.

[0032] Step S21: Set up multi-level UAV flight control monitoring points and perform binarization mapping; select the flight state feature representation layer in the UAV flight control decision model; for example, the flight state feature representation layer Dense_1 is used to monitor whether the UAV flight control decision model correctly perceives environmental features and makes flight control decisions; the flight action decision layer Dense_2 is used to monitor whether the UAV flight control decision model generates correct flight action mappings as flight control monitoring nodes. For the first... Output vector of the monitoring layer A binary activation map of the UAV flight control monitoring points is generated using a threshold function, specifically as follows: ; in, For the current flight status data sample in the th The first monitoring layer The binary activation state of each neuron is used to count the activation frequency in step S22 and to construct a standard flight action activation map and calculate the Hamming distance in step S23. For the first Monitoring layer The output vector of the flight state feature representation layer for each neuron; This is an indicator function; it takes the value 1 when the formula satisfies the condition, and takes the value 0 when the condition is not met. For the first Monitoring layer Number of neurons; For flight control monitoring layer index; This is an index of neurons within the flight control monitoring layer.

[0033] Step S22: Perform grouping and frequency statistics of standard flight maneuver modes; during the offline high-fidelity flight simulation phase, collect... A normal flight state data sample is generated, and the flight state data sample is grouped according to the flight action mode in step S12 to form a group corresponding to a specific flight action mode. Flight status dataset ; Targeting flight maneuver modes and the designated flight control monitoring layer Calculate each hidden layer neuron The activation frequency in the flight action mode is: ; in, For the first Monitoring layer One neuron in flight action mode The activation frequency is below; For the h-th flight state data sample Monitoring layer The activation state value of each neuron; This is a flight state dataset corresponding to a specific flight maneuver mode; For flight action modes The total number of flight state samples included; This belongs to the flight status dataset The sample number index.

[0034] Step S23: Generate a standard flight action activation map based on the flight control logic evaluation, and calculate the reliability evaluation parameters for the degree of flight action deviation; for each type of flight action mode, determine the standard flight action activation map, specifically as follows: ; in, For flight action modes Lower monitoring layer The Middle The standard activation binary state of each neuron; This is a conditional function; For the first Monitoring layer The activation frequency of each neuron; The value is 0.5, which is the standard flight action activation parameter in this example.

[0035] The set of neurons commonly activated by more than 50% of the normal flight state data samples was solidified into a standard activation neuron map for this flight maneuver mode; reliability assessment parameters for determining the degree of deviation from flight maneuvers were determined. for: ; ; in, Reliability assessment parameters for the degree of deviation from flight maneuvers; Hamming distance for the flight state feature representation layer; Hamming distance for the flight maneuver decision-making layer; The weighting coefficients for the flight state feature representation layer; Weighting coefficients for flight action decision-making; The result is the Hamming distance calculation. This is a function to count the number of activation states of 1 in a binary vector; This represents the activation feature vector of the currently flying sample in the monitoring layer. The standard flight motion activation map vector for the corresponding motion mode, which is fixed in the offline phase; The total number of neurons involved in the calculation in the flight control monitoring layer, used to normalize the Hamming distance; This is the neuron activation state vector.

[0036] like Figure 4 The diagram illustrates the data-driven process of solidifying hidden layer neuron feature paths and generating a neural network map: the right side shows the grouping and statistical counting of sample data; the middle section shows the setting of multi-level monitoring points, using the ReLU activation function and a binarization threshold function to generate individual binary activation fingerprints; the bottom section shows the majority voting logic, where over 50% of the normal flight state data samples jointly activate and establish a standard cognitive loop, which is then solidified into a standard activation neuron network map for standard flight action modalities. A standard flight action activation network map for hidden layer neurons was constructed. For the Dense_1 flight state feature representation layer (128 neurons) and the Dense_2 flight action decision layer (128 neurons) within the deep reinforcement learning DQN model, 600 normal flight state data samples under specific flight action modalities were collected in offline simulations. Standard flight action activation maps were extracted by calculating activation frequencies greater than 0.5, and the results are shown in Table 2.

[0037] Table 2. Statistical table of standard hidden layer activation maps for each action mode. The above results are consistent with the sparse activation of ReLU network activation function. The flight action decision layer Dense_2 is sparser than the flight state feature representation layer Dense_1. Moreover, the number of activated neurons varies in each modality, such as Yaw, which constitutes the unique cognitive features of this modality.

[0038] like Figure 8 This shows the Hamming distance distribution of the Dense_1 feature representation layer in 600 flight state samples, with the horizontal axis representing the Hamming distance of the feature layer. The range of values ​​is The larger the value, the further the current feature activation mode deviates from the standard flight action activation map. The vertical axis represents the frequency of flight state samples. The average Hamming distance of the flight state feature representation layer Dense_1 is 0.396. This figure reflects the diversity and volatility of the feature layer's perception when the UAV faces various physical states, such as various pitch angles and angles of attack combinations.

[0039] like Figure 9 This diagram illustrates the Hamming distance distribution of the Dense_2 flight action decision layer, with the horizontal axis representing the Hamming distance of the decision layer. ; This represents the frequency of flight state samples. Compared to the feature layer, the distribution of the decision layer shifts to the left, the mean decreases to 0.148, and the distribution range is more convergent, mainly concentrated in... to This demonstrates that the flight control system has stronger stability at the decision level; even if there are certain fluctuations in the input characteristics, the generated underlying control actions can still be highly similar to the baseline logic.

[0040] like Figure 10 The propagation relationship from perceptual bias to action bias was quantified using a scatter plot, with the horizontal axis representing the Hamming distance of the feature layer. The vertical axis represents the Hamming distance of the decision-making level. The dashed line on the diagonal of the coordinate system in the figure represents the deviation line, and all observation points are located below the diagonal line. This distribution characteristic indicates that there is an error compression mechanism inside the flight control system, that is, a large feature perception deviation will not lead to an equal magnitude of action output deviation. This characteristic physically corresponds to the robustness of the control law, enabling the UAV to make logical action choices even in complex environments.

[0041] like Figure 11The differences in logic consistency across various control dimensions were compared. The horizontal axis represents four typical flight maneuver modes: pitch, roll, yaw, and throttle. The vertical axis represents the average Hamming distance. The height of the bars in the graph represents the deviation from the mean, and the error bars represent the standard deviation. The results show that the roll mode has the lowest average deviation and the smallest fluctuation, indicating that the control logic in this dimension is the most robust. While the pitch mode has a moderate mean, the error bars are relatively long, reflecting that the pitch control logic has flexibility and adjustment space under various flight altitude or speed deviations.

[0042] Step S3: Perform multi-level UAV flight control based on dual reliability assessment and adaptive threshold, plan the UAV flight path, and enable the UAV to perform flight missions; this step utilizes the UAV flight control logic consistency parameters obtained in step S1. The reliability assessment parameters of the deviation from the flight maneuver obtained in step S2 This study employs a dual cross-validation approach for UAV flight control decisions, considering both input feature dependency logic and the internal cognitive path of the model. The optimal adaptive control threshold, obtained through data-driven learning based on the offline working characteristic curve ROC, is used to evaluate the consistency parameters of the UAV flight control logic. Reliability assessment parameters for deviation from flight maneuvers This invention employs a multi-level dynamic control strategy, including independent judgment, trust execution, action restriction, or isolation takeover, to dynamically adjust control commands corresponding to physical quantities such as pitch angle, roll angle, altitude deviation, and heading angle. During online operation, the invention uses the logical consistency of input features and the cognitive state of the deep network as two parallel flight monitoring channels, implementing flight status control through their respective data-driven thresholds.

[0043] Step S31: Optimize the dual-channel adaptive control threshold based on data-driven methods; To avoid the performance limitations of traditional traffic management RTA systems due to overly strict threshold settings, this invention establishes a data-driven optimal threshold learning mechanism.

[0044] During the offline phase, the consistency parameters of the UAV flight control logic are used respectively. Reliability assessment parameters for deviation from flight maneuvers As the predictor variable, with the execution of flight missions as the true label, the Operating Characteristic Curve (ROC) is determined, and the optimal UAV flight warning threshold for each variable is obtained, specifically: ; ; in, Consistency parameters for UAV flight control logic The corresponding optimal isolation threshold; Parameters for reliability assessment of flight maneuver deviation The corresponding optimal isolation threshold; To identify the probability of logically abnormal flight states, candidate isolation threshold variables are used. As a parameter for consistency of UAV flight control logic The probability of correctly identifying a logically abnormal flight state when judging whether it is dangerous; The probability of misclassifying a normal flight state as an abnormal flight state is represented by a candidate isolation threshold variable. Consistency parameters for UAV flight control logic The probability of misjudging a normal flight state as an abnormal flight state when determining the threshold for danger; To correctly identify the probability of logically abnormal flight states, candidate isolation threshold variables are used. Reliability assessment parameters for the degree of deviation from flight maneuvers The probability of correctly identifying a logically abnormal flight state when judging whether it is dangerous; The probability of misclassifying a normal flight state as an abnormal flight state is represented by a candidate isolation threshold variable. Reliability assessment parameters for the degree of deviation from flight maneuvers The probability of misjudging a normal flight state as an abnormal flight state when determining the threshold for danger; This is a function to find the maximum value. The candidate isolation threshold variable has a value range of 1. .

[0045] The execution threshold is set to the 95th percentile of the parameter values ​​in the respective normal flight state sample distribution. Specifically, let the normal flight state sample set be... Then the drone flight control logic consistency parameters Reliability assessment parameters for deviation from flight maneuvers The corresponding execution threshold is: ; ; in, Consistency parameters for UAV flight control logic The execution threshold; Parameters for flight maneuver response reliability assessment The execution threshold; To take the lower 5th percentile of the parameter values ​​corresponding to the normal flight state sample; For the flight control logic consistency parameters corresponding to sample p under normal flight conditions; The parameters are the flight action response reliability assessment parameters corresponding to the sample p under normal flight conditions.

[0046] Execution threshold The settings improve the success rate of most normal flight commands, maintaining system availability; isolation threshold This marks the critical point at which the actual abnormal gains are maximized, and at this critical point, takeover must be implemented to carry out the flight mission.

[0047] like Figure 5 The figure shown illustrates a data-driven optimal control threshold learning mechanism. The top and left sides display the input data, historical experience data, and reliability assessment parameters required for offline learning. The middle section shows the data-driven offline learning process, which uses ROC curves combined with a benefit function that maximizes multiple considerations, including true positive rate (TPR), false positive rate (FPR), true negative rate (TNR), and false negative rate (FNR), to calculate the optimal threshold. The bottom output shows the automatically learned and set adaptive execution threshold and isolation threshold, overcoming the limitation of traditional traffic management RTA systems where overly strict threshold settings restrict system performance.

[0048] Step S32: Perform three-level monitoring and traffic management (RTA) dynamic control based on dual verification; integrate the consistency parameters of the UAV flight control logic from the input end. Reliability assessment parameters for deviation from flight maneuvers A dynamic degradation strategy is implemented based on the learned execution threshold and isolation threshold: Trust execution: If the drone flight control logic parameters are consistent Reliability assessment parameters for the degree of deviation from flight maneuvers All are above the execution threshold The system determines that the drone's flight control decision logic and flight control actions are highly consistent, and executes the drone's flight control action commands.

[0049] Isolation and takeover: If the drone flight control logic parameters are consistent Reliability assessment parameters for the degree of deviation from flight maneuvers Any parameter in the data is below the optimal isolation threshold. If the system determines that the flight control system has a logical abnormality or cognitive illusion, it will cut off the signal transmission link and switch to the underlying backup controller for forced takeover.

[0050] Action restrictions: If the drone flight control logic consistency parameters Reliability assessment parameters for the degree of deviation from flight maneuvers All are not lower than the optimal isolation threshold And not all are higher than the execution threshold This indicates a slight deviation in the drone's flight control decision-making logic. While retaining the commands, the flight control system proportionally reduces the amplitude of flight control actions, such as by 50%.

[0051] This invention analyzes the Hamming distance distribution and dual-validation quadrant in embodiments of the invention; in 500 independent tests targeting the internal activation state, the average Hamming distance of the flight state feature representation layer is... The value was 0.28, with a standard deviation of 0.12, indicating a flight action decision-making level. The deviation was 0.31, with a standard deviation of 0.14. Among them, the pitch mode had the lowest deviation at 0.26, while the yaw mode had the highest at 0.31. The consistency parameters of the human-machine flight control logic were then considered. Reliability assessment parameters for the degree of deviation from flight maneuvers Two-dimensional quadrant analysis was performed, and the correlation coefficient r between the two was obtained. The distribution of traffic management RTA decisions is 0.72, as shown in the table below: Table 3. Statistical distribution of traffic management RTA decisions based on 500 test samples. Compared to traffic management RTA that only uses a single feature input, the introduction of neural double verification increased the isolation takeover rate from about 10% to 16.8%, successfully intercepting a large number of high-risk hidden failures where the inputs appeared normal but the internal cognition was abnormal.

[0052] like Figure 6 The diagram illustrates a dual verification and dynamic control mechanism that integrates input logic and internal model cognitive deviations: the left side shows the activation extraction, logic consistency calculation, and offline standard flight action activation map comparison during online runtime, calculating the Hamming distance; the middle section shows the overall reliability assessment parameters generated based on the Hamming distance, reflecting the internal action deviations of the model; the right side shows the optimized optimal threshold, implementing trust execution and action restrictions based on logic consistency and overall reliability, such as reducing action amplitude by 50% or isolating and taking over, switching to the underlying backup controller, a three-level dynamic degradation strategy, blocking pre-emptive actions before physical failure occurs.

[0053] This invention provides an overview of the UAV flight process and verifies the confusion matrix. In this case, 800 test samples were generated (70% normal flight envelope, 30% boundary samples) for flight testing. Test results show that normal flight state data samples accounted for 84.3% (674 cases), while abnormal flight state data samples accounted for 15.7% (126 cases). The confusion matrix results show that approximately 550 true negatives were correct predictions, and approximately 80 true positives were correct predictions of danger, effectively intercepting a large number of dangerous boundary actions.

[0054] like Figure 12 The figure shows the consistency parameters based on the UAV flight control logic. Reliability assessment parameters for deviation from flight maneuvers The traffic management RTA decision space is divided into execution threshold and isolation threshold, where the isolation threshold is... and The execution thresholds are 0.375 and 0.412. and The values ​​are 0.668 and 0.791.

[0055] like Figure 13 The figure shows the time series of the consistency parameters of human-machine flight control logic and the reliability evaluation parameters of the deviation degree of flight actions under pitch violation scenarios. Two line types are used to represent the warning threshold of 0.6 and the danger threshold of 0.45, respectively. The overall reliability decreases before the actual pitch angle exceeds the limit, and a warning is triggered 8.80 seconds before the limit is exceeded; this indicates that the present invention can output control limitation signals in advance when risk accumulation occurs in the pitch channel, buying time for subsequent degraded control.

[0056] like Figure 14 The figure shows the time series of reliability assessment parameters for the consistency of human-machine flight control logic and the degree of deviation of flight maneuvers under angle-of-attack violation scenarios. The figure uses different line types to distinguish between the two types of assessment parameters and the threshold boundary. The system generates a risk warning 6.20 seconds before the angle of attack exceeds the limit, and subsequently both types of parameters continue to decrease and enter the danger zone; this indicates that during the formation of stall-related risks, the dual assessment parameters can simultaneously characterize input logic anomalies and internal cognitive deviations.

[0057] like Figure 15 The figure shows the evolution of physical parameters in multiple violation scenarios. The roll angle and dynamic pressure Q are represented by different line types, with the roll angle envelope boundary being ±30 degrees. The first violation stage is mainly manifested by the roll angle exceeding the limit, and the second violation stage is mainly manifested by the abnormal increase in dynamic pressure. This demonstrates that the present invention can map continuous abnormal events to the corresponding physical state evolution process.

[0058] like Figure 16 The figure shows the distribution of runtime traffic management (RTA) decisions based on 500 real-world tests. Various textures or fill styles are used to distinguish between three types of results: trusted execution, action restriction, and isolation takeover. This reflects the overall decision structure of the invention under large-sample testing, providing a statistical basis for subsequent decision space analysis.

[0059] The embodiments of the present invention verify the beneficial effects of the UAV flight control method based on neuron activation maps and gradient sensitivity through the following three aspects.

[0060] (1) Early warning capability. Based on the above... Figures 13-15 As shown, experimental results demonstrate that the proposed neuron activation monitoring method can provide effective early warning before flight envelope violations occur. In all three scenarios, the early warning time exceeds 6 seconds, with an average of 7.33 seconds, providing sufficient time for the runtime traffic management RTA system to prevent critical envelope violations. Human-machine flight control logic consistency parameters... The indicator decreased 5-10 seconds before the actual violation occurred, verifying the effectiveness of the neuron activation monitoring method; human-machine flight control logic consistency parameters Reliability assessment parameters for the degree of deviation from flight maneuvers The combination of these two technologies provides redundant monitoring, and both exhibit a correlated decreasing trend in all scenarios, reducing the false alarm rate.

[0061] (2) Continuous early warning capability in multiple violation scenarios. (The above text...) Figure 15 This demonstrates the system's ability to handle consecutive violations, with the first violation occurring between steps 40 and 80, at which point the roll angle exceeds 30 degrees. The value dropped below 0.6; the second violation occurred between steps 85 and 130, with dynamic pressure Q exceeding 80, affecting the consistency parameters of the drone flight control logic. The value dropped below 0.45 again, successfully identifying two violations and providing a distinguishing warning signal, demonstrating robustness in complex, multi-fault scenarios.

[0062] (3) Analysis of the characteristics of the flight control decision space. Figure 12 Demonstrates drone flight control logic consistency parameters and The joint distribution is as follows: Execution area, UAV flight control logic consistency parameters Reliability assessment parameters for flight maneuver deviation The concentration of samples here indicates that Deep Reinforcement Learning (DQN) operates under normal conditions.

[0063] Isolation zone, drone flight control logic consistency parameters or reliability assessment parameters for the degree of deviation from flight maneuvers In high-risk situations, immediate runtime traffic management (RTA) control is required.

[0064] Warning zones are areas other than normal and danger zones; they indicate a transitional state that requires monitoring and may limit decision-making.

[0065] The second aspect of this invention proposes a UAV flight control system based on a UAV flight control method using neuron activation maps and gradient sensitivity, comprising: a data-level processing gradient calculation module, a hidden layer neuron state extraction bit operation module, and a traffic decision module.

[0066] The data-level processing gradient calculation module is a computational component for handling the input and correlation of underlying physical signals. It transforms the continuous-time stream data returned from various hardware interfaces into a uniformly formatted input matrix and simultaneously analyzes the influence weights of various external physical parameters on the final issued commands. At the hardware computation level, it quantifies the actual interference of external environmental disturbances on the system command output, providing a quantitative basis for subsequent screening of abnormal data dependencies. The internal structure and function of each unit within the data-level processing gradient calculation module are as follows: The multi-channel signal synchronization and noise reduction unit is responsible for receiving the raw electrical signals from the inertial measurement unit, positioning module, and airspeed sensor, among other multi-channel hardware interfaces. It performs timestamp alignment and buffer management on the input data streams of each frequency and uses a low-pass filtering algorithm to remove abnormal high-frequency noise pulses, ensuring high fidelity in the data transmitted to the main control processor. The dimensionless elimination and matrix mapping unit receives the denoised multi-dimensional physical measurement values ​​and calls preset extreme value mapping parameters to proportionally scale the values ​​with various physical dimensions to a uniform dimensionless numerical range. This operation aims to eliminate the truncation error caused by differences in the underlying numerical amplitudes during subsequent partial derivative calculations. While the main control processor executes the forward instruction derivation cycle, the partial derivative matrix solving unit starts an automatic differentiation background process. It calculates the partial derivatives of the current instruction array with respect to the input state matrix, generating a sensitivity array that reflects the degree of parameter dependence. By monitoring the step amplitude of the values ​​in this array, it identifies in advance whether the control algorithm is approaching an unstable divergence boundary.

[0067] The hidden layer neuron state extraction bitwise operation module is a component for sampling and rapidly comparing the internal operating states of deep learning network structures. It departs from the conventional approach of only monitoring the final output instructions. By directly extracting the data array of intermediate computation layers and performing dimensionality reduction encoding, it performs high-frequency detection of the normality of the network's internal computational paths with extremely low processor clock cycle overhead. Internal structure and function of each unit: During the main control algorithm compilation and deployment phase, the network intermediate layer data extraction unit configures software-level data extraction points behind designated fully connected computation nodes. Within each control cycle of system operation, it directly reads and copies the floating-point numerical array output by that specific layer, obtaining the internal computational features of the algorithm processing the current data stream. The array dimensionality reduction and discretization encoding unit receives the extracted floating-point computation array and uses a preset truncation function to forcibly convert its internal continuous variables into discrete state values ​​composed of zeros and ones. This operation compresses high-dimensional features into a fixed-length binary encoded sequence, significantly reducing memory usage and bus bandwidth consumption in subsequent comparison operations. The bit logic operation difference measurement unit calls the currently generated binary encoded sequence and performs a hardware-level XOR logic operation with the reference encoded sequence pre-programmed in the system's read-only memory. By counting the specific number of non-zero state bits in the result sequence, a numerical index that can quantify the degree of difference between the current internal calculation path and the standard path is obtained, and this index is passed to the subsequent traffic decision module for final decision-making.

[0068] The traffic decision module is used for scheduling UAV control permissions and executing signals. Its significance lies in breaking the limitations of statically set boundaries on the availability of the control system and establishing dynamic operational boundaries based on offline data statistical characteristics. By jointly evaluating the matching degree of external data output from the front-end components and the difference in internal network state, the current control commands are precisely divided into different risk levels, and corresponding physical-level signal interventions are executed, maximizing the operating efficiency of the main control algorithm while maintaining the aircraft's attitude. The internal structure and functions of each unit of the traffic decision module are as follows: an offline feature statistical dual-boundary calibration unit, a dual-channel parameter synchronization interval determination unit, and a hierarchical response bus scheduling unit. The offline feature statistical dual-boundary calibration unit is responsible for processing historical flight test logs before system deployment. This unit uses statistical programs to draw classification performance distribution charts, finds the extreme point that maximizes the difference between the correct interception rate and the false alarm rate, and directly solidifies the quantized value corresponding to this extreme point as the system's hard isolation boundary. Simultaneously, this unit statistically analyzes the distribution of evaluation parameters for normal flight state samples, extracting specific quantile values ​​covering most normal operating conditions as the system's soft execution boundary, thereby improving the pass rate of control commands under normal flight conditions. During the online operation phase of the dual-channel parameter synchronization interval determination unit, the data matching parameters from the input end and the state difference parameters within the model are read synchronously. These two independent-dimensional parameters are then rapidly compared in absolute value with the soft execution boundary and hard isolation boundary pre-written into the non-volatile memory by the calibration unit to determine the risk level of the actuation command issued by the current main control computer. Based on the interval determination result, the hierarchical response bus scheduling unit directly controls the signal transmission channel of the aircraft's underlying servo mechanism, specifically including three control logics. The first is uninterrupted release: when both parameters are greater than the soft execution boundary, the main control link is deemed to be in good condition, the data bus remains unobstructed, and the original command is forwarded to the underlying driver at full amplitude. The second is signal attenuation limitation: when any parameter falls into the transition interval between the above two boundaries, the bus signal limiter is activated, and the original actuation command is attenuated and scaled by a specific ratio before being sent, suppressing potential excessive deflection. The third is physical isolation and takeover. When either of the two parameters falls below the hard isolation boundary, it is determined that the main control algorithm has seriously deviated from the calculation. The system immediately triggers the bus hard interrupt mechanism to block the algorithm's output path and activates the independently powered bottom-level backup control board to forcibly take over the aircraft's attitude maintenance task.

[0069] The drone flight control system can detect dangerous tendencies generated by the model due to internal cognitive deviations and trigger control 15 to 20 control cycles before the drone violates its physical envelope.

[0070] The beneficial effects of the embodiments of the present invention are as follows: The embodiments of the present invention propose a UAV flight control method and system based on neuron activation maps and gradient sensitivity, which can identify early signs of abnormal feature mappings within the model before the physical flight state of the UAV deteriorates substantially, providing sufficient time margin for takeover and degraded control; it can avoid the large-scale sampling iteration required by traditional proxy models, control the consumption of computing resources and processing latency, and strictly meet the constraints of the high-frequency control cycle of the airborne flight control computer; it effectively reduces the redundancy false alarm and false interception rate of the system under complex boundary situations, and effectively improves the task execution efficiency of the decision algorithm while maintaining the flight state of the UAV.

[0071] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A UAV flight control method based on neuron activation maps and gradient sensitivity, characterized in that: It includes: S1: Set the UAV flight state vector and flight maneuver response parameters Quantitative sensitivity of UAV flight state characteristic parameters ; Gaussian mixture models are used to mine the flight control logic of UAVs; decision logic parameters for flight states are selected based on the sensitivity of flight state characteristic parameters. The key feature set corresponding to the flight action mode is obtained. Determine the logical consistency parameters of the drone. Assess the consistency of the drone flight control logic; S2: The model's internal flight actions are represented by monitoring semantically information-dense hidden layers. Flight state data samples are grouped according to flight action modalities, and the neurons in each hidden layer are identified. Activation frequency in flight action mode Extract the activation vectors of neurons in a specified hidden layer within the UAV flight control decision model, and generate a standard flight action activation map through binarization; calculate the Hamming distance using the standard flight action activation map corresponding to the flight action mode constructed by the flight control logic evaluation, and then generate reliability evaluation parameters for the degree of flight action deviation. Construct a data-driven activation map of hidden layer neurons for standard flight actions; S3: Determine the UAV flight control logic consistency parameters obtained in step S1 Corresponding optimal isolation threshold The reliability assessment parameters of the degree of deviation of flight maneuvers obtained in step S2 Corresponding optimal isolation threshold Double cross-validation is performed on the UAV flight control decisions; based on the optimal adaptive control threshold obtained through data-driven learning from the working characteristic curve, the consistency parameters of the UAV flight control logic are adjusted accordingly. Reliability assessment parameters for deviation from flight maneuvers Independent judgment is made, a multi-level dynamic control strategy is implemented, and control commands are dynamically adjusted to achieve multi-level drone flight control.

2. The UAV flight control method based on neuron activation maps and gradient sensitivity according to claim 1, characterized in that: Step S1 is as follows: S11: Set the UAV flight state vector Specifically, this includes: pitch angle Angle of attack Height deviation Pitch rate q, roll angle Roll rate m and heading angle and flight maneuver response parameters Quantitative sensitivity of UAV flight state characteristic parameters ; S12: Yes Sensitivity vector of feature parameters for each flight state sample Establish a Gaussian mixture model and use the Gaussian mixture model (GMM) to explore the flight control logic of the UAV. S13: Selecting decision logic parameters for flight status based on the sensitivity of flight status characteristic parameters. The key feature set corresponding to the flight action mode is obtained. Determine the logical consistency parameters of the drone. .

3. The UAV flight control method based on neuron activation maps and gradient sensitivity according to claim 2, characterized in that: Sensitivity of UAV flight state characteristic parameters in step S11 for: ； in, For the first Flight status characteristic parameters of a drone Sensitivity; The sign for partial derivatives; For the first Values ​​of flight status characteristic parameters of each UAV; The optimal action response parameters output by the UAV flight control system; This is the index of the flight state characteristic parameters of the first UAV; This represents the total number of characteristic parameters of the UAV's flight status. This is a sensitivity constant to avoid situations where the denominator is 0; For the first The covariance matrix of each UAV flight action mode; For the first Values ​​of flight status characteristic parameters of each UAV; This is the index of the flight status characteristic parameters of the second UAV.

4. The UAV flight control method based on neuron activation maps and gradient sensitivity according to claim 2, characterized in that: UAV logical consistency parameters in step S13 for: ； in, For consistency parameters of UAV flight control logic; For auxiliary parameters of flight control logic; For the first Sensitivity to individual UAV flight status characteristic parameters; This is a set of key characteristic parameters for flight control. For the first A set of key characteristic parameters for flight control; These are the decision logic parameters for flight status.

5. The UAV flight control method based on neuron activation maps and gradient sensitivity according to claim 1, characterized in that: Step S2 is as follows: S21: Set up multi-level UAV flight control monitoring points, select the flight state feature representation layer in the UAV flight control decision model, and use the threshold function to generate the binary activation map of the UAV flight control monitoring points; S22: Group the flight status data samples according to flight maneuver modes to form groups corresponding to specific flight maneuver modes. Flight status dataset ; Identify each hidden layer neuron Activation frequency in flight action mode ; S23: Generate a standard flight maneuver activation map based on flight control logic evaluation, and calculate the reliability evaluation parameters for the degree of flight maneuver deviation. .

6. The UAV flight control method based on neuron activation maps and gradient sensitivity according to claim 5, characterized in that: Step S21 is as follows: ； in, For the current flight status data sample in the th The first monitoring layer The binary activation state of a neuron; For the first Monitoring layer The output vector of the flight state feature representation layer for each neuron; For indicator functions; For the first Monitoring layer Number of neurons; For flight control monitoring layer index; This is an index of neurons within the flight control monitoring layer.

7. The UAV flight control method based on neuron activation maps and gradient sensitivity according to claim 5, characterized in that: The standard flight maneuver activation map in step S23 is as follows: ； in, For flight action modes Lower monitoring layer The Middle The standard activation binary state of each neuron; This is a conditional function; For the first Monitoring layer The activation frequency of each neuron; These are the activation parameters for standard flight maneuvers.

8. The UAV flight control method based on neuron activation maps and gradient sensitivity according to claim 5, characterized in that: Reliability assessment parameters for the degree of deviation of flight maneuvers in step S23 for: ; ; in, Reliability assessment parameters for the degree of deviation from flight maneuvers; Hamming distance for the flight state feature representation layer; Hamming distance for the flight maneuver decision-making layer; The weighting coefficients for the flight state feature representation layer; Weighting coefficients for flight action decision-making; The result is the Hamming distance calculation. This is a function to count the number of activation states of 1 in a binary vector; This represents the activation feature vector of the currently flying sample in the monitoring layer. The standard flight motion activation map vector for the corresponding motion mode, which is fixed in the offline phase; This represents the total number of neurons involved in the calculation within the flight control monitoring layer. This is the neuron activation state vector.

9. The UAV flight control method based on neuron activation maps and gradient sensitivity according to claim 1, characterized in that: Step S3 is as follows: S31: Obtain UAV flight control logic consistency parameters Corresponding optimal isolation threshold Reliability assessment parameters for deviation from flight maneuvers Corresponding optimal isolation threshold ; Based on the flight status sample set Determine the consistency parameters of UAV flight control logic Reliability assessment parameters for deviation from flight maneuvers The corresponding execution threshold; optimize the dual-channel adaptive control threshold based on data-driven approaches; S32: Unmanned Aerial Vehicle (UAV) flight control logic consistency parameters at the fusion input. Reliability assessment parameters for deviation from flight maneuvers Based on dual verification, three levels of monitoring are implemented for dynamic traffic management and control.

10. A UAV flight control system for the UAV flight control method based on neuron activation maps and gradient sensitivity as described in any one of claims 1 to 9, characterized in that: It includes: The module includes a data-level processing gradient calculation module, a hidden layer neuron state extraction bit operation module, and a traffic decision module. The data-level processing gradient calculation module connects the multi-channel signal synchronization and noise reduction unit in the UAV flight control process to the inertial measurement unit, positioning module, and airspeed sensor; performs timestamp alignment and cache management on the input data streams of each frequency, and removes abnormal high-frequency noise pulses through a low-pass filtering algorithm to ensure that the data transmitted to the main control processor has high fidelity. The hidden layer neuron state extraction bit operation module samples and compares the internal operating state of the deep learning network structure at high speed, directly extracts the data array of the intermediate operation layer of the network and performs dimensionality reduction encoding. The hidden layer neuron state extraction bit operation module includes: a network intermediate layer data truncation unit, an array dimensionality reduction discretization encoding unit, and a bit logic operation difference measurement unit; the network intermediate layer data truncation unit reads and copies the floating-point numerical array output by the specific layer, and the internal operation characteristics when processing the current data stream; the array dimensionality reduction discretization encoding unit receives the truncated floating-point operation array, and uses a preset truncation function to force the continuous variables into discrete state values ​​composed of zeros and ones; the bit logic operation difference measurement unit calls the currently generated binary encoding sequence to quantify the numerical index that represents the degree of difference between the current internal operation path and the standard path; The traffic decision module is used for scheduling UAV control authority and executing signals. Specifically, it includes: an offline feature statistics dual-boundary calibration unit, a dual-channel parameter synchronization interval determination unit, and a hierarchical response bus scheduling unit. The offline feature statistics dual-boundary calibration unit statistically analyzes the distribution of evaluation parameters for samples in normal flight states to improve the pass rate of control commands under normal flight states. The dual-channel parameter synchronization interval determination unit determines the risk level of the actuation commands issued by the current main control computer. The hierarchical response bus scheduling unit directly controls the signal transmission channels of the aircraft's underlying servo mechanisms based on the interval determination results.

Images