Multivariable flight data anomaly detection and positioning method based on reconstruction model

Abstract

A multivariable flight data anomaly detection and positioning method based on a reconstruction model comprises the following steps: a data preprocessing step: collecting flight data of an unmanned aerial vehicle, dividing original input data into a training set and a test set according to a preset proportion, and performing normalization processing on the data; a model construction step: constructing a neural network reconstruction model based on DCANNs; the automatic encoder module compresses input data into a potential variable through an encoder, and then a decoder decodes z back to an original data space to obtain reconstructed data; a self-adaptive anomaly detection step: using the training set to reconstruct the residual error to calculate an anomaly threshold, and comparing the anomaly score of the test set with the threshold to realize anomaly detection; an abnormal positioning step: positioning abnormal parameters by calculating and analyzing statistical characteristics of residual errors; the invention aims to solve the problems that a traditional deep learning model is insufficient in feature extraction, a fixed statistical threshold cannot adapt to flight data dynamic change, and a specific source of an abnormal parameter cannot be positioned.

Description

Technical Field

[0001] The invention belongs to the technical field of data processing and data analysis, and relates to a multivariate flight data anomaly detection and positioning method based on a reconstruction model. Background Art

[0002] In the aviation field, with the widespread use of aircraft such as drones and civil aircraft, effective monitoring of flight data has become increasingly critical. As a key source of information reflecting the operating status of an aircraft, flight data covers multivariate information such as speed, acceleration, attitude angle, magnetometer, gyroscope, etc., and has the characteristics of strong time series and complex correlation between variables. Accurate anomaly detection and positioning can detect potential aircraft failures in a timely manner, which is of far-reaching significance for ensuring flight safety, improving aircraft reliability and optimizing maintenance strategies.

[0003] In actual production and use scenarios, flight data anomaly detection and location face many difficulties. On the one hand, most existing methods have deficiencies in the feature extraction link. For example, the common LSTM-based model is prone to information forgetting or insufficient retention when processing long-term dependencies of long-time series data such as flight data. It is difficult to accurately capture the complex correlation patterns in the data, resulting in poor feature extraction effects, which seriously affects the accuracy and generalization ability of the model, and it is difficult to detect anomalies stably and reliably under complex flight conditions. On the other hand, current research focuses on single parameter anomaly detection, ignoring the actual situation that flight anomalies often involve the coordinated changes of multiple parameters, lacking the ability to comprehensively control the overall abnormal situation, and easily missing key abnormal information, and unable to comprehensively and accurately evaluate the operating status of the aircraft. At the same time, traditional anomaly detection methods usually rely on fixed statistical thresholds. Given the complex and changeable flight environment and the strong dynamics of flight data, such thresholds are difficult to adjust adaptively, resulting in frequent false alarms and missed alarms under different flight conditions, which is difficult to meet the actual flight safety assurance needs. Existing research focuses on the anomaly recognition performance of the model, but rarely explores the root causes of anomalies in depth. In practical applications, it is not sufficient to only realize anomaly detection. The key lies in accurately locating the specific parameters that cause the anomaly, which provides an important basis for the operation and maintenance decision-making of the aircraft.

[0004] After searching, it was found that some studies are committed to improving the model architecture to improve performance. For example, patent CN117851946A proposes an anomaly detection method for multivariate flight time series data. It first divides the historical time series data into multiple subsequences and inputs them into the RNN network training, then processes them through the CNN filter layer and the covector filter layer, and fuses the processing results to obtain a prediction model. Finally, the model is used to determine the anomaly of the real-time data subsequence. The advantage of this method is that it integrates multi-source information and different processing technologies to a certain extent, trying to enhance the model's understanding and processing capabilities of data. However, its shortcomings are also obvious. The inherent defects of the RNN model itself in processing long sequences have not been fundamentally solved, and it may still be limited in the processing of complex data relationships. The calculation process is relatively complex, which may affect the detection efficiency. At the same time, the solution uses traditional statistical thresholds and is fixed, which cannot effectively respond to the dynamic change detection needs of flight data. In addition, the patent focuses on the anomaly detection capability of the model, but fails to effectively provide valuable insights into the source of the anomaly. Once an anomaly occurs, it is not only necessary to accurately identify and detect it, but also to further locate the specific parameter source that causes the abnormal behavior, so as to provide decision support for subsequent operation and maintenance.

[0005] After comprehensive consideration, the present invention intends to adopt a new architecture design, by building a reconstruction model based on DCANNs, using multiple one-dimensional convolutions (Conv1D) and attention layers to extract local patterns and key features, and realize accurate data reconstruction; at the same time, it combines the dynamic threshold calculation method based on extreme value theory (EVT) for adaptive anomaly detection, and locates abnormal parameters with the help of residual statistical features; this method is expected to overcome the shortcomings of the existing technology, and with its unique structural advantages and dynamic threshold mechanism, effectively improve the accuracy, reliability and adaptability of multivariate flight data anomaly detection and positioning, and effectively ensure the safe and stable operation of the aircraft. Summary of the invention

[0006] The present invention provides a multivariate flight data anomaly detection and positioning method based on a reconstruction model, which solves the problems of insufficient feature extraction of traditional deep learning models, inability of fixed statistical thresholds to adapt to dynamic changes in flight data, and inability to locate the specific source of abnormal parameters.

[0007] In order to solve the above problems, the technical solution adopted by the invention is: A multivariate flight data anomaly detection and location method based on a reconstruction model, characterized in that it comprises the following steps: S01 Data processing stage: collect UAV flight data and perform normalization preprocessing on the raw data; divide the raw data into training set and test set in proportion. The training set contains only normal data, and the test set generates abnormal data by injecting deviation anomalies, drift anomalies, and stuck anomalies into normal data; S02 Model construction and training phase: Construct the DCANNs reconstruction model, which includes a local feature extraction module based on one-dimensional convolution, a key feature learning module based on the attention mechanism, and a data representation module based on the autoencoder; use the training set data to train the model to minimize the reconstruction error of the autoencoder; S03 anomaly detection stage: calculate the reconstruction residuals of the training set and the test set, and use the reconstruction residuals of the training set to initialize the initial threshold of the method based on extreme value theory; estimate the shape parameter and scale parameter of the generalized Pareto distribution by maximizing the likelihood function to obtain the final anomaly threshold; compare the anomaly score of the test set data with the threshold to determine whether the data is abnormal; S04 anomaly location stage: Based on the test set residual, the residual sum of each variable is calculated, and the residual sum is used as the abnormality score of the variable. The abnormal parameters are located by comparing the abnormality scores.

[0008] The principle of this scheme is: In the data preprocessing step, the original flight data are normalized and mapped to the same numerical range, eliminating the impact of dimensions between different variables.

[0009] In the model construction step, the DCANNs-based neural network reconstruction model makes full use of the characteristics of different modules. The Conv1D module extracts local patterns through sliding windows and can capture local features in flight data. The attention mechanism module captures global dependencies by dynamically assigning weights, enabling the model to focus on important data features. The autoencoder module compresses and decodes data, learns the intrinsic representation of the data, and minimizes the reconstruction error to obtain effective data representation.

[0010] In the adaptive anomaly detection step, the dynamic threshold calculation method based on extreme value theory (EVT) can dynamically calculate the anomaly detection threshold according to the reconstructed residuals of the training set and the test set, thereby improving the anomaly detection performance.

[0011] In the anomaly location step, the abnormal parameters are located by calculating the statistical characteristics of the residuals. The residuals are calculated based on the difference between the original data and the reconstructed data, and then the residual sum of each variable is calculated as the abnormality score. The higher the score, the more likely the variable is an abnormal parameter, thus achieving accurate location of the anomaly.

[0012] The beneficial effects of this program: The model uses multiple one-dimensional convolutional layers to effectively extract local patterns in drone flight data. These convolutional layers can improve the feature extraction ability of complex flight data by capturing potential correlations in time series data. At the same time, the attention mechanism is introduced in the model, which can automatically identify and focus on the most important features or time segments, thereby enhancing the learning ability of key features. This mechanism enables the model to effectively focus on the parts that have an important impact on drone flight data, avoid information redundancy, and further improve the potential representation learning ability of multivariate flight data. In this way, this method can not only reconstruct flight data more accurately, but also adapt to drone data of different types and environments, with strong robustness and generalization ability. Therefore, this model shows higher accuracy and stability when processing complex and changeable drone flight data.

[0013] By performing residual analysis on drone flight data, the difference between normal flight status and abnormal behavior is effectively captured. The model first reconstructs the input data and obtains the residual by calculating the reconstruction error. In order to improve the accuracy of anomaly detection, a dynamic threshold generation mechanism is proposed, which can dynamically adjust the threshold according to the volatility and complexity of the data, so as to more accurately identify potential anomalies. This method shows strong robustness and flexibility in a variety of flight environments, which helps to improve the safety and reliability of drone systems. In terms of anomaly location, by analyzing the statistical characteristics of the residuals and locating the characteristic variables where the anomaly occurs, the source of the anomaly can be determined more simply and efficiently, which can effectively reduce the time and cost of manpower investigation.

[0014] The use of one-dimensional convolutional neural networks and attention mechanisms effectively reduces the computational complexity of the reasoning process. Through the one-dimensional convolutional layer, the model can quickly extract local patterns in the flight data, while the attention mechanism helps the model focus on key features and reduce the interference of irrelevant information. This design not only makes the model structure more concise, but also ensures that reasoning can be completed quickly when processing large-scale flight data. This method shows a high response speed in real-time monitoring application scenarios, ensuring the real-time nature of the anomaly detection method.

[0015] Furthermore, in S01, the formula for normalizing the original data x is: ,in is the normalized data, and are the maximum and minimum values ​​of the original data, respectively, and the normalized value is calculated for each variable; Furthermore, in S01, the ratio is 7:3, that is, the original input data is divided into a training set and a test set at a ratio of 70% and 30%, and the abnormal data in the test set is generated by an abnormal injection method. The injected abnormality types include deviation abnormality, drift abnormality and stuck abnormality. The abnormal injection formula is as follows: Deviation anomaly:

[0016] in is a constant, is the original flight data; Drift anomaly: ,in is the drift rate, t is the time period; Stuck exception: ,in is a constant.

[0017] Further, in said S02, in the local feature extraction module based on one-dimensional convolution, for the input data , after Conv1D operation, the output , The Lth element of , is the convolution kernel and , K is the length of the convolution kernel, yes L elements, is the Kth element of the convolution kernel.

[0018] Further, in S02, in the key feature learning module based on the attention mechanism, given the input data X and the query vector q, The weight of

[0019] In the formula is a scoring function, soft is used to Converted to the softmax function of probability distribution, yes Then, and The weighted sum is used to obtain the information that the model focuses on, expressed as Atten, as shown in the following formula: .

[0020] Further, in the S02, in the data representation module based on the autoencoder, the encoder of the autoencoder (AE) compresses the input data D into a latent variable z ,in is the weight of the encoder, is the encoder offset. is the activation function, and the decoder decodes z into reconstructed data ,in is the decoder weight, is the bias of the decoder, and the training goal of AE is to make D and The reconstruction error between them is as small as possible, and the loss function can be defined as follows:

[0021] Further, in S03, the residual is reconstructed using the training set As the anomaly score S, the threshold th is initialized. The part exceeding the threshold can be written as S-th, which satisfies the generalized Pareto distribution, and its accumulation function in and σ are the shape parameter and scale parameter of GPD, respectively, which can be obtained by maximizing the likelihood function

[0022] estimate and σ, is the total number of peaks greater than th, the final threshold ,

[0023] Where q represents the predefined extreme point quartile, n represents the total number of samples, if the abnormal score of the i-th data point of the v-th variable in the test set Greater than the threshold value of this point , then the point is considered abnormal, otherwise it is normal.

[0024] 8. In S04, for the test set The cth residual of v variables is defined as

[0025] In the formula , and They are The cth test residual, original value and reconstructed value of the vth variable in , the residual and

[0026] As the abnormality score of this variable, The higher it is, the more likely variable v is an outlier parameter.

[0027] Furthermore, the experimental data used in the above method includes simulated data and real UAV flight data. The simulated data is derived from the XTDrone simulation platform based on PX4, ROS and Gazebo, and the real data uses the Thor Flight 98 dataset published by the Minnesota UAV Laboratory.

[0028] Furthermore, in S03, accuracy (Acc), true positive rate (TPR), recall rate (recall), precision rate (precision), false positive rate (FPR) and F1 score are used as evaluation indicators; in the anomaly localization task, ALC and NLC are used as evaluation indicators. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 Adaptive anomaly detection and localization framework for UAV flight data based on neural network reconstruction model; Figure 2 The DCANNs neural network autoencoder model designed for this purpose; Figure 3 Schematic diagram of Conv1D operation; Figure 4 It is a schematic diagram of AE structure; Figure 5 To simulate the UAV data and the real UAV data curve; Figure 6 The abnormal parameter curve, abnormal amplitude and abnormal interval of the real test set and the simulated test set after the anomaly is injected; Figure 7 Visualize the anomaly detection results of DCANNs-ADL for simulated and real UAV flight data; Figure 8 Visualization of anomaly localization results for DCANNs-ADL, LSTM-AE, DAGMM, TranAD, and USAD. DETAILED DESCRIPTION

[0030] Embodiment 1, a multivariate flight data anomaly detection and location method based on a reconstruction model, characterized in that it comprises the following steps: S01 Data processing stage: collect UAV flight data and perform normalization preprocessing on the raw data; divide the raw data into training set and test set in proportion. The training set contains only normal data, and the test set generates abnormal data by injecting deviation anomalies, drift anomalies, and stuck anomalies into normal data; S02 Model construction and training phase: Construct the DCANNs reconstruction model, which includes a local feature extraction module based on one-dimensional convolution, a key feature learning module based on the attention mechanism, and a data representation module based on the autoencoder; use the training set data to train the model to minimize the reconstruction error of the autoencoder; S03 anomaly detection stage: calculate the reconstruction residuals of the training set and the test set, and use the reconstruction residuals of the training set to initialize the initial threshold of the method based on extreme value theory; estimate the shape parameter and scale parameter of the generalized Pareto distribution by maximizing the likelihood function to obtain the final anomaly threshold; compare the anomaly score of the test set data with the threshold to determine whether the data is abnormal; S04 anomaly location stage: Based on the test set residual, the residual sum of each variable is calculated, and the residual sum is used as the abnormality score of the variable. The abnormal parameters are located by comparing the abnormality scores.

[0031] In the data preprocessing step, the original flight data are normalized and mapped to the same numerical range, eliminating the impact of dimensions between different variables.

[0032] In the model construction step, the DCANNs-based neural network reconstruction model makes full use of the characteristics of different modules. The Conv1D module extracts local patterns through sliding windows and can capture local features in flight data. The attention mechanism module captures global dependencies by dynamically assigning weights, enabling the model to focus on important data features. The autoencoder module compresses and decodes data, learns the intrinsic representation of the data, and minimizes the reconstruction error to obtain effective data representation.

[0033] In the adaptive anomaly detection step, the dynamic threshold calculation method based on extreme value theory (EVT) can dynamically calculate the anomaly detection threshold according to the reconstructed residuals of the training set and the test set, thereby improving the anomaly detection performance.

[0034] In the anomaly location step, the abnormal parameters are located by calculating the statistical characteristics of the residuals. The residuals are calculated based on the difference between the original data and the reconstructed data, and then the residual sum of each variable is calculated as the abnormality score. The higher the score, the more likely the variable is an abnormal parameter, thus achieving accurate location of the anomaly.

[0035] The model uses multiple one-dimensional convolutional layers to effectively extract local patterns in drone flight data. These convolutional layers can improve the feature extraction ability of complex flight data by capturing potential correlations in time series data. At the same time, the attention mechanism is introduced in the model, which can automatically identify and focus on the most important features or time segments, thereby enhancing the learning ability of key features. This mechanism enables the model to effectively focus on the parts that have an important impact on drone flight data, avoid information redundancy, and further improve the potential representation learning ability of multivariate flight data. In this way, this method can not only reconstruct flight data more accurately, but also adapt to drone data of different types and environments, with strong robustness and generalization ability. Therefore, this model shows higher accuracy and stability when processing complex and changeable drone flight data.

[0036] By performing residual analysis on drone flight data, the difference between normal flight status and abnormal behavior is effectively captured. The model first reconstructs the input data and obtains the residual by calculating the reconstruction error. In order to improve the accuracy of anomaly detection, a dynamic threshold generation mechanism is proposed, which can dynamically adjust the threshold according to the volatility and complexity of the data, so as to more accurately identify potential anomalies. This method shows strong robustness and flexibility in a variety of flight environments, which helps to improve the safety and reliability of drone systems. In terms of anomaly location, by analyzing the statistical characteristics of the residuals and locating the characteristic variables where the anomaly occurs, the source of the anomaly can be determined more simply and efficiently, which can effectively reduce the time and cost of manpower investigation.

[0037] The use of one-dimensional convolutional neural networks and attention mechanisms effectively reduces the computational complexity of the reasoning process. Through the one-dimensional convolutional layer, the model can quickly extract local patterns in the flight data, while the attention mechanism helps the model focus on key features and reduce the interference of irrelevant information. This design not only makes the model structure more concise, but also ensures that reasoning can be completed quickly when processing large-scale flight data. This method shows a high response speed in real-time monitoring application scenarios, ensuring the real-time nature of the anomaly detection method.

[0038] In S01, the formula for normalizing the original data x is: ,in is the normalized data, and are the maximum and minimum values ​​of the original data, respectively, and the normalized value is calculated for each variable; Different variables in drone flight data, such as speed, acceleration, attitude angle, etc., usually have different dimensions and orders of magnitude. Separate normalization can eliminate these differences and enable different variables to be compared and analyzed at the same scale. For example, the speed unit may be meters per second, and the acceleration unit is meters per second squared. If they are not normalized separately, the small changes in the acceleration data may be masked by the larger values ​​of the speed data, making it difficult for the model to accurately capture the characteristics and laws of different variables during the learning process. After normalization, the model can treat each variable equally and improve the ability to process multivariate data. Calculating the normalized values ​​separately can better retain these characteristics, so that the model can more accurately capture the normal and abnormal patterns of each variable during the learning process. For example, a variable has a small normal range, and abnormal changes may cause large fluctuations in the value; another variable has a large normal range, and a small change in the value may represent anomalies. Separate normalization allows the model to learn according to the characteristics of different variables, detect and locate anomalies more accurately, and thus improve the accuracy of the model in anomaly detection and positioning tasks. The drone flight environment is complex and changeable, and the flight data is susceptible to noise interference. Normalizing each variable separately can reduce the impact of noise on the model to a certain extent. Because normalization is based on the characteristics of each variable itself, even if there is noise in the data, it can be mapped to a reasonable range, reducing the interference of noise on model judgment, so that the model can still maintain good performance when facing different degrees of noise, and enhance the robustness of the model.

[0039] In S01, the ratio is 7:3, that is, the original input data is divided into a training set and a test set at a ratio of 70% and 30%. The abnormal data in the test set is generated by an abnormal injection method. The injected abnormal types include deviation abnormality, drift abnormality and stuck abnormality. The abnormal injection formula is as follows: Deviation anomaly:

[0040] in is a constant, is the original flight data; Drift anomaly: ,in is the drift rate, t is the time period; Stuck exception: ,in is a constant.

[0041] UAVs may encounter various abnormal conditions in actual flight. Deviation anomalies simulate the situation where the data deviates from the normal range due to the instantaneous deviation of the sensor or the sudden failure of the system, such as the deviation of the sensor due to electromagnetic interference. Drift anomalies correspond to data trend anomalies caused by the gradual aging of the system and the slow change of component performance, such as the gradual decrease of battery power causing some parameters to change slowly. Stuck anomalies reflect the failure of the sudden failure of the component that causes the data to remain unchanged, such as the data is no longer updated after the sensor fails. Through these three types of anomalies, the problems that may occur in real flight can be fully simulated, and abnormal data close to the actual situation can be provided for model training, making the trained model more practical in actual applications. At the same time, different types of anomalies have great differences in data characteristics and change rules. Deviation anomalies are discrete mutations, drift anomalies are continuous gradients, and stuck anomalies are fixed anomalies. In the process of learning these abnormal data with different characteristics, the model can enhance its adaptability and recognition ability to various abnormal situations, and can more accurately detect and locate abnormalities when facing complex and diverse real anomalies.

[0042] In the S02, in the local feature extraction module based on one-dimensional convolution, for the input data , after Conv1D operation, the output , The Lth element of , is the convolution kernel and , K is the length of the convolution kernel, yes L elements, is the Kth element of the convolution kernel. The UAV flight data is multidimensional data with time series characteristics, which contains a lot of local correlation information. Through this convolution operation, Conv1D can slide on the data sequence with a fixed-size convolution kernel, and each operation focuses on a local window of the data. For example, when processing data that changes over time, such as flight speed and acceleration, the convolution kernel can capture the change pattern of these parameters in a short period of time, such as the sharp change in speed in a short period of time, the fluctuation law of acceleration, etc. This effective capture of local patterns helps to extract key information reflecting the flight status of the UAV, providing an important basis for subsequent anomaly detection and positioning. In addition, the characteristics of weight sharing and fixed parameter count make the model adaptable to input data of different lengths and reduce the risk of overfitting. Regardless of how the length of the input data changes, Conv1D can extract local features in the same way, thereby improving the generalization ability of the model. For example, in practical applications, the length of data generated by different flight missions may be different, but the model can effectively process and accurately extract features. This enables the model to maintain stable performance and more reliably detect and locate anomalies when facing multivariate flight data in various flight scenarios.

[0043] In S02, in the key feature learning module based on the attention mechanism, given the input data X and the query vector q, The weight of

[0044] In the formula is a scoring function, soft is used to Converted to the softmax function of probability distribution, yes Then, and The weighted sum is used to obtain the information that the model focuses on, expressed as Atten, as shown in the following formula:

[0045] The flight data of drones contains multiple parameters, and the importance of each parameter in judging the flight status varies under different flight conditions. The model can automatically measure the importance of each input data by calculating weights with the softmax function. For example, during the takeoff and landing phases of a drone, attitude angle data is crucial to determining flight safety. The attention mechanism will assign higher weights to attitude angle-related data, allowing the model to pay more attention to these key information, thereby accurately capturing key features related to flight anomalies and improving the accuracy of anomaly detection and positioning. The drone flight environment is complex and changeable, and the important features of flight data at different times will also change dynamically. The attention mechanism can adjust weights in real time and flexibly allocate attention according to the current flight conditions. When encountering airflow interference, the importance of data such as speed and acceleration will change. The attention mechanism can capture this change in time and reallocate attention to ensure that the model always focuses on the most representative features, enhance the model's adaptability to complex dynamic environments, and effectively respond to anomaly detection tasks in various flight scenarios.

[0046] In S02, in the data representation module based on the autoencoder, the encoder of the autoencoder (AE) compresses the input data D into a latent variable z ,in is the weight of the encoder, is the encoder offset. is the activation function, and the decoder decodes z into reconstructed data ,in is the decoder weight, is the bias of the decoder, and the training goal of AE is to make D and The reconstruction error between them is as small as possible, and the loss function can be defined as follows:

[0047] Since the training goal of the autoencoder is to minimize the reconstruction error between the original data D and the reconstructed data, during the training process, the model will learn to remove the noise in the input data. If the flight data is mixed with noise during the acquisition or transmission process, the autoencoder will tend to restore the real flight data signal when reconstructing the data, thereby suppressing the noise. The trained autoencoder can convert the noisy original data into purer data that can better reflect the real flight status of the drone, improve the data quality, and provide a more reliable data basis for anomaly detection and positioning. The loss function is constructed by calculating the reconstruction error between the original data D and the reconstructed data, which provides an effective basis for anomaly detection. Under normal flight conditions, the autoencoder can accurately reconstruct the data with a small reconstruction error; when the data is abnormal, the reconstruction error will increase significantly. Based on this feature, by setting a suitable threshold, when the reconstruction error exceeds the threshold, it is judged that the data is abnormal.

[0048] In S03, the residual is reconstructed using the training set As the anomaly score S, the threshold th is initialized. The part exceeding the threshold can be written as S-th, which satisfies the generalized Pareto distribution, and its accumulation function

[0049] in and σ are the shape parameter and scale parameter of GPD, respectively, which can be obtained by maximizing the likelihood function

[0050] estimate and σ, is the total number of peaks greater than th, the final threshold ,

[0051] Where q represents the predefined extreme point quartile, n represents the total number of samples, if the abnormal score of the i-th data point of the v-th variable in the test set Greater than the threshold value of this point , then the point is considered abnormal, otherwise it is normal.

[0052] The flight data of UAVs changes dynamically with the flight environment and tasks. The traditional fixed threshold cannot adapt to such changes, which easily leads to a high false alarm rate. This method uses the training set to reconstruct the residual initialization threshold th, and dynamically adjusts the final threshold based on the GPD estimated shape parameter γ and scale parameter σ When the flight data fluctuates or changes in complexity, γ and σ change accordingly, so that the threshold can reflect the real-time characteristics of the data. For example, under the interference of complex electromagnetic environment, the flight data fluctuations intensify, and the threshold will be adjusted accordingly to ensure the accuracy of anomaly detection and reduce false positives and negatives. In S04, for the test set The cth residual of v variables is defined as

[0053] In the formula , and They are The cth test residual, original value and reconstructed value of the vth variable in , the residual and

[0054] As the abnormality score of this variable, The higher it is, the more likely variable v is an outlier parameter.

[0055] By calculating the residual sum as the abnormality score, the deviation of each variable from the normal state is simply and intuitively quantified. When the variable is in a normal state, its reconstructed value will be very close to the original value, the residual will be small, and the corresponding residual sum will also be small; conversely, if the variable is abnormal, the difference between the reconstructed value and the original value will increase, and the residual sum will also increase accordingly. Taking the acceleration variable in the UAV flight data as an example, the difference between the reconstructed value and the original value of the acceleration data during normal flight is very small, and the residual sum is close to zero; when the acceleration sensor fails or the aircraft is affected by abnormal external forces, the deviation between the reconstructed value and the original value of the acceleration data increases, and the residual sum rises rapidly. The degree of abnormality is clear at a glance, which is convenient for quickly identifying abnormal variables; the residual sum as an abnormality score can accurately point out which variable or variables are abnormal. The residual sums of different variables are calculated independently of each other. Even if there are complex situations where multiple variables are abnormal at the same time, the variable with the highest probability of abnormality can be quickly located by comparing the residual sum size of each variable. For example, when the attitude control of a drone is abnormal, multiple variables such as attitude angle and angular velocity may be abnormal at the same time. By calculating the residual sum of each variable, it is possible to clearly distinguish which attitude angle or angular velocity variable has a more prominent abnormality.

[0056] The experimental data used in the above method includes simulated data and real UAV flight data. The simulated data comes from the XTDrone simulation platform based on PX4, ROS and Gazebo. The real data uses the Thor Flight 98 data set published by the Minnesota UAV Laboratory. The simulated data generated by the XTDrone simulation platform based on PX4, ROS and Gazebo can accurately control the generation of various flight conditions and abnormal situations. Researchers can set different flight scenarios at will according to experimental requirements, such as different weather conditions, flight missions, and various types of abnormal injections, and these experiments can be repeated many times. For example, if you want to test the model's ability to detect specific sensor deviation anomalies under strong wind interference, you can repeatedly set the same strong wind parameters and deviation anomaly parameters in the simulation platform, and obtain multiple sets of data for model training and evaluation to ensure the reliability and repeatability of the experimental results.

[0057] In S03, the accuracy rate (Acc), true positive rate (TPR), recall rate (recall), precision rate (precision), false positive rate (FPR) and F1 score are used as evaluation indicators; in the anomaly localization task, ALC and NLC are used as evaluation indicators. The accuracy rate (Acc) can intuitively reflect the overall proportion of normal and abnormal samples correctly identified by the model, and present the performance of the model as a whole. However, relying solely on the accuracy rate may be misleading due to problems such as data imbalance. The true positive rate (TPR, equivalent to the recall rate recall) focuses on the proportion of correctly detected samples in actual abnormal samples, which reflects the model's ability to capture abnormal samples. The precision rate (precision) measures the proportion of truly abnormal samples in the samples judged as abnormal by the model, which can prevent the model from misjudging too many normal samples as abnormal. The false positive rate (FPR) indicates the proportion of normal samples that are incorrectly marked as abnormal, and complements the precision rate to evaluate the model from the perspective of misjudgment. The F1 score combines the precision rate and the recall rate, balances the impact of the two, and can more comprehensively and objectively evaluate the performance of the model in the anomaly detection task. For example, in the anomaly detection of drone flight data, if the model Acc is high but the FPR is also high, it means that although the overall correct judgment ratio is high, there are many misjudgments. Combining these indicators can clearly identify model problems; in the anomaly positioning task, ALC directly measures the model's ability to accurately find abnormal parameters, and NLC reflects the accuracy of the model in distinguishing normal parameters from abnormal parameters. These two indicators can accurately evaluate the effect of the model in locating anomalies in multivariate flight data. In complex flight data, there may be multiple parameter anomalies. ALC can intuitively reflect the probability of the model accurately pointing out abnormal parameters, and NLC can reflect the model's misjudgment of normal parameters, helping to judge the reliability of model positioning.

[0058] Embodiment 2: 1 Methodology Framework like Figure 1 As shown in the figure, in order to realize adaptive anomaly detection and localization of multivariate UAV flight data, the DCANNs-ADL framework is proposed. The framework mainly includes three links: data preparation, model training and testing, and anomaly detection and localization. In the data preparation link, the UAV flight data is collected and preprocessed. In the model training and testing link, the advantages of one-dimensional convolutional neural network and attention mechanism in local pattern and global feature extraction and autoencoder in data representation are fully utilized. The constructed DCANNs autoencoder model can fully extract and learn the key features of the data, thereby realizing accurate data reconstruction. In the anomaly detection and localization link, a dynamic threshold strategy is adopted to reduce the false alarm rate and improve the anomaly detection performance of the model in complex dynamic environments. At the same time, the residual statistical features are used to characterize the abnormal degree of flight parameters, so as to realize the rapid and accurate localization of abnormal parameters.

[0059] 1.1 Neural Network Reconstruction Model Architecture Figure 2 The designed neural network model is presented. The model first uses a one-dimensional convolutional layer to capture the local features of the UAV flight data. These local features reflect the key flight and environmental information collected by the UAV during flight, such as speed, acceleration, attitude angle, etc. Through the convolutional layer, the model can effectively extract local patterns in time series data and reveal potential correlations and short-term trends in the data. At the same time, the model also introduces an embedded attention mechanism to capture the global dependencies of the data. The attention layer assigns dynamic weights to each sampling point, allowing the model to focus on the most representative features in the flight data, thereby better capturing the global patterns in long time series. In this way, the model can not only process local features, but also consider longer-distance dependencies in the time series, improving the ability to understand complex UAV flight data.

[0060] In addition, the model also uses autoencoders to extract and learn key features of the data. In this process, the autoencoder not only compresses the high-dimensional representation of the data, but also effectively filters the noise in the input data, reducing the impact of noise on the data reconstruction results. Through this noise reduction process, the model is able to reconstruct the flight data more accurately and enhance the fidelity of the original signal. Finally, by combining one-dimensional convolution, attention mechanism and autoencoder technology, the model demonstrates excellent ability in the process of flight data reconstruction, thereby achieving high-precision data reconstruction.

[0061] Local feature extraction module based on one-dimensional convolution Figure 3 As shown in the figure, one-dimensional convolution (Conv1D) is used to capture local patterns in time series data, using sliding windows to effectively extract local relevant features. Its weight sharing mechanism between convolution kernels ensures computational efficiency and produces a fixed parameter count that is independent of the input length. This lightweight structure minimizes the risk of over-coupling and ensures scalability across different input lengths.

[0062] Given input data , after Conv1D operation, the output can be obtained , as shown below:

[0063] In the formula yes The lth element of is the convolution kernel and , K is the length of the convolution kernel, yes The l+k-1th element of is the kth element of the convolution kernel.

[0064] (2) Key feature learning module based on attention mechanism The attention mechanism extracts information from the input data, enabling the model to assign attention weights at different time steps or spatial locations, thereby processing the data more efficiently. Given the input data X and the query vector q, The weight can be expressed as follows:

[0065] Where S(∙) is the scoring function, soft is the softmax function used to convert S(∙) into a probability distribution. yes Then, and The weighted sum is used to obtain the information that the model focuses on, expressed as Atten, as shown in the following formula:

[0066] (3) Data representation module based on autoencoder like Figure 4 As shown in Figure 2, the autoencoder (AE) extracts the main features of the data by constructing an encoder and a decoder, compresses the input data into a low-dimensional representation, and then reconstructs it. AE performs well in denoising and data degradation and is widely used in image processing, natural language processing, and time-varying data analysis.

[0067] Given the input data D, it is first compressed by the encoder into a latent variable z, which can be expressed as follows:

[0068] In the formula is the weight of the encoder, is the bias of the encoder. f(∙) is the activation function. The decoder then decodes z back to the original data space to obtain the reconstructed data , which can be expressed as follows:

[0069] In the formula is the decoder weight, is the bias of the decoder. The training goal of AE is to make D and The reconstruction error between them is as small as possible, and the loss function can be defined as follows:

[0070] Where loss is the loss function, Represents the square of the Euclidean norm of a vector.

[0071] 1.2 Adaptive Anomaly Detection During model training and testing, the reconstruction residuals of the training set and test set can be calculated by the following formula:

[0072]

[0073] In the formula , , and are the training set and test set and their sample lengths respectively. and as well as and They are and The reconstructed residual (as the anomaly score S) and the reconstructed data. Current research mainly uses fixed statistical thresholds, which are difficult to adapt to changes in drone flight status and system conditions, resulting in a high false alarm rate. To solve this problem, the stream peak-over-threshold (SPOT) method based on extreme value theory is adopted. SPOT does not require any assumptions about the data and has good scalability when processing multidimensional data. Determining the anomaly threshold according to extreme value theory only depends on the distribution of extreme values. The part exceeding the threshold can be written as (S-th), which satisfies the generalized Pareto distribution (GPD). Its cumulative function is defined as follows

[0074] Where th is the initial threshold for retrieving the peak anomaly score, which is As anomaly score Initialize. and σ are the shape parameter and scale parameter of GPD, respectively, which can be estimated by maximizing the likelihood function as follows:

[0075] In the formula is the total number of peaks greater than th. Then, we can get and the estimated value of σ and The estimated value of .

[0076]

[0077] Where q represents the predefined extreme point quartile, n represents the total number of samples. ∆ is the final threshold The difference between the initial threshold th, that is, Rewriting formula (19) yields The expression is defined as follows:

[0078] Finally, by Anomaly detection is achieved by comparing it with the anomaly score S. Specifically, if the anomaly score of the i-th data point of the v-th variable in the test set is Greater than the threshold value of this point , then the point is marked as 1, i.e. abnormal, vice versa, marked as 0, as shown in the following formula:

[0079] 1.3 Abnormal Location In practical applications, flight data is generated in the form of data streams, which makes it often difficult to determine which parameters exhibit abnormalities. In addition to detecting potential anomalies, locating abnormal parameters is also crucial for implementing corresponding measures in a timely manner. The cth residual of the v variables in is defined as follows:

[0080] In the formula , and They are The cth test residual, original value, and reconstructed value of the vth variable in . , the residual sum of the vth variable can be calculated according to the following formula:

[0081] In the formula is the residual sum of the vth variable, serving as the abnormality score of the variable. The higher it is, the more likely variable v is an outlier parameter, and vice versa.

[0082] 2 Experimental data description Figure 5 As shown in Figure 1, the experimental data in this paper includes simulated data and real UAV flight data. The simulation data comes from the XTDrone simulation platform based on PX4, ROS and Gazebo. The real data uses the ThorFlight98 dataset released by the Minnesota UAV Laboratory. The simulation data has 22,340 sampling points for each parameter, and the sampling frequency is 10 Hz. The real data has 21,015 sampling points for each parameter, and the sampling frequency is 50 Hz. 12,000 sampling points were selected as experimental data, which include 9 flight parameters from magnetometer, accelerometer and gyroscope sensors. Table 1 and Figure 5Descriptions and original curves of these parameters are shown respectively.

[0083] Table 1 Experimental parameter description

[0084] 3 Data partitioning and exception injection Figure 6 As shown in Figure 2, before model training, the original input x needs to be divided into a training set and a test set in order to train and evaluate the model. and ,in Contains only normal data, Contains abnormal data. Considering the difficulty and risk of obtaining real abnormal data, the abnormal data is generated by the abnormal injection method. In order to simulate the abnormal conditions in the UAV flight data as much as possible, different types of abnormalities are injected, including deviation abnormality, drift abnormality and stuck abnormality, as shown in the following formula:

[0085]

[0086]

[0087] Where α and β are constants, y(t) is the original flight data, φ is the drift rate, and t is the time period. Figure 4 The anomaly parameter curves, anomaly amplitudes, and anomaly intervals of the real test set and the simulated test set after anomaly injection are shown.

[0088] 4 Data Preprocessing For the original data x, use the following formula to preprocess it:

[0089] in is the normalized data, and are the maximum and minimum values ​​of x respectively. The normalized data of the vth variable It can be calculated according to the following formula:

[0090] In the formula and They are The maximum and minimum values ​​of .

[0091] 5 Evaluation criteria For the anomaly detection task, accuracy (Acc), true positive rate (TPR), recall rate (recall), precision rate (precision), false positive rate (FPR) and F1 score are used as evaluation indicators, as shown in the following formula. The larger the value of Acc, TPR (recall rate), precision and F1 score, and the smaller the value of FPR, the better the performance of the model.

[0092]

[0093]

[0094]

[0095]

[0096]

[0097] In the formula, TP represents the correctly identified normal samples, FP represents the incorrectly labeled abnormal samples, TN represents the correctly identified abnormal samples, and FN represents the incorrectly classified normal samples. For the abnormality localization task, ALC and NLC are designed as evaluation indicators, and the calculation formula is as follows:

[0098]

[0099] Where ALC and NLC represent the correct positioning accuracy of abnormal parameters and normal parameters respectively.

[0100] 6 Experimental results 6.1 Comparison with Baseline Methods To comprehensively evaluate the effectiveness of the proposed method, advanced models for multi-source anomaly detection in the same and related but different fields, including LSTM-AE, TranAD, DAGMM and USAD, are used as comparative baseline methods. 1) LSTM-AE is a hybrid neural network method based on LSTM and AE. This method uses LSTM and AE to capture the spatiotemporal characteristics of multivariate UAV flight data and uses the reconstruction error as the anomaly score. 2) TranAD is a Transformer-based multivariate anomaly detection method. It uses the multi-head attention mechanism of Transformer to capture the global characteristics of multivariate data, thereby achieving fast inference. 3) DAGMM combines Gaussian mixture models (GMM) with deep autoencoders to generate low-dimensional representations. This method performs joint optimization in an end-to-end manner while using a separate estimation network to facilitate the learning of the hybrid model. 4) USAD is an anomaly detection method that reconstructs data using two autoencoders and adversarial training. This method uses the reconstruction error of two autoencoders to calculate the anomaly score.

[0101] 6.2 Anomaly Detection Figure 7 As shown in Figure 2, the visual anomaly detection results of DCANNs-ADL on simulated and real UAV flight data. DCANNs-ADL can clearly distinguish between normal and abnormal patterns and effectively identify abnormal data. To quantify the anomaly detection results, Table 2 lists the anomaly detection results of DCANNs-ADL, LSTM-AE, TranAD, DAGMM, and USAD. The best indicators are marked in bold. DCANNs-ADL performs best on most indicators of simulated and real UAV flight data. In particular, it can accurately detect all abnormal data, as shown by FPR=0.00000 and Precision=1.00000. In addition, DCANNs-ADL achieves the highest Acc value and F1 value of 0.99481 and 0.99728, and 0.99623 and 0.99803 on simulated and real UAV flight data, respectively. Considering all indicators, although the average TPR (Recall) of DCANNs-ADL is slightly lower than that of TranAD, its average FPR, Acc, Precision and F1 are 0.00000, 0.99552, 1.00000 and 0.99765 respectively, which are better than LSTM-AE, DAGMM, TranAD and USAD. These results verify the excellent performance of DCANNs-ADL in anomaly detection.

[0102] The average values ​​of TPR (Recall), Acc, Pci-sion, and F1 of LSTM-AE on simulated and real UAV flight data all exceed 0.97000. However, its average FPR is higher, which is 0.19607. This is mainly because LSTM-AE uses LSTM to capture features and ignores the learning of global information. In addition, LSTM-AE may lose some useful information when selectively retaining historical information. TranAD learns the dependency information between different time steps through the attention mechanism, thereby more comprehensively capturing the temporal trend of UAV time series data. Although the average FPR value of TranAD is 0.27857, which is relatively high, its overall performance is worse than LSTM-AE. In particular, TranAD has the highest average TPR (Recall) value of 0.99613. However, TranAD may amplify noise in the inference stage. Therefore, the overall performance of TranAD is not as good as DCANNs-ADL. The average values ​​of TPR (Recall), FPR, Acc, Precision, and F1 of DAGMM on simulated and real UAV flight data are 0.98665, 0.25964, 0.97600, 0.98828, and 0.98741, respectively. Its overall performance is comparable to that of LSTM-AE and TranAD. This is because the input data of DAGMM only contains a single observation, not a historical time series window. However, time dependency is crucial for multivariate time series anomaly detection. Therefore, DAGMM is not suitable for time dependency modeling. With proper regularization, USAD can give full play to the advantages of AE and achieve good overall performance. However, USAD will also amplify noise in the inference stage like TranAD. Therefore, the overall performance of USAD is not as good as that of DCANNs-ADL.

[0103] Table 2 Anomaly detection results of DCANNs-ADL and baseline methods.

[0104]

[0105] Table 3 lists the time spent per epoch in the model training process of these methods. In real and simulated UAV flight data, DCANNs-ADL has the shortest single training time, which is 0.18413 seconds and 0.18490 seconds respectively. This is because DCANNs-ADL adopts one-dimensional convolutional layer and attention layer. The one-dimensional convolutional layer effectively reduces the number of parameter layers by sharing weights and local connections, thereby improving the computational efficiency of the model. The attention layer can selectively focus on learning key features, further reducing unnecessary computational overhead. Specifically, compared with LSTM-AE, TranAD, DAGMM and USAD, the average training time per epoch of DCANNs-ADL is only 0.18451 seconds, which is 92.99%, 97.31%, 99.12% and 99.86% less respectively. These results further verify the excellent computational efficiency of the proposed method.

[0106] Table 3. Training time (seconds) per epoch for the above methods.

[0107]

[0108] 6.3 Abnormal Location Figure 8 The anomaly localization results of DCANNs-ADL, LSTM-AE, DAGMM, TranAD, and USAD are shown. Figure 8 For example, the left sub-figure shows the residual matrix diagram of the test set, where the x-axis represents the multivariate flight parameter serial number and the y-axis represents the reconstructed residual of the test set. The right sub-figure shows a histogram of the anomaly location results, where the x-axis represents multiple variable flight parameters and the y-axis represents the sum of residuals. As mentioned earlier, the larger the value, the higher the probability that the vth variable corresponds to an abnormal parameter. "×" indicates that the abnormal parameter is located incorrectly, and "√" indicates that the abnormal parameter is located correctly. From Figure 6 As can be seen from (a) and (b), DCANNs-ADL accurately locates all abnormal parameters in real data and also performs well in simulated data.

[0109] Table 4 lists the anomaly localization results of the above methods. DCANNs-ADL has the best anomaly localization result, with ALC and NLC values ​​of 1.000 and 1.000, and 0.800 and 0.750 in real and simulated UAV flight data, respectively. TranAD obtained the second best result, with ALC and NLC values ​​of 0.750 and 0.800, and 0.800 and 0.750 in real and simulated UAV flight data, respectively. This is mainly because DCANNs-ADL and TranAD use attention mechanisms to capture the global dependencies of input data, so as to better distinguish and locate abnormal parameters and normal parameters. In contrast, the overall performance of LSTM-AE, DAGMM and USAD is relatively poor. Compared with LSTM-AE, TranAD, DAGMM and USAD, the average values ​​of ALC and NLC of DCANNs-ADL are improved by 33.33% and 34.62%, 16.13% and 12.90%, 33.33% and 34.62%, and 33.33% and 34.62%, respectively. The experimental results show that DCANNs-ADL has excellent consistency in anomaly detection and localization.

[0110] Table 4 Anomaly localization results of DCANNs-ADL and baseline methods.

[0111]

[0112] The above are only embodiments of the present invention. Common knowledge such as the known specific structures and characteristics in the scheme is not described in detail here. Ordinary technicians in the relevant field are aware of all the common technical knowledge in the technical field to which the invention belongs before the application date or priority date, can obtain all the existing technologies in the field, and have the ability to apply conventional experimental means before that date. Ordinary technicians in the relevant field can improve and implement this scheme in combination with their own abilities under the enlightenment given by this application. Some typical known structures or known methods should not become obstacles for ordinary technicians in the relevant field to implement this application. It should be pointed out that for those skilled in the art, several deformations and improvements can be made without departing from the structure of the present invention, which should also be regarded as the scope of protection of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A multivariate flight data anomaly detection and location method based on reconstruction model, characterized in that: The following steps are involved: S01 Data processing stage: collect UAV flight data and perform normalization preprocessing on the raw data; divide the raw data into training set and test set in proportion. The training set contains only normal data, and the test set generates abnormal data by injecting deviation anomalies, drift anomalies, and stuck anomalies into normal data; S02 Model construction and training phase: Construct the DCANNs reconstruction model, which includes a local feature extraction module based on one-dimensional convolution, a key feature learning module based on the attention mechanism, and a data representation module based on the autoencoder; use the training set data to train the model to minimize the reconstruction error of the autoencoder; S03 anomaly detection stage: calculate the reconstruction residuals of the training set and the test set, and use the reconstruction residuals of the training set to initialize the initial threshold of the method based on extreme value theory; The shape parameter and scale parameter of the generalized Pareto distribution are estimated by maximizing the likelihood function to obtain the final anomaly threshold. Compare the anomaly score of the test set data with the threshold to determine whether the data is abnormal; S04 anomaly location stage: Based on the test set residual, the residual sum of each variable is calculated, and the residual sum is used as the abnormality score of the variable. The abnormal parameters are located by comparing the abnormality scores.

2. The multivariate flight data anomaly detection and location method based on reconstruction model according to claim 1 is characterized in that: In S01, the formula for normalizing the original data x is: ,in is the normalized data, and are the maximum and minimum values ​​of the original data, respectively, and the normalized value is calculated for each variable.

3. The multivariate flight data anomaly detection and location method based on reconstruction model according to claim 1, characterized in that: In S01, the ratio is 7:3, that is, the original input data is divided into a training set and a test set at a ratio of 70% and 30%. The abnormal data in the test set is generated by an abnormal injection method. The injected abnormal types include deviation abnormality, drift abnormality and stuck abnormality. The abnormal injection formula is as follows: Deviation anomaly: in is a constant, is the original flight data; Drift anomaly: ,in is the drift rate, t is the time period; Stuck exception: ,in is a constant.

4. The multivariate flight data anomaly detection and location method based on reconstruction model according to claim 1, characterized in that: In the S02, in the local feature extraction module based on one-dimensional convolution, for the input data , after Conv1D operation, the output , The Lth element of , is the convolution kernel and , K is the length of the convolution kernel, yes L elements, is the Kth element of the convolution kernel.

5. The multivariate flight data anomaly detection and location method based on reconstruction model according to claim 1, characterized in that: In S02, in the key feature learning module based on the attention mechanism, given the input data X and the query vector q, The weight of In the formula is a scoring function, soft is used to Converted to the softmax function of probability distribution, yes The weight of and The weighted sum is used to obtain the information that the model focuses on, expressed as Atten, as shown in the following formula: 。 6. The multivariate flight data anomaly detection and location method based on reconstruction model according to claim 1, characterized in that: In S02, in the data representation module based on the autoencoder, the encoder of the autoencoder (AE) compresses the input data D into a latent variable z ,in is the weight of the encoder, is the encoder bias, is the activation function, and the decoder decodes z into reconstructed data ,in is the decoder weight, is the bias of the decoder, and the training goal of AE is to make D and The reconstruction error between them is as small as possible, and the loss function can be defined as follows: 。 7. The multivariate flight data anomaly detection and location method based on reconstruction model according to claim 1, characterized in that: In S03, the residual is reconstructed using the training set As the anomaly score S, the threshold th is initialized. The part exceeding the threshold can be written as S-th, which satisfies the generalized Pareto distribution, and its accumulation function in and σ are the shape parameter and scale parameter of GPD, respectively, which can be obtained by maximizing the likelihood function estimate and σ, is the total number of peaks greater than th, the final threshold , Where q represents the predefined extreme point quartile, n represents the total number of samples, if the abnormal score of the i-th data point of the v-th variable in the test set Greater than the threshold value of this point , then the point is considered abnormal, otherwise it is normal.

8. The multivariate flight data anomaly detection and location method based on reconstruction model according to claim 1, characterized in that: In S04, for the test set The cth residual of v variables is defined as In the formula , and They are The cth test residual, original value and reconstructed value of the vth variable in , the residual and As the abnormality score of this variable, The higher it is, the more likely variable v is an outlier parameter.

9. The multivariate flight data anomaly detection and location method based on reconstruction model according to claim 1, characterized in that: The experimental data used in the above method includes simulated data and real UAV flight data. The simulated data comes from the XTDrone simulation platform based on PX4, ROS and Gazebo, and the real data uses the Thor Flight 98 dataset published by the Minnesota UAV Laboratory.

10. The multivariate flight data anomaly detection and location method based on reconstruction model according to claim 1, characterized in that: In the S03, accuracy (Acc), true positive rate (TPR), recall rate (recall), precision rate (precision), false positive rate (FPR) and F1 score are used as evaluation indicators; in the anomaly localization task, ALC and NLC are used as evaluation indicators.

Images