Deep Learning-Based UAV Atmospheric Data Anomaly Detection and Correction System

Abstract

This invention relates to the field of data detection technology, and more particularly to a deep learning-based system for detecting and correcting anomalies in UAV atmospheric data. The proposed scheme involves constructing an air mass reference coordinate system based on the UAV's attitude and relative wind direction, and reprojecting the original measurement data to obtain an atmospheric data sequence with a stable reference direction. Subsequently, a time-varying frequency mask is generated to constrain the location of rotor disturbance energy during time-frequency analysis. Then, based on the multi-channel frequency mask, the time-frequency coefficient matrix is ​​decomposed into multiple scales to extract aerodynamic disturbance estimation signals, background atmospheric signals, and turbulent structure signals. These three types of signals are input into a pre-set deep learning model to identify and correct atmospheric data anomalies. This application can distinguish between rotor disturbances and the actual atmospheric structure, improving the quality and reliability of atmospheric observation data.

Description

Technical Field

[0001] This application relates to the field of data detection technology, and in particular to a deep learning-based system for detecting and correcting anomalies in UAV atmospheric data. Background Technology

[0002] Unmanned aerial vehicles (UAVs) have been widely used for rapid acquisition of low-altitude atmospheric elements due to their high mobility and flexible deployment. However, current technologies generally suffer from insufficient validity and stability in atmospheric measurement data acquired by UAVs. First, during flight missions, the periodic disturbances generated by the rotor itself are superimposed on the real atmospheric signal, introducing a large amount of interference from non-atmospheric sources into the measurement results. Existing methods generally struggle to accurately distinguish between aircraft disturbances and natural turbulence signals in high-frequency and harmonic components, leading to structural distortion in the final observation results. Second, atmospheric data acquired by multi-source sensors under complex attitude changes often exhibit inconsistent directions and scales, reducing the correlation between measurement results and further weakening the reliability of the data. In addition, traditional data filtering and anomaly detection strategies are mostly based on fixed thresholds or static frequency domain rules, which are insufficiently adaptable to rapidly changing conditions during flight. This means that even after correction, outliers may still remain or excessive smoothing may occur, affecting the accurate inversion of atmospheric parameters.

[0003] To address the above issues, this application presents a deep learning-based system for detecting and correcting anomalies in UAV atmospheric data. Summary of the Invention

[0004] The technical problem addressed in this application is to overcome the shortcomings of existing technologies by providing a deep learning-based system for detecting and correcting anomalies in UAV atmospheric data. The system constructs an air mass reference coordinate system based on the UAV's attitude and relative wind direction, and reprojects the original measurement data to obtain an atmospheric data sequence with a stable reference direction. Subsequently, it combines motor speed, attitude angular velocity, and throttle changes to construct a flight condition intensity index, generating a time-varying frequency mask to constrain the localization of rotor disturbance energy during time-frequency analysis. Then, based on the multi-channel frequency mask, the time-frequency coefficient matrix is ​​decomposed into multiple scales to extract aerodynamic disturbance estimation signals, background atmospheric signals, and turbulent structure signals. Finally, these three types of signals are input into a pre-set deep learning model to identify and correct atmospheric data anomalies. This application can distinguish between rotor disturbances and the actual atmospheric structure, improving the quality and reliability of UAV-borne atmospheric observation data.

[0005] To achieve the above objectives, this application provides the following technical solution:

[0006] A deep learning-based UAV atmospheric data anomaly detection and correction system, the system comprising:

[0007] The data preprocessing module is used to collect raw measurement data through the UAV, construct an air mass reference coordinate system based on the attitude of the UAV and the relative wind direction, reproject the raw measurement data onto the air mass reference coordinate system, and generate an atmospheric data sequence.

[0008] The data decomposition module is used to obtain the motor speed of the UAV, calculate the corresponding frequency mask, and perform time-frequency analysis on the atmospheric data sequence to obtain the time-frequency coefficient matrix. Based on the frequency mask, the time-frequency coefficient matrix is ​​decomposed into a multi-scale decomposition to obtain the aerodynamic disturbance estimation signal, the background atmospheric signal, and the turbulent structure signal.

[0009] The data detection module is used to input the aerodynamic disturbance estimation signal, background atmospheric signal and turbulent structure signal as input parameters into a preset deep learning model, and process the input parameters through the deep learning model to obtain atmospheric data anomaly detection results and corresponding correction data.

[0010] The data preprocessing module includes:

[0011] The wind direction decomposition unit performs attitude transformation on the body coordinate system based on the pitch angle, roll angle and yaw angle acquired by the UAV in real time, and obtains the decomposition vector of the relative wind direction in the body coordinate system;

[0012] The coordinate construction unit is used to construct an air mass reference coordinate system with the direction of the decomposed vector as the first reference axis, the normal direction of the first reference axis on the horizontal plane as the second reference axis, and the vertical direction of the first reference axis as the third reference axis, based on the relative flow direction, the horizontal lateral direction, and the vertical direction.

[0013] The data reprojection unit is used to reproject the original measurement data according to the air mass reference coordinate system, and generate an atmospheric data sequence according to the time order of each physical quantity after reprojection. The reprojection transforms the change in observation direction corresponding to the movement of the UAV body to the air mass reference coordinate system to generate a unified reference direction that moves with the air mass.

[0014] The data preprocessing module is configured with a reprojection strategy, which includes reprojection logic, and the reprojection logic is configured within the data reprojection unit.

[0015] The original measurement data includes wind speed vector, temperature, humidity, air pressure, and particulate matter concentration. The reprojection logic is configured as follows:

[0016] The wind speed vector is oriented by a preset attitude transformation matrix, and the projection components of the corrected wind speed vector on the first reference axis, the second reference axis and the third reference axis are calculated. The three projection components are weighted and summed to obtain the wind speed component.

[0017] Based on the reprojection direction of the wind speed component, the measurement data of temperature, humidity, air pressure and particulate matter concentration are oriented to obtain the observed components.

[0018] The continuity correction is performed based on the changing trend of the observed components at adjacent time points, wherein the continuity correction is implemented by an interpolation smoothing algorithm.

[0019] The data decomposition module includes:

[0020] The frequency mask calculation unit generates frequency mask parameters to distinguish between rotor disturbance frequency bands and non-disturbance frequency bands based on the UAV motor speed and flight condition information;

[0021] The time-frequency analysis unit is used to perform time-frequency transformation on the atmospheric data sequence to generate a time-frequency coefficient matrix that characterizes different time and frequency components;

[0022] A multi-scale decomposition unit is used to perform multi-scale decomposition on the time-frequency coefficient matrix under the constraints of the frequency mask parameters.

[0023] The data decomposition module includes a data decomposition strategy, which includes mask calculation logic and decomposition logic. The mask calculation logic is configured within the frequency mask calculation unit, and the decomposition logic is configured within the multi-scale decomposition unit.

[0024] The mask calculation logic is configured as follows:

[0025] The motor speeds of the UAV at different times are obtained, a speed sequence is generated, and the corresponding basic rotor frequency sequence and speed change rate are calculated based on the speed sequence.

[0026] The intensity index used to characterize the flight condition is calculated based on the rotational speed change rate and the pre-collected attitude angular velocity, throttle command change, and airspeed change of the UAV.

[0027] Based on the intensity index, a frequency mask centered on the basic rotor frequency and its corresponding harmonic in the basic rotor frequency sequence is determined.

[0028] Determining a frequency mask centered on the fundamental rotor frequency and its corresponding harmonics in the fundamental rotor frequency sequence based on the intensity index includes:

[0029] Based on the aforementioned basic rotor frequency sequence, determine the basic rotor frequency and the center frequency of the fundamental rotor frequency harmonics at each time point;

[0030] The target bandwidth is determined within a preset bandwidth range based on the intensity index. When the intensity index is less than a preset index threshold, a first bandwidth is selected. When the intensity index is greater than the preset index threshold, a second bandwidth is selected. The frequency coverage range of the first bandwidth is less than the frequency coverage range of the second bandwidth.

[0031] Centered on the center frequency, a corresponding frequency range is determined based on the target bandwidth, and weight values ​​representing the rotor disturbance probability are allocated within the corresponding frequency range.

[0032] A frequency mask that varies over time is constructed based on the center frequency, target bandwidth, and weight values.

[0033] The decomposition logic is configured as follows:

[0034] The time-frequency coefficient matrix is ​​marked as a multi-channel time-frequency coefficient matrix according to the channel division, wherein the number of channels is the same as the number of sensors;

[0035] Based on the frequency mask and the spatial layout relationship between each sensor and the rotor, a channel weight is assigned to each channel, and the frequency mask is expanded into a multi-channel frequency mask defined in the channel dimension, time dimension and frequency dimension according to the channel weight.

[0036] Under the constraint of the multi-channel frequency mask, the multi-channel time-frequency coefficient matrix is ​​decomposed into background atmospheric coefficient tensor, aerodynamic disturbance coefficient tensor and turbulent structure coefficient tensor, wherein the algebraic sum of each coefficient tensor at each channel, each time and each frequency point is equal to the original time-frequency coefficient.

[0037] The background atmospheric coefficient tensor is subjected to low-rank and smooth constraints in the channel dimension and time dimension, the aerodynamic disturbance coefficient tensor is subjected to mask weighted sparsity constraints, and the turbulent structure coefficient tensor is subjected to broad-spectrum energy distribution constraints.

[0038] By performing time-frequency inverse transforms on the constrained background atmospheric coefficient tensor, aerodynamic disturbance coefficient tensor, and turbulent structure coefficient tensor, the aerodynamic disturbance estimation signal, background atmospheric signal, and turbulent structure signal are obtained.

[0039] Under the constraint of the multi-channel frequency mask, the multi-channel time-frequency coefficient matrix is ​​decomposed into a background atmospheric coefficient tensor, an aerodynamic disturbance coefficient tensor, and a turbulent structure coefficient tensor, including:

[0040] Based on the multi-channel frequency mask, different mask weights are assigned to the time-frequency coefficient matrices of each channel, each time, and each frequency point. The corresponding aerodynamic disturbance confidence, background field confidence, and turbulence confidence are calculated according to the mask weights. The mask weights increase sequentially according to the difference in Euclidean distance between the frequency point and the center frequency.

[0041] In each channel, a decomposition weight vector is constructed based on the aerodynamic disturbance confidence, background field confidence, and turbulence confidence to characterize the background atmospheric component, aerodynamic disturbance component, and turbulence component, and the time-frequency coefficients are weighted and decomposed using the decomposition weight vector.

[0042] Consistency constraints are applied to the allocated time-frequency coefficients in the channel dimension, time dimension, and frequency dimension. If the decomposition error in the channel dimension, time dimension, and frequency dimension is greater than or equal to the decomposition threshold of any dimension, the decomposition weight vector is updated according to the decomposition error of the corresponding dimension, and the time-frequency coefficients are redistributed until all are less than the decomposition threshold.

[0043] After the time-frequency coefficients at each time-frequency point satisfy the aforementioned consistency constraint, the background atmospheric component, aerodynamic disturbance component, and turbulent component corresponding to each time-frequency point are aggregated according to their channel, time, and frequency position to obtain the background atmospheric coefficient tensor, aerodynamic disturbance coefficient tensor, and turbulent structure coefficient tensor.

[0044] The deep learning model includes:

[0045] The feature extraction layer is used to extract feature information from the aerodynamic disturbance estimation signal, the background atmospheric signal, and the turbulent structure signal, respectively.

[0046] The feature fusion layer is used to fuse various feature information through an attention mechanism, assign dynamic weights to each feature according to a frequency mask, process the fused features, and generate a comprehensive feature vector.

[0047] The decision layer is used to input the comprehensive feature vector into a preset two-layer fully connected neural network, output a multi-dimensional detection vector through a linear activation function, and correct the multi-dimensional detection vector through a Monte Carlo method, outputting atmospheric data anomaly detection results and corresponding corrected data.

[0048] The process involves outputting a multidimensional detection vector using a linear activation function, and then correcting the multidimensional detection vector using a Monte Carlo method to output atmospheric data anomaly detection results and corresponding corrected data, including:

[0049] The detection components in the multidimensional detection vector are perturbed and sampled, and multiple sets of sampled detection vectors are generated based on a preset Gaussian distribution.

[0050] For each sampled detection vector, the rate of change of response on the background atmospheric signal, the estimated aerodynamic disturbance signal, and the turbulent structure signal is calculated respectively. If the rate of change of response is within the preset physical range, the corresponding detection component is marked as a valid sample corresponding to the real atmospheric change; otherwise, it is marked as an invalid sample.

[0051] For the detection components that are marked as valid samples, the original output is maintained. For the detection components that are marked as invalid samples, the expected value of the detection component is calculated as the corrected detection vector based on the corresponding perturbation distribution, and the corrected detection vector is output as the correction result.

[0052] Compared with the prior art, the beneficial effects of this application are:

[0053] This application introduces an air mass reference coordinate system before data acquisition, ensuring that multi-source observations maintain a consistent directional representation under different attitude conditions, thereby reducing observation offsets caused by aircraft motion. Furthermore, by constructing a dynamic frequency mask using rotational speed sequences and flight condition intensity indices, rotor disturbances are controllably constrained in the time-frequency domain, improving the separation accuracy between aircraft disturbances and natural turbulence. Based on this, a multi-channel constrained time-frequency decomposition method is employed, enabling background atmosphere, aerodynamic disturbances, and turbulent structure signals to be independently expressed in tensor form. This provides deeper, more physically semantic input to the deep model, enhancing the reliability of atmospheric measurement data. Attached Figure Description

[0054] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0055] Figure 1 An exemplary application scenario diagram provided for an embodiment of this application;

[0056] Figure 2 A schematic diagram of the modules of the UAV atmospheric data anomaly detection and correction system based on deep learning provided in the embodiments of this application;

[0057] Figure 3 This is a flowchart illustrating the deep learning-based method for detecting and correcting anomalies in UAV atmospheric data provided in this application embodiment. Detailed Implementation

[0058] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0059] The term "embodiment" as used herein means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0060] With drones gradually becoming a routine means of atmospheric sounding and environmental monitoring, this application does not focus on how to simply install more sensors, but rather on a fundamental problem that has long troubled those skilled in the art:

[0061] Atmospheric data collected by UAVs in complex aerodynamic environments are already affected by aerodynamic disturbances, thermal disturbances, and non-ideal characteristics of sensors. The challenge is to identify and correct these measurement anomalies while maintaining the true atmospheric structure.

[0062] With the widespread use of multi-rotor, small fixed-wing, and compound-wing UAVs in scenarios such as boundary layer vertical profile observation, typhoon peripheral wind field detection, urban canyon pollution distribution monitoring, and local wind field assessment of highways or ports, it is difficult to simply categorize the anomalies that frequently appear in the output of airborne atmospheric sensors as real extreme weather or measurement errors.

[0063] In exemplary techniques, common practices typically fall into two categories:

[0064] One type is a quality control method based on empirical thresholds, sliding windows, and smoothing filters. By setting upper and lower limits and rate of change thresholds for time series such as wind speed and temperature, points that exceed the empirical range are judged as abnormal and interpolated or filtered.

[0065] Another approach involves introducing a conventional deep learning model, directly inputting the raw or simply preprocessed sensor sequences into the neural network, which then learns abnormal patterns and provides normal / abnormal or corrected values.

[0066] Understandably, these two types of methods can be effectively applied in static or weakly disturbed scenarios, but they share common challenges in typical UAV application environments:

[0067] On the one hand, non-atmospheric factors such as rotor downwash, airframe attitude maneuvering, and heat dissipation of airborne electronic equipment are highly coupled with real atmospheric processes such as turbulent bursts, local wind shear, and frontal passage in the time and frequency domains. Empirical thresholds can easily mistake real small-scale structures for noise. On the other hand, although end-to-end deep networks have fitting capabilities, they often confuse airframe disturbances with real atmospheric changes in the absence of physical constraints. Overfitting to individual scenarios during training may also lead to failure in extreme weather or new scenarios. It is difficult for those skilled in the art to explain the model's judgment criteria at specific outliers.

[0068] Especially in multi-rotor platforms, the rotor's fundamental frequency and its harmonics form a significant narrowband structure in atmospheric measurement signals. When flight missions involve rapid climbs, dives, sharp turns, or strong wind control, the drastic changes in motor speed and attitude cause these narrowband frequencies to drift rapidly or even broaden over time. Simultaneously, the UAV's relative wind direction constantly changes, causing the effective sampling direction of the sensors to continuously twist between the airframe and Earth coordinate systems. Traditional time-series analysis based on fixed coordinates struggles to maintain consistency in spatial physical meaning. In such scenarios, simply filtering on the time axis or clipping a fixed frequency band in the spectrum cannot cleanly remove aerodynamic disturbances and easily destroys the already sparse information about the true atmospheric structure, leading to subsequent deep learning models passively learning in a physically corrupted signal space.

[0069] Understandably, the root of the problem lies not in whether deep learning models are good enough, but in the fact that traditional solutions generally lack a physically meaningful data classification system specifically for UAV atmospheric sounding scenarios.

[0070] It can eliminate observation direction disturbances introduced by the aircraft attitude at the coordinate level, and can also use the rotor aerodynamic characteristics to structurally decompose the signal at the frequency and scale levels, so as to separate the aircraft disturbance, background atmosphere and turbulent structure in a physically interpretable way before entering the deep learning model.

[0071] Based on this approach, the technical solution proposed in this application does not rely on a specific aircraft model, fixed sensor layout, or a certain type of weather system. Instead, it addresses the widespread UAV atmospheric three-dimensional observation scenarios characterized by rotor aerodynamic disturbances, intense attitude maneuvers, and multi-source coupling of observation variables. By constructing an air mass reference coordinate system that varies with the relative incoming flow, the original measurement data is reprojected, allowing atmospheric variables to be expressed in a unified reference system that moves with the air mass. Furthermore, by combining an adaptive frequency mask derived from motor speed and flight conditions, the reprojected signal is decomposed into multiple scales in the time and frequency domain, decoupling the narrowband energy that is highly correlated with the rotor frequency from the low-frequency background field and the mid-to-high-frequency turbulence structure.

[0072] Furthermore, the deep learning model in this application does not need to directly deal with the original time series that is mixed with the disturbance of the organism and the actual atmospheric changes. Instead, it receives the aerodynamic disturbance estimation signal, background atmospheric signal and turbulent structure signal that have been divided according to physical roles. By jointly modeling the characteristics of these three types of signals, it focuses on identifying the correctable measurement anomalies caused by sensor failure, attitude abnormality, instantaneous occlusion, etc., while maintaining the uncorrected output for the real extreme processes that conform to the characteristics of atmospheric dynamics.

[0073] It should be noted that the anomaly detection and correction mechanism proposed in this application is not customized for a specific type of UAV or a specific type of observation task, and its applicable application scenarios include, but are not limited to:

[0074] Scenarios involving vertical profile detection in the low boundary layer with significant airflow effects; scenarios where rotor disturbances coexist with natural turbulence when performing missions near strong convection, the outer periphery of typhoons, or complex terrain; and scenarios where atmospheric data can only be obtained from a limited perspective in areas with limited visibility, such as between urban high-rises, canyons, or the sea, and where sensors are easily obstructed by the aircraft structure.

[0075] refer to Figure 1 , Figure 1 This is an exemplary application scenario diagram provided for an embodiment of this application.

[0076] Figure 1 The diagram shows how a drone collects data using sensor components, obtains raw measurement data, and then sends the raw measurement data to a ground data station.

[0077] It is understood that the sensor components in this application may include, but are not limited to, any type of airborne sensor used to measure wind speed, temperature, humidity, air pressure, particulate matter concentration or other atmospheric parameters. The number of sensors, installation method and layout can be flexibly set according to the specific UAV platform and mission requirements, and this application does not limit them.

[0078] also, Figure 1 The ground data station shown can be understood as any type of data processing terminal used to receive, store, and process various atmospheric observation data reported by UAVs. In specific implementations, the ground data station may include at least one processor, which is configured with the deep learning-based UAV atmospheric data anomaly detection and correction system of this application, used to perform functions such as raw data preprocessing, aerodynamic disturbance decomposition, and anomaly detection and correction. The processor may be an application-specific integrated circuit (ASIC), a programmable logic device (FPGA), a digital signal processor (DSP), a general-purpose central processing unit (CPU), or other hardware devices with data processing capabilities, or it may be a distributed processing system composed of multiple processing units, which is not limited in this application.

[0079] refer to Figure 2 , Figure 2 A schematic diagram of the modules of the deep learning-based UAV atmospheric data anomaly detection and correction system provided in the embodiments of this application.

[0080] In one example, this application provides a deep learning-based UAV atmospheric data anomaly detection and correction system, the system comprising:

[0081] The data preprocessing module is used to collect raw measurement data through the UAV, construct an air mass reference coordinate system based on the attitude of the UAV and the relative wind direction, reproject the raw measurement data onto the air mass reference coordinate system, and generate an atmospheric data sequence.

[0082] The data decomposition module is used to obtain the motor speed of the UAV, calculate the corresponding frequency mask, and perform time-frequency analysis on the atmospheric data sequence to obtain the time-frequency coefficient matrix. Based on the frequency mask, the time-frequency coefficient matrix is ​​decomposed into a multi-scale decomposition to obtain the aerodynamic disturbance estimation signal, the background atmospheric signal, and the turbulent structure signal.

[0083] The data detection module is used to input the aerodynamic disturbance estimation signal, background atmospheric signal and turbulent structure signal as input parameters into a preset deep learning model, and process the input parameters through the deep learning model to obtain atmospheric data anomaly detection results and corresponding correction data.

[0084] The data preprocessing module includes:

[0085] The wind direction decomposition unit performs attitude transformation on the body coordinate system based on the pitch angle, roll angle and yaw angle acquired by the UAV in real time, and obtains the decomposition vector of the relative wind direction in the body coordinate system;

[0086] The coordinate construction unit is used to construct an air mass reference coordinate system with the direction of the decomposed vector as the first reference axis, the normal direction of the first reference axis on the horizontal plane as the second reference axis, and the vertical direction of the first reference axis as the third reference axis, based on the relative flow direction, the horizontal lateral direction, and the vertical direction.

[0087] The data reprojection unit is used to reproject the original measurement data according to the air mass reference coordinate system, and generate an atmospheric data sequence according to the time order of each physical quantity after reprojection. The reprojection transforms the change in observation direction corresponding to the movement of the UAV body to the air mass reference coordinate system to generate a unified reference direction that moves with the air mass.

[0088] The data preprocessing module is configured with a reprojection strategy, which includes reprojection logic, and the reprojection logic is configured within the data reprojection unit.

[0089] The data decomposition module includes:

[0090] The frequency mask calculation unit generates frequency mask parameters to distinguish between rotor disturbance frequency bands and non-disturbance frequency bands based on the UAV motor speed and flight condition information;

[0091] The time-frequency analysis unit is used to perform time-frequency transformation on the atmospheric data sequence to generate a time-frequency coefficient matrix that characterizes different time and frequency components;

[0092] A multi-scale decomposition unit is used to perform multi-scale decomposition on the time-frequency coefficient matrix under the constraints of the frequency mask parameters.

[0093] The data decomposition module includes a data decomposition strategy, which includes mask calculation logic and decomposition logic. The mask calculation logic is configured within the frequency mask calculation unit, and the decomposition logic is configured within the multi-scale decomposition unit.

[0094] In some optional implementations, embodiments of this application provide a method for detecting and correcting anomalies in UAV atmospheric data based on deep learning.

[0095] refer to Figure 3 , Figure 3 This is a flowchart illustrating the deep learning-based method for detecting and correcting anomalies in UAV atmospheric data provided in this application embodiment.

[0096] Figure 3 The methods shown include:

[0097] S1: Collect raw measurement data using a drone, construct an air mass reference coordinate system based on the drone's attitude and relative wind direction, reproject the raw measurement data onto the air mass reference coordinate system, and generate an atmospheric data sequence.

[0098] In this embodiment, the original measurement data includes, but is not limited to, real-time observations such as wind speed vector, temperature, humidity, air pressure, and particulate matter concentration. Since UAVs often experience attitude changes such as pitch, roll, and yaw during mission flights, and the relative wind direction continuously changes in the aircraft's coordinate system, directly using the aircraft's system sampling values ​​for analysis would lead to coupling between the real atmospheric signal and the changes in the observed direction. This embodiment first calculates the decomposed vector of the relative wind direction in the aircraft's coordinate system based on real-time attitude information, and then constructs an air mass reference coordinate system based on the incoming direction, horizontal lateral direction, and vertical direction. This allows the original measurement results to be reprojected onto a unified reference direction that moves with the air mass, thus achieving a physical decoupling of the background atmospheric signal, turbulent structure, and aircraft disturbance at the coordinate level.

[0099] Those skilled in the art will understand that this reprojection method can minimize directional errors caused by attitude maneuvers, thus establishing a stable, continuous, and physically meaningful data foundation for subsequent time-frequency analysis.

[0100] S2: Obtain the rotational speed of the UAV motor, calculate the corresponding frequency mask, and perform time-frequency analysis on the atmospheric data sequence to obtain the time-frequency coefficient matrix;

[0101] In this embodiment, real-time changes in motor speed are used to determine the rotor's fundamental frequency and its harmonics. In a multi-rotor structure, rotor disturbances in the airflow exhibit a significant narrow band characteristic, drifting or widening depending on flight conditions. Direct spectral or time-frequency analysis of atmospheric data would result in rotor disturbances being superimposed on high-frequency components of natural turbulence, making effective separation difficult under traditional filtering conditions. This embodiment constructs a flight condition intensity index based on the rate of change of rotational speed, as well as variations in the UAV's attitude angular velocity and throttle commands. Based on this, the bandwidth of the frequency mask is dynamically adjusted to accurately cover the region where rotor disturbance energy is located. Under this mechanism, the mask automatically widens under strong maneuvering conditions and automatically converges under stable cruise conditions, making the coefficient matrix obtained from time-frequency analysis more physically controllable.

[0102] It should be noted that the specific data form of the atmospheric data sequence in this application can be understood as a multidimensional physical quantity sequence arranged in chronological order. The types of variables, sampling frequency, and number of channels included can vary depending on the UAV mission configuration, as long as they meet the minimum requirements for obtaining the time-frequency distribution characteristics based on the sequence. This application does not impose any special limitations on this. Specific time-frequency analysis methods can employ continuous wavelet transform, short-time Fourier transform, or other time-frequency transformation methods suitable for processing non-stationary signals. Those skilled in the art can choose the specific implementation method according to actual needs, and this application will not elaborate on this.

[0103] Furthermore, the time-frequency coefficient matrix obtained through time-frequency analysis can be understood as a two-dimensional or multi-channel three-dimensional representation structure after expanding the atmospheric data sequence in the time and frequency dimensions. Each coefficient point carries energy information at the corresponding time and frequency, which can reflect the common distribution of the background atmospheric gradual change trend, the narrow band structure of rotor aerodynamic disturbance, and the multi-scale random characteristics of turbulence.

[0104] S3: Based on the frequency mask, the time-frequency coefficient matrix is ​​decomposed into multi-scale components to obtain the aerodynamic disturbance estimation signal, the background atmospheric signal, and the turbulent structure signal;

[0105] In this embodiment, time-frequency coefficients from different sensors or different physical quantities are jointly modeled in the channel dimension, time dimension, and frequency dimension. Using a weighted field formed by an adaptive frequency mask, the energy of the rotor frequency band is mainly allocated to the aerodynamic disturbance channel, the low-frequency slowly varying energy is allocated to the background atmosphere channel, and the mid-to-high frequency multi-scale random structure is mapped to the turbulent structure channel. Simultaneously, low-rank constraints are applied to the background component to maintain spatial consistency; mask-weighted sparsity constraints are applied to the aerodynamic disturbance to highlight its dominant frequency characteristics; and broad-spectrum energy distribution constraints are applied to the turbulent component to better conform to the spectral shape of the natural turbulence model. This multi-scale decomposition not only achieves mathematically reversible decomposition but also physically decouples the background atmosphere, turbulence, and body disturbances, allowing subsequent deep learning models to perform anomaly identification in a defined physical semantic space, rather than dealing with chaotic raw quantities.

[0106] S4: Input the aerodynamic disturbance estimation signal, background atmospheric signal and turbulent structure signal as input parameters into a preset deep learning model, and process the input parameters through the deep learning model to obtain atmospheric data anomaly detection results and corresponding correction data;

[0107] In this embodiment, the deep learning model does not directly process the raw measurement data. Instead, it receives three types of feature signals after physical decoupling, giving the model input inherent attributes such as background trend, turbulent randomness, and narrowband aerodynamic disturbance. This embodiment introduces a feature extraction layer to encode the three types of signals separately, and uses an attention mechanism in the feature fusion layer to dynamically adjust the weights of each signal based on a frequency mask, enabling the model to adaptively focus on more reliable physical components under different flight conditions. Subsequently, the decision layer outputs a multi-dimensional detection vector, and verifies whether the detection results conform to atmospheric dynamic characteristics through a Monte Carlo perturbation sampling mechanism. Components judged as measurement anomalies are corrected, while the original output remains unchanged for natural turbulent abrupt changes or real weather phenomena.

[0108] Those skilled in the art will understand that attitude angles and attitude angular velocities can be obtained from the inertial navigation unit or flight control attitude calculation output; motor speeds can be obtained from the speed estimates returned by the electronic speed controller (ESC) or from speed sensors; throttle command changes can be obtained from the flight control unit's control of the ESC; and airspeed changes can be obtained from the airspeed meter or from GPS / IMU fusion estimation. The above data are recorded with a unified timestamp and aligned to the same sampling period using linear interpolation or spline interpolation. When missing samples exist, they are filled in using nearest neighbor hold or local interpolation to ensure continuous input required for subsequent frequency mask construction.

[0109] Before delving into the specific operational logic, it is necessary to explain the unique characteristics of the atmospheric data processing link in light of the long-standing measurement uncertainties in UAV atmospheric observations.

[0110] Unlike fixed ground-based meteorological measurements, UAVs move continuously in the airspace with unsteady attitudes, resulting in a much more complex airflow structure around their fuselage compared to stationary observation environments. Particularly in multi-rotor platforms, factors such as rotor downwash, arm obstruction, local vortex structures, and equipment self-heating directly alter the boundary layer morphology near the sensors. This causes the sensor readings to include both the intrinsic characteristics of the external atmosphere and perturbations from the coupling of the aircraft's motion and attitude. For applications aiming to reconstruct the true atmospheric field through UAV observations, relying solely on simple smoothing or empirical thresholds is insufficient for effective differentiation, a well-recognized and prevalent technical challenge in this field.

[0111] Furthermore, UAV measurement data is typically collected simultaneously in the form of multiple variables, including wind speed, air pressure, temperature, humidity, and other extended atmospheric elements. Although these variables are physically related, they are affected by differences in their respective sensor installation locations, sampling rates, and degrees of disturbance. Data collected at the same time often exhibit different levels of disturbance and time-varying sensitivities. If these variables are not structurally interpreted during data preprocessing and anomaly identification, subsequent modeling processes may become confused at the underlying logic level, or even misinterpret disturbances from different sources as consistent meteorological upheavals, thereby reducing the reliability of the overall detection chain.

[0112] For the reasons mentioned above, the overall design of the processing logic in this application tends to first structure the raw information before inputting the data into the deep learning model. This allows the body disturbances, background atmospheric trends, and turbulent randomness to form a certain hierarchical relationship in their expression, thereby enabling subsequent deep learning inferences to unfold based on clearer physical semantics. This hierarchical approach is not a reconstruction of the real atmospheric structure, but rather a reasonable decomposition of coupled signals, allowing disturbances from different sources to exhibit relatively independent and identifiable feature patterns in mathematical space.

[0113] Based on this, by performing multi-scale characterization, energy convergence and divergence behavior analysis, and cross-variable consistency modeling on signals from different sources, the model can not only capture whether the values ​​themselves deviate when performing anomaly detection, but also determine whether the deviation may be caused by real atmospheric phenomena by combining the physical scene of the signal.

[0114] For example, when high-frequency disturbances occur in the turbulent structure but the background component remains continuous, it is more likely to originate from sensor noise or machine vibration; conversely, when both the turbulence and background components show a synchronous gradual change trend, it is more likely to be related to changes in the spatial structure of the atmosphere itself. Through this logical differentiation, the accuracy of identifying measurement anomalies can be improved without mistakenly omitting real atmospheric changes.

[0115] Next, we will further elaborate on the technical aspects of the method in this application regarding the generation of atmospheric data sequences.

[0116] In one example, generating the atmospheric data sequence includes:

[0117] S1.1: Based on the pitch angle, roll angle and yaw angle acquired in real time by the UAV, the attitude transformation of the body coordinate system is performed to obtain the decomposition vector of the relative wind direction in the body coordinate system;

[0118] Specifically, during flight, the attitude of a UAV constantly changes with control commands, environmental disturbances, or mission maneuvers, causing the relative relationship between the UAV's coordinate system and the ground reference system to be in a state of continuous dynamic adjustment. Since atmospheric measurement sensors are generally fixed to the UAV's structure, their sampling direction is always affected by the UAV's attitude. If the wind speed or other vector physical quantities collected by the sensors are directly regarded as stable directional quantities, it will cause coupling between wind direction information and the UAV's attitude, thus making the original measurement values ​​unable to accurately reflect the true direction of the incoming airflow. To ensure that atmospheric data has consistent spatial semantics in subsequent processing, the UAV's coordinate system needs to be corrected according to its real-time attitude, enabling it to map the observation direction under the instantaneous attitude to a stable reference relationship.

[0119] S1.2: Using the direction of the decomposed vector as the first reference axis, the normal direction of the first reference axis on the horizontal plane as the second reference axis, and the vertical direction of the first reference axis as the third reference axis, construct an air mass reference coordinate system with the relative inflow direction, horizontal lateral direction and vertical direction as the reference.

[0120] Specifically, the body coordinate system is a coordinate system that changes in real time with the drone's attitude, and its coordinate directions lack stable reference meaning. If the body coordinate system is directly used to express airflow direction or turbulent disturbances, the same atmospheric structure will exhibit different directional projections under different attitudes, making it difficult for subsequent atmospheric process analysis and deep learning models to learn based on a unified data semantics. To enable atmospheric variables to be expressed under a stable directional reference, a reference coordinate system bound to the atmosphere itself needs to be introduced, which can adaptively change with the direction of airflow, giving the measurement data continuity in describing physical processes within this coordinate system.

[0121] In this embodiment, the normalized relative wind direction decomposition vector is set as the direction vector of the first reference axis. To construct the second reference axis, the first reference axis is projected onto the horizontal plane and its normal direction is calculated, enabling it to represent lateral airflow or lateral disturbances, thus giving the coordinate system the ability to describe crosswinds. The third reference axis is obtained by constructing a direction vector that is orthogonal to both the first and second reference axes, enabling it to represent the vertical movement or change of the air mass. To ensure that the three directions maintain strict orthogonality, they can be orthogonalized again after calculating the third reference axis, so that the resulting air mass reference coordinate system conforms to the requirements of a right-handed coordinate system. This coordinate system can be dynamically updated as the incoming flow direction changes, thus ensuring that atmospheric variables maintain a consistent reference direction throughout different flight phases.

[0122] S1.3: The original measurement data is reprojected according to the air mass reference coordinate system, and atmospheric data sequence is generated by arranging the physical quantities after reprojection in time order. The reprojection is performed by converting the change in observation direction corresponding to the movement of the UAV body to the air mass reference coordinate system to generate a unified reference direction that moves with the air mass.

[0123] Specifically, while gas motion is continuous, the attitude and flight trajectory of a UAV change rapidly, causing the original measurements to appear as vectors with constantly shifting directions in the UAV's coordinate system. If data without standardized orientation is used directly in the analysis, the same atmospheric process will exhibit discontinuous directional jumps in the data sequence, affecting the stability of time-frequency analysis and deep learning models. To ensure that atmospheric variables are expressed around the true direction of air mass motion, all physical quantities need to be projected onto an air mass reference coordinate system, binding their directions to the atmospheric motion itself and independent of the UAV's motion. Therefore, wind speed, temperature gradient, and other vector components observed in the UAV's coordinate system are projected and transformed according to the orientation of the reference coordinate system, so that their components on the three reference axes reflect the true changes in airflow structure.

[0124] In this embodiment, the raw measurement data includes wind speed vector, temperature, humidity, air pressure, and particulate matter concentration.

[0125] In one example, reprojecting the original measurement data according to the air mass reference coordinate system includes:

[0126] The wind speed vector is oriented by a preset attitude transformation matrix, and the projection components of the corrected wind speed vector on the first reference axis, the second reference axis and the third reference axis are calculated. The three projection components are weighted and summed to obtain the wind speed component.

[0127] Understandably, when correcting the direction of wind speed vectors, traditional three-dimensional wind field measurement techniques typically employ attitude angle-driven direction transformations to convert the wind direction observed by the sensor from the aircraft coordinate system to a more stable reference coordinate system. This process essentially relies on rotating the spatial vector using an attitude transformation matrix to eliminate directional disturbances caused by changes in the aircraft's attitude. The attitude transformation matrix is ​​generally constructed by rotating the pitch, roll, and yaw angles along the coordinate axes in a specific order. This embodiment also follows this mature existing technological path, mapping the wind speed vector to a target coordinate system independent of the aircraft's instantaneous attitude through continuous rotation. During projection calculations, the wind speed vector is projected onto the three principal axes of the reference coordinate system, ensuring that each component represents the atmospheric motion trend in its corresponding direction. This approach can be considered a vector decomposition method commonly used in existing wind direction observation fields, ensuring orthogonality among the three calculated components and making wind speed variations independent in different directions.

[0128] Based on the reprojection direction of the wind speed component, the measurement data of temperature, humidity, air pressure and particulate matter concentration are oriented to obtain the observed components.

[0129] In this embodiment, although temperature, humidity, and air pressure do not possess vector attributes in existing sensor systems, their sampling results are still affected by sensor orientation, ventilation flow, and installation location. Therefore, after reprojecting the wind speed, conventional methods adjust the scalar observations for directional consistency by incorporating changes in wind direction, ensuring that their time series corresponds to sampling behavior of the same atmospheric volume.

[0130] This application calibrates the response time of temperature, humidity, and air pressure by changing the principal direction of the wind speed vector, thereby eliminating the problem of sampling surface changes caused by the rotation of the aircraft. This method is a common processing technique in small UAV observation systems, aiming to ensure consistency in sampling time, spatial representativeness, and sampling direction among multiple physical quantities, and will not be elaborated upon here.

[0131] The continuity correction is performed based on the changing trend of the observed components at adjacent time points, wherein the continuity correction is implemented by an interpolation smoothing algorithm;

[0132] It should be noted that after orientation correction is completed, due to factors such as flight attitude, propulsion system vibration, and sensor noise, the reprojected data often exhibits discontinuous jumps within a short period of time.

[0133] In this embodiment, by calculating the changing trends of each physical quantity over continuous time slices after reprojection, the data is made differentiable and continuous on a shorter time scale, thereby avoiding unnecessary high-frequency spurious signals in subsequent time-frequency analysis. This type of continuity correction is a basic operation in the field of data processing, which makes the data more consistent with the gradual characteristics of real atmospheric processes in the time domain, and will not be elaborated upon here.

[0134] Next, the technical content of the frequency mask in this application will be further elaborated.

[0135] Understandably, in this application, the frequency mask is used to isolate the energy bands associated with the rotor in order to distinguish the disturbances introduced by the airframe from the actual atmospheric turbulence structure.

[0136] It should be noted that in the unsteady flight environment of multi-rotor UAVs, conventional masking methods often fail to provide stable signal separation capabilities in actual missions.

[0137] On the one hand, the rotor fundamental frequency of a multi-rotor platform exhibits significant time drift characteristics as mission thrust, load changes, and attitude adjustments occur. Traditional masking methods based on fixed narrow-band windows are difficult to maintain effective coverage of the frequency position during changes, and are prone to missed detections or offsets during high-speed maneuvers or wind-resistant hovering.

[0138] On the other hand, in complex wind fields, the frequency range of rotor disturbances will broaden or narrow with airflow interference and motor transient response, making it impossible for traditional equal-width masks to adapt to the changes in spectrum morphology in sync, which in turn affects the ability of subsequent decomposition steps to identify aerodynamic disturbance energy.

[0139] Furthermore, conventional frequency masks are often constructed based on a single frequency feature, while atmospheric data contains multiple physical quantities, and their responses in the frequency domain vary considerably.

[0140] For example, wind speed is significant in rotor frequency doubling structures, while the energy intensity of temperature or air pressure at the same frequency band is usually weaker. If a uniform mask is used directly, it may lead to inconsistent processing of different physical quantities in the frequency domain, destroying the structural correspondence between multivariable data after masking. Furthermore, the perturbation coupling of each physical quantity in the UAV platform is time-varying and has a nonlinear trend. Traditional frequency masks do not consider this intervariable relationship, which often results in the rotor perturbation of one physical quantity being masked in actual data processing, while another physical quantity retains a spurious signal in the same frequency band, causing inconsistencies in energy distribution among multi-source measurements.

[0141] In engineering applications, the update frequency of the frequency mask is also a key factor affecting its reliability. Traditional masks are mostly based on a fixed refresh period or a static frequency point. However, under complex mission conditions, the motor speed of UAVs changes at high frequencies, and the drift speed of the rotor frequency position is significantly faster than the mask refresh period. This makes it impossible for the mask to accurately correspond to the energy region in the actual signal, resulting in a large number of unmasked aerodynamic disturbance components in the time-frequency coefficient matrix. At the same time, if the mask is updated too frequently, it is easy to cause the frequency domain boundary to jitter drastically over time, introducing additional discontinuities and interferences, which manifest as pseudo-turbulence or pseudo-structural jumps in the multi-scale decomposition process.

[0142] In one example, obtaining the drone motor speed and calculating the corresponding frequency mask includes:

[0143] S2.1: Obtain the motor speed of the UAV at different times, generate a speed sequence, and calculate the corresponding basic rotor frequency sequence and speed change rate based on the speed sequence;

[0144] Specifically, in atmospheric measurement tasks using multi-rotor UAVs, the frequency of periodic disturbances generated by the rotor is directly related to the motor speed, which exhibits significant time-varying characteristics due to load changes, attitude adjustments, wind speed disturbances, and control commands. If the speed value is only read at a single moment for subsequent frequency positioning, misjudgments often occur when frequency drift is rapid, preventing the frequency mask from accurately covering the true interference frequency band. Therefore, it is necessary to construct a time-varying speed sequence through continuous sampling to reflect the dynamic changes in the motor's operating state, and to deduce the rotor's fundamental frequency sequence and its harmonic positions. Furthermore, the amplitude and rate of change of the speed can reveal the operating conditions of the UAV, such as whether it is hovering against the wind, rapidly climbing, or gliding. Therefore, the rate of change of the speed has direct physical significance in subsequently determining the intensity of the operating conditions and the mask bandwidth.

[0145] In this embodiment, the flight controller acquires instantaneous speed data of multiple motors at a fixed sampling period, and weights and fuses the speeds of different motors to obtain a final speed sequence that represents the average rotor speed fluctuation of the entire platform. To ensure the continuity of the sequence, a time synchronization mechanism is used to align the data of each motor, avoiding inconsistencies caused by sampling delays. After obtaining the speed sequence, the speed values ​​are converted into basic rotor frequencies through structured lookup tables or empirical calibration relationships, and a basic rotor frequency sequence is generated according to the same time index. Then, by calculating and normalizing the speed difference between adjacent time points, the speed change rate, which reflects the degree of transient speed change, is obtained, enabling it to describe the strength of the motor's dynamic response within a short time window.

[0146] In one optional implementation, the base rotor frequency is obtained from the rotational speed sequence through calibration mapping; the harmonic center frequency is extended in integer multiples, and the harmonic order can be set to 3–6, or set to the maximum order not exceeding the highest frequency in the time-frequency analysis. The target bandwidth is selected within a preset range: a first bandwidth is used under stable operating conditions to form a compact frequency band coverage, and a second bandwidth is used under intense operating conditions to broaden the coverage spectrum; the bandwidth is smoothed on the time axis using a sliding window to avoid boundary jitter. The weights are set with high values ​​at the center frequency band, gradually decreasing at the frequency band edges, and a transition guard band is set outside the center frequency to make the weighted suppression of the time-frequency coefficients continuously change.

[0147] S2.2: Calculate the intensity index to characterize the flight condition based on the rotational speed change rate and the pre-collected attitude angular velocity, throttle command change and airspeed change of the UAV.

[0148] Specifically, when UAVs perform atmospheric sounding missions, changes in motor speed are only a partial manifestation of the strength of aerodynamic disturbances to the airframe. The complexity of actual operating conditions is often also affected by factors such as attitude angular velocity, throttle variation, and airspeed variation.

[0149] For example, when a drone is in a rapid roll or yaw adjustment, even if the rotational speed does not change much, the frequency harmonic structure generated by the airflow disturbance of the rotor will still broaden. It is necessary to comprehensively model the changes of multiple physical quantities to obtain operating condition indicators that can represent the complexity of flight dynamics, so that the frequency mask can make adaptive adjustments to the actual interference level.

[0150] In this embodiment, the angular velocity data provided by the attitude sensor, the command changes output by the throttle controller, and the instantaneous airspeed changes from the airspeed measurement unit are all centralized on a unified time axis. Through normalization, scale alignment, and time synchronization, dynamic changes from different sources can be received by the same operational condition judgment mechanism. When calculating the intensity index of the flight condition, the intensity index reflects whether the UAV is in a stable flight, making minor adjustments, or undergoing high-dynamic maneuvers by comprehensively evaluating the amplitude, rate of change, and directional consistency of the above variables. Essentially, the intensity index is a dynamic reference quantity obtained from multi-sensor fusion, and its magnitude directly reflects the intensity of aerodynamic interference and the possibility of spectral broadening.

[0151] S2.3: Determine a frequency mask centered on the basic rotor frequency and its corresponding harmonics in the basic rotor frequency sequence based on the intensity index;

[0152] Specifically, traditional frequency masks typically use a fixed bandwidth set around a fundamental frequency, but this approach is prone to failure in the unsteady environment of multi-rotor drones.

[0153] Under high-maneuverability conditions, the fundamental frequency position drifts rapidly, and a fixed mask cannot effectively cover it; under low-dynamic conditions, a fixed bandwidth will block too much high-frequency atmospheric information. Therefore, the mask bandwidth needs to be closely related to the flight conditions, so that the mask can dynamically change with the fundamental rotor frequency and its harmonics. The frequency mask not only needs to cover the fundamental frequency, but also needs to be extended according to the broadening of the harmonic structure; otherwise, rotor disturbances at higher harmonics will remain in the original signal, affecting subsequent decomposition.

[0154] In one example, determining a frequency mask centered on the fundamental rotor frequency and its corresponding harmonics in the fundamental rotor frequency sequence based on the intensity index includes:

[0155] Based on the aforementioned basic rotor frequency sequence, determine the basic rotor frequency and the center frequency of the fundamental rotor frequency harmonics at each time point;

[0156] The target bandwidth is determined within a preset bandwidth range based on the intensity index. When the intensity index is less than a preset index threshold, a first bandwidth is selected. When the intensity index is greater than the preset index threshold, a second bandwidth is selected. The frequency coverage range of the first bandwidth is less than the frequency coverage range of the second bandwidth.

[0157] Centered on the center frequency, a corresponding frequency range is determined based on the target bandwidth, and weight values ​​representing the rotor disturbance probability are allocated within the corresponding frequency range.

[0158] A frequency mask that varies over time is constructed based on the center frequency, target bandwidth, and weight values.

[0159] In this embodiment, the target bandwidth is selected within a preset bandwidth range based on the intensity index. For the low intensity range, a first bandwidth is selected to create a compact coverage structure of the mask near the fundamental frequency, preserving as much true atmospheric mid-frequency and high-frequency information as possible. For the high intensity range, a second bandwidth is selected to create an extended coverage of the mask around the fundamental frequency and its harmonics, thus adapting to the spectral broadening of the harmonic structure. In practical applications, the bandwidth range can be determined empirically based on historical mission data, while the dynamic changes in the intensity index allow for real-time bandwidth adjustment, giving the mask both flexibility and stability.

[0160] It is important to note that the selection of the first and second bandwidths in this application can be achieved by analyzing the statistical intervals of rotor disturbance frequency distribution in historical flight missions. Specifically, this can be quantified based on the spectral energy ratio or the half-power width in the frequency domain. In engineering practice, energy density analysis can be performed on time-frequency maps collected from multiple flights. By calculating the inflection point of the energy decay curve in the neighborhood of the base rotor frequency, the typical concentration interval of rotor disturbance under stable operating conditions can be obtained, which can be used as the suggested range for the first bandwidth. In high-maneuverability conditions, the diffusion degree of the harmonic structure on the time axis can be observed, and the upper limit of the second bandwidth can be determined by statistically broadening the energy envelope boundary. Furthermore, to ensure that the bandwidth setting is not affected by individual mission anomalies, a quantile strategy can be used during the statistical process to screen the energy distribution, so that the first bandwidth corresponds to the energy concentration segment at lower disturbance levels, while the second bandwidth corresponds to the common range of spectral broadening under high disturbance conditions, thus ensuring the robustness and mission adaptability of the bandwidth allocation.

[0161] Furthermore, with the center frequency as the axis, disturbance probability weights are assigned to each frequency point within the target bandwidth coverage area, so that the weight values ​​are maximized near the center of the frequency band and gradually decrease at the edges. The weight sequence can be constructed using a piecewise linear or exponential structure to accommodate energy attenuation patterns under different operating conditions. After construction, the center frequency, target bandwidth, and weight values ​​are combined in a time series to form a mask matrix that dynamically changes over time, i.e., a frequency mask, ensuring spatial continuity and temporal stability in the time-frequency domain. In use, the mask matrix applies weighted suppression to the time-frequency coefficients, reducing the impact of rotor interference regions according to the weight changes, rather than completely shielding them.

[0162] Next, we will further elaborate on the technical content of the multi-scale decomposition method in this application.

[0163] In one example, the time-frequency coefficient matrix is ​​decomposed into multiple scales based on the frequency mask to obtain an aerodynamic disturbance estimation signal, a background atmospheric signal, and a turbulent structure signal, including:

[0164] S3.1: The time-frequency coefficient matrix is ​​marked as a multi-channel time-frequency coefficient matrix according to the channel division, wherein the number of channels is the same as the number of sensors;

[0165] Specifically, during UAV atmospheric observations, different sensors typically exhibit time delays, differences in noise composition, and biases in the sensitive direction in their responses to the same atmospheric disturbance. Simply merging the time-frequency coefficients of all sensors leads to aliasing, making it impossible to maintain consistency across physical quantities in subsequent decomposition. Therefore, the original time-frequency coefficient matrix needs to be explicitly divided into multiple channels before decomposition, with each channel corresponding to a single sensor or a single set of physical quantities. This allows the decomposition process to preserve sensor independence along the channel dimension while restoring physical consistency after decomposition.

[0166] In this embodiment, the time-frequency coefficients from different sensors such as wind speed, temperature, humidity, and air pressure are labeled with channels according to their sampling sources. Each channel records the time-frequency energy changes of the sensor across the entire time and frequency dimensions. To maintain index consistency between channels, a unified time axis and a unified frequency division are used, and the time-frequency graphs of different sensors are interpolated and aligned to ensure that the matrices have a strict correspondence in three-dimensional space.

[0167] S3.2: Based on the frequency mask and the spatial layout relationship between each sensor and the rotor, assign channel weights to each channel, and expand the frequency mask into a multi-channel frequency mask defined in the channel dimension, time dimension and frequency dimension according to the channel weights;

[0168] Specifically, the different mounting positions of various sensors on the airframe structure result in varying intensities of rotor disturbance received. For example, a wind speed sensor closer to the rotor disk receives stronger harmonic energy than a sensor located at the end of the arm; temperature sensors, due to their slower response time, often exhibit weaker signal amplitudes in the harmonic region. If sensor layout is not considered in multi-channel decomposition, some channels may over-absorb rotor energy, while others may be excessively attenuated. Therefore, it is necessary to assign adaptive weights to each channel based on the spatial relationship between the sensor and the rotor, allowing the frequency mask to have a differentiated effect on each sensor channel, thus ensuring that the energy distribution during decomposition better conforms to physical laws.

[0169] In this embodiment, by collecting UAV structural geometry data, the positions of sensors closer to the rotor and more aligned with the propeller surface are defined as high-weight channels; while sensors farther from the rotor or facing the crosswind direction are set as low-weight channels. Then, the frequency mask is extended along the channel dimension according to the channel weight, so that the mask has a stronger shielding effect in the high-weight channels and retains more of the real atmospheric structure in the low-weight channels, making the influence of the mask spatially correlated along the channel dimension.

[0170] S3.3: Under the constraint of the multi-channel frequency mask, the multi-channel time-frequency coefficient matrix is ​​decomposed into background atmospheric coefficient tensor, aerodynamic disturbance coefficient tensor and turbulent structure coefficient tensor, wherein the algebraic sum of each coefficient tensor at each channel, each time and each frequency point is equal to the original time-frequency coefficient.

[0171] Specifically, atmospheric signals are composed of background atmospheric trends, turbulent structures, and aircraft disturbances, but these three components have different scales, distributions, and correlations in the time-frequency domain. Directly decomposing multi-channel time-frequency coefficients without constraints can lead to energy overlap, component mixing, or unclear physical meanings. Therefore, it is necessary to combine multi-channel masks to confine the main energy of rotor disturbances to the frequency range corresponding to the mask, ensuring a clear energy attribution relationship in the tensor decomposition process and strengthening the physical independence of the three types of components from a mathematical structure perspective.

[0172] In one example, under the constraint of the multi-channel frequency mask, the multi-channel time-frequency coefficient matrix is ​​decomposed into a background atmospheric coefficient tensor, an aerodynamic disturbance coefficient tensor, and a turbulent structure coefficient tensor, including:

[0173] Based on the multi-channel frequency mask, different mask weights are assigned to the time-frequency coefficient matrices of each channel, each time, and each frequency point. The corresponding aerodynamic disturbance confidence, background field confidence, and turbulence confidence are calculated according to the mask weights. The mask weights increase sequentially according to the difference in Euclidean distance between the frequency point and the center frequency.

[0174] In each channel, a decomposition weight vector is constructed based on the aerodynamic disturbance confidence, background field confidence, and turbulence confidence to characterize the background atmospheric component, aerodynamic disturbance component, and turbulence component, and the time-frequency coefficients are weighted and decomposed using the decomposition weight vector.

[0175] Consistency constraints are applied to the allocated time-frequency coefficients in the channel dimension, time dimension, and frequency dimension. If the decomposition error in the channel dimension, time dimension, and frequency dimension is greater than or equal to the decomposition threshold of any dimension, the decomposition weight vector is updated according to the decomposition error of the corresponding dimension, and the time-frequency coefficients are redistributed until all are less than the decomposition threshold.

[0176] After the time-frequency coefficients at each time-frequency point satisfy the aforementioned consistency constraint, the background atmospheric component, aerodynamic disturbance component, and turbulent component corresponding to each time-frequency point are aggregated according to their channel, time, and frequency position to obtain the background atmospheric coefficient tensor, aerodynamic disturbance coefficient tensor, and turbulent structure coefficient tensor.

[0177] Understandably, when performing three-component decomposition of multi-channel atmospheric data, traditional methods often treat time-frequency coefficients as independent points and assign them using thresholding or energy-based methods, without considering the differentiated effects of frequency masks at different channels, time slices, and frequency points. This can easily lead to problems such as component mixing, cross-channel energy mismatch, and blurred boundaries between turbulence and body disturbances. This embodiment uses a time-frequency weight field composed of multi-channel frequency masks to give the energy distribution of each time-frequency point in different channels a structural tendency, thereby making tensor decomposition mathematically physically interpretable. Specifically, by linking the weight changes of the mask in the frequency dimension to the distance from the center frequency, the mask weight at each frequency point exhibits a concentrated suppression trend in the narrow band region and gradually weakens away from the center frequency, so that the time-frequency coefficients can naturally converge towards aerodynamic disturbances or natural atmospheric components according to the spectral morphology.

[0178] In this embodiment, after the mask weight field is constructed, to prevent the energy at each time-frequency point from being allocated through a simple threshold or fixed ratio, this embodiment calculates three types of confidence levels using mask weights, giving each time-frequency point a probabilistic weight distribution that is more likely to belong to the background atmosphere, more likely to belong to rotor disturbances, and more likely to belong to turbulent structures. Subsequently, within each channel, a decomposed weight vector is constructed based on the confidence level for the time-frequency coefficients, so that different frequency positions have a decomposition trend consistent with the physical process.

[0179] For example, in the narrow band region where the mask peak is located, the confidence of rotor disturbance naturally increases, allowing the aerodynamic disturbance component to occupy a higher weight in this region; while in the low-frequency stable range, the background confidence is high, making the background trend continuous in the time dimension; for the high-frequency region, the turbulence confidence is dominant, allowing small-scale structures to be aggregated into the turbulence tensor.

[0180] After allocation, to avoid cross-dimensional inconsistencies caused by a single allocation, this embodiment performs consistency checks on the three types of components in the channel dimension, frequency dimension, and time dimension respectively, ensuring that each time-frequency point maintains an energy conservation relationship in three-dimensional space. When a deviation in a certain dimension exceeds a set threshold, the decomposition weight vector of the corresponding dimension is updated, so that the energy gradually approaches the stable solution after iterative updates, ultimately forming a tensor with structural stability and clear physical meaning for the three types of components.

[0181] Furthermore, to ensure that the final background, disturbance, and turbulence tensors are fully represented in a three-dimensional structure, this embodiment aggregates the components at each time-frequency point at the corresponding channel, time slice, and frequency position after decomposition and stabilization, giving the three types of components a complete three-dimensional representation in tensor form. Through the aggregation operation, the background atmospheric tensor exhibits cross-channel consistency and temporal trend stability; the aerodynamic disturbance tensor forms a narrow-band distribution at the center frequency and its harmonics, consistent with the rotor characteristics; and the turbulence tensor exhibits random and broad-spectrum energy distribution characteristics in the high-frequency region.

[0182] S3.4: Apply low-rank and smooth constraints on the channel dimension and time dimension to the background atmospheric coefficient tensor, apply mask weighted sparsity constraints to the aerodynamic disturbance coefficient tensor, and apply broad-spectrum energy distribution constraints to the turbulence structure coefficient tensor.

[0183] Specifically, in the natural atmosphere, the background atmosphere exhibits temporal continuity and consistency across sensor directions; rotor disturbances have a narrow and sparse structure; and turbulence displays broadband, irregular, and high-frequency randomness. Without imposing these physical constraints on the tensor dimension, the decomposition results, while mathematically valid, become physically ambiguous. Therefore, it is necessary to transform physical characteristics into constraints and introduce them into the tensor structure to make the decomposition results more consistent with real atmospheric dynamics.

[0184] In this embodiment, the background atmospheric tensor is constrained by a low-rank channel dimension to ensure consistency across different sensor channels; simultaneously, a time-dimensional smoothing constraint is applied to ensure the trend term exhibits continuous change over time. The aerodynamic disturbance tensor is subject to sparsity constraint based on the mask, concentrating rotor interference energy in a specific narrow band region and gradually attenuating at higher harmonics. The turbulent structure tensor is subjected to a broad-spectrum constraint, resulting in a relatively uniform energy distribution in the high-frequency region and preventing pseudo-structures from appearing in the frequency band.

[0185] S3.5: Perform time-frequency inverse transform on the constrained background atmospheric coefficient tensor, aerodynamic disturbance coefficient tensor, and turbulent structure coefficient tensor to obtain the aerodynamic disturbance estimation signal, background atmospheric signal, and turbulent structure signal;

[0186] In this embodiment, an inverse transform method corresponding to the forward analysis is used to restore the background tensor, disturbance tensor, and turbulence tensor to time series, respectively. During the inverse transform, the consistency of amplitude and phase is maintained to ensure that the background atmospheric signal exhibits a smooth trend, the aerodynamic disturbance signal exhibits a periodic narrow-band structure, and the turbulence signal exhibits random high-frequency characteristics.

[0187] In one optional implementation, the time-frequency coefficients of each channel obtained from time-frequency analysis are complex coefficients, containing amplitude and phase information. To ensure reversible recovery to the time-domain signal after decomposition, the decomposition process employs a weighted distribution of amplitude while maintaining phase stability.

[0188] The three types of components at the same time frequency share the original phase, and only the amplitude is split and written into the coefficient tensors of background, disturbance and turbulence respectively.

[0189] The multi-channel time-frequency coefficient matrix can be understood as a three-dimensional tensor structure:

[0190] The channel dimension corresponds to the sensor or physical quantity channel, the time dimension corresponds to the center time index of the time-frequency window, and the frequency dimension corresponds to the discrete frequency points. Different channels may have differences in sampling frequency or time-frequency resolution. Before decomposition, the time-frequency plots of each channel are unified to the same time axis and frequency grid.

[0191] The time axis can use a uniform step size (e.g., consistent with the time-frequency window step size), and the frequency axis can use a uniform frequency interval. For misaligned channels, linear interpolation is used to map their coefficients to a uniform grid; for missing moments, adjacent moment interpolation or nearest neighbor preservation is used to fill in the gaps, ensuring a one-to-one correspondence between the three-dimensional indices.

[0192] For any harmonic center frequency at any given time, first determine the target bandwidth, and then assign mask weights to each frequency point within that bandwidth. The engineering meaning of mask weights is: the closer a frequency point is to the center bandwidth, the more likely it is to belong to the rotor disturbance region, thus assigning higher weights to the disturbance components in subsequent confidence calculations.

[0193] For ease of implementation, the mask weights can be constructed using piecewise linear methods:

[0194] Within the coverage frequency band, the weight gradually increases from the edge of the band to the center, and the weight outside the band is zero. When there are multiple octave center points at the same time, the maximum value or weighted sum of the weights of each octave mask at that frequency point is taken (in engineering, the maximum value can be taken to ensure coverage).

[0195] Next, a specific and complete example will be used to illustrate the entire process. The example is only for illustrating the feasibility of the calculation. The number of frequency points and parameters can be adjusted according to the actual task and do not represent the actual values.

[0196] Parameter settings: Center frequency at a certain moment: 120 Hz (which can be understood as the fundamental frequency or the center point of a certain harmonic); Target bandwidth: 20 Hz, half bandwidth is 10 Hz; Frequency point set: 110, 116, 120, 124, 130, 150 Hz;

[0197] Two channels: Channel 1 is closer to the drone's propeller disk, with a channel weight of 1.0; Channel 2 is farther from the propeller disk, with a channel weight of 0.7.

[0198] This application also provides a specific and easy-to-understand prior rule (which can be reproduced through lookup-based engineering rules):

[0199] Low frequency (≤5 Hz): background is too high, turbulence is too low (in this example, the frequency points are all above this range, so it is not used);

[0200] Mid-frequency (5–30 Hz): Background and turbulence equalization (not used in this example);

[0201] High frequency (>30 Hz): Background prior is 0.1, turbulence prior is 0.7 (both frequencies in this example are satisfied);

[0202] Taking channel 1 as an example, the mask weights are first calculated. Within the coverage band, the weights are calculated according to the linear rule that the closer to the center, the greater the weight.

[0203] 110 Hz: 10 Hz from the center, located at the edge, weight is 0;

[0204] 116 Hz: 4 Hz from the center, center proximity is 6 / 10, weight is 0.6;

[0205] 120 Hz: 0 Hz from the center, weight 1.0;

[0206] 124 Hz: 4 Hz from the center, weighted at 0.6;

[0207] 130 Hz: 10 Hz from the center, weight is 0;

[0208] 150 Hz: 30 Hz from the center, out of band, weight 0;

[0209] Channel 2 is then multiplied by a channel weight of 0.7 and truncated to zero or one:

[0210] 116 Hz: 0.6 × 0.7, specifically 0.42;

[0211] 120 Hz: 1.0 × 0.7, specifically 0.7;

[0212] 124 Hz: 0.6 × 0.7, specifically 0.42;

[0213] The rest remain at 0.

[0214] Secondly, the three types of confidence scores are calculated and normalized into a decomposed weight vector.

[0215] For example, the mask weights are directly used as the perturbation confidence; the out-of-band degree is used as the basis for the non-perturbation confidence, and then multiplied by the background prior and the turbulence prior respectively to obtain the background confidence and the turbulence confidence. The out-of-band degree can be represented by "1 - mask weights".

[0216] Taking channel 1 with a frequency of 116 Hz as an example: mask weight 0.6, out-of-band severity 0.4, high-frequency prior background 0.1, and turbulence 0.7, then:

[0217] Perturbation confidence level: 0.6;

[0218] Background confidence level: 0.4 × 0.1, specifically 0.04;

[0219] Turbulence confidence level: 0.4 × 0.7, specifically 0.28;

[0220] The sum of the three is 0.92, and normalization yields the decomposed weight vector:

[0221] Perturbation weight: 0.6 / 0.92≈0.652;

[0222] Background weight: 0.04 / 0.92≈0.043;

[0223] Turbulence weight: 0.28 / 0.92≈0.304;

[0224] Taking channel 1 and frequency 120 Hz as an example: with a mask weight of 1.0 and out-of-band degree of 0, then:

[0225] Perturbation confidence level: 1.0;

[0226] Background confidence level: 0;

[0227] Turbulence confidence level: 0;

[0228] After normalization, the perturbation weight is 1, and the rest are 0;

[0229] Taking channel 1 and frequency 150 Hz as an example: with mask weight 0 and out-of-band degree 1, then:

[0230] Perturbation confidence level: 0;

[0231] Background confidence level: 1×0.1, specifically 0.1;

[0232] Turbulence confidence level: 1 × 0.7, specifically 0.7;

[0233] After normalization:

[0234] Background weight: 0.1 / 0.8 = 0.125;

[0235] Turbulence weight: 0.7 / 0.8 = 0.875;

[0236] Perturbation weight: 0.

[0237] Finally, the complex coefficients are split according to phase preservation and amplitude distribution: Assume that the complex coefficients of channel 1 at this moment (expressed in terms of amplitude and phase) are: 116 Hz: amplitude 3.0, phase 30°, 120 Hz: amplitude 5.0, phase 20° and 150 Hz: amplitude 2.0, phase -10°.

[0238] The amplitudes of the components are then divided into three categories as follows (the phases of all components remain the same):

[0239] 116 Hz: Disturbance amplitude 3.0×0.652≈1.956 (phase 30°); Background amplitude 3.0×0.043≈0.129 (phase 30°); Turbulence amplitude 3.0×0.304≈0.912 (phase 30°).

[0240] 120 Hz: Disturbance amplitude 5.0×1=5.0 (phase 20°); background and turbulence are 0.

[0241] 150 Hz: Background amplitude 2.0 × 0.125 = 0.25 (phase -10°); Turbulence amplitude 2.0 × 0.875 = 1.75 (phase -10°); Disturbance is 0.

[0242] The same rule is applied to split the coefficients at each frequency point in channel 2, resulting in three types of coefficient points corresponding to each channel-time-frequency position. These points are then written into the background atmospheric coefficient tensor, aerodynamic disturbance coefficient tensor, and turbulent structure coefficient tensor by index, thus completing one initial tensor allocation driven by confidence.

[0243] In one example, the deep learning model includes:

[0244] The feature extraction layer is used to extract feature information from the aerodynamic disturbance estimation signal, the background atmospheric signal, and the turbulent structure signal, respectively.

[0245] The feature fusion layer is used to fuse various feature information through an attention mechanism, assign dynamic weights to each feature according to a frequency mask, process the fused features, and generate a comprehensive feature vector.

[0246] The decision layer is used to input the comprehensive feature vector into a preset two-layer fully connected neural network, output a multi-dimensional detection vector through a linear activation function, and correct the multi-dimensional detection vector through a Monte Carlo method, outputting atmospheric data anomaly detection results and corresponding corrected data.

[0247] In another example, the method outputs a multidimensional detection vector through a linear activation function, and then corrects the multidimensional detection vector using a Monte Carlo method to output atmospheric data anomaly detection results and corresponding corrected data, including:

[0248] The detection components in the multidimensional detection vector are perturbed and sampled, and multiple sets of sampled detection vectors are generated based on a preset Gaussian distribution.

[0249] For each sampled detection vector, the rate of change of response on the background atmospheric signal, the estimated aerodynamic disturbance signal, and the turbulent structure signal is calculated respectively. If the rate of change of response is within the preset physical range, the corresponding detection component is marked as a valid sample corresponding to the real atmospheric change; otherwise, it is marked as an invalid sample.

[0250] For the detection components that are marked as valid samples, the original output is maintained. For the detection components that are marked as invalid samples, the expected value of the detection component is calculated as the corrected detection vector based on the corresponding perturbation distribution, and the corrected detection vector is output as the correction result.

[0251] In some optional implementations, the deep learning model is trained offline under supervised supervision. The training samples consist of aligned aerodynamic disturbance estimation signals, background atmospheric signals, and turbulent structure signals from the same flight mission. A sliding window is used to form the input segment. Within each window, normalization and scale alignment are performed on each physical quantity channel. Normalization can be achieved by scaling based on the window mean and scale factor, or by robust scaling based on mission statistics. The frequency masks within the corresponding window are aggregated in the time dimension into a mask intensity sequence as prior input. In this embodiment, the training labels are constructed for "correctable measurement anomalies":

[0252] One type of label is used to characterize the confidence level of anomalies, which can be obtained from the consistency check of multi-source redundant sensors, ground calibration records, physical consistency constraints of the same flight segment, or manual verification and labeling;

[0253] Another type of label is used to characterize the correction target, which can be provided by reference sensors, fused estimates, or quality-controlled baseline sequences, and uses the "difference between the target value and the original value" as the correction increment supervision signal.

[0254] To enhance generalization ability, perturbations can be applied to the input segments during the training phase, including random missing test masking, amplitude jitter, slight time scaling, and band-limited noise injection, so that the model can maintain stable output under different operating conditions and different noise levels.

[0255] In terms of network configuration, the feature extraction layer can extract local temporal features from the three inputs using several layers of one-dimensional convolution or temporal coding modules, and output a feature sequence of uniform dimension. The feature fusion layer uses an attention module to fuse the three features, where the mask intensity sequence obtained by frequency mask aggregation is mapped to attention bias or gating coefficients, so that the attention allocation has dynamic priors for rotor frequency-adjustable neighboring features. The number of attention heads can be selected from 2 to 8, and the feature dimension can be selected from 64 to 256. Random deactivation can be introduced to support Monte Carlo uncertainty assessment in the inference stage. The decision layer uses a two-layer fully connected network to output a multi-dimensional detection vector, which includes at least two components: "correctable anomaly confidence" and "correction increment". The training objective is jointly optimized using multi-task loss, where the anomaly confidence uses cross-entropy or focus loss to handle sample imbalance, and the correction increment uses L1 / L2 or Huber loss to suppress outliers. At the same time, a time continuity regularization term can be added to constrain the smooth change of the correction increment of adjacent windows, and a mask consistency constraint can be added to avoid the correction mainly falling on the turbulent broad spectrum structure. The optimization strategy can use Adam or AdamW, with an initial learning rate of 1e-4 to 1e-3 and a batch size of 16 to 256. Early stopping and learning rate annealing should be performed using a validation set. Data partitioning is recommended by "flight mission / segment" dimension for training, validation, and test sets to avoid overestimating performance due to leaks in the same flight segment. After training, during the inference phase, random deactivation is retained, and forward computation is repeated multiple times to obtain the detection vector set. The mean is used as the final output, and the variance is used as the uncertainty. When the uncertainty exceeds a threshold, the correction increment is suppressed or weighted and fused to reduce the risk of mistaking real atmospheric mutations as correctable anomalies.

[0256] In one example, the input to the deep learning model is not the original measurement sequence, but rather a three-way sequence consisting of the aerodynamic disturbance estimation signal, the background atmospheric signal, and the turbulent structure signal aligned on the same time axis.

[0257] Specifically, to enable the model to stably handle data streams with different sampling rates or task durations, the three sequences are truncated into sample segments using a sliding window, and normalization is performed on each physical quantity within the window. This normalization can employ linear scaling based on the window mean and scale factor, or robust scaling based on historical statistics, to ensure that the amplitude distributions at different flight stages remain within a comparable range. To avoid edge distortion caused by window segmentation, the window output is backfilled on the time axis using an overlapping fusion method, i.e., a weighted average or confidence-weighted fusion is applied to the overlapping regions, ensuring temporal continuity between the final detection result and the corrected sequence.

[0258] Those skilled in the art will understand that the above window length and step size can be configured according to the dynamic characteristics of the task, as long as they meet the minimum requirements of being able to cover the rotor disturbance period and typical time scale of turbulence.

[0259] In this embodiment, the feature extraction layer generates feature representations for each of the three inputs, and each feature representation contains at least two types of information:

[0260] One type is the time-domain variation information of the sequence (e.g., local amplitude variation trend, adjacent sampling difference, short window statistics);

[0261] Another type is characterization information related to time-frequency structure (such as component energy distribution, in-band and out-of-band energy ratio, and multi-scale fluctuation measurement of turbulent components obtained from the aforementioned decomposition).

[0262] The feature fusion layer also includes an attention mechanism to complete the fusion across components and physical quantities, and injects the frequency mask into the attention calculation in a dynamic weighted prior manner, so that the model can adjust the attention intensity of different features under different operating conditions.

[0263] In some optional implementations, the frequency mask is aggregated into a mask intensity sequence over time within the current window, and further mapped to attention bias or gating coefficients. When the mask intensity is high, higher participation weights are assigned to relevant features from aerodynamic disturbance components, and attention weights are increased for feature channels adjacent to rotor overtones. When the mask intensity is low, more attention weights are allocated to background atmospheric and turbulent structure-related features to reduce suppression of the real atmospheric structure. By using the mask as a constraint source for attention weights, the fused integrated feature vector simultaneously contains three types of information in physical semantics: aircraft disturbance state, background trend state, and turbulent structure state. The proportions of these three types of information under different flight conditions can be adaptively adjusted with changes in the mask.

[0264] It is understood that the attention mechanism itself can be implemented using self-attention, multi-head attention, or cross-attention methods well-known to those skilled in the art, and this application does not limit the specific network structure of the attention. This embodiment focuses on introducing a frequency mask as a dynamic weight prior into the attention calculation, so that the attention allocation can be controllably changed with the strength of rotor disturbance and the degree of frequency band coverage.

[0265] For details on how to inject attention, please refer to common implementation paths in this field such as attention bias, gated attention, and prior guided attention. Several exemplary implementation methods are given below so that those skilled in the art can reproduce them.

[0266] For example, the frequency masks are aggregated along the time dimension within the current window to obtain a mask intensity sequence corresponding to the time step, and this sequence is mapped to an attention bias term, specifically including:

[0267] For each time step, the frequency band weights covered by the mask at that time step are summed or averaged to obtain the mask intensity for that time step. Then, the mask intensity is scaled up and truncated to a preset range according to a preset ratio, which is used as a bias for different time steps or different feature channels in the attention calculation. This makes it easier for time steps with higher mask intensity to obtain greater attention weight in attention allocation, thereby making the perturbation features adjacent to the rotor frequency doubling more prominent in the fusion process.

[0268] In another implementation, the frequency mask is mapped to gating coefficients to perform gating modulation on the attention output or attention input, specifically including:

[0269] The mask intensity sequence is used to generate gating coefficients through a set of monotonic mapping rules. These gating coefficients can be applied to the fusion inlet of aerodynamic disturbance components, background atmospheric features, and turbulent features, respectively. When the mask intensity is high, the gating coefficient for aerodynamic disturbance components increases while the gating coefficients for background and turbulence decrease; when the mask intensity is low, the gating coefficients are adjusted in the opposite direction, thus allowing for a dynamically adjustable participation ratio of features from different sources within the same attention module. To avoid abrupt feature changes caused by gating switching, the gating coefficients can be smoothly updated between adjacent windows, ensuring continuity in weight changes.

[0270] In addition, the multidimensional detection vector includes at least: a correctable measurement anomaly confidence component for each observed physical quantity, and a correction increment component corresponding to that physical quantity;

[0271] The correctable measurement anomalies are used to distinguish between anomalies such as sensor drift, short-term jitter, and attitude coupling errors that can be corrected by algorithms, while abrupt changes caused by real atmospheric processes are not considered as objects of correction.

[0272] The Monte Carlo method is used to evaluate and correct the uncertainty of the above detection vectors. Specifically, it can be implemented by random deactivation or random perturbation sampling during the inference stage.

[0273] While keeping the input window unchanged, forward inference is repeated multiple times to obtain multiple sets of detection vectors. The mean of the sets is used as the final detection vector, and the dispersion of the sets is used as the uncertainty measure. When the uncertainty exceeds a preset threshold, the output correction increment is suppressed or reduced-weighted fusion to avoid forcibly correcting abrupt signals from uncertain sources. Only when the anomaly confidence is high and the uncertainty is low is the correction increment applied to the original observations to generate corrected data.

[0274] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A deep learning-based UAV atmospheric data anomaly detection and correction system, characterized in that, The system includes: The data preprocessing module is used to collect raw measurement data through the UAV, construct an air mass reference coordinate system based on the attitude of the UAV and the relative wind direction, reproject the raw measurement data onto the air mass reference coordinate system, and generate an atmospheric data sequence. The data decomposition module is used to obtain the motor speed of the UAV, calculate the corresponding frequency mask, and perform time-frequency analysis on the atmospheric data sequence to obtain the time-frequency coefficient matrix. Based on the frequency mask, the time-frequency coefficient matrix is ​​decomposed into a multi-scale decomposition to obtain the aerodynamic disturbance estimation signal, the background atmospheric signal, and the turbulent structure signal. The data detection module is used to input the aerodynamic disturbance estimation signal, background atmospheric signal and turbulent structure signal as input parameters into a preset deep learning model, and process the input parameters through the deep learning model to obtain atmospheric data anomaly detection results and corresponding correction data; The data decomposition module includes: The frequency mask calculation unit generates frequency mask parameters to distinguish between rotor disturbance frequency bands and non-disturbance frequency bands based on the UAV motor speed and flight condition information; The time-frequency analysis unit is used to perform time-frequency transformation on the atmospheric data sequence to generate a time-frequency coefficient matrix that characterizes different time and frequency components; A multi-scale decomposition unit is used to perform multi-scale decomposition on the time-frequency coefficient matrix under the constraints of the frequency mask parameters. The data decomposition module includes a data decomposition strategy, which includes mask calculation logic and decomposition logic. The mask calculation logic is configured within the frequency mask calculation unit, and the decomposition logic is configured within the multi-scale decomposition unit. The mask calculation logic is configured as follows: The motor speeds of the UAV at different times are obtained, a speed sequence is generated, and the corresponding basic rotor frequency sequence and speed change rate are calculated based on the speed sequence. The intensity index used to characterize the flight condition is calculated based on the rotational speed change rate and the pre-collected attitude angular velocity, throttle command change, and airspeed change of the UAV. Based on the intensity index, a frequency mask centered on the basic rotor frequency and its corresponding harmonic in the basic rotor frequency sequence is determined. The decomposition logic is configured as follows: The time-frequency coefficient matrix is ​​marked as a multi-channel time-frequency coefficient matrix according to the channel division, wherein the number of channels is the same as the number of sensors; Based on the frequency mask and the spatial layout relationship between each sensor and the rotor, a channel weight is assigned to each channel, and the frequency mask is expanded into a multi-channel frequency mask defined in the channel dimension, time dimension and frequency dimension according to the channel weight. Under the constraint of the multi-channel frequency mask, the multi-channel time-frequency coefficient matrix is ​​decomposed into background atmospheric coefficient tensor, aerodynamic disturbance coefficient tensor and turbulent structure coefficient tensor, wherein the algebraic sum of each coefficient tensor at each channel, each time and each frequency point is equal to the original time-frequency coefficient. The background atmospheric coefficient tensor is subjected to low-rank and smooth constraints in the channel dimension and time dimension, the aerodynamic disturbance coefficient tensor is subjected to mask weighted sparsity constraints, and the turbulent structure coefficient tensor is subjected to broad-spectrum energy distribution constraints. By performing time-frequency inverse transforms on the constrained background atmospheric coefficient tensor, aerodynamic disturbance coefficient tensor, and turbulent structure coefficient tensor, the aerodynamic disturbance estimation signal, background atmospheric signal, and turbulent structure signal are obtained.

2. The deep learning-based UAV atmospheric data anomaly detection and correction system according to claim 1, characterized in that, The data preprocessing module includes: The wind direction decomposition unit performs attitude transformation on the body coordinate system based on the pitch angle, roll angle and yaw angle acquired by the UAV in real time, and obtains the decomposition vector of the relative wind direction in the body coordinate system; The coordinate construction unit is used to construct an air mass reference coordinate system with the direction of the decomposed vector as the first reference axis, the normal direction of the first reference axis on the horizontal plane as the second reference axis, and the vertical direction of the first reference axis as the third reference axis, based on the relative flow direction, the horizontal lateral direction, and the vertical direction. The data reprojection unit is used to reproject the original measurement data according to the air mass reference coordinate system, and generate an atmospheric data sequence according to the time order of each physical quantity after reprojection. The reprojection transforms the change in observation direction corresponding to the movement of the UAV body to the air mass reference coordinate system to generate a unified reference direction that moves with the air mass. The data preprocessing module is configured with a reprojection strategy, which includes reprojection logic, and the reprojection logic is configured within the data reprojection unit.

3. The deep learning-based UAV atmospheric data anomaly detection and correction system according to claim 2, characterized in that, The original measurement data includes wind speed vector, temperature, humidity, air pressure, and particulate matter concentration. The reprojection logic is configured as follows: The wind speed vector is oriented by a preset attitude transformation matrix, and the projection components of the corrected wind speed vector on the first reference axis, the second reference axis and the third reference axis are calculated. The three projection components are weighted and summed to obtain the wind speed component. Based on the reprojection direction of the wind speed component, the measurement data of temperature, humidity, air pressure and particulate matter concentration are oriented to obtain the observed components. The continuity correction is performed based on the changing trend of the observed components at adjacent time points, wherein the continuity correction is implemented by an interpolation smoothing algorithm.

4. The deep learning-based UAV atmospheric data anomaly detection and correction system according to claim 1, characterized in that, Determining a frequency mask centered on the fundamental rotor frequency and its corresponding harmonics in the fundamental rotor frequency sequence based on the intensity index includes: Based on the aforementioned basic rotor frequency sequence, determine the basic rotor frequency and the center frequency of the fundamental rotor frequency harmonics at each time point; The target bandwidth is determined within a preset bandwidth range based on the intensity index. When the intensity index is less than a preset index threshold, a first bandwidth is selected. When the intensity index is greater than the preset index threshold, a second bandwidth is selected. The frequency coverage range of the first bandwidth is less than the frequency coverage range of the second bandwidth. Centered on the center frequency, a corresponding frequency range is determined based on the target bandwidth, and weight values ​​representing the rotor disturbance probability are allocated within the corresponding frequency range. A frequency mask that varies over time is constructed based on the center frequency, target bandwidth, and weight values.

5. The deep learning-based UAV atmospheric data anomaly detection and correction system according to claim 1, characterized in that, Under the constraint of the multi-channel frequency mask, the multi-channel time-frequency coefficient matrix is ​​decomposed into a background atmospheric coefficient tensor, an aerodynamic disturbance coefficient tensor, and a turbulent structure coefficient tensor, including: Based on the multi-channel frequency mask, different mask weights are assigned to the time-frequency coefficient matrices of each channel, each time, and each frequency point. The corresponding aerodynamic disturbance confidence, background field confidence, and turbulence confidence are calculated according to the mask weights. The mask weights increase sequentially according to the difference in Euclidean distance between the frequency point and the center frequency. In each channel, a decomposition weight vector is constructed based on the aerodynamic disturbance confidence, background field confidence, and turbulence confidence to characterize the background atmospheric component, aerodynamic disturbance component, and turbulence component, and the time-frequency coefficients are weighted and decomposed using the decomposition weight vector. Consistency constraints are applied to the allocated time-frequency coefficients in the channel dimension, time dimension, and frequency dimension. If the decomposition error in the channel dimension, time dimension, and frequency dimension is greater than or equal to the decomposition threshold of any dimension, the decomposition weight vector is updated according to the decomposition error of the corresponding dimension, and the time-frequency coefficients are redistributed until all are less than the decomposition threshold. After the time-frequency coefficients at each time-frequency point satisfy the aforementioned consistency constraint, the background atmospheric component, aerodynamic disturbance component, and turbulent component corresponding to each time-frequency point are aggregated according to their channel, time, and frequency position to obtain the background atmospheric coefficient tensor, aerodynamic disturbance coefficient tensor, and turbulent structure coefficient tensor.

6. The deep learning-based UAV atmospheric data anomaly detection and correction system according to claim 1, characterized in that, The deep learning model includes: The feature extraction layer is used to extract feature information from the aerodynamic disturbance estimation signal, the background atmospheric signal, and the turbulent structure signal, respectively. The feature fusion layer is used to fuse various feature information through an attention mechanism, assign dynamic weights to each feature according to a frequency mask, process the fused features, and generate a comprehensive feature vector. The decision layer is used to input the comprehensive feature vector into a preset two-layer fully connected neural network, output a multi-dimensional detection vector through a linear activation function, and correct the multi-dimensional detection vector through a Monte Carlo method, outputting atmospheric data anomaly detection results and corresponding corrected data.

7. The deep learning-based UAV atmospheric data anomaly detection and correction system according to claim 6, characterized in that, The process involves outputting a multidimensional detection vector using a linear activation function, and then correcting the multidimensional detection vector using a Monte Carlo method to output atmospheric data anomaly detection results and corresponding corrected data, including: The detection components in the multidimensional detection vector are perturbed and sampled, and multiple sets of sampled detection vectors are generated based on a preset Gaussian distribution. For each sampled detection vector, the rate of change of response on the background atmospheric signal, the estimated aerodynamic disturbance signal, and the turbulent structure signal is calculated respectively. If the rate of change of response is within the preset physical range, the corresponding detection component is marked as a valid sample corresponding to the real atmospheric change; otherwise, it is marked as an invalid sample. For the detection components that are marked as valid samples, the original output is maintained. For the detection components that are marked as invalid samples, the expected value of the detection component is calculated as the corrected detection vector based on the corresponding perturbation distribution, and the corrected detection vector is output as the correction result.

Images