Unmanned aerial vehicle historical data loss compensation system

Abstract

The invention discloses an unmanned aerial vehicle historical data loss compensation system, and particularly relates to the field of unmanned aerial vehicle data processing. Comprising a data acquisition module, a judgment condition establishment module, a judgment condition I module, a judgment condition II module, a data state evaluation module, a data initialization module, a GAN generation module, a Transform modeling module, a compensation verification optimization module and a data output module. By introducing a judgment condition establishment module and a data state evaluation module, a pre-evaluation mechanism is set in the scheme before a compensation process, a judgment condition I and a judgment condition II are established by utilizing a data integrity proportion and a time interval distribution deviation, and whether data are abnormal or not is quickly screened; by means of the design, redundant calculation of data without significant anomalies is effectively avoided, and system resources and processing time are saved.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle (UAV) data processing, and more specifically, to a UAV historical data loss compensation system. Background Art

[0002] UAV historical data loss compensation refers to the process of modeling, predicting, and compensating for missing data caused by equipment failures, signal interference, or environmental restrictions during flight through algorithms, with the aim of reconstructing the integrity of historical data and ensuring the accuracy of subsequent analysis and mission execution.

[0003] In the prior art, UAV data loss compensation systems usually directly model and compensate for missing data, but lack a pre-evaluation and screening link, resulting in the system consuming a large amount of resources to execute complex compensation algorithms even when the degree of data loss is relatively small or there are no significant abnormalities. This design without screening easily causes redundant waste of computing resources and time. Summary of the Invention

[0004] To overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a UAV historical data loss compensation system. By introducing a determination condition establishment module and a data status evaluation module, a pre-evaluation mechanism is set up before the compensation process in the solution. Determination condition 1 and determination condition 2 are established using the data integrity ratio and the time interval distribution deviation to quickly screen whether the data is abnormal, so as to solve the problems raised in the above background art.

[0005] To achieve the above object, the present invention provides the following technical solution: A UAV historical data loss compensation system includes a data acquisition module, a determination condition establishment module, a determination condition 1 module, a determination condition 2 module, a data status evaluation module, a data initialization module, a GAN generation module, a Transformer modeling module, a compensation verification and optimization module, and a data output module.

[0006] The data acquisition module, when the UAV is running, acquires real-time UAV time-series consistency data; the UAV time-series consistency data includes the data integrity ratio and the data time interval distribution deviation.

[0007] The determination condition establishment module establishes determination condition 1 and determination condition 2 based on the UAV time-series consistency data. If either determination condition 1 or determination condition 2 determines an abnormality, a data status evaluation model is established for the historical data of the UAV.

[0008] The determination condition 1 module is used to establish determination condition 1; Determination condition 1: If the data integrity ratio is less than the system-predefined data integrity ratio threshold, it is determined as abnormal.

[0009] The determination condition two module is used to establish determination condition two; determination condition two: if the distribution deviation of the data time interval is greater than the system - preset data time - interval distribution deviation threshold, it is determined as abnormal;

[0010] The data status evaluation module evaluates the historical data of the UAV through the data status evaluation model, calculates the data status evaluation value based on the data status evaluation model. If the data status evaluation value is less than the data status evaluation threshold, it is determined as an abnormal loss state;

[0011] The data initialization module is used to classify the data loss state in the abnormal loss state, initialize the combined model based on the generative adversarial network (GAN) and the Transformer, and configure network parameters adapted to different states;

[0012] The GAN generation module models the potential distribution of the lost data through the generation ability of the generative adversarial network (GAN); the generator of the generative adversarial network (GAN) generates preliminary compensation data according to the characteristics of the historical data, and the discriminative network evaluates the similarity between the generated data and the real data to improve the authenticity of the generated data in an adversarial optimization manner;

[0013] The Transformer modeling module is used to align or fuse the preliminary compensation data generated by the generative adversarial network (GAN) with the original data, and then perform global sequence modeling through the Transformer model; based on the multi - head attention mechanism of the Transformer, it captures the long - term dependencies and local context information of the time - series data;

[0014] The compensation verification and optimization module is used to verify the compensation data generated by the combination of the generative adversarial network (GAN) and the Transformer, and measure the compensation effect through error analysis; if the verification result does not meet the standard, adjust the training parameters of the generative adversarial network (GAN) or the modeling range of the Transformer for iterative optimization until the compensation effect expected by the system is achieved;

[0015] The data output module is used to integrate and output the compensation data with the complete historical data.

[0016] In a preferred embodiment, the data status evaluation model includes time - dimension integrity evaluation, space - dimension distribution evaluation, and feature - dimension consistency evaluation;

[0017] The time - dimension integrity evaluation includes data integrity ratio, data time - interval distribution deviation, and data missing rate in a time period;

[0018] The space - dimension distribution evaluation includes space coverage integrity, regional data missing distribution, and neighborhood data correlation;

[0019] Feature dimension consistency evaluation includes feature distribution deviation, multi-feature association consistency, and environmental condition adaptability.

[0020] In a preferred embodiment, calculate the data integrity ratio in the UAV timing consistency data; formulate the data integrity ratio as P int ;

[0021]

[0022] where M represents a total of M valid data segments; ω k is the weight factor of the kth data segment; δ k is the time length of the kth data segment; T max is the end time during the UAV operation; T min is the start time during the UAV operation; T max -T min represents the entire operation time range of the UAV;

[0023] Calculate the data time interval distribution deviation in the UAV timing consistency data; formulate the data time interval distribution deviation as Δ int ;

[0024]

[0025] where τ(t) is the data sampling interval at time t; τ ref (t) is the reference sampling interval at time t; β is the deviation sensitivity index.

[0026] In a preferred embodiment, the time dimension integrity evaluation includes the data integrity ratio P int , the data time interval distribution deviation Δ int , and the data missing rate R in the time period miss ;

[0027]

[0028] where N is the number of missing data segments; ν j is the importance weight of the jth data missing segment; D j is the time length of the jth data missing segment; T end -T start is the total time range of the task.

[0029] In a preferred embodiment, the spatial dimension distribution evaluation includes spatial coverage integrity C space , regional data missing distribution D region and neighborhood data correlation R neigh ;

[0030]

[0031] Where P is the number of effective coverage areas; η k is the importance weight of the k-th area; A k is the effective coverage area of the k-th area; A total is the total coverage area; Q is the total number of areas; ζ m is the importance weight of the m-th area; F m is the data missing ratio of the m-th area; R is the number of data point pairs in the neighborhood; φ n is the importance weight of the n-th neighborhood data pair; S n is the similarity metric value of the n-th neighborhood data pair.

[0032] In a preferred embodiment, the feature dimension consistency evaluation includes the feature distribution deviation D feat , the multi-feature association consistency C assoc and the environmental condition adaptability A env ;

[0033]

[0034]

[0035] Where S is the total number of feature variables; γ r is the importance weight of the r-th feature; x r is the actual value of the r-th feature; x ref,r is the reference value of the r-th feature; α is the deviation index; U is the number of feature association pairs; δ t is the importance weight of the t-th feature association pair; ρ t is the correlation coefficient of the t-th feature association pair; z min , z max are respectively the lower limit of the environmental variable and the upper limit of the environmental variable; χ(z) is the effectiveness weight of task execution under the environmental condition; ξ(z) is the actual contribution degree of the environmental condition z to the task.

[0036] In a preferred embodiment, a data status evaluation model is established; the data status evaluation value is designated as S eval ;

[0037]

[0038] Where w t represents the weight of the time dimension in the data status evaluation model; w s represents the weight of the space dimension in the data status evaluation model; w f represents the weight of the feature dimension in the data status evaluation model; λ tThe weight for controlling the sensitivity in the time dimension; λ f The weight for controlling the sensitivity in the feature dimension.

[0039] In a preferred embodiment, in the abnormal loss state, the characteristics of data loss are clarified through classification, providing a basis for subsequent model configuration;

[0040]

[0041] where C state is the classification result of the current data loss; K is the total number of all possible loss state categories; φ k is the importance weight of category k; M k is the number of features included in category k; g j,k (F) is the matching degree of the j-th feature under category k, calculated based on the similarity of feature F.

[0042] In a preferred embodiment, the generative adversarial network GAN is used to model the latent distribution of the lost data;

[0043] In the generative adversarial network GAN, through adversarial optimization, the generator G simulates the distribution of the lost data, and the discriminator D evaluates the similarity between the generated data and the real data;

[0044] The optimization objective of the generative adversarial network GAN generator:

[0045]

[0046] The optimization objective of the generative adversarial network GAN discriminator:

[0047]

[0048] where G * represents the parameter solution determined by the generator under the optimization objective; z ∼ P z represents the distribution of the random noise z; G(z; Θ G ) represents the compensated data generated by the generator, where Θ G are the parameters of the generator; T(F) is the target distribution, defined by the historical data feature F; β is the bias sensitivity index;

[0049] where is the loss function of the discriminator; P real is the distribution of the real data; x ∼ P real is the distribution P that the real data sample x follows real ; D(x) is the output value of the discriminator D for the real data sample x; G(z) is the pseudo-data sample generated by the generative adversarial network GAN generator based on the input noise z;

[0050] The pseudo-data sample G(z) generated by the generative adversarial network GAN and the original data X orig are fused, and the Transformer model is used to model the long-term dependencies between data;

[0051] X fuse = α·X orig +(1 - α)·G(z)

[0052] where X fuse is the fused data; α is the data fusion coefficient;

[0053] Establish the multi-head attention mechanism of the Transformer;

[0054]

[0055] where Q, K, and V are the query matrix, key matrix, and value matrix respectively; d k is the dimension of the key matrix; softmax is used to generate attention weights; the matrix operation represents calculating the similarity between the query and the key.

[0056] In a preferred embodiment, the error verification formula is:

[0057]

[0058] where ε is the error value; X target (t) is the value of the target data at time t; γ is the error sensitivity index; t start , t end respectively represent the upper and lower limits of the integral; X fuse (t) is the value of the generated data after fusion at time t;

[0059] The optimization formula is:

[0060]

[0061] where Θ * are the optimized model parameters; is the regularization term; λ is the regularization weight;

[0062] The data integration formula is:

[0063] X final = β 1 ·X fuse + β 2 ·X history

[0064] where X final is the final integrated data; X history is the historical data; β1 is the integration weight coefficient for compensating data; β 2 is the integration weight coefficient for historical data.

[0065] Technical effects and advantages of the present invention:

[0066] 1. By introducing a determination condition establishment module and a data status evaluation module, a pre-evaluation mechanism is set before the compensation process in the solution. Using the data integrity ratio and the time interval distribution deviation to establish determination condition 1 and determination condition 2, quickly screening whether the data is abnormal. This design effectively avoids redundant calculations for data with no significant abnormalities, saving system resources and processing time;

[0067] 2. Using the data status evaluation model to comprehensively evaluate the historical data of the unmanned aerial vehicle, comprehensively calculating the data status evaluation value through three dimensions of time, space, and features, quantifying the severity of data loss and clarifying whether it belongs to an abnormal loss state;

[0068] 3. By introducing the generative adversarial network GAN, the adversarial optimization feature is utilized in the latent distribution modeling of the lost data; the generator generates compensation data according to the characteristics of the historical data, and the discriminator evaluates the similarity between the generated data and the real data to ensure the authenticity and high quality of the generated data;

[0069] 4. Combining the compensation data generated by GAN and the original data, performing global sequence modeling through the Transformer model, using the multi-head attention mechanism to capture the long-term dependencies and local context information of the time-series data, enhancing the time-series consistency and global relevance of the compensation data;

[0070] 5. The compensation verification optimization module verifies the generated compensation data through error analysis. After discovering the deviation between the model generation and the actual requirements, iteratively optimize by adjusting the parameters of GAN and Transformer until the expected effect is achieved, thereby ensuring the accuracy of the compensation result. Brief Description of the Drawings

[0071] Figure 1 is the flowchart of the present invention. Detailed Embodiments

[0072] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0073] Refer to the accompanying drawings of the specification Figure 1, a historical data loss compensation system for an unmanned aerial vehicle (UAV) according to an embodiment of the present invention, includes a data acquisition module, a determination condition establishment module, a determination condition one module, a determination condition two module, a data status evaluation module, a data initialization module, a GAN generation module, a Transformer modeling module, a compensation verification and optimization module, and a data output module;

[0074] The data acquisition module, during the operation of the UAV, acquires the UAV's time-series consistency data in real time; the UAV's time-series consistency data includes the data integrity ratio and the data time interval distribution deviation;

[0075] The determination condition establishment module establishes determination condition one and determination condition two based on the UAV's time-series consistency data. If determination condition one or determination condition two is determined to be abnormal, a data status evaluation model for the UAV's historical data is established;

[0076] The determination condition one module is used to establish determination condition one; determination condition one: if the data integrity ratio is less than the system's preset data integrity ratio threshold, it is determined to be abnormal;

[0077] The determination condition two module is used to establish determination condition two; determination condition two: if the data time interval distribution deviation is greater than the system's preset data time interval distribution deviation threshold, it is determined to be abnormal;

[0078] The data status evaluation module evaluates the UAV's historical data through the data status evaluation model, calculates the data status evaluation value based on the data status evaluation model. If the data status evaluation value is less than the data status evaluation threshold, it is determined to be an abnormal loss state;

[0079] The data initialization module is used to classify the data loss state in the abnormal loss state, initialize a combined model based on the generative adversarial network (GAN) and the Transformer, and configure network parameters adapted to different states;

[0080] The GAN generation module models the potential distribution of the lost data through the generative ability of the generative adversarial network (GAN); the generator of the generative adversarial network (GAN) generates preliminary compensation data according to the characteristics of the historical data, and the discriminator network evaluates the similarity between the generated data and the real data to improve the authenticity of the generated data in an adversarial optimization manner;

[0081] The Transformer modeling module is used to align or fuse the preliminary compensation data generated by the generative adversarial network (GAN) with the original data, and then perform global sequence modeling through the Transformer model; based on the multi-head attention mechanism of the Transformer, capture the long-term dependencies and local context information of the time-series data;

[0082] The compensation verification optimization module is used to verify the compensation data generated by the combination of the Generative Adversarial Network (GAN) and the Transformer, and measure the compensation effect through error analysis. If the verification result does not meet the standard, adjust the training parameters of the GAN or the modeling range of the Transformer for iterative optimization until the expected compensation effect of the system is achieved.

[0083] The data output module is used to integrate and output the compensation data with the complete historical data.

[0084] It should be noted that by obtaining the time-series consistency data of the UAV in real time, anomalies in data integrity and time intervals can be quickly identified. Early screening is achieved through simple judgments of determination condition 1 and determination condition 2, thereby reducing redundant processing of anomaly-free data. After being determined as abnormal, the data status evaluation model is further used to comprehensively evaluate the historical data. Combining the data integrity ratio and the distribution deviation of time intervals, the severity of data loss is quantified and it is determined whether it is in an abnormal loss state, ensuring that data anomalies can be captured as efficiently as possible and targeted compensation measures can be taken in complex tasks.

[0085] By classifying the data loss status, ensure that the most suitable compensation strategy is adopted for different types of losses (such as continuous time loss, concentrated space loss or random loss) to improve the flexibility of the model. Use the generator and discriminator of the GAN to model the potential distribution of the lost data and generate high-quality compensation data, and improve the authenticity of data generation through adversarial optimization. After fusing the compensation data generated by the GAN with the original data, introduce the multi-head attention mechanism of the Transformer to fully capture the long-term dependencies and local context relationships of the time-series data, thereby enhancing the global consistency of the compensation data. In the verification stage, measure the compensation effect through error analysis, find the deviation between the model generation and the actual requirements, and use the parameter adjustment and range optimization of the GAN and the Transformer to iteratively improve the compensation accuracy. Finally, through data integration, fuse the compensation data with the complete historical data, which not only makes up for the data loss but also ensures the consistency with the existing data, thus outputting high-quality compensation results to meet the system expectations.

[0086] The data status evaluation model includes time dimension integrity evaluation, space dimension distribution evaluation, and feature dimension consistency evaluation.

[0087] The time dimension integrity evaluation includes data integrity ratio, data time interval distribution deviation, and data missing rate in time periods.

[0088] The space dimension distribution evaluation includes space coverage integrity, regional data missing distribution, and neighborhood data correlation.

[0089] The evaluation of feature dimension consistency includes feature distribution deviation, multi-feature association consistency, and environmental condition adaptability.

[0090] Calculate the data integrity ratio in the UAV timing consistency data; assume the data integrity ratio is P int ;

[0091]

[0092] Among them, the data integrity ratio P int represents the weighted proportion of valid data points within the total time range and is the quantification result of data integrity; M represents that there are a total of M valid data segments, and each segment has undergone sampling analysis; ω k is the weight factor of the k-th data segment, reflecting the importance of different data segments. For example, the weight of the critical mission area is higher; δ k is the time length of the k-th data segment, that is, the duration of valid data in this segment; T max is the end time during the UAV operation; T min is the start time during the UAV operation; T max -T min represents the entire operation time range of the UAV, which is used to normalize the numerator to a ratio; in the formula is used to sum up the time lengths δ k of all valid data segments after weighting by the weight ω k to ensure that the data ratio reflects the importance of the mission critical area; the above formula is calculated through weighted sum and normalization, reflecting the integrity and relative importance of valid data in the time dimension and is suitable for the case where the segment weights are different;

[0093] Calculate the data time interval distribution deviation in the UAV timing consistency data; assume the data time interval distribution deviation is Δ int ;

[0094]

[0095] Among them, the data time interval distribution deviation Δ int is used to quantify the fluctuation degree of the sampling time interval and is an index of time dimension stability; T min is the start time during the UAV operation, the lower limit of the integral; T max is the end time during the UAV operation, the upper limit of the integral; τ(t) is the data sampling interval at time t, representing the actual sampling interval value; τ ref (t) is the reference sampling interval at time t, representing the ideal sampling interval value; β is the deviation sensitivity index, usually β>1, which controls the amplification degree of the deviation and is used to emphasize larger fluctuations;

[0096] The above formula adopts a weighted integral form of deviation, emphasizing the impact of deviation on time stability. Through normalization, the influence of the sampling interval dimension is eliminated, making it applicable to comparisons of different tasks; in the formula represents the deviation of the sampling time interval |τ(t) - τ ref (t)| weighted (by the β exponent) and integrated over the entire time range to calculate the total deviation; represents the integral after weighting the ideal reference sampling interval τ ref (t), providing a normalization benchmark to ensure the proportionality of the deviation.

[0097] The time dimension integrity assessment includes the data integrity ratio P int , the data time interval distribution deviation Δ int , and the data missing rate R miss ;

[0098]

[0099] where R miss is the data missing rate for the time period, representing the weighted proportion of the lost data segments; N is the number of lost data segments; ν j is the importance weight of the j-th data missing segment; D j is the time length of the j-th data missing segment; T end -T start is the total time range of the task.

[0100] The spatial dimension distribution assessment includes the spatial coverage integrity C space , the regional data missing distribution D region , and the neighborhood data correlation R neigh ;

[0101]

[0102] where C space is the spatial coverage integrity, representing the weighted proportion of the effectively covered area; P is the number of effectively covered areas; η k is the importance weight of the k-th area; A k is the effectively covered area of the k-th area; A total is the total covered area; D region is the regional data missing distribution, measuring the unevenness of data missing in each area; Q is the total number of areas; ζ m is the importance weight of the m-th area; F m is the data missing proportion of the m-th area; R neigh is the neighborhood data correlation, representing the average correlation of adjacent data points; R is the number of data point pairs in the neighborhood; φ nis the importance weight of the nth neighborhood data pair; S n is the similarity metric value of the nth neighborhood data pair.

[0103] Feature dimension consistency evaluation includes feature distribution deviation D feat , multi-feature association consistency C assoc and environmental condition adaptability A env ;

[0104]

[0105] where D feat is the feature distribution deviation, quantifying the weighted deviation between the feature value and the reference value; S is the total number of feature variables; γ r is the importance weight of the rth feature; x r is the actual value of the rth feature; x ref,r is the reference value of the rth feature; α is the deviation index for adjusting sensitivity; C assoc is the multi-feature association consistency, measuring the association consistency between features; U is the number of feature association pairs; δ t is the importance weight of the tth feature association pair; ρ t is the correlation coefficient of the tth feature association pair; A env is the environmental condition adaptability, indicating the task adaptability degree of the environmental conditions; z min , z max are respectively the lower limit of the environmental variable and the upper limit of the environmental variable; χ(z) is the effectiveness weight of task execution under environmental conditions; ξ(z) is the actual contribution degree of environmental condition z to the task; dz represents the infinitesimal change of environmental variable z, which is the differential element of the integration operation.

[0106] Establish a data status evaluation model; the proposed data status evaluation value is S eval ;

[0107]

[0108] where w t represents the weight of the time dimension in the data status evaluation model; w s represents the weight of the space dimension in the data status evaluation model; w f represents the weight of the feature dimension in the data status evaluation model; λ t is the weight for controlling the sensitivity of the time dimension, making it flexible in different tasks; λ f is the weight for controlling the sensitivity of the feature dimension, adjusting the influence intensity of the feature dimension under different tasks;

[0109] For the exponential scaling part in the formula Used for smoothing the evaluation values of each dimension to avoid the interference of extreme values on the overall result; The evaluation values in the time dimension are mapped to the interval (0, 1) using an exponential function (S-shaped curve) to avoid the interference of extreme values on the overall evaluation, and at the same time emphasize the cases with larger deviations; P int -Δ int -R miss Partially consider the three evaluation indicators in the time dimension through subtraction, P int The larger it is, the higher the data integrity, so it has a positive contribution to the overall evaluation, Δ int and R miss The larger they are, the higher the data interval deviation and data missing rate, so they have a negative contribution to the overall evaluation; the use of the exponential function can amplify the impact of small deviations on the final evaluation value, help identify potential problems, and avoid numerical overflow;

[0110] In the formula, C space +R neigh Used to emphasize the positive contribution of spatial coverage integrity and neighborhood data correlation to the overall evaluation, C space The larger it is, the more complete the spatial range covered by the UAV mission, R neigh The larger it is, the stronger the correlation between adjacent data points and the higher the data quality; the subtraction part -D in the formula region Used to reflect the negative impact of data missing distribution on the overall evaluation. The higher the regional data missing ratio, the lower the contribution to the spatial evaluation;

[0111] In the formula The evaluation values in the feature dimension are smoothly mapped to (0, 1) through an exponential function to avoid the extreme interference of abnormal feature values on the overall evaluation; D feat +C assoc +A env Integrates the three indicators in the feature dimension, the feature distribution deviation D feat , the larger the deviation, the higher the inconsistency of the feature distribution, and it has a negative contribution to the overall evaluation; the multi-feature association consistency C assoc , the higher the value, the stronger the association between features, and it has a positive contribution to the overall evaluation; the environmental condition adaptability A env , the higher the value, the more suitable the environmental conditions are for the task, and it has a positive contribution to the overall evaluation.

[0112] In the abnormal loss state, clarify the characteristics of data loss through classification and provide a basis for subsequent model configuration;

[0113]

[0114] Among them, C stateIt represents the classification result of the current data loss, indicating the most suitable loss category; K is the total number of all possible loss status categories, including but not limited to time loss, spatially concentrated loss, randomly distributed loss, etc.; φ k is the importance weight of category k, used to reflect the priority of different categories for task requirements; M k is the number of features included in category k; g j,k (F) is the matching degree of the j-th feature under category k, calculated based on the similarity of feature F; φ in the formula k reflects the category priority through weights. For example, key regions are processed preferentially; g j,k (F) is the matching degree score of each category through specific features, comprehensively evaluating the matching degree between the data and the classification target; the above formula uses the weight φ k weighted feature matching score g j,k (F) to ensure that the classification process focuses on the key categories of the task.

[0115] Use the generative adversarial network GAN to model the latent distribution of the lost data;

[0116] In the generative adversarial network GAN, through adversarial optimization, the generator G simulates the distribution of the lost data, and the discriminator D evaluates the similarity between the generated data and the real data;

[0117] The optimization objective of the generator of the generative adversarial network GAN:

[0118]

[0119] The optimization objective of the discriminator of the generative adversarial network GAN:

[0120]

[0121] where G * represents the parameter solution determined by the generator under the optimization objective, that is, the optimal parameters of the generator. The optimization objective is to make the generated data close to the target distribution; z ∼ P z represents the distribution of the random noise z, usually a Gaussian distribution G(z; Θ G ) represents the compensated data generated by the generator, where Θ G are the parameters of the generator; T(F) is the target distribution, defined by the historical data feature F, used to reflect the latent distribution of the data; β is the bias sensitivity index, amplifying the influence of larger biases;

[0122] where is the loss function of the discriminator, used to optimize the discriminator's ability to distinguish between real data and generated data; P real is the distribution of the real data, used for the training of the adversarial generator; is the symbol of mathematical expectation, representing the calculation of the expected value of a random variable; x ~ P real is the distribution P that the real data sample x follows real , that is, the real distribution of the data set, P real is the real data distribution, indicating that the discriminator samples real data from it; z ~ P z is the distribution P that the input noise z of the generator follows z , usually a Gaussian distribution or a uniform distribution, P z is the noise distribution, indicating that the generator samples the input z from it; D(x) is the output value of the discriminator D for the real data sample x, representing the probability that x is real data, ranging from [0, 1]; G(z) is the pseudo data sample generated by the generative adversarial network GAN generator based on the input noise z; 1 - D(G(z)) is the output value of the discriminator D for the generated data G(z), representing the probability that G(z) is pseudo data;

[0123] The pseudo data sample G(z) generated by the generative adversarial network GAN and the original data X orig are fused, and the Transformer model is used to model the long-term dependencies between data;

[0124] X fuse = α·X orig +(1 - α)·G(z)

[0125] where X fuse is the fused data for subsequent modeling; α is the data fusion coefficient, used to control the proportion of the original data and the generated data; in the formula, α·X orig represents retaining the reliable part of the original data to avoid relying entirely on the generated data; in the formula, (1 - α)·G(z) represents the compensatory data generated by the generator to make up for the lost part;

[0126] Establish the multi-head attention mechanism of the Transformer

[0127]

[0128] where Q, K, and V are the query matrix, key matrix, and value matrix respectively, extracted from the fused data X fuse ; d k is the dimension of the key matrix, a normalization factor, used to avoid excessive numerical values; softmax is used to generate attention weights to measure the correlation between data; the matrix operation represents calculating the similarity between the query and the key to capture the long-term dependencies between data.

[0129] In verification and optimization, the quality of the generated data is evaluated through error analysis during verification, and the model parameters are adjusted by minimizing the error during optimization;

[0130] The error verification formula is as follows:

[0131]

[0132] where ε is the error value, used to measure the deviation between the generated data and the target data; X target (t) is the value of the target data at time t; γ is the error sensitivity index, used to amplify larger errors; t start , t end respectively represent the upper and lower limits of the integral, that is, the start time and end time of the data verification period; X fuse (t) is the value of the fused generated data at time t; in the formula, |X fuse (t) - X target (t)| γ represents the absolute error at time t. After being amplified by the deviation sensitivity index γ, it emphasizes larger deviations; represents the cumulative value of the error within the entire time range, and the integral is used to calculate the deviation over continuous time;

[0133] The optimization formula is as follows:

[0134]

[0135] where Θ * is the optimized model parameter, including the parameter set of the generator G Transformer model. The goal is to find the optimal parameter configuration by minimizing the error and regularization constraints; is the regularization term, used to prevent overfitting of the model. It improves the generalization ability by restricting the size or complexity of the model parameters, usually using or sparse regularization is represented; λ is the regularization weight, used to control the strength of regularization; represents finding the parameter Θ that minimizes the objective function;

[0136] Finally, the generated data is integrated with the historical data to generate the final compensation result;

[0137] The data integration formula is as follows:

[0138] X final = β 1 ·X fuse + β 2 ·X history

[0139] where X final is the final integrated data; X history is the historical data; β 1 is the integration weight coefficient of the compensation data; β2 is the integration weight coefficient for historical data.

[0140] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. A UAV historical data loss compensation system, comprising a data acquisition module, a judgment condition establishment module, a judgment condition one module, a judgment condition two module, a data state evaluation module, a data initialization module, a GAN generation module, a Transformer modeling module, a compensation verification optimization module, and a data output module, characterized in that: The data acquisition module acquires the UAV timing consistency data in real time when the UAV is running; the UAV timing consistency data includes data integrity ratio and data time interval distribution deviation; The judgment condition establishment module establishes judgment condition 1 and judgment condition 2 based on the UAV time series consistency data. If judgment condition 1 or judgment condition 2 is judged to be abnormal, a data status evaluation model is established for the historical data of the UAV. The judgment condition one module is used to establish judgment condition one; judgment condition one: if the data integrity ratio is less than the data integrity ratio threshold preset by the system, it is judged as abnormal; The judgment condition 2 module is used to establish the judgment condition 2; judgment condition 2: if the data time interval distribution deviation is greater than the data time interval distribution deviation threshold preset by the system, it is judged as abnormal; The data status evaluation module evaluates the historical data of the drone through the data status evaluation model, and calculates the data status evaluation value based on the data status evaluation model. If the data status evaluation value is less than the data status evaluation threshold, it is determined to be an abnormal loss state; The data initialization module is used to classify the data loss state when it is in an abnormal loss state, initialize the combined model based on the generative adversarial network GAN and the transformer Transformer, and configure the network parameters that adapt to different states; The GAN generation module models the potential distribution of missing data through the generation capability of the Generative Adversarial Network (GAN); The generator of the Generative Adversarial Network (GAN) generates preliminary compensation data based on the characteristics of historical data, and the discriminant network evaluates the similarity between the generated data and the real data, thereby improving the authenticity of the generated data in an adversarial optimization manner; The Transformer modeling module is used to align or fuse the preliminary compensated data generated by the Generative Adversarial Network (GAN) with the original data, and then perform global sequence modeling through the Transformer model; The transformer-based multi-head attention mechanism captures the long-term dependencies and local context information of time series data; The compensation verification and optimization module is used to verify the compensation data generated by the combination of the generative adversarial network GAN and the transformer Transformer, and measure the compensation effect through error analysis; if the verification result does not meet the standard, adjust the training parameters of the generative adversarial network GAN or the modeling range of the transformer Transformer for iterative optimization until the expected compensation effect of the system is achieved; The data output module is used to integrate and output the compensation data with the complete historical data.

2. The UAV historical data loss compensation system according to claim 1, characterized in that: The data status assessment model includes time dimension integrity assessment, space dimension distribution assessment, and feature dimension consistency assessment; The time dimension integrity assessment includes data integrity ratio, data time interval distribution deviation, and time period data missing rate; The distribution evaluation of spatial dimension includes spatial coverage completeness, regional data missing distribution, and neighborhood data correlation; The consistency assessment of feature dimensions includes feature distribution deviation, multi-feature association consistency, and environmental condition adaptability.

3. The UAV historical data loss compensation system according to claim 2, characterized in that: Calculate the data integrity ratio in the UAV time series consistency data; the proposed data integrity ratio is P int ; Where M indicates that there are a total of M valid data segments; ω k is the weight factor of the kth data segment; δ k is the time length of the kth data segment; T max The end time of the UAV operation; T min is the start time of the UAV operation; T max -T min Indicates the entire operating time range of the drone; Calculate the data time interval distribution deviation in the UAV time series consistency data; the proposed data time interval distribution deviation is Δ int ; Where τ(t) is the data sampling interval at time t; τ ref (t) is the reference sampling interval at time t; β is the deviation sensitivity index.

4. The UAV historical data loss compensation system according to claim 3 is characterized by: The time dimension integrity assessment includes the data integrity ratio P int , Data time interval distribution deviation Δ int , missing rate of time period data R miss ; Where N is the number of lost data segments; ν j is the importance weight of the jth data missing segment; D j is the length of the jth data missing segment; T end -T start The total time span of the task.

5. The UAV historical data loss compensation system according to claim 4, characterized in that: Spatial dimension distribution evaluation includes spatial coverage completeness C space , Regional data missing distribution D region and neighborhood data correlation R neigh ; Where P is the number of effective coverage areas; η k is the importance weight of the kth region; A k is the effective coverage area of ​​the kth region; A total is the total coverage area; Q is the total number of regions; ζ m is the importance weight of the mth region; F m is the data missing ratio of the mth region; R is the number of data point pairs in the neighborhood; φ n is the importance weight of the nth neighborhood data pair; S n is the similarity measure of the nth neighborhood data pair.

6. The UAV historical data loss compensation system according to claim 5, characterized in that: The feature dimension consistency evaluation includes the feature distribution deviation D feat , Multi-feature association consistency C assoc Adaptability to environmental conditions A env ; Where S is the total number of characteristic variables; γ r is the importance weight of the rth feature; x r is the actual value of the rth feature; x ref,r is the reference value of the rth feature; α is the deviation index; U is the number of feature association pairs; δ t is the importance weight of the t-th feature association pair; ρ t is the correlation coefficient of the tth feature association pair; z min , z max They are the lower limit and upper limit of the environment variable respectively; χ(z) is the effectiveness weight of task execution under environmental conditions; ξ(z0 is the actual contribution of environmental condition z to the task.

7. The UAV historical data loss compensation system according to claim 6, characterized in that: Establish a data status assessment model; propose a data status assessment value of S eval ; where w t represents the weight of the time dimension in the data state evaluation model; w s represents the weight of the spatial dimension in the data state evaluation model; w f Represents the weight of the feature dimension in the data state evaluation model; λ t The weight used to control the sensitivity of the time dimension; λ f Weights used to control the sensitivity of feature dimensions.

8. The UAV historical data loss compensation system according to claim 7, characterized in that: In the case of abnormal loss, the characteristics of data loss are clarified through classification, and a basis is provided for subsequent model configuration; Among them C state is the classification result of the current data loss; K is the total number of all possible loss status categories; φ k is the importance weight of category k; M k is the number of features contained in category k; g j,k (F) is the matching degree of the jth feature under category k, calculated based on the similarity of feature F.

9. The UAV historical data loss compensation system according to claim 8, characterized in that: Modeling the latent distribution of missing data using Generative Adversarial Networks (GANs); Generate adversarial networks (GANs) use adversarial optimization, where the generator G simulates the distribution of missing data and the discriminator D evaluates the similarity between generated data and real data. The optimization goal of the Generative Adversarial Network GAN generator is: The optimization goal of the Generative Adversarial Network GAN discriminator is: Among them G * represents the parameter solution determined by the generator under the optimization objective; z~P z represents the distribution of random noise z; G(z; Θ G ) represents the compensation data generated by the generator, where Θ G is the parameter of the generator; T(F) is the target distribution, which is defined by the historical data feature F; β is the deviation sensitivity index; in is the loss function of the discriminator; P real is the distribution of real data; x~P real is the distribution P obeyed by the real data sample x real ; D(x) is the output value of the discriminator D for the real data sample x; G(z) is the pseudo data sample generated by the generator of the generative adversarial network GAN based on the input noise z; Generate the pseudo data sample G(z) generated by the adversarial network GAN and the original data X orig Fusion, the Transformer model is used to model long-term dependencies between data; X fuse =α·X orig +(1-α)·G(z) Where X fuse is the fused data; α is the data fusion coefficient; Establish a multi-head attention mechanism for Transformer; Where Q, K, and V are query matrix, key matrix, and value matrix respectively; d k is the dimension of the key matrix; softmax is used to generate attention weights; matrix operation QK T Indicates calculating the similarity between query and key.

10. The UAV historical data loss compensation system according to claim 9, characterized in that: The error verification formula is: Where ε is the error value; X target (t) is the value of the target data at time t; γ is the error sensitivity index; t start , t end Respectively represent the upper and lower limits of the integral; X fuse (t) is the generated data value after fusion at time t; The optimization formula is: where Θ * are the optimized model parameters; is the regularization term; λ is the regularization weight; The data integration formula is: X final =β1·X fuse +β2·X history Where X final is the final data after integration; X history is the historical data; β1 is the integration weight coefficient of the compensation data; β2 is the integration weight coefficient of historical data.

Images