Unmanned aerial vehicle multi-sensor flight data recovery method based on group structure conditional diffusion

Abstract

The application provides a kind of unmanned aerial vehicle multi-sensor flight data recovery method based on group structure condition diffusion, by obtaining the original flight data collected by unmanned aerial vehicle multi-sensor system, and the original flight data is preprocessed, the unmanned aerial vehicle flight time series sample to be recovered is obtained;Partial observation flight data is obtained based on observation mask;Variable group structure prior is constructed;Flight state condition information corresponding to the unmanned aerial vehicle flight time series sample to be recovered is obtained;Partial observation flight data, observation mask, variable group structure prior and flight state condition information are input into group structure condition diffusion recovery model, and the unmanned aerial vehicle multi-sensor flight data recovered by group structure condition diffusion recovery model is output;The application can improve the recovery quality of unmanned aerial vehicle multi-sensor flight data in random point loss, continuous fragment loss and variable level loss scene, and can improve the accuracy, structure consistency and flight state dynamic rationality of unmanned aerial vehicle flight data recovery.

Description

Technical Field

[0001] This invention relates to a method for recovering multi-sensor flight data of unmanned aerial vehicles (UAVs) based on group structure condition diffusion, belonging to the field of UAV data processing technology. Background Technology

[0002] With the continuous development of drone technology, drones have been widely used in scenarios such as inspection and mapping, agricultural remote sensing, and disaster monitoring. However, drone flight data is prone to loss during the collection, transmission, and storage process: on the one hand, drones usually operate in open, dynamic, and complex environments, and are easily affected by factors such as unstable communication links, environmental obstruction, and electromagnetic interference; on the other hand, the drone's own sensors may also experience observation gaps due to equipment aging, short-term failures, or abnormal sampling.

[0003] Missing UAV flight data is complex. This lack can manifest not only as random points of absence at individual points in time, but also as structured gaps across various flight phases, such as missing consecutive segments, unobservable single sensor channels, or simultaneous absence of multiple relevant variables. For UAV intelligent systems that rely on flight data for status monitoring, anomaly identification, trajectory prediction, or safety assessment, missing data can lead to biased analysis results and even affect the reliability of subsequent flight decisions.

[0004] In recent years, diffusion-generative models have demonstrated strong generative modeling capabilities in time series data recovery tasks, addressing the problem of missing time series data. Compared to traditional deterministic recovery methods, diffusion models can generatively recover missing portions through progressive denoising and, to some extent, express the uncertainties in the recovery process. However, existing diffusion-based time series recovery methods are mostly geared towards general multivariate time series interpolation tasks, typically lacking a modeling mechanism for variable group structure in UAV multi-sensor flight data, and also lacking an effective way to jointly model flight state condition information, observation mask constraints, and sensor variable group relationships. Therefore, in the scenario of UAV multi-sensor flight data missing recovery, existing technologies still struggle to simultaneously achieve complex missing data recovery accuracy, preservation of variable group internal structure, and dynamic rationality under different flight states. Summary of the Invention

[0005] The purpose of this invention is to overcome the problems of insufficient utilization of variable group structure, insufficient perception of flight state conditions, and low reliability of recovery results in complex missing scenarios in existing UAV multi-sensor flight data recovery methods, and to provide a UAV multi-sensor flight data recovery method based on group structure condition diffusion.

[0006] The technical solution of this invention is:

[0007] A method for recovering multi-sensor flight data of unmanned aerial vehicles (UAVs) based on group structure condition diffusion includes the following steps:

[0008] S1. Acquire the raw flight data collected by the UAV multi-sensor system, and preprocess the raw flight data to obtain the flight time sequence sample of the UAV to be recovered;

[0009] S2. Construct an observation mask based on the missing data of the UAV flight time series samples to be recovered, and obtain partial observation flight data based on the observation mask;

[0010] S3. Based on the sensor source, physical meaning, statistical attributes, flight control function or local correlation of each variable in the flight time series sample of the UAV to be recovered, construct a variable group structure prior, and divide the variables in the flight time series sample of the UAV to be recovered into multiple variable groups based on the variable group structure prior.

[0011] S4. Obtain the flight status condition information corresponding to the flight time sequence sample of the UAV to be recovered;

[0012] S5. Input some observed flight data, observation mask, variable group structure priors, and flight state condition information into the pre-trained group structure conditional diffusion recovery model. The group structure conditional diffusion recovery model includes a noise estimation network and a data recovery module. During the back diffusion process of the group structure conditional diffusion recovery model:

[0013] The noise estimation network obtains the predicted noise for the current diffusion step based on the noisy flight sample corresponding to the current diffusion step, partial observed flight data, observation mask, prior variable group structure, flight state condition information and diffusion step information.

[0014] The data recovery module performs backdiffusion update based on the noisy flight sample and predicted noise corresponding to the current diffusion step, and keeps the data at the observed positions unchanged based on the observation mask, only iteratively recovering the data at the missing positions to obtain the noisy flight sample corresponding to the next diffusion step; after completing the backdiffusion process, it outputs the recovered UAV multi-sensor flight data.

[0015] Further, in step S1, the raw flight data is preprocessed to obtain the flight time sequence sample of the UAV to be recovered, specifically as follows:

[0016] Acquire raw flight data from at least two types of acquisition units, including the UAV's global positioning module, inertial measurement unit, altimeter, velocity sensor, attitude sensor, heading sensor, environmental perception module, or flight control system.

[0017] The raw flight data is processed by at least one of the following methods: timestamp alignment, sampling frequency unification, outlier sampling removal, missing location marking, and data standardization, to obtain preprocessed continuous flight data.

[0018] The preprocessed continuous flight data is reconstructed using a sliding window to obtain a flight time series sample of the UAV to be recovered, which has a preset time length and preset variable dimensions.

[0019] Further, in step S2, an observation mask is constructed based on the missing data in the flight time series samples of the UAV to be recovered, and partial observation flight data is obtained based on the observation mask. Specifically,

[0020] The flight time sequence sample of the drone to be recovered is denoted as... The flight time series sample X of the UAV to be recovered includes multiple time slices and multiple flight data variables;

[0021] Construct an observation mask with the same time and variable dimensions as the flight time series sample X of the UAV to be recovered. ;

[0022] When the data of the UAV flight time series sample X to be recovered at the corresponding time slice and corresponding variable position is observable, the observation mask will be applied. The element at the corresponding position is set to the first value;

[0023] When the data for the drone flight time series sample to be recovered is missing at the corresponding time slice and the corresponding variable position, the observation mask will be used. The element at the corresponding position is set to the second value;

[0024] Based on observation mask Treatment of recovery of drone flight time series samples After masking, partial observation flight data was obtained. .

[0025] Furthermore, in step S3, the prior structure of the variable group includes multiple variable groups, which include at least two of the following: position variable group, velocity variable group, attitude variable group, angular velocity variable group, acceleration variable group, altitude variable group, heading variable group, and environmental observation variable group.

[0026] Each variable group includes at least one flight data variable. Different variable groups are used to characterize a set of variables in UAV multi-sensor flight data that have different sensor sources, physical meanings, statistical scales, flight control functions, or local correlations.

[0027] Furthermore, in step S5, the noise estimation network includes a state and diffusion step encoding module, a two-stream group structure encoding module, at least one two-path residual block, and a group-level output header.

[0028] State and Diffusion Step Encoding Module: Used to encode flight state condition information to obtain flight state representation, including local flight state representation and global flight state representation, and to encode diffusion step information to obtain diffusion step representation;

[0029] The dual-stream group structure encoding module performs group-level encoding on the noisy flight samples and observation masks by the noisy branch to obtain the noisy branch global group structure representation; and performs group-level encoding on some observed flight data, observation masks, and local flight state representations by the conditional branch to obtain the conditional branch global group structure representation; after fusing the noisy branch global group structure representation, the conditional branch global group structure representation, the group identity embedding, and the diffusion step embedding, the initial hidden representation is obtained and output.

[0030] Dual-path residual block: For the initial hidden representation of the input, long-term time-series dependency features of UAV flight data are extracted using bidirectional temporal paths, and local dynamic features are extracted from local interaction paths; group-level gating coefficients are generated, and hidden states and skip-connection outputs are generated from the group-level gating coefficients, long-term time-series dependency features and local dynamic features;

[0031] Group-level output head: Used to aggregate the jump-connected outputs of dual-path residual blocks to obtain aggregated features. According to the prior structure of variable groups, the aggregated features are split into multiple group-level sub-representations. Group-level prediction noise is generated through the output head corresponding to each variable group, and the group-level prediction noises are concatenated to output the prediction noise corresponding to the current diffusion step.

[0032] Furthermore, in the state and diffusion step encoding module, the flight state condition information is encoded to obtain a flight state representation, including a local flight state representation and a global flight state representation. The diffusion step information is encoded to obtain a diffusion step representation, specifically as follows:

[0033] When the flight state condition information is a sample-level flight state vector, the sample-level flight state vector is mapped to a global flight state representation, and then copied along the time dimension to obtain a local flight state representation.

[0034] When the flight state condition information is a time-varying flight state sequence, the time-varying flight state sequence is time-series mapped to obtain a local flight state representation, and the local flight state representation is time-dimensionally aggregated to obtain a global flight state representation;

[0035] Embedding encoding is performed on the current diffusion step to obtain the diffusion step representation;

[0036] Among them, the local flight state is used to characterize the flight state conditions corresponding to each time slice, the global flight state is used to characterize the overall operation stage of the current flight sample, and the diffusion step is used to characterize the denoising stage of the current back diffusion process.

[0037] Furthermore, the dual-stream group structure coding module includes noisy branches, conditional branches, and a fusion mapping layer.

[0038] Noisy branch: used to divide the noisy flight sample and observation mask corresponding to the current diffusion step into multiple group-level noisy inputs according to the variable group structure prior, and encode multiple group-level noisy inputs respectively to obtain multiple group-level noisy embeddings, which are then concatenated to obtain the noisy branch global group structure representation;

[0039] Conditional branch: It is used to divide and concatenate some observed flight data, observation mask and local flight state representation according to the variable group structure prior to obtain multiple group-level conditional inputs, and encode multiple group-level conditional inputs to obtain multiple group-level conditional embeddings. After concatenation, the global group structure representation of the conditional branch is obtained.

[0040] Fusion Mapping Layer: The noisy branch global group structure representation, the conditional branch global group structure representation, the group identity embedding consisting of a set of learnable parameters, and the diffusion step embedding obtained by mapping the diffusion step representation are fused to obtain the initial hidden representation for subsequent noise estimation.

[0041] Furthermore, in the noise estimation network, the dual-path residual block includes a bidirectional temporal path, a local interactive path, and a group-aware flight state gating module.

[0042] Bidirectional temporal path: used to perform bidirectional sequence feature extraction on the input hidden representation to obtain long temporal dependency features in UAV multi-sensor flight data;

[0043] Local interaction path: Under the constraint of the variable group structure prior, the input hidden representation is split into multiple group-level sub-representations, and local features within each group-level sub-representation are extracted. The local features within the group are then fused with the local flight state representation to obtain condition-guided cross-group interaction features.

[0044] Group-aware flight state gating module: used to generate group-level gating coefficients based on the diffusion step representation and the global flight state representation. The group-level gating coefficients include bidirectional timing path gating coefficients, local interaction path gating coefficients, and jump-connection output gating coefficients. It also adjusts the outputs of the bidirectional timing path and the local interaction path based on the group-level gating coefficients to generate hidden representations and jump-connection outputs for subsequent group-level output head aggregation.

[0045] Furthermore, in step S5, the training process of the pre-trained group structure conditional diffusion recovery model includes:

[0046] Acquire complete UAV flight time sequence samples, observation masks, and flight status condition information, and obtain partial observation flight data based on the observation masks;

[0047] Sample diffusion steps from the set of diffusion steps, and sample noise tensors from a preset noise distribution;

[0048] Noisy flight samples are constructed based on complete UAV flight time series samples, diffusion steps, and noise tensors;

[0049] Noisy flight samples, partial observation flight data, observation mask, diffusion step, and flight state condition information are input into the noise estimation network to obtain the predicted noise.

[0050] Calculate the noise prediction loss between the predicted noise and the noise tensor on the set of missing locations;

[0051] The parameters of the noise estimation network are updated based on the noise prediction loss.

[0052] Furthermore, in step S5, the reverse diffusion process of the group structure conditional diffusion recovery model is specifically as follows:

[0053] Initialize the back diffusion start point from the preset noise distribution;

[0054] Perform reverse diffusion updates sequentially in descending order of diffusion steps;

[0055] In each back-diffusion step, the observed positions in the current noisy flight sample are replaced with the corresponding partial observed flight data based on the observation mask to obtain the replaced noisy flight sample.

[0056] The replaced noisy flight sample is input into the trained noise estimation network to obtain the predicted noise corresponding to the current diffusion step.

[0057] Update missing location data in the current noisy flight sample based on predicted noise;

[0058] After all the reverse diffusion steps are completed, the recovered UAV multi-sensor flight data is obtained.

[0059] The beneficial effects of this invention are:

[0060] I. This method for recovering UAV multi-sensor flight data based on group structure conditional diffusion models the recovery of UAV multi-sensor flight data as a group structure conditional diffusion generation process. It can perform stepwise generative recovery of missing positions under the constraints of observed flight data. Compared to recovery methods that rely solely on interpolation, statistical estimation, or deterministic reconstruction, it is better suited to complex missing scenarios such as random point missing data, continuous segment missing data, and sensor channel missing data. This method can improve the recovery quality of UAV multi-sensor flight data in scenarios with random point missing data, continuous segment missing data, and variable-level missing data. Under the condition of partial observation of UAV multi-sensor flight data, it can perform conditional generative recovery of missing positions by combining prior sensor variable group structure information and flight state condition information, thereby improving the accuracy, structural consistency, and dynamic rationality of UAV flight data recovery.

[0061] Second, this invention introduces a priori variable group structure, and groups and models each variable in the UAV flight data according to the sensor source, physical meaning, flight control function and local correlation. This enables the recovery process to maintain the local consistency within the variable groups such as position, speed, attitude, inertial measurement, altitude and heading, and reduces the disruption of sensor variable coordination relationships caused by uniform tiling modeling.

[0062] Third, this invention introduces flight state condition information and uses local and global flight state representations to conditionally guide the noise estimation process, enabling the recovery model to perform differentiated recovery based on different flight states such as hovering, level flight, ascent, descent, turning, acceleration, or deceleration, thereby improving the consistency between the recovery results and the actual dynamic laws of the corresponding flight stages.

[0063] Fourth, this invention extracts the features to be recovered from noisy flight samples and the conditional constraint features from some observed flight data simultaneously through dual-stream group structure encoding, enabling the noise estimation network to simultaneously perceive the location to be recovered, the observed location, and the flight state conditions, thereby improving the noise prediction capability under complex missing conditions.

[0064] Fifth, this invention simultaneously models long-term temporal dependencies and local dynamic features within variable groups in UAV flight data through dual-path residual blocks, and adaptively adjusts the output weights of different variable groups, different paths and different diffusion stages through a group-aware flight state gating module, thereby enhancing the recovery model's ability to express the dynamic changes of UAV multi-sensor flight data.

[0065] VI. In the training phase, this invention calculates noise prediction loss only for missing locations. In the recovery phase, it keeps the observed location data unchanged and performs backdiffusion update only for the missing locations. This allows the model training and inference process to focus on the data locations that actually need to be recovered, which can reduce the risk of unnecessary modification of observed flight data and improve the reliability and usability of the recovery results. Attached Figure Description

[0066] Figure 1 This is a flowchart illustrating the UAV multi-sensor flight data recovery method based on group structure condition diffusion according to an embodiment of the present invention.

[0067] Figure 2 This is a schematic diagram illustrating the prior construction of the multi-sensor flight data variable group structure of the UAV in the embodiment;

[0068] Figure 3 This is a schematic diagram of the overall framework of the group structure conditional diffusion recovery model in the embodiment;

[0069] Figure 4 This is a schematic diagram illustrating the state and diffusion step encoding module in the embodiment;

[0070] Figure 5 This is a schematic diagram illustrating the dual-stream group structure encoding module in the embodiment;

[0071] Figure 6 This is a schematic diagram illustrating the bidirectional temporal path, local interaction path, and group-aware flight state gating module in the dual-path residual block of the embodiment;

[0072] Figure 7 This is a schematic diagram of the group-level output head and the prediction noise generation process in the embodiment;

[0073] Figure 8 This is a schematic diagram of the training process of the group structure conditional diffusion recovery model in the embodiment;

[0074] Figure 9 This is a schematic diagram of the UAV multi-sensor flight data recovery process in the embodiment. Detailed Implementation

[0075] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0076] The embodiment provides a method for recovering multi-sensor flight data of unmanned aerial vehicles based on group structure conditional diffusion, such as Figure 1 This includes the following steps:

[0077] S1. Acquire the raw flight data collected by the UAV's multi-sensor system, and preprocess the raw flight data to obtain the flight time sequence sample of the UAV to be recovered.

[0078] In one specific embodiment, the raw flight data may include multiple flight records, each flight record including multiple time slices, and each time slice including multiple state acquisition data. State acquisition data may include the UAV's position, altitude, speed, attitude, angular velocity, acceleration, heading angle, flight control inputs, state estimates, environmental observations, or flight state labels. Multiple flight records may originate from flight data of the same UAV in different missions, or from data collected by the same or different UAV models under the same or different flight environments.

[0079] In one optional embodiment, preprocessing may include one or more of the following: timestamp alignment, sampling frequency unification, invalid field removal, outlier sample removal, missing location marking, data standardization, field filtering, flight record sorting, and sliding window reconstruction. Specifically, the original flight data can be sorted according to flight identifiers and timestamps to ensure that the data in the same flight record maintains its original chronological order; timestamp alignment can be performed on data collected by different sensors to form multivariate time series samples on a unified time axis; flight variables with different scales can be standardized to reduce the impact of variable scale differences on model training and recovery; and a sliding window can be used to divide continuous flight records into fixed-length UAV flight time series samples.

[0080] In one specific embodiment, let the length of the sliding window be... The sliding step size is For lengths not less than The flight records are converted from the original continuous flight records into multivariate flight time series samples of uniform length. , Represents the composition of real numbers OK A set of column matrices. Flight records of length less than L may be excluded from sample construction, or they may be processed using padding methods depending on specific application requirements.

[0081] S2. Construct an observation mask based on the missing data of the UAV flight time series samples to be recovered, and obtain partial observation flight data based on the observation mask.

[0082] In one specific embodiment, the observation mask is used to identify whether the data at each time slice and variable position in the UAV flight time series sample is observable. The observation mask is denoted as... ,and .when The lth time slice, the Each variable corresponds to an element At that time, it indicates the drone flight sample Data is observable at this location; when When the value is missing, it indicates that data is missing at that location. (Based on the observation mask) This allows us to obtain some observational flight data. ,in This indicates element-wise multiplication. Missing data can originate from sensor malfunctions, communication interruptions, sampling anomalies, or data packet loss during the actual data acquisition process, or it can be simulated during the training phase using a preset missing data pattern.

[0083] In one optional embodiment, data missing scenarios can include random point missing, consecutive segment missing, or variable-level missing. Random point missing indicates that the missing locations are discretely distributed across different time slices and different variable locations; consecutive segment missing indicates that the same variable or multiple variables are unobservable over a continuous time range; variable-level missing includes single sensor channel missing and simultaneous missing of multiple related variables, where single sensor channel missing indicates that a certain sensor variable is unobservable throughout the entire sample time window, while simultaneous missing of multiple related variables indicates that a group of variables with local correlations are simultaneously missing in part or all of the time range. By using these different missing modes, the model can adapt to the complex missing scenarios commonly found in real UAV flight data.

[0084] S3. Based on the sensor source, physical meaning, statistical attributes, flight control function or local correlation of each variable in the flight time series sample of the UAV to be recovered, construct a priori variable group structure, and divide the variables in the flight time series sample of the UAV to be recovered into multiple variable groups based on the priori variable group structure.

[0085] Reference Figure 2 In one specific embodiment, the variable group structure prior is used to describe the local structural relationships between different variables in the UAV multi-sensor flight data. The variable group structure prior is denoted as... ,in Indicates the first Group of variables, Indicates the number of variable groups. Each variable group... It includes at least one flight variable, and different variable groups can be mutually exclusive, and all variable groups can cover all flight variables to be recovered. Variable group structure prior. This is used to instruct the model to perform group-level modeling according to the local structure of the UAV flight variables during the recovery process.

[0086] In one specific embodiment, a variable group structure prior can be automatically generated based on variable names. The variable group structure prior may include position variable groups, attitude variable groups, velocity variable groups, angular velocity variable groups, acceleration variable groups, and variable groups such as altitude / heading / environment. The position variable group may include position variables... , , Attitude variable group can include attitude variables , , The velocity variable group can include velocity variables. , , The angular velocity variable group can include angular velocity variables. , , The acceleration variable group can include acceleration variables. , , The altitude / heading / environment variable groups can include altitude variables, heading variables, environmental observation variables, or other variables not covered by the aforementioned rules but related to the UAV's flight state. This grouping method organizes variables with similar physical meanings and local correlations into the same variable group, enabling the recovery model to maintain the coordination between UAV flight variables within the group-level space.

[0087] S4. Obtain the flight status condition information corresponding to the flight time sequence sample of the UAV to be recovered.

[0088] In one specific embodiment, flight state condition information is used to characterize the operational state or mission stage of the UAV in a corresponding time slice or corresponding flight sample. The flight state condition information is denoted as... If the flight state condition information is a sample-level state vector, then It can represent the current flight stage or mission state of the entire UAV flight sample; if the flight state condition information is a time-varying state sequence, then... This can represent the flight state of the UAV at each time slice. Flight state condition information can be in the form of one-hot state labels or continuous state vectors. By introducing flight state condition information, the group structure condition diffusion recovery model can be guided to perform differentiated data recovery under different flight states, avoiding the mixing of flight dynamics laws under different states such as hovering, ascent, descent, turning, level flight, acceleration, or deceleration in the model.

[0089] S5. Input some observed flight data, observation mask, variable group structure priors, and flight state condition information into the pre-trained group structure conditional diffusion recovery model. The group structure conditional diffusion recovery model includes a noise estimation network and a data recovery module. During the back diffusion process of the group structure conditional diffusion recovery model:

[0090] The noise estimation network obtains the predicted noise for the current diffusion step based on the noisy flight sample corresponding to the current diffusion step, some observed flight data, observation mask, variable group structure prior, flight state condition information and diffusion step information;

[0091] The data recovery module performs backdiffusion update based on the noisy flight sample and predicted noise corresponding to the current diffusion step, and keeps the data at the observed positions unchanged based on the observation mask, only iteratively recovering the data at the missing positions to obtain the noisy flight sample corresponding to the next diffusion step; after completing the backdiffusion process, it outputs the recovered UAV multi-sensor flight data.

[0092] Reference Figure 3 The group structure conditional diffusion recovery model is used to progressively generatively recover missing locations in UAV flight time series samples under observation constraints. The group structure conditional diffusion recovery model uses partial observed flight data. Observation mask Prior knowledge of variable group structure Flight status information and diffusion step information As conditional information, the noise of the current diffusion step is predicted through a noise estimation network.

[0093] The group-structured conditional diffusion recovery model models the recovery of UAV multi-sensor flight data as a conditional diffusion generation process. Unlike recovery methods that directly use deterministic functions to output missing values, this embodiment uses the inverse denoising process of the diffusion model to gradually recover the missing positions, enabling the model to generate data for the missing positions under the constraints of observation conditions and flight state conditions.

[0094] In one specific embodiment, the conditional diffusion recovery model includes a forward diffusion process and a backward diffusion process. The forward diffusion process is used to progressively add noise to the complete UAV flight samples during the training phase to obtain noisy flight samples for the corresponding diffusion steps; the backward diffusion process is used to progressively denoise from the noise distribution during the recovery phase, and to recover the missing positions by combining conditional information at each step.

[0095] Set condition information Including some observation flight data Observation mask Flight status information and variable group structure prior Conditional diffusion recovery model expects to learn the conditional distribution. In order to provide information under given conditions In this case, a recovery sample consistent with the original UAV flight data is generated. This conditional distribution can be expanded according to the intermediate state of diffusion as follows:

[0096]

[0097]

[0098] in, Indicates by parameters A defined probability distribution The learnable parameters of the noise estimation network in the group structure conditional diffusion recovery model are represented. This represents a flight sample in a zero-noise state, serving as the initial flight sample during forward diffusion and corresponding to the complete UAV flight time sequence sample during the training phase; during back diffusion... Indicates the initial noise sample The final denoised flight sample obtained after stepwise denoising. This represents the initial noise sample in the back diffusion process; Indicates the first Noisy flight samples corresponding to each diffusion step; Indicates by the first The noisy flight sample obtained by reverse updating the previous diffusion step in each diffusion step; Indicates by The resulting diffusion state sequence; Indicates by The intermediate diffusion state sequence; Represents the sequence of intermediate diffusion states Integrate points; Indicates the total number of diffusion steps; Indicates the current diffusion step; Indicates from the first diffusion step to the second. A series of multiplication operations for each diffusion step; Indicates the first In each diffusion step, in the current noisy flight sample and condition information Generate the flight sample of the previous diffusion step under given conditions. The conditional reverse transition distribution.

[0099] During the forward diffusion process, let the complete UAV flight sample be... Initial diffusion sample The forward diffusion process can be represented by adding noise stepwise according to a Markov chain, as follows:

[0100]

[0101]

[0102] in, This represents the probability distribution during the forward diffusion process; Indicates that in the given first Noisy flight samples corresponding to each diffusion step In the case of generating the first Noisy flight samples corresponding to each diffusion step The forward transition distribution; Indicates a Gaussian distribution; This represents the mean of the Gaussian distribution; This represents the covariance matrix of the Gaussian distribution; Indicates the first The noise scheduling coefficients corresponding to each diffusion step; Represents the identity matrix.

[0103] Furthermore, the forward diffusion process can be written in closed form:

[0104]

[0105]

[0106] in, This represents the mean of the Gaussian distribution; This represents the covariance matrix of the Gaussian distribution; This represents the identity matrix. In one specific embodiment, ,and ,in, Indicates the first Signal retention coefficients corresponding to each diffusion step Indicates the first The noise scheduling coefficients corresponding to each diffusion step Indicates from the first diffusion step to the second. The cumulative product of the signal retention coefficients for each diffusion step. Indicates to obtain And towards The noise added is typically sampled from a standard Gaussian distribution. When the diffusion steps... When large enough, It will approach ,and It will also gradually converge to a standard Gaussian distribution, that is, it has .

[0107] In the reverse diffusion process, the group structure conditional diffusion recovery model is based on conditional information. and current noisy flight samples estimate The reverse diffusion process can be represented as:

[0108]

[0109] in, Indicates the first The conditional backward transition distribution corresponding to each backward diffusion step; Represents the learnable parameters of the noise estimation network; Indicates the first Noisy flight samples corresponding to each diffusion step; Indicates by the first The flight sample of the previous diffusion step is obtained by reverse updating the diffusion step in each diffusion step; Indicates a Gaussian distribution; Indicates by parameters A defined conditional mean function is used to determine the current noisy flight sample. Current diffusion step and condition information Generate the mean updated by backdiffusion; This represents the covariance matrix of the Gaussian distribution; Indicates the first The variance coefficients corresponding to each reverse diffusion step; Represents the identity matrix.

[0110] In one specific embodiment, the mean can be learned. :

[0111]

[0112] in, Indicates by parameters A defined back-diffusion conditional mean function is used to generate the first... The mean of each backward diffusion step; Indicates the first The signal retention coefficients corresponding to each diffusion step; Indicates from the first diffusion step to the second. The cumulative product of the signal retention coefficients for each diffusion step; This represents the noisy flight sample corresponding to the current diffusion step in the noise estimation network. Current diffusion step and condition information The output prediction noise under the condition; and For the signal preservation coefficient and its cumulative value Determined back diffusion update coefficients.

[0113] Therefore, the backdiffusion update can be expressed as:

[0114]

[0115] in, Indicates the first The noise standard deviation corresponding to each reverse diffusion step; This represents the random noise introduced during the backsampling process, and ; This indicates that the mean is 0 and the covariance matrix is ​​the identity matrix. The standard Gaussian distribution. In actual recovery, the settings can be configured according to application requirements. Random noise terms can also be reduced or removed in the last few diffusion steps. This is to achieve more stable recovery results.

[0116] In one specific embodiment, the noise estimation network includes a state and diffusion step encoding module, a two-stream group structure encoding module, a two-path residual block, and a group-level output head. These structures collectively achieve structural awareness, observation awareness, condition awareness, and temporal dependency modeling of UAV flight data.

[0117] Reference Figure 4 The state and diffusion step encoding module is used to process flight state condition information. and diffusion step Encoding is performed to obtain local flight state representation, global flight state representation, and diffusion step representation. Since the sensor dynamics of the UAV differ under different flight states, and the diffusion model faces different noise intensities at different denoising stages, relying solely on the current noisy flight samples... Accurate noise estimation is difficult to obtain. By encoding flight state condition information and diffusion step information, the noise estimation network can simultaneously perceive the current flight dynamic conditions and the current denoising stage. Among them, the local flight state representation is used as input to the two-stream group structure encoding module, the global flight state representation and diffusion step representation are used as input to the group-aware flight state gating module of the dual-path residual block, and the diffusion step representation is also used to generate diffusion step embeddings and participate in the feature fusion of the two-stream group structure encoding module.

[0118] In one specific embodiment, if flight status information It is a sample-level flight state vector, and ,in, If the original flight state feature dimension is represented, then it can be represented by a linear mapping. Projecting onto the state embedding space yields the global flight state representation. :

[0119]

[0120] in, and These are the learnable weight matrix and bias vector, respectively. Since the sample-level flight state vector does not include a time dimension, the global flight state can be represented... Copy along the time dimension to the length The local flight state representation is obtained. This operation broadcasts the overall state semantics of the current flight sample to each time slice, providing a shared local conditional background for subsequent noise estimation.

[0121] In another specific embodiment, if flight status condition information It is a time-varying flight state sequence, and It can then be mapped to a local flight state representation through one-dimensional convolution, linear mapping, or other temporal coding methods. :

[0122]

[0123] in, This represents the convolution operation. and These are the learnable convolutional kernel and the bias, respectively. This operation maps the original flight state features to a unified embedding space on a time-slice basis without changing the time length. Subsequently, a global flight state representation can be obtained through further operations. :

[0124]

[0125] in, This represents the global flight state, used to characterize the overall operational phase of the current flight sample; This represents the local flight state and is used to characterize the flight state conditions corresponding to each time slice. This represents the learnable weight parameters used to map the local flight state representation to the global flight state representation; This represents the corresponding learnable bias parameter; This represents a pooling operation, which can be global average pooling, max pooling, or attention pooling.

[0126] In one specific embodiment, the diffusion step The diffusion step representation can be obtained through sinusoidal position encoding, learnable embeddings, or multilayer perceptrons. When using sinusoidal position encoding, it can be represented as:

[0127]

[0128]

[0129] in, Indicates the diffusion step. The first in One portion, Indicates the embedding dimension of the diffusion step. Indicates the position index, and , This represents the component index in the diffusion step representation. Subsequently, the initial diffusion step embedding can be mapped to a diffusion step representation suitable for use in the noise estimation backbone using a two-layer perceptron. :

[0130]

[0131] in, It is a nonlinear activation function. The diffusion step is used to characterize the current denoising stage of the backdiffusion process, enabling the noise estimation network to perform corrections at different scales based on different noise intensities.

[0132] Reference Figure 5 The dual-stream group structure encoding module performs group-level encoding on the noisy flight samples and observation masks by the noisy branch to obtain the noisy branch global group structure representation; and performs group-level encoding on some observed flight data, observation masks and local flight state representations by the conditional branch to obtain the conditional branch global group structure representation; after fusing the noisy branch global group structure representation, the conditional branch global group structure representation, the group identity embedding and the diffusion step embedding, the initial hidden representation is obtained and output.

[0133] The dual-stream group structure encoding module is used to extract the features to be recovered from noisy flight samples and the conditional constraint features from some observed flight data at the entrance of the noise estimation network. By explicitly utilizing the variable group structure prior, the dual-stream group structure encoding module enables the group structure conditional diffusion recovery model to have structure awareness, observation awareness, and flight state condition awareness capabilities before entering the denoising backbone.

[0134] In one specific embodiment, the dual-stream group structure coding module includes a noisy branch, a conditional branch, and a fusion mapping layer. The noisy branch is used to process the noisy flight samples of the current diffusion step. and observation mask M. Based on the prior structure of the variable set. It can Divide into multiple group-level noisy subsequences, It is divided into group-level observation masks corresponding to each group-level noisy subsequence. For the first... There are n sets of variables:

[0135]

[0136] in, Indicates the first Noisy subsequences corresponding to each variable group This represents the corresponding group-level observation mask. The group-level observation mask is used to explicitly identify whether the variable group is observable at each time slice and at each variable location.

[0137] For the A set of variables can be used to define the noisy subsequences in that set. and group-level observation mask Concat the concatenation along the feature dimension to obtain the group-level noisy input. :

[0138]

[0139] Subsequently, group-level noisy embedding is obtained through a group-level noise encoder. :

[0140] .

[0141] In an alternative embodiment, the noisy branch includes at least one group-level noise encoder. Group-level noise encoder It consists of at least one set of one-dimensional convolutional layers and activation functions, used to perform feature compression, local temporal pattern extraction, and group-level embedding mapping within corresponding variable groups:

[0142]

[0143] in, It can be GELU, SiLU, ReLU, or other nonlinear activation functions. This represents the convolution operation. , and , This represents the learnable parameters.

[0144] Conditional branches are used to process part of the observation flight data. Observation mask and local flight status representation Based on the prior structure of the variable group. It can Divided into multiple group-level observation subsequences, Divide into corresponding group-level observation masks. For the first... A group of variables can be used to represent group-level observation subsequences. Group-level observation mask and local flight status representation By concatenating the data, we obtain the group-level conditional input. :

[0145]

[0146] Subsequently, group-level conditional embeddings are obtained through a group-level conditional encoder. :

[0147]

[0148] In an alternative embodiment, conditional branching includes at least one group-level conditional encoder. Group-level condition encoder It can also be composed of at least one set of one-dimensional convolutional layers and activation functions. This conditional branch reflects not only which locations have been observed and which locations need to be recovered, but also the conditions under which the recovery is performed.

[0149]

[0150] in, It can be GELU, SiLU, ReLU, or other nonlinear activation functions. , and , This represents the learnable parameters.

[0151] After obtaining the group-level noisy embeddings and group-level conditional embeddings for each variable group, the dual-stream group structure encoding module concatenates all group-level noisy embeddings in the order of the variable groups to obtain the noisy branch global group structure representation. Embed all group-level conditions and concatenate them in the order of variable groups to obtain the global group structure representation of the conditional branches. :

[0152]

[0153] .

[0154] Furthermore, learnable group identity embeddings are introduced. and the diffusion step embedding obtained by mapping the diffusion step representation The fusion mapping layer combines it with and Fusion yields the basic fusion representation. :

[0155]

[0156] Among them, group identity embedding It consists of a set of learnable parameters used to identify different groups of variables. Diffusion step embedding. This is used to inject information from the current denoising stage into the initial hidden representation. The embedding of the diffusion step is adjusted using a linear mapping. By repeating this process along the time dimension, we can obtain... .

[0157] Subsequently, local integration and cross-group correction are performed through a fusion layer to obtain the initial hidden representation. :

[0158]

[0159] Wherein, the initial hidden representation As input to subsequent dual-path residual blocks. It can be GELU, SiLU, ReLU, or other nonlinear activation functions. , and , This represents the learnable parameters. Through the above design, the model can simultaneously obtain the information to be recovered from the current noisy flight sample, the observation constraint information from the partial observation flight data, the group identity information from the prior of the variable group structure, and the contextual information from the diffusion step and flight state conditions before the denoising backbone begins.

[0160] Reference Figure 6 The dual-path residual block is used to simultaneously extract long-term temporal dependencies and local dynamic features within variable groups from UAV flight data in the backbone of the noise estimation network. The initial hidden representation of the input is obtained by the dual-path residual block using a bidirectional temporal path to extract long-term temporal dependencies from the UAV flight data, and by using a local interaction path to extract local dynamic features. Group-level gating coefficients are generated, including bidirectional temporal path gating coefficients, local interaction path gating coefficients, and jump-connection output gating coefficients. The hidden state and jump-connection output are then generated from the group-level gating coefficients, long-term temporal dependencies, and local dynamic features. The dual-path residual block includes a bidirectional temporal path, a local interaction path, and a group-aware flight state gating module. Through the bidirectional temporal path and the local interaction path, the noise estimation network can simultaneously perceive the global temporal evolution of UAV flight data, local patterns within sensor variable groups, and conditional correlations between different variable groups. The group-aware flight state gating module generates group-level gating coefficients based on the diffusion step representation and the global flight state representation. It then uses these group-level gating coefficients to adjust the weights of the bidirectional temporal path output, the local interactive path output, and the jump-connection output on each variable group. This enables the noise estimation network to adaptively adjust the data recovery strength and feature fusion method of each variable group according to different denoising stages and different flight states.

[0161] In one specific embodiment, the first Each dual-path residual block receives the hidden state of the output from the previous layer. A bidirectional temporal path is used to extract global temporal dependencies in flight samples along the time dimension. This path may include a bidirectional sequence feature extractor, which may employ one or more of the following: a state-space sequence model, a Mamba model, a recurrent neural network, a gated recurrent unit, a long short-term memory network, a temporal convolutional network, or a Transformer encoder. In a preferred embodiment, the bidirectional sequence feature extractor is implemented using a Mamba-2 model, such as... Figure 6 The forward and reverse Mamba-2 modules are used to extract flight dynamics dependencies over a longer time range with lower sequence complexity.

[0162] In one specific embodiment, the bidirectional temporal path first performs normalization processing on the input hidden representation:

[0163]

[0164] in, Indicates the first The input hidden representation of a dual-path residual block; This indicates that the input is hidden. The normalized hidden representation obtained after normalization; The layer index represents the two-path residual block; in a preferred embodiment, the normalization process can use root mean square normalization, i.e. .

[0165] Then Input the forward Mamba-2 module and the reverse Mamba-2 module respectively. For the state-space model-based Mamba-2 module, you can... The input features are split into multiple time slices along the time dimension, and the first slice is... The input features corresponding to each time slice are denoted as follows: Its core calculation can be abstracted as:

[0166]

[0167]

[0168] in, Indicates the time slice index; , Indicates the first The hidden states corresponding to each time slice; Indicates the first Output features corresponding to each time slice; , , and Indicates the relationship with the first Each time slice inputs the relevant state-space parameters.

[0169] The forward Mamba-2 module processes the hidden states in the original time sequence and outputs forward temporal features. The reverse Mamba-2 module processes the hidden state according to the reversed time sequence and outputs the reversed temporal features. , is represented as:

[0170]

[0171]

[0172] in, This represents the normalized hidden representation; This indicates the sequence feature extractor used in the bidirectional time-series path. In a preferred embodiment, the sequence feature extractor is implemented using the Mamba-2 model. This represents a tensor inversion operation along the time dimension.

[0173] Subsequently, the forward temporal features and reverse time series features Averaging and fusion are performed, and a bidirectional temporal path output is obtained through convolutional mapping layers, activation functions, random deactivation, and residual connections:

[0174]

[0175]

[0176] in, This represents the convolution operation; Indicates the learnable weight parameters; This represents the corresponding learnable bias parameter; Indicates the characteristics of intermediate residuals; Represents a nonlinear activation function; Indicates a random deactivation operation; Indicates the first The timing characteristics of the outputs of the bidirectional timing paths in a dual-path residual block. Used in residual connections. Indicates the first The original input hidden representation of each bidirectional residual block is added to the features after normalization, temporal feature extraction, and mapping to preserve input features and stabilize the training process of multi-layer residual blocks. The bidirectional temporal path is used to capture motion trends, state transitions, and dynamic evolution patterns across a long time range in UAV flight data.

[0177] Local interaction paths are used to extract local dynamic features within a group under the constraints of the variable group structure, and to perform cross-group feature interaction in conjunction with local flight state representations. Specifically, in the first... In a dual-path residual block, the local interactive path receives the input hidden representation. And hide the input representation After normalization, the data is divided into multiple group-level sub-representations based on the prior structure of the variable groups. For each group-level sub-representation, the local interaction path can perform local convolution, depthwise separable convolution, pointwise convolution, or other intra-group feature extraction operations to extract local dynamic features within the variable group. For the first... A group of variables, local dynamic characteristics within the group It can be represented as:

[0178]

[0179] in, Indicates input hiding The first result obtained after normalization and prior partitioning according to the variable group structure Group-level sub-representations; Indicates the first The depthwise convolution parameters corresponding to each variable group; Indicates the first Pointwise convolution parameters corresponding to each set of variables; Indicates the first The depthwise convolution bias parameters corresponding to each set of variables; Indicates the first Pointwise convolution bias parameters corresponding to each set of variables; Represents a nonlinear activation function; This indicates a convolution operation.

[0180] By combining the local features of each group, a group-level local feature representation can be obtained. Subsequently, the group-level local feature representation and the local flight state representation are combined. Concatenate the data and implement condition-guided cross-group interactions through a cross-group mapping layer:

[0181]

[0182]

[0183] in, Indicates the first The group-level local feature representation obtained by combining the intra-group local dynamic features of each variable group in a dual-path residual block; Indicates feature concatenation operation; and This represents the learnable weight parameters in the cross-group mapping layer; and This represents the learnable bias parameters in the cross-group mapping layer; Represents a nonlinear activation function; This represents the conditionally guided cross-group interaction features obtained after fusing local flight state representations. Indicates the first The input hidden representation of a dual-path residual block; Indicates a random deactivation operation; Indicates the first The local interaction feature representation of the local interaction path output in each dual-path residual block. The residual connection uses... Indicates the first The original input hidden representation of each dual-path residual block is added to the conditionally guided cross-group interaction features extracted from the local interaction paths to preserve input features and stabilize the training process of multi-layer residual blocks.

[0184] Therefore, local interaction paths can not only extract local temporal features within a variable group, but also adjust the feature interactions between different variable groups by leveraging local flight state conditions.

[0185] The group-aware flight state gating module is used to represent the diffusion step. and global flight status representation Generate group-level gating coefficients. Specifically, this can be done by... and After performing the feature concatenation operation (Concat), the input is given to a multilayer perceptron (MLP) to obtain the gating coefficients corresponding to the bidirectional temporal path, the local interaction path, and the skip connection output, respectively.

[0186]

[0187]

[0188] in, Indicates the first The gated pre-activation vectors corresponding to each dual-path residual block; , and These represent the gated pre-activation components corresponding to the bidirectional timing path, the local interaction path, and the jump-connection output, respectively; Indicates will , and Concatenate according to feature dimensions; This represents a non-linear activation function, preferably the Sigmoid function; , and These are group-level gating coefficients corresponding to the bidirectional temporal path, local interaction path, and jump-connection output, respectively. These gating coefficients can be broadcast along the feature dimensions within the variable group to make their dimensions consistent with the hidden state dimensions, thereby adjusting the weights of the bidirectional temporal path output, local interaction path output, and jump-connection output on each variable group.

[0189] After obtaining the gating coefficients, the bidirectional temporal path output and the local interaction path output can be weighted and fused to obtain the features. :

[0190]

[0191] in, This represents element-wise multiplication. and This represents the gated vector after broadcasting along the feature dimension within the group. Subsequently, the main residual branch is obtained through post-processing mapping. :

[0192]

[0193] in, This represents the convolution operation. This represents the learnable weight parameters in the post-processing mapping; This represents the learnable bias parameters in the post-processing mapping; Represents a nonlinear activation function; This indicates a random deactivation operation.

[0194] No. The hidden state of the output of each two-path residual block is represented as follows:

[0195]

[0196] in, Indicates the convolution operation; This represents the learnable weight parameters used to map the main residual branch features to the hidden state space; This represents the corresponding learnable bias parameter; This represents the residual update term obtained by mapping the features of the main residual branches;

[0197] 2 scaling factor It can stabilize characteristic amplitudes when multiple residuals are stacked. Furthermore, the residual block can generate skip-connected outputs for subsequent group-level output header aggregation.

[0198]

[0199] in, Indicates the convolution operation; Indicates the first The skip-connected output generated by a dual-path residual block; Indicates the first The gating vector is obtained by broadcasting the gating coefficients of the skip-connection output gating coefficients along the feature dimension inside the variable group in a dual-path residual block; This represents element-wise multiplication; This represents the learnable weight parameters used to generate the jump-connected output; This represents the corresponding learnable bias parameter; This represents the jump-connected output candidate features obtained after mapping the main residual branch features.

[0200] Through the above gating design, the model can adaptively adjust the global temporal dependency, the interaction of local variable groups, and the contribution of jump outputs to each variable group according to the current denoising stage and the overall flight state of the UAV, thereby improving the pertinence of noise estimation and the dynamic rationality of the recovery results.

[0201] Reference Figure 7The group-level output header is used to map the deep hidden representation extracted from the dual-path residual block back to the original UAV flight variable space and generate the full prediction noise for the current diffusion step.

[0202] In one specific embodiment, if the noise estimation network includes For each dual-path residual block, the skipped outputs of all residual blocks can be aggregated and normalized first.

[0203]

[0204] in, This represents aggregated features. By aggregating the skip-connected outputs of multiple residual blocks, both shallow local pattern information and deep high-order semantic information can be preserved simultaneously, reducing the excessive dependence of the final noise prediction on the last hidden state.

[0205] Since the hidden dimension can be divided into multiple group-level subspaces according to the prior structure of the variable groups, the aggregated features can be... Decomposed into multiple group-level sub-representations along the feature dimension:

[0206]

[0207] in, Indicates the first Each variable group corresponds to a group-level sub-representation. For any variable group, the corresponding prediction noise can be generated using an independent group-level output header. :

[0208]

[0209] in, , and corresponding bias , For the first The parameters are learned independently for each group of variables. Using different output heads for different groups of variables can reduce the risk of excessive mixing of different physical variable groups in the output stage, thus maintaining the consistency of the variable group structure in the output stage.

[0210] Repeat the above process until group-level prediction noise for all variable groups has been generated. Then, concatenate the group-level prediction noise according to the variable group order to obtain the prediction noise for the current diffusion step:

[0211]

[0212] Complete prediction noise Noisy flight samples Having the same time length and variable dimensions, it can be used for subsequent backdiffusion updates.

[0213] Reference Figure 8 This embodiment describes the training process of the pre-trained group structure conditional diffusion recovery model. The goal of the training phase is to learn a conditional noise predictor for missing locations, enabling it to accurately predict the noise added to the missing locations during forward diffusion, given partial observation flight data, observation mask, variable group structure priors, and flight state condition information.

[0214] In one specific embodiment, the training set used to train the group structure conditional diffusion recovery model It can be represented as ,in; Indicates the index of the time-series flight samples of UAVs in the training set; This represents the total number of time-series flight samples of UAVs in the training set; Represents the first in the training set A complete UAV flight time sequence sample; Indicates and The corresponding observation mask; Indicates and The corresponding flight status condition information. The training process may include the following steps:

[0215] T1: Initialize noise estimation network parameters Set the number of training epochs and the number of diffusion steps. .

[0216] T2: Obtain complete UAV flight time series samples from the training set. Observation mask and flight status information and based on Obtained partial observation flight data .

[0217] T3: From the diffusion step set Uniform sampling diffusion step and from the standard Gaussian distribution Medium sampling noise tensor ε.

[0218] T4: Based on sampling diffusion step Complete flight sample and noise tensor Construct the current noisy flight sample :

[0219] .

[0220] T5: Noisy Flight Sample Partial observation flight data Observation mask Flight status information Prior knowledge of variable group structure and diffusion step Input noise estimation network , to obtain predicted noise :

[0221] .

[0222] T6: Determine the set of missing locations , And only on the set of missing locations, calculate the loss between predicted noise and true noise. ;in, Indicates the time slice index; Indicates the index of flight data variables; This represents the position of a time slice variable in a drone flight time series sample; Indicates the first A complete drone flight time sequence sample The corresponding observation mask; Indicates observation mask The Middle The first time slice, the first Mask elements at the locations of flight data variables; This indicates that data for the corresponding time slice and the corresponding variable position is missing; The predicted loss for diffuse noise is calculated as follows:

[0223]

[0224] in, The noise estimation network is represented in the first... The predicted noise output in the nth diffusion step is at the 1st... The first time slice, the first The values ​​at the locations of each flight data variable; Indicates the first The true noise corresponding to the nth diffusion step is at the nth diffusion step. The first time slice, the first The values ​​at the locations of each flight data variable; Represents the set of missing locations Sum all missing positions.

[0225] T7: Based on loss Perform gradient descent to update the noisy estimated network parameters. Alternatively, Adam, AdamW, SGD, or other gradient optimization algorithms can be used.

[0226] In one specific embodiment, the training phase can employ a preset number of training epochs, a preset learning rate, and a preset number of diffusion steps for model training. For example, 400 training epochs, a learning rate of 0.001, and 200 diffusion steps can be configured as parameters. The above parameters are merely one optional implementation and do not constitute a limitation on the scope of protection of this invention.

[0227] By calculating the noise prediction loss only at the missing locations, the model training process can focus on the locations that actually need to be recovered, avoiding the noise prediction of observed locations from dominating the model optimization direction, thereby improving the model's adaptability to missing location recovery tasks in complex missing scenarios.

[0228] Reference Figure 9 This embodiment describes the process of recovering multi-sensor flight data from an unmanned aerial vehicle (UAV). During the recovery phase, the input to the group structure conditional diffusion recovery model is the portion of the observed flight data to be recovered. Observation mask Flight status information and variable group structure prior The output is the recovered complete UAV multi-sensor flight data. .

[0229] In one specific embodiment, the flight data recovery process may include the following steps.

[0230] R1: From the standard Gaussian distribution Initialize the reverse diffusion start point .

[0231] R2: Perform reverse diffusion update in descending order of diffusion steps, i.e., let from Decrease to 0.

[0232] R3: In each back-diffusion step, first based on the observation mask Perform observation position replacement so that the current noisy flight sample remains the corresponding partial observation flight data at the previously observed position:

[0233]

[0234] in, Used to select missing locations This step is used to provide the true observation values ​​for observed locations. It ensures that the backdiffusion process does not modify known flight data, but only iteratively recovers missing locations.

[0235] R4: Replace the noisy flight sample Partial observation flight data Observation mask Flight status information Prior knowledge of variable group structure and diffusion step Input the trained noise estimation network to obtain the predicted noise for the current diffusion step. .

[0236] R5: Based on predicted noise Perform a reverse diffusion update to obtain In one specific embodiment, the update method can be adopted as follows:

[0237]

[0238] in, It can be sampled from a standard Gaussian distribution. Optionally, in the last few diffusion steps or deterministic recovery settings, it can be reduced or omitted. item.

[0239] R6: Repeat R3 to R5 until all diffusion steps are completed.

[0240] R7: Based on the final result and some observation flight data Output the recovered UAV multi-sensor flight data :

[0241] .

[0242] Through the above-described recovery process, this invention consistently constrains the observed flight data to a known state during the sampling and recovery phase, and only performs backdiffusion updates on missing locations. This reduces the risk of unnecessary modification to the observed flight data and improves the usability of the recovery results for subsequent UAV status monitoring, anomaly detection, and flight safety analysis tasks. The recovered UAV multi-sensor flight data can be used for UAV status monitoring, anomaly detection, trajectory prediction, flight safety assessment, or subsequent intelligent analysis tasks, enhancing the support capabilities of the recovered data for status monitoring, anomaly detection, and flight safety analysis tasks.

[0243] The example further illustrates this using UAV flight telemetry data:

[0244] In one specific embodiment, the UAV platform can be equipped with a global positioning module, an inertial measurement unit, a flight control module, and an environmental perception module. The raw flight log may include timestamps, position information, velocity information, attitude information, angular velocity information, acceleration information, heading information, and flight status labels. First, the flight log is sorted by timestamps, and different sensor fields are time-aligned; then, selected flight variables are standardized, and a fixed-length flight time series sample is constructed using a sliding window.

[0245] In one specific embodiment, flight state condition information may include state labels such as idling hover, ascent, turn, descent, level flight, acceleration, deceleration, and mission switching, or it may include continuous state vectors output by the flight control system or state estimation algorithm. If the state labels are discrete, they can be converted into one-hot vectors; if the flight state condition information is continuous, it can be directly used as the state condition input of the model; if the flight state condition information changes over time, it can be organized into a time-varying flight state sequence.

[0246] During the training phase, observation masks can be generated based on real missing records or simulated missing patterns. Simulated missing patterns can include random point missing, continuous segment missing, and variable channel missing. By introducing multiple missing patterns during the training phase, the noise estimation network can learn to recover missing flight data under different missing conditions using observation constraints, variable set structure priors, and flight state condition information.

[0247] During the recovery phase, when data is missing from the actual drone flight logs, an observation mask can be constructed first. and some observation flight data Then, a priori variable group structure is constructed based on field names, sensor sources, or physical meanings. And obtain the corresponding flight status information. Subsequently, the above information is input into the trained group structure conditional diffusion recovery model; finally, the model outputs the recovered complete UAV multi-sensor flight data. .

[0248] In one optional embodiment, the recovered UAV multi-sensor flight data can be further input into a UAV anomaly detection model, trajectory prediction model, state estimation model, or flight safety assessment model. Because the recovery process maintains the observed positions unchanged and performs generative recovery at the missing positions using group structures and flight state conditions, the recovered UAV multi-sensor flight data can better preserve the local coordination relationships between UAV flight variables and the dynamic rationality under different flight states.

[0249] It should be noted that the specific network structures of each module in the above embodiments can be replaced or combined according to actual application requirements. For example, the state encoder can use linear layers, one-dimensional convolution, recurrent neural networks, Transformer encoders, or state-space models; the group-level encoder can use one-dimensional convolution, temporal convolution, linear mapping, multilayer perceptrons, or attention modules; the bidirectional temporal path can use Mamba-2, other state-space sequence models, recurrent neural networks, temporal convolutional networks, or Transformer encoders; and the local interaction path can use depthwise separable convolution, ordinary convolution, graph convolution, or attention mechanisms. Any implementation that can achieve the same or similar functionality within the group structure conditional diffusion recovery framework of this invention should be considered an alternative implementation of this invention.

[0250] It should also be noted that the specific prior division of the variable group structure in the above embodiments should not be construed as a limitation on the scope of protection of this invention. Under different UAV platforms, different flight missions, and different sensor configurations, the variable groups can be adjusted according to the sensor source, physical meaning, statistical attributes, flight control functions, or local correlations. The variable groups can be manually preset or determined based on variable correlation, domain knowledge, sensor structure, or data-driven methods.

[0251] Furthermore, the number of diffusion steps, training epochs, learning rate, hidden dimension, number of variable groups, sliding window length, optimizer type, and missing data ratio in the above embodiments are all optional parameters. Those skilled in the art can adjust these parameters based on the scale of UAV flight data, sampling frequency, missing data severity, computational resources, and real-time requirements. These parameters should not be construed as limiting the scope of protection of this invention.

[0252] In summary, this invention constructs a group structure conditional diffusion recovery model by using the observation mask, variable group structure prior, and flight state condition information of UAV multi-sensor flight data as conditional information. It then utilizes state and diffusion step encoding, dual-stream group structure encoding, dual-path residual blocks, group-aware flight state gating, and group-level output heads to achieve conditional noise prediction. During the back-diffusion process, it keeps the observed positions unchanged and only iteratively recovers the missing positions, thereby improving the accuracy, structural consistency, and dynamic rationality of UAV multi-sensor flight data recovery under complex missing conditions.

Claims

1. A method for recovering multi-sensor flight data of unmanned aerial vehicles based on group structure conditional diffusion, characterized in that: Includes the following steps, S1. Acquire the raw flight data collected by the UAV multi-sensor system, and preprocess the raw flight data to obtain the flight time sequence sample of the UAV to be recovered; S2. Construct an observation mask based on the missing data of the UAV flight time series samples to be recovered, and obtain partial observation flight data based on the observation mask; S3. Based on the sensor source, physical meaning, statistical attributes, flight control function or local correlation of each variable in the flight time series sample of the UAV to be recovered, construct a variable group structure prior, and divide the variables in the flight time series sample of the UAV to be recovered into multiple variable groups based on the variable group structure prior. S4. Obtain the flight status condition information corresponding to the flight time sequence sample of the UAV to be recovered; S5. Input some observed flight data, observation mask, variable group structure priors, and flight state condition information into the pre-trained group structure conditional diffusion recovery model. The group structure conditional diffusion recovery model includes a noise estimation network and a data recovery module. During the back diffusion process of the group structure conditional diffusion recovery model: The noise estimation network obtains the predicted noise for the current diffusion step based on the noisy flight sample corresponding to the current diffusion step, some observed flight data, observation mask, variable group structure prior, flight state condition information and diffusion step information; The data recovery module performs backdiffusion update based on the noisy flight sample and predicted noise corresponding to the current diffusion step, and keeps the data at the observed positions unchanged based on the observation mask, only iteratively recovering the data at the missing positions to obtain the noisy flight sample corresponding to the next diffusion step; after completing the backdiffusion process, it outputs the recovered UAV multi-sensor flight data.

2. The method for recovering UAV multi-sensor flight data based on group structure conditional diffusion as described in claim 1, characterized in that: In step S1, the raw flight data is preprocessed to obtain the flight time sequence sample of the UAV to be recovered, specifically as follows: Acquire raw flight data from at least two types of acquisition units, including the UAV's global positioning module, inertial measurement unit, altimeter, velocity sensor, attitude sensor, heading sensor, environmental perception module, or flight control system. The raw flight data is processed by at least one of the following methods: timestamp alignment, sampling frequency unification, outlier sampling removal, missing location marking, and data standardization, to obtain preprocessed continuous flight data. The preprocessed continuous flight data is reconstructed using a sliding window to obtain a flight time series sample of the UAV to be recovered, which has a preset time length and preset variable dimensions.

3. The method for recovering UAV multi-sensor flight data based on group structure conditional diffusion as described in claim 1, characterized in that: In step S2, an observation mask is constructed based on the missing data in the flight time series samples of the UAV to be recovered, and partial observation flight data is obtained based on the observation mask. Specifically, The flight time series sample of the UAV to be recovered is denoted as X. The flight time series sample X of the UAV to be recovered includes multiple time slices and multiple flight data variables. Construct an observation mask M that has the same time dimension and variable dimension as the flight time series sample X of the UAV to be recovered; When the data of the UAV flight time series sample X to be recovered is observable at the corresponding time slice and the corresponding variable position, the element at the corresponding position in the observation mask M is set to the first value; When the data of the UAV flight time series sample to be recovered is missing at the corresponding time slice and the corresponding variable position, the element at the corresponding position in the observation mask M is set to the second value; Based on the observation mask M, the flight time series sample X of the UAV to be recovered is masked to obtain partial observation flight data. .

4. The method for recovering UAV multi-sensor flight data based on group structure conditional diffusion as described in any one of claims 1-3, characterized in that: In step S3, the prior structure of the variable group includes multiple variable groups, which include at least two of the following: position variable group, velocity variable group, attitude variable group, angular velocity variable group, acceleration variable group, altitude variable group, heading variable group, and environmental observation variable group. Each variable group includes at least one flight data variable. Different variable groups are used to characterize a set of variables in UAV multi-sensor flight data that have different sensor sources, physical meanings, statistical scales, flight control functions, or local correlations.

5. The method for recovering UAV multi-sensor flight data based on group structure conditional diffusion as described in any one of claims 1-3, characterized in that: In step S5, the noise estimation network includes a state and diffusion step encoding module, a two-stream group structure encoding module, at least one two-path residual block, and a group-level output header. State and Diffusion Step Encoding Module: Used to encode flight state condition information to obtain flight state representation, including local flight state representation and global flight state representation, and to encode diffusion step information to obtain diffusion step representation; The dual-stream group structure encoding module performs group-level encoding on the noisy flight samples and observation masks by the noisy branch to obtain the noisy branch global group structure representation; and performs group-level encoding on some observed flight data, observation masks, and local flight state representations by the conditional branch to obtain the conditional branch global group structure representation; after fusing the noisy branch global group structure representation, the conditional branch global group structure representation, the group identity embedding, and the diffusion step embedding, the initial hidden representation is obtained and output. Dual-path residual block: The initial hidden representation of the input is used to extract long-term temporal dependency features of UAV flight data through bidirectional temporal paths, and local dynamic features are extracted by local interaction paths; Generate group-level gating coefficients, and generate hidden states and skip outputs from group-level gating coefficients, long-term time-dependent features, and local dynamic features; Group-level output head: Used to aggregate the jump-connected outputs of dual-path residual blocks to obtain aggregated features. According to the prior structure of variable groups, the aggregated features are split into multiple group-level sub-representations. Group-level prediction noise is generated through the output head corresponding to each variable group, and the group-level prediction noises are concatenated to output the prediction noise corresponding to the current diffusion step.

6. The method for recovering UAV multi-sensor flight data based on group structure conditional diffusion as described in claim 5, characterized in that: In the state and diffusion step encoding module, flight state condition information is encoded to obtain flight state representations, including local and global flight state representations. Diffusion step information is encoded to obtain diffusion step representations. Specifically, When the flight state condition information is a sample-level flight state vector, the sample-level flight state vector is mapped to a global flight state representation, and then copied along the time dimension to obtain a local flight state representation. When the flight state condition information is a time-varying flight state sequence, the time-varying flight state sequence is time-series mapped to obtain a local flight state representation, and the local flight state representation is time-dimensionally aggregated to obtain a global flight state representation; Embedding encoding is performed on the current diffusion step to obtain the diffusion step representation; Among them, the local flight state is used to characterize the flight state conditions corresponding to each time slice, the global flight state is used to characterize the overall operation stage of the current flight sample, and the diffusion step is used to characterize the denoising stage of the current back diffusion process.

7. The method for recovering UAV multi-sensor flight data based on group structure conditional diffusion as described in claim 5, characterized in that: The two-stream group structure coding module includes noisy branching, conditional branching, and a fusion mapping layer. Noisy branch: used to divide the noisy flight sample and observation mask corresponding to the current diffusion step into multiple group-level noisy inputs according to the variable group structure prior, and encode multiple group-level noisy inputs respectively to obtain multiple group-level noisy embeddings, which are then concatenated to obtain the noisy branch global group structure representation; Conditional branch: It is used to divide and concatenate some observed flight data, observation mask and local flight state representation according to the variable group structure prior to obtain multiple group-level conditional inputs, and encode multiple group-level conditional inputs to obtain multiple group-level conditional embeddings. After concatenation, the global group structure representation of the conditional branch is obtained. Fusion Mapping Layer: The noisy branch global group structure representation, the conditional branch global group structure representation, the group identity embedding consisting of a set of learnable parameters, and the diffusion step embedding obtained by mapping the diffusion step representation are fused to obtain the initial hidden representation for subsequent noise estimation.

8. The method for recovering UAV multi-sensor flight data based on group structure conditional diffusion as described in claim 7, characterized in that: In the noise estimation network, the dual-path residual block includes a bidirectional temporal path, a local interactive path, and a group-aware flight state gating module. Bidirectional temporal path: used to perform bidirectional sequence feature extraction on the input hidden representation to obtain long temporal dependency features in UAV multi-sensor flight data; Local interaction path: Under the constraint of the variable group structure prior, the input hidden representation is split into multiple group-level sub-representations, and local features within each group-level sub-representation are extracted. The local features within the group are then fused with the local flight state representation to obtain condition-guided cross-group interaction features. Group-aware flight state gating module: used to generate group-level gating coefficients based on the diffusion step representation and the global flight state representation. The group-level gating coefficients include bidirectional timing path gating coefficients, local interaction path gating coefficients, and jump-connection output gating coefficients. It also adjusts the outputs of the bidirectional timing path and the local interaction path based on the group-level gating coefficients to generate hidden representations and jump-connection outputs for subsequent group-level output head aggregation.

9. The method for recovering UAV multi-sensor flight data based on group structure conditional diffusion as described in any one of claims 1-3, characterized in that: In step S5, the training process of the pre-trained group structure conditional diffusion recovery model includes: Acquire complete UAV flight time sequence samples, observation masks, and flight status condition information, and obtain partial observation flight data based on the observation masks; Sample diffusion steps from the set of diffusion steps, and sample noise tensors from a preset noise distribution; Noisy flight samples are constructed based on complete UAV flight time series samples, diffusion steps, and noise tensors; Noisy flight samples, partial observation flight data, observation mask, diffusion step, and flight state condition information are input into the noise estimation network to obtain the predicted noise. Calculate the noise prediction loss between the predicted noise and the noise tensor on the set of missing locations; The parameters of the noise estimation network are updated based on the noise prediction loss.

10. The method for recovering UAV multi-sensor flight data based on group structure conditional diffusion as described in any one of claims 1-3, characterized in that: In step S5, the reverse diffusion process of the group structure conditional diffusion recovery model is specifically as follows: Initialize the back diffusion start point from the preset noise distribution; Perform reverse diffusion updates sequentially in descending order of diffusion steps; In each back-diffusion step, the observed positions in the current noisy flight sample are replaced with the corresponding partial observed flight data based on the observation mask to obtain the replaced noisy flight sample. The replaced noisy flight sample is input into the trained noise estimation network to obtain the predicted noise corresponding to the current diffusion step. Update missing location data in the current noisy flight sample based on predicted noise; After all the reverse diffusion steps are completed, the recovered UAV multi-sensor flight data is obtained.

Images