A method for high-precision anomaly detection and data recovery of UAV flight data

Abstract

The present invention relates to the technical field of unmanned aerial vehicle (UAV) detection, and particularly relates to a method for high-precision abnormal detection and data recovery of UAV flight data. The method includes the following steps: S1. Construct a multivariate regression model based on TCN-KANs. The TCN-KANs model includes an input layer, three residual blocks, and a KAN layer. Among them, the residual blocks extract features through convolutional layers and residual connections, perform non-linear transformation on the input features, extract potential patterns, and generate a final output. The KAN layer serves as the final output layer of TCN-KANs; S2. Minimize the error between the predicted value and the true value; S3. Calculate the residual according to the prediction result of the model. After smoothing the residual, determine the abnormal threshold by analyzing the statistical characteristics of the residual; S4. During the real-time monitoring and operation control of the UAV flight, generate a recovery value through the model to perform data recovery after detecting an abnormality.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle (UAV) detection, and particularly to a method for high-precision anomaly detection and data recovery of UAV flight data. Background Art

[0002] With the continuous expansion of the application fields of UAVs and the rapid growth of the market scale, people's attention and research on them have been increasing. A UAV is a complex physical closed-loop control system that faces challenges such as the lack of real-time observation by a pilot and the ability to make quick decisions. This makes the demand for the safety and reliability of UAVs become even more urgent. UAVs rely on sensors such as GPS, gyroscopes, and accelerometers to collect information related to the flight state of the UAV. These information reflect performance indicators such as the flight trajectory, attitude change, and motion state of the UAV, providing an important data basis for its health management and operation and maintenance. By deeply analyzing these data, real-time monitoring and evaluation of the UAV operation state can be achieved, thereby significantly improving the reliability and safety of operations. In this context, anomaly detection of flight data has become one of the important means for UAV state monitoring. By processing and analyzing flight data, situations deviating from the normal flight mode can be detected in a timely manner. Once an anomaly is identified, further by implementing anomaly recovery or mitigation measures, the UAV flight path or control instructions can be actively adjusted to ensure the safe operation of the UAV. Therefore, actively carrying out research on UAV flight data anomaly detection and recovery has important practical value and broad application prospects.

[0003] Currently, many studies on UAV flight data anomaly detection have been carried out in the prior art. From the perspective of methodology, they are mainly divided into traditional machine learning-based methods and deep learning-based methods, including K-means, SVM, KPCA, KNN, decision tree (DT), random forest (RF), Kalman filter (KF), as well as LSTM, BiLSTM, 1D CNN-LSTM, etc. However, there are still the following problems:

[0004] 1. Although these traditional machine learning-based methods are widely used in the anomaly or fault detection of UAV flight data, they often show limited flexibility and adaptability when dealing with complex high-dimensional data. Secondly, these methods often require manual feature engineering, which is not only time-consuming and laborious but may also lead to the omission of important features. In addition, for application scenarios that emphasize the integrity and reliability of data after anomaly detection, reasonable recovery of the detected abnormal data is required. However, most of the above methods lack analytical redundancy and cannot effectively achieve abnormal data recovery.

[0005] 2. Deep learning-based methods often use LSTM to model flight data, which may lead to the problem of gradient vanishing or explosion when dealing with long-sequence flight data, thus affecting the performance of the model. Although LSTM-CNN performs well in short-term dependencies, the fixity of the convolution kernel and local features limit its ability to process complex high-dimensional and non-linear time-series data, affecting the overall performance of the model. Therefore, although these methods have achieved remarkable results in anomaly and fault detection, when dealing with flight data with long-term dependencies and complex time-series structures, there are still many challenges in building an accurate prediction model. This requires the model to not only have efficient feature extraction capabilities but also excellent time-series representation capabilities to accurately capture long-term dependencies and thus significantly improve the overall performance.

[0006] 3. Existing deep learning methods usually use the fully connected layer of the multi-layer perceptron (MLP) to further process the extracted features. However, due to the use of fixed activation functions and linear transformations at the nodes in MLP, it lacks sufficient flexibility and adaptability in representing complex time-series data and modeling non-linear relationships, resulting in the underutilization of the features extracted by the model and thus potentially reducing the model performance. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a method for high-precision anomaly detection and data recovery of UAV flight data, which can improve the flight safety and data reliability of UAVs.

[0008] The basic solution provided by the present invention: includes the following steps: A method for high-precision anomaly detection and data recovery of UAV flight data, including the following steps:

[0009] S1. Construct a multivariate regression model based on TCN-KANs. The TCN-KANs model includes an input layer, three residual blocks, and a KAN layer. Use KANs as the final output layer of TCN-KANs. The residual blocks extract features through convolutional layers and residual connections, perform non-linear transformation on the input features, extract potential patterns, and generate the final output;

[0010] S2. Minimize the error between the predicted value and the true value;

[0011] S3. Calculate the residuals according to the prediction results of the model. After smoothing the residuals, determine the anomaly threshold by analyzing the statistical characteristics of the residuals;

[0012] S4. During the real-time monitoring and operation control of the UAV, generate a recovery value through the model to perform data recovery after detecting an anomaly.

[0013] Further, the S1 includes the following steps:

[0014] S11. Set the first residual block to include two extended causal convolutional layers with a kernel size of 3, an expansion factor of 1, and 16 filters. In the second residual block, the kernel size is 3, the expansion factor is 2, and the number of filters is 32. In the third residual block, the kernel size is 3, the expansion factor is 4, and the number of filters is 64;

[0015] The outputs after passing through the three residual blocks are as follows:

[0016]

[0017] where , and represent the outputs of the first, second, and third residual blocks respectively, and represents the output features after being processed by the mapping function TCN(⋅);

[0018] S12. Transform through the KAN layer, which is a non - linear transformation of the input features to extract potential patterns and generate the final output, as shown in the following formula:

[0019]

[0020] where represents the final prediction result, and represents the mapping function.

[0021] Furthermore, the S2 includes the following steps:

[0022] S21. Update the weights by calculating the gradients of the network to minimize the error between the predicted value and the true value until the predetermined model accuracy requirement is reached, and use the mean squared error MSE as the loss function:

[0023]

[0024] In the formula, and are the true value and the predicted value of the th data point respectively, and is the sample length.

[0025] Furthermore, the S3 includes the following steps:

[0026] S31. Calculate the residuals according to the prediction results of the model, and input the training set and the test set into the model respectively to obtain:

[0027]

[0028]

[0029] Among them, and are respectively and predicted values, is a mapping function;

[0030] S32. Training residuals and test residuals :

[0031]

[0032]

[0033] Among them and are respectively the training set and the test set of the target variable;

[0034] S33. The exponentially weighted moving average method EWMA is introduced to smooth and :

[0035]

[0036] Among them, is the smoothed value of the th residual, is an adjustable weight parameter, is the original residual, and ;

[0037] S34. The SPOT method based on extreme value theory is used to obtain the dynamic threshold, and the part exceeding the threshold is expressed as :

[0038]

[0039] Among them, is the initial threshold initialized by , and γ and σ are the shape parameter and the scale parameter respectively;

[0040] S35. The method of maximizing the likelihood function is used to estimate the γ and σ parameters:

[0041]

[0042]

[0043] Among them is the sum of the sample data of the peaks greater than ;

[0044] and are updated parameters obtained by the maximum likelihood estimation method, where q represents a predefined extreme value point quantile, is the total number of samples. is the difference value between the final threshold and ; ;

[0045] S36. Obtained according to S35:

[0046]

[0047] S37. According to S33 and S36, perform anomaly detection:

[0048]

[0049] where is the smoothed test residual of the th sample, is threshold, 1 indicates anomaly, 0 indicates normal.

[0050] Furthermore, the S4 includes the following steps:

[0051] S41. Take the prediction result in S31 as the recovery value to correct data deviation.

[0052] The principle and advantages of the present invention are as follows:

[0053] 1. A drone flight data anomaly detection and recovery framework TCKANs-ADR is proposed. This framework integrates advanced deep learning network architectures and anomaly detection strategies, providing a more accurate and robust solution for the field to address challenges such as complex time series data modeling, time series feature extraction, and the influence of random noise.

[0054] 2. A multivariate regression method based on TCN and KANs (TCN-KANs) is proposed. This method inherits the advantages of TCN in processing long sequence data and extracting time series features. At the same time, KANs is introduced into the field of drone flight data anomaly detection and recovery for the first time, and it cleverly replaces the fully connected layer in TCN, thereby enhancing the modeling ability for complex spatio-temporal correlation flight data and improving the accuracy and robustness of model prediction.

[0055] 3. A method for calculating dynamic thresholds based on extreme value theory and residual smoothing is proposed. This method first smooths the residuals using the exponentially weighted moving average method to overcome random noise interference. Then, by analyzing the extreme value distribution of the residual historical data, the threshold is dynamically adjusted to adapt to the changes in flight data, thereby improving the performance of model anomaly detection. Description of the Drawings

[0056] Figure 1 Schematic diagram of an embodiment of a method for high-precision anomaly detection and data recovery of UAV flight data according to the present invention;

[0057] Figure 2 Schematic diagram of the TCN-KANs model structure of an embodiment of a method for high-precision anomaly detection and data recovery of UAV flight data according to the present invention;

[0058] Figure 3 Schematic diagram of the curve of partial parameters of flight data of an embodiment of a method for high-precision anomaly detection and data recovery of UAV flight data according to the present invention;

[0059] Figure 4 Schematic diagram of flight data after injecting anomalies in an embodiment of a method for high-precision anomaly detection and data recovery of UAV flight data according to the present invention;

[0060] Figure 5 Visualization diagram of anomaly detection results of an embodiment of a method for high-precision anomaly detection and data recovery of UAV flight data according to the present invention;

[0061] Figure 6 Visualization diagram of data recovery results of an embodiment of a method for high-precision anomaly detection and data recovery of UAV flight data according to the present invention; Detailed Description of the Invention

[0062] The following is a further detailed description through specific embodiments:

[0063] The embodiment is basically as shown in the Figure 1 drawings:

[0064] A method for high-precision anomaly detection and data recovery of UAV flight data includes the following steps:

[0065] S1. Construct a multivariate regression model based on TCN-KANs. The TCN-KANs model includes an input layer, three residual blocks, and a KAN layer. The KANs is used as the final output layer of the TCN-KANs. The residual blocks extract features through convolutional layers and residual connections, perform non-linear transformation on the input features, extract potential patterns, and generate the final output.

[0066] Specifically, accurate anomaly detection depends on good model prediction ability. TCN demonstrates superior performance to LSTM and CNN in time series prediction. Thanks to its causal convolution and dilated convolution characteristics, it can effectively capture long-term dependencies and avoid the problems of gradient vanishing and explosion. In addition, the parallel computing ability of TCN significantly improves the training speed, making it more efficient in processing large-scale data. Although these advantages have made TCN attract much attention in time series prediction tasks, the multi-layer perceptron (MLP) it mostly uses as the output layer may not fully exploit the spatio-temporal features extracted by TCN. Due to the characteristics of UAV flight data such as high dimensionality, non-linearity, and dynamic changes, MLP may not be able to effectively capture these complex data features, which may limit the improvement of the model's anomaly detection and prediction performance. However, the emergence of KANs provides a possibility to solve this problem. Therefore, a multivariate regression model based on TCN-KANs as shown in Figure 2 is constructed in this application. The aim is to more effectively capture the complex dynamic features of UAV flight data and achieve accurate prediction.

[0067] The designed TCN-KANs model consists of an input layer, three residual blocks, and a KAN layer. Different from the traditional TCN, the fully connected layer in TCN is abandoned, and KANs is used as the final output layer of TCN-KANs. The residual blocks effectively extract features through convolutional layers and residual connections to enhance the model's processing ability for time series data. The first residual block consists of two dilated causal convolutional layers. The kernel size is set to 3, the dilation factor is set to 1, and the number of filters is set to 16. In the second residual block, the kernel size is set to 3, the dilation factor is set to 2, and the number of filters is set to 32. Finally, in the third residual block, the kernel size is also set to 3, the dilation factor is set to 4, and the number of filters is set to 64. For multivariate flight data , it is first initially feature-extracted by TCN.

[0068] S1 includes the following steps:

[0069] S11. Set the first residual block to include two dilated causal convolutional layers with a kernel size of 3, a dilation factor of 1, and a number of filters of 16. In the second residual block, the kernel size is 3, the dilation factor is 2, and the number of filters is 32. In the third residual block, the kernel size is 3, the dilation factor is 4, and the number of filters is 64;

[0070] The output after passing through the three residual blocks is expressed as follows:

[0071]

[0072] Where , and respectively represent the outputs of the first, second, and third residual blocks, represents the output features after being processed by the mapping function TCN(⋅);

[0073] S12. Apply a transformation through the KAN layer, a non-linear transformation of the input features, to extract latent patterns and generate the final output, as shown in the following formula:

[0074]

[0075] where represents the final prediction result, represents the mapping function.

[0076] S2. Minimize the error between the predicted value and the true value.

[0077] The S2 includes the following steps:

[0078] The TCN-KANs model is trained based on the backpropagation algorithm. This algorithm updates the weights by calculating the gradients of the network to minimize the predicted value and the true value between the errors until the predetermined model accuracy requirement is met.

[0079] S21. Update the weights by calculating the gradients of the network to minimize the predicted value and the true value between the errors until the predetermined model accuracy requirement is met, using the mean squared error MSE as the loss function:

[0080]

[0081] In the formula and are respectively the true value and the predicted value of the th data point, is the sample length.

[0082] In this application, the training process of the TCN-KANs model is efficiently carried out by using the fast-converging Adam optimizer to update the network weights and combining with the regularization technique, so as to obtain superior performance and high accuracy in real applications.

[0083] S3. Calculate the residuals according to the prediction results of the model. After smoothing the residuals, determine the anomaly threshold by analyzing the statistical characteristics of the residuals.

[0084] S3 includes the following steps:

[0085] S31. Calculate the residuals based on the prediction results of the model. Input the training set and the test set into the model to obtain:

[0086]

[0087]

[0088] where and are the predicted values of and respectively, and is the mapping function;

[0089] S32. Train the residual and the test residual :

[0090]

[0091]

[0092] where and are the training set and the test set of the target variable respectively;

[0093] Due to the existence of random noise in the UAV flight data, the residuals fluctuate violently, which interferes with the anomaly detection. Therefore, the exponentially weighted moving average (EWMA) method is introduced to and for smoothing to suppress the influence of random noise.

[0094] S33. Introduce the exponentially weighted moving average method EWMA to and for smoothing:

[0095]

[0096] where is the smoothed value of the th residual, is the adjustable weight parameter, is the original residual, and ;

[0097] In addition, to adapt the anomaly detection threshold to the performance of the UAV in a dynamically changing flight environment, the SPOT method based on extreme value theory is further introduced to obtain the dynamic threshold. This method effectively avoids the strict assumptions of traditional statistical thresholds on data distribution, such as normality. In addition, the SPOT method has good scalability and can adaptively adjust the threshold under different environmental conditions, thus improving the anomaly detection ability and flight safety of the UAV. The SPOT method obtains the threshold according to the distribution of extreme values.

[0098] S34. Obtain the dynamic threshold using the SPOT method based on extreme value theory, and represent the part exceeding the threshold as :

[0099]

[0100] where is the initial threshold initialized by , and γ and σ are the shape parameter and scale parameter respectively;

[0101] S35. Use the method of maximizing the likelihood function to estimate the γ and σ parameters:

[0102]

[0103]

[0104] where is the sum of the sample data of the peaks greater than ;

[0105] and are the updated parameters obtained by the maximum likelihood estimation method, q represents the predefined extreme point quantile, is the total number of samples. is the difference value between the final threshold and , ;

[0106] S36. Obtain according to S35:

[0107]

[0108] S37. According to S33 and S36, perform anomaly detection:

[0109]

[0110] where is the -th sample of the smoothed test residual , is The threshold value, where 1 indicates an anomaly and 0 indicates normal.

[0111] S4. During the real-time flight monitoring and operation control of the UAV, when an anomaly is detected by the model, a recovery value is generated for data recovery.

[0112] In the scenario of real-time flight monitoring and operation control of the UAV, effective data recovery can quickly restore the system to the normal state after an anomaly is detected, thus maintaining the stability and reliability of the UAV system. Based on the expected output of the normal state by the model, that is, the prediction result in Equation (5) is used as the recovery value to quickly correct the data deviation and reduce the flight risk caused by anomalies.

[0113] S4 includes the following steps:

[0114] S41. Take the prediction result in S31 as the recovery value to correct the data deviation.

[0115] It also includes the following steps:

[0116] S5. Generate experimental data and anomaly data to verify the reliability of the model.

[0117] The experimental data comes from the actual flight data independently collected by a real composite-wing UAV, which is transmitted to the ground station in real time through a data radio to achieve real-time recording and storage of the data. The parameters collected include those from sensors such as barometers, accelerometers, gyroscopes, and magnetometers, including but not limited to airspeed, roll angle, eastward velocity, longitude, latitude, pitch angle, altitude, etc. Sixteen flight parameters related to UAV attitude control were selected as experimental data. These parameter names are GPS_Alt, GPS_Lat, GPS_Lon, V_east, V_north, Vz, AirSpeed, Yaw, Pitch, Roll, Acc_X, Acc_Y, Acc_Z, wX, wY, and wZ, and their meanings and units are shown in Table 1. A segment of the above parameters with a length of 12000 was intercepted as the model input.

[0118] Table 1

[0119]

[0120] Figure 3 Shows partial parameter curves of the flight data used.

[0121] Since it is difficult and costly to obtain actual anomaly data, the anomaly injection method was adopted to obtain anomaly data. Taking Acc_Z as the monitoring parameter, bias and drift anomalies were injected into it according to Equation (21).

[0122]

[0123] where y(t) and are normal and abnormal flight data respectively, is a constant, is a function of t. In this paper, is set to 0.1. To represent , we take equally spaced values in the interval [0.08, 0.12]. For the flight parameter Acc_Z, we injected 2000 abnormal points with an abnormality rate of 0.167. The injection of abnormal points starts from the 10000th sample point and continues until the end of the data. This means that in the entire data sequence, the first 10000 sample points are normal, while the part after the 10000th sample point contains artificially introduced abnormal data. Figure 4 shows the original and injected biased and abnormal flight data.

[0124] In this application, the maximum - minimum method is used to normalize the parameters shown in Table 1. Maximum - minimum normalization is a commonly used data pre - processing technique, whose purpose is to unify parameters with different dimensions and ranges to the same standard for subsequent analysis and to avoid the impact of different dimensions on the model performance. Specifically, this method converts each parameter value into a value within the range [0, 1] through the following formula:

[0125]

[0126] where, represents the normalized data, represents the original data, and represent the maximum and minimum values of

[0127] This application uses accuracy (Acc), true positive rate (TPR), recall rate (Rec), precision (Pre), false positive rate (FPR), and F1 - score as evaluation metrics for anomaly detection, as shown in formulas (16) - (20). Mean absolute error (MAE) and MSE are used as evaluation metrics for data recovery:

[0128]

[0129]

[0130]

[0131]

[0132]

[0133]

[0134]

[0135] Among them, TP and TN represent correctly identified normal samples and abnormal samples, and FP and FN represent normal samples and abnormal samples misjudged as abnormal samples and normal samples, and respectively represent the original value and predicted value of the i-th sample, and N represents the total number of samples. The larger the TPR, Acc, Pre, and F1 values, and the smaller the FPR, MAE, and MSE values, the better the performance of the model.

[0136] To verify the effectiveness of TCKANs-ADR, it is compared with a variety of benchmark regression methods, including Support Vector Regression (SVR), LSTM, BiLSTM, 1D CNN-LSTM, and 1D CBiAM. These methods cover advanced regression techniques based on traditional machine learning and deep learning, thus providing a comprehensive perspective for comparison. Figure 5 Shows the visualization results of anomaly detection of TCKANs-ADR. It can be seen that although there are a few FNs for different anomaly types in TCKANs-ADR, there are relatively high TNs and few FPs for the vast majority of anomaly points. This indicates that the model has good performance and stability in distinguishing normal and abnormal data. To quantitatively evaluate the anomaly detection performance of TCKANs-ADR and the above benchmark methods, Table 2 shows the results of each method in the anomaly detection task. Benefiting from the advantages of TCN and KANs, as well as the residual smoothing method and dynamic threshold strategy, TCKANs-ADR achieves the optimal anomaly detection results compared with the benchmark methods. For bias anomalies, the TPR, FPR, Acc, Pre, and F1 values of TCKANs-ADR are 90.69%, 1.00%, 95.30%, 98.64%, and 94.50% respectively. For drift anomalies, the TPR, FPR, Acc, Pre, and F1 values of TCKANs-ADR are 86.00%, 0.15%, 93.69%, 99.78%, and 92.38% respectively. These results confirm the stability and reliability of TCKANs-ADR in processing different types of abnormal data.

[0137] Table 2

[0138]

[0139] The FPR, Acc, Pre, and F1 values of SVR in deviation and drift anomaly detection are 8.85% and 3.90%, 79.31% and 82.06%, 85.36% and 92.97%, and 73.48% and 76.16% respectively. However, the true TPR value of SVR is only 64.50%, which indicates that although the model performs well in distinguishing normal and abnormal samples, a considerable number of true normal samples are not detected, potentially leading to the neglect of potential problems. LSTM also performs poorly in the detection of deviation and drift anomalies, with its FPR values reaching as high as 56.28% and 63.73% respectively, and the Acc, Pre, and F1 values are also low. This may be due to the limited ability of LSTM to capture dependencies when dealing with non-stationary flight data. At the same time, the lack of a suitable noise processing strategy leads to a decrease in the detection accuracy of LSTM. Although BiLSTM also outperforms LSTM in overall performance, its performance in terms of FPR, Acc, Pre, and F1 values is still not ideal. This may still be due to the interference of random noise and the insufficient feature extraction ability of BiLSTM, thus affecting its anomaly detection performance.

[0140] 1D CNN-LSTM combines the feature extraction of 1D CNN and the temporal dependence modeling ability of LSTM, showing relatively good detection performance. In the detection of deviation anomalies, the TPR, Acc, Pre, and F1 values of 1D CNN-LSTM are 81.69%, 86.69%, 87.54%, and 84.51% respectively, but its FPR value is relatively high, reaching 9.30%. For drift anomalies, although the overall performance of 1D CNN-LSTM decreases, it still maintains relatively high Acc and Pre values, which are 80.83% and 90.34% respectively, and at the same time the FPR is significantly reduced to 5.45%. Similarly, 1D CBiAM, which combines the advantages of multiple deep learning methods, also shows good anomaly detection performance overall. Compared with LSTM, BiLSTM, and 1D CNN-LSTM, 1D CBiAM achieves the lowest FPR values, which are 2.90% and 4.85% respectively, and realizes relatively high Acc, Pre, and F1 values. However, the TPR values of 1D CBiAM are relatively low, which are 69.69% and 71.56% for deviation and drift anomalies respectively. This indicates that although 1D CBiAM performs well in terms of FPR and other evaluation metrics, there is still significant room for improvement in terms of TPR.

[0141] Tables 3 and 4 list the differences in performance metrics of TCKANs-ADR compared with the baseline methods under two anomaly detection tasks respectively. Whether it is partial anomaly or drift anomaly, TCKANs-ADR has significantly improved in multiple performance metrics compared with the baseline methods. Specifically, in the partial anomaly detection task, compared with the baseline methods, TCKANs-ADR's performance in TPR, Acc, Pre, and F1 is respectively 9.00% to 26.19%, 8.61% to 39.04%, 3.58% to 48.07%, and 9.98% to 35.11% higher. In addition, TCKANs-ADR is significantly lower than the baseline methods in terms of FPR, and the FPR ranges from -55.28% to -1.90%. For drift anomaly, TCKANs-ADR compared with the baseline methods is higher in TPR, Acc, Pre, and F1 by 13.81% to 22.31%, 9.03% to 41.46%, 6.81% to 52.23%, and 11.80% to 35.05% respectively, while it is lower in FPR by -63.58% to -3.75%. These results further prove the effectiveness and superiority of the anomaly detection performance of TCKANs-ADR.

[0142] Table 3

[0143]

[0144] Table 4

[0145]

[0146] Figure 6 shows the data recovery visualization results of TCKANs-ADR and the baseline methods. It can be seen that, for example, methods such as LSTM, BiLSTM, 1D CNN-LSTM, and 1D CBiAM have obvious recovery biases for the data recovery results of different anomaly types. In contrast, TCKANs-ADR can still fit the data trend and distribution and recover the data more precisely.

[0147] Table 5 lists the recovery results of the above methods. It can be seen that LSTM performs the worst, with its MAE and MSE values for bias and drift being 0.041376 and 0.002549, and 0.052706 and 0.003923 respectively. This is mainly because complex flight data often has strong non - linear and time - varying characteristics, and LSTM may be restricted by its structure in modeling these characteristics. Compared with LSTM, BiLSTM and 1D CNN - LSTM have a certain degree of improvement in the MAE and MSE values for bias and drift anomalies, but they are still relatively high. This indicates that these models still face challenges when dealing with flight data with strong non - linear and time - varying characteristics. Although 1D CBiAM performs well in the recovery of bias anomalies, with values of 0.033561 and 0.001761 respectively, it has relatively high MAE and MSE values for drift anomalies, which are 0.042438 and 0.002674 respectively. SVR has good overall data recovery performance, with its MAE and MSE values for bias and drift anomalies being 0.032566 and 0.001742, and 0.032548 and 0.001740 respectively. Compared with the baseline method, TCKANs - ADR performs best in dealing with bias and drift anomalies, with the lowest MAE and MSE values, which are 0.031322 and 0.001619 for bias anomaly and 0.031317 and 0.001635 for drift anomaly respectively. This excellent performance benefits from the advantages of TCKANs - ADR in modeling complex non - linear relationships and feature extraction, enabling it to better fit the distribution changes of flight data.

[0148] Table 5

[0149]

[0150] To further explore the differences in data recovery between different anomaly types, the difference values of various methods are listed in Table 6. The ∆MAE and ∆MSE of LSTM are still the highest, being 0.011330 and 0.001374 respectively. This indicates that LSTM is most affected by the switching of different anomaly models and cannot effectively adapt to changes in anomaly patterns, resulting in significant fluctuations in its recovery performance under different anomaly conditions. In contrast, the ∆MAE and ∆MSE values of BiLSTM, 1D CNN-LSTM, and 1D CBiAM are relatively low, but they are still affected to a certain extent by the switching of different anomaly patterns. For example, the ∆MAE and ∆MSE values of 1D CNN-LSTM are 0.011152 and 0.001172 respectively. The overall data recovery performance of SVR is good, especially its ∆MSE value is the lowest, being 0.000002. Although the ∆MSE value of TCKANs-ADR is slightly lower than that of SVR, being 0.000016, its ∆MAE value is the lowest, being 0.000005. This shows that while maintaining the best anomaly detection performance, TCKANs-ADR has high adaptability and stability to the switching of different anomaly patterns. Overall, TCKANs-ADR can not only effectively identify and recover abnormal data but also maintain a consistent recovery effect when facing different anomaly types, demonstrating stronger robustness.

[0151] Table 6

[0152]

[0153] The above are only embodiments of the present invention. Specific structures and common knowledge such as characteristics that are well-known in the art are not described in detail here. Those of ordinary skill in the art know all the common general technical knowledge in the technical field to which the invention belongs before the application date or the priority date, can know all the existing technologies in this field, and have the ability to apply the conventional experimental means before this date. Those of ordinary skill in the art can, under the inspiration given in this application and in combination with their own abilities, improve and implement this solution. Some typical well-known structures or well-known methods should not become obstacles for those of ordinary skill in the art to implement this application. It should be noted that for those skilled in the art, without departing from the structure of the present invention, several deformations and improvements can still be made, and these should also be regarded as the protection scope of the present invention, and these will not affect the implementation effect of the present invention and the practicality of the patent. The protection scope required by this application should be based on the content of its claims, and the specific implementation manners described in the specification can be used to interpret the content of the claims.

Claims

1. A high-precision abnormal detection and data recovery method for UAV flight data, characterized in that: Including the following steps: S1. Construct a multiple regression model based on TCN-KANs. The TCN-KANs model includes an input layer, three residual blocks, and a KAN layer. The KAN layer is used as the final output layer of TCN-KANs. The residual blocks extract features through convolutional layers and residual connections, perform non-linear transformation on the input features, extract potential patterns, and generate the final output; S2. Minimize the error between the predicted value and the true value; S3. Calculate the residuals according to the prediction results of the model. After smoothing the residuals, determine the anomaly threshold by analyzing the statistical characteristics of the residuals; S4. During the real-time monitoring and operation control of the drone, generate a recovery value through the model to perform data recovery after detecting an anomaly; The S1 includes the following steps: S11. Set the first residual block to include two dilated causal convolutional layers with a kernel size of 3, a dilation factor of 1, and a filter number of 16. In the second residual block, the kernel size is 3, the dilation factor is 2, and the filter number is 32. In the third residual block, the kernel size is 3, the dilation factor is 4, and the filter number is 64; The output after passing through the three residual blocks is expressed as follows: Among them , and respectively represent the outputs of the first, second, and third residual blocks, represents the output features after being processed by the mapping function TCN(⋅); S12. Transform through the KAN layer, which is a non-linear transformation of the input features, extract the latent patterns and generate the final output as shown in the following equation: through the KAN layer, which is a non-linear transformation of the input features, extract the latent patterns and generate the final output as shown in the following equation: wherein represents the final prediction result, represents the mapping function; The S3 includes the following steps: S31. Calculate the residuals based on the prediction results of the model, and input the training set and the test set into the model respectively, and obtain: Among them, and are respectively and predicted values of is the mapping function; S32. Training residuals and test residuals : wherein and are the training set and the test set of the target variable, respectively; S33. The exponential weighted moving average method EWMA method is introduced to and perform smoothing processing: Among them, is the smoothed value of the th residual, is an adjustable weight parameter, is the original residual, and ; S34. Obtain a dynamic threshold by using the SPOT method based on the extreme value theory for the peak flow exceeding the threshold, and represent the part exceeding the threshold as : Among them, the initial threshold is initialized by γ and σ are the shape parameter and the scale parameter respectively; S35. Use the method of maximizing the likelihood function to estimate the γ and σ parameters: wherein is the sum of sample data greater than the peak value; and are updated parameters obtained by the maximum likelihood estimation method, where q represents a predefined extreme value quantile, is the total number of samples, is the final threshold and the difference value of ; S36. Obtained according to S35: S37. According to S33 and S36, perform anomaly detection: wherein is the smoothed test residual of the th sample, is the threshold value, 1 indicates abnormality, and 0 indicates normal.

2. The high-precision abnormal detection and data recovery method for UAV flight data according to claim 1, characterized in that: The S2 includes the following steps: S21. Update the weights by calculating the gradients of the network to minimize the predicted value and the true value between the errors until the predetermined model accuracy requirement is reached, and the mean squared error MSE is used as the loss function: where and are the true value and the predicted value of the th data point respectively, and is the sample length.

3. A high-precision anomaly detection and data recovery method for UAV flight data according to claim 2, characterized in that: The S4 includes the following steps: S41. Use the prediction result in S31 as the recovery value to correct data deviation.

4. A high-precision abnormal detection and data recovery method for UAV flight data according to claim 3, characterized in that: Also include the following steps: S5. Generate experimental data and anomaly data to verify the reliability of the model.

Images