A multivariate flight data anomaly detection and location method based on reconstruction model

Abstract

A multivariate flight data anomaly detection and location method based on a reconstruction model comprises the following steps: a data preprocessing step: collecting unmanned aerial vehicle (UAV) flight data, dividing the original input data into a training set and a test set according to a preset ratio, and normalizing the data; a model building step: building a neural network reconstruction model based on DCANNs; an autoencoder module compressing the input data into latent variables through an encoder, and then decoding z back into the original data space by a decoder to obtain reconstructed data; an adaptive anomaly detection step: calculating an anomaly threshold using the reconstructed residuals of the training set, and achieving anomaly detection by comparing the anomaly score of the test set with the threshold; and an anomaly location step: locating anomaly parameters by calculating and analyzing the statistical features of the residuals. The present invention aims to solve the problems of insufficient feature extraction in traditional deep learning models, the inability of fixed statistical thresholds to adapt to dynamic changes in flight data, and the inability to locate the specific source of anomaly parameters.

Description

Technical Field

[0001] The present 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 sector, with the widespread adoption of aircraft such as drones and civilian aircraft, effective monitoring of flight data has become increasingly critical. As a key source of information reflecting an aircraft's operational status, flight data encompasses multiple variables such as velocity, acceleration, attitude angle, magnetometer, and gyroscope. It also exhibits strong temporal characteristics and complex inter-variable correlations. Accurate anomaly detection and location can promptly identify potential aircraft failures, which is crucial for ensuring flight safety, improving aircraft reliability, and optimizing maintenance strategies.

[0003] In actual production and use scenarios, flight data anomaly detection and localization face numerous challenges. Firstly, most existing methods suffer from deficiencies in feature extraction. For example, the common LSTM-based model is prone to forgetting or insufficiently retaining information when processing long-term dependencies in long-term data series, such as flight data. This makes it difficult to accurately capture complex correlation patterns in the data, resulting in poor feature extraction, severely impacting model accuracy and generalization, and making it difficult to stably and reliably detect anomalies under complex flight conditions. Secondly, current research often focuses on detecting anomalies in a single parameter, ignoring the fact that flight anomalies often involve the coordinated changes of multiple parameters. This lacks comprehensive control over the overall anomaly situation, easily misses key anomaly information, and prevents a comprehensive and accurate assessment of aircraft operational status. Furthermore, traditional anomaly detection methods typically rely on fixed statistical thresholds. Given the complex and ever-changing flight environment and the highly dynamic nature of flight data, such thresholds are difficult to adjust adaptively, resulting in frequent false positives and missed negatives under varying flight conditions, making them inadequate for meeting actual flight safety requirements. Existing research has focused on the model's anomaly recognition performance, but rarely delve into the root causes of the anomalies. In practical applications, simply detecting anomalies is not sufficient; the key lies in accurately locating the specific parameters that cause the anomaly, which provides an important basis for aircraft operation and maintenance decisions.

[0004] A search revealed that some research focuses on improving model architecture to enhance performance. For example, patent CN117851946A proposes an anomaly detection method for multivariate flight time series data. This method first divides historical time series data into multiple subsequences and feeds them into an RNN network for training. The data is then processed through a CNN filter layer and a covector filter layer. The results are then integrated to form a prediction model, which is then used to determine anomalies in real-time data subsequences. This method's advantage lies in its ability to integrate multiple sources of information and different processing techniques, enhancing the model's ability to understand and process data. However, its drawbacks are also significant. The inherent limitations of the RNN model in processing long sequences have not been fundamentally addressed, and it may still be limited in processing complex data relationships. The computational complexity also compromises detection efficiency. Furthermore, this solution utilizes traditional fixed statistical thresholds, making it ineffective for detecting dynamic changes in flight data. Furthermore, while the patent focuses on the model's anomaly detection capabilities, it fails to provide valuable insights into the source of the anomaly. Once an anomaly occurs, accurate identification and detection are crucial, as is further locating the specific parameter causing the abnormal behavior to support subsequent operational and maintenance decisions.

[0005] After comprehensive consideration, the present invention proposes to adopt a new architectural design. By constructing a reconstruction model based on DCANNs, multiple one-dimensional convolutions (Conv1D) and attention layers are used to extract local patterns and key features to achieve accurate data reconstruction. At the same time, a dynamic threshold calculation method based on extreme value theory (EVT) is combined for adaptive anomaly detection, and abnormal parameters are located with the help of residual statistical features. This method is expected to overcome the shortcomings of existing technologies. With its unique structural advantages and dynamic threshold mechanism, it can effectively improve the accuracy, reliability and adaptability of anomaly detection and positioning in multivariate flight data, and effectively ensure the safe and stable operation of aircraft. Summary of the Invention

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

[0008] A multivariate flight data anomaly detection and location method based on a reconstruction model, characterized by comprising the following steps:

[0009] S01 Data Processing Stage: Collect UAV flight data and perform normalization preprocessing on the raw data; divide the raw data into training and test sets 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 the normal data.

[0010] S02 Model Construction and Training Phase: Construct a 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 an autoencoder. The model is trained using the training set data to minimize the reconstruction error of the autoencoder.

[0011] S03 Anomaly Detection Phase: Calculate the reconstruction residuals of the training set and test set, and use the training set reconstruction residuals 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;

[0012] S04 Anomaly Location Stage: Based on the test set residuals, 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.

[0013] The principle of this solution is:

[0014] In the data preprocessing step, the original flight data are normalized and mapped to the same numerical range, eliminating the dimensional effects of different variables.

[0015] During the model construction step, the DCANNs-based neural network reconstruction model fully utilizes the characteristics of different modules. The Conv1D module extracts local patterns through a sliding window, capturing local features in the flight data. The attention mechanism module dynamically assigns weights to capture global dependencies, enabling the model to focus on important data features. The autoencoder module compresses and decodes data to learn the intrinsic representation of the data, minimizing reconstruction error to obtain an effective data representation.

[0016] 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.

[0017] In the anomaly location step, the statistical characteristics of the residuals are calculated to locate abnormal parameters. The residuals are calculated based on the difference between the original data and the reconstructed data. The residual sum of each variable is then calculated as the anomaly score. The higher the score, the more likely the variable is an abnormal parameter, thus achieving accurate anomaly location.

[0018] The beneficial effects of this program:

[0019] This model employs multiple one-dimensional convolutional layers to effectively extract local patterns from drone flight data. These convolutional layers improve feature extraction capabilities for complex flight data by capturing latent correlations within time series data. Furthermore, the model incorporates an attention mechanism that automatically identifies and focuses on the most important features or time segments, thereby enhancing its ability to learn key features. This mechanism enables the model to effectively focus on components that significantly impact drone flight data, avoiding information redundancy and further improving its ability to learn latent representations for multivariate flight data. In this way, this method not only reconstructs flight data more accurately but also adapts to drone data from diverse types and environments, demonstrating strong robustness and generalization capabilities. Consequently, this model demonstrates improved accuracy and stability when processing complex and variable drone flight data.

[0020] 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 calculates the residual by calculating the reconstruction error. To improve the accuracy of anomaly detection, a dynamic threshold generation mechanism is proposed. It can dynamically adjust the threshold based on the volatility and complexity of the data, thereby more accurately identifying potential anomalies. This method demonstrates strong robustness and flexibility in a variety of flight environments, helping to improve the safety and reliability of drone systems. In terms of anomaly localization, by analyzing the statistical characteristics of the residuals and locating the characteristic variables of the anomaly, the source of the anomaly can be determined more simply and efficiently, effectively reducing the time and cost of manual investigation.

[0021] The use of one-dimensional convolutional neural networks and the attention mechanism effectively reduces the computational complexity of the inference process. Through the one-dimensional convolutional layer, the model can quickly extract local patterns in 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 inference can be completed quickly when processing large-scale flight data. This enables this method to demonstrate a high response speed in real-time monitoring application scenarios and ensures the real-time performance of the anomaly detection method.

[0022] Furthermore, in S01, the formula for normalizing the original data x is:

[0023]

[0024] in is the normalized data, and are the maximum and minimum values ​​of the original data, and the normalized value is calculated for each variable;

[0025] 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%. The abnormal data in the test set is generated by the abnormal injection method. The injected abnormality types include deviation abnormality, drift abnormality and stuck abnormality. The abnormal injection formula is as follows:

[0026] Deviation anomaly:

[0027] in is a constant, is the original flight data, t is the time period;

[0028] Drift anomaly: ,in is the drift rate;

[0029] Stuck exception: ,in is a constant.

[0030] Furthermore, in the S02, in the local feature extraction module based on one-dimensional convolution, for the input data , output after Conv1D operation , No. The elements are , is the convolution kernel and , K is the length of the convolution kernel, yes No. elements, is the convolution kernel elements.

[0031] Furthermore, in the S02, in the key feature learning module based on the attention mechanism, given the input data X and query vector q, The weight is

[0032]

[0033] In the formula is the scoring function, soft is used to Converted to a probability distribution softmax function, yes Then, and Perform weighted summation to obtain the information that the model focuses on, expressed as Atten , as shown below:

[0034] .

[0035] Furthermore, in the data representation module based on the autoencoder in S02, the encoder of the autoencoder AE compresses the input data D into latent variables ,in are the encoder weights, is the encoder offset. is the activation function, the decoder will Decoding to reconstructed data ,in are the decoder weights, 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:

[0036]

[0037] Furthermore, in S03, the residual is reconstructed using the training set As anomaly score S Initialization threshold th , the part exceeding the threshold can be written as S-th Satisfies the generalized Pareto distribution, and its cumulative function

[0038]

[0039] in and σ They are the shape parameter and scale parameter of GPD, which can be obtained by maximizing the likelihood function

[0040] estimate and σ , is greater than th The total number of peaks, the final threshold ,

[0041]

[0042] in q represents the predefined extreme point quartiles, n Indicates the total number of samples. If v The variable i Anomaly score for each data point Greater than the threshold value of this point , then the point is considered abnormal, otherwise it is normal.

[0043] In S04, for the test set The first v variables c The residual is defined as In the formula 、 and They are Middle v The variable c test residuals, original values ​​and reconstructed values, v The residuals of the variables and

[0044] As the abnormality score of this variable, The higher the variable v The more likely it is an abnormal parameter.

[0045] Furthermore, the experimental data used in the above method includes simulated data and real drone flight data. The simulated data comes from the XTDrone simulation platform based on PX4, ROS and Gazebo, and the real data uses the ThorFlight98 dataset published by the Minnesota UAV Laboratory.

[0046] 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, and ALC and NLC represent the correct positioning accuracy of abnormal parameters and normal parameters, respectively. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 Adaptive anomaly detection and localization framework for UAV flight data based on neural network reconstruction model;

[0048] Figure 2 The DCANNs neural network autoencoder model designed for this purpose;

[0049] Figure 3 Schematic diagram of Conv1D operation;

[0050] Figure 4 It is a schematic diagram of AE structure;

[0051] Figure 5 For the simulated drone data and real drone data curves;

[0052] 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;

[0053] Figure 7Visualize the anomaly detection results of DCANNs-ADL for simulated and real UAV flight data;

[0054] Figure 8 Visualization of anomaly localization results for DCANNs-ADL, LSTM-AE, DAGMM, TranAD, and USAD. DETAILED DESCRIPTION

[0055] Example 1, a multivariate flight data anomaly detection and location method based on a reconstruction model, characterized by comprising the following steps:

[0056] S01 Data Processing Stage: Collect UAV flight data and perform normalization preprocessing on the raw data; divide the raw data into training and test sets 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 the normal data.

[0057] S02 Model Construction and Training Phase: Construct a 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 an autoencoder. The model is trained using the training set data to minimize the reconstruction error of the autoencoder.

[0058] S03 Anomaly Detection Phase: Calculate the reconstruction residuals of the training set and test set, and use the training set reconstruction residuals 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;

[0059] S04 Anomaly Location Stage: Based on the test set residuals, 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.

[0060] In the data preprocessing step, the original flight data are normalized and mapped to the same numerical range, eliminating the dimensional effects of different variables.

[0061] During the model construction step, the DCANNs-based neural network reconstruction model fully utilizes the characteristics of different modules. The Conv1D module extracts local patterns through a sliding window, capturing local features in the flight data. The attention mechanism module dynamically assigns weights to capture global dependencies, enabling the model to focus on important data features. The autoencoder module compresses and decodes data to learn the intrinsic representation of the data, minimizing reconstruction error to obtain an effective data representation.

[0062] In the adaptive anomaly detection step, the dynamic threshold calculation method based on extreme value theory (EVT) can dynamically calculate the anomaly detection threshold based on the reconstructed residuals of the training set and the test set, thereby improving the anomaly detection performance.

[0063] In the anomaly location step, the statistical characteristics of the residuals are calculated to locate abnormal parameters. The residuals are calculated based on the difference between the original data and the reconstructed data. The residual sum of each variable is then calculated as the anomaly score. The higher the score, the more likely the variable is an abnormal parameter, thus achieving accurate anomaly location.

[0064] This model employs multiple one-dimensional convolutional layers to effectively extract local patterns from drone flight data. These convolutional layers improve feature extraction capabilities for complex flight data by capturing latent correlations within time series data. Furthermore, the model incorporates an attention mechanism that automatically identifies and focuses on the most important features or time segments, thereby enhancing its ability to learn key features. This mechanism enables the model to effectively focus on components that significantly impact drone flight data, avoiding information redundancy and further improving its ability to learn latent representations for multivariate flight data. In this way, this method not only reconstructs flight data more accurately but also adapts to drone data from diverse types and environments, demonstrating strong robustness and generalization capabilities. Consequently, this model demonstrates improved accuracy and stability when processing complex and variable drone flight data.

[0065] 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 calculates the residual by calculating the reconstruction error. To improve the accuracy of anomaly detection, a dynamic threshold generation mechanism is proposed. It can dynamically adjust the threshold based on the volatility and complexity of the data, thereby more accurately identifying potential anomalies. This method demonstrates strong robustness and flexibility in a variety of flight environments, helping to improve the safety and reliability of drone systems. In terms of anomaly localization, by analyzing the statistical characteristics of the residuals and locating the characteristic variables of the anomaly, the source of the anomaly can be determined more simply and efficiently, effectively reducing the time and cost of manual investigation.

[0066] The use of one-dimensional convolutional neural networks and the attention mechanism effectively reduces the computational complexity of the inference process. Through the one-dimensional convolutional layer, the model can quickly extract local patterns in 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 inference can be completed quickly when processing large-scale flight data. This enables this method to demonstrate a high response speed in real-time monitoring application scenarios and ensures the real-time performance of the anomaly detection method.

[0067] 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, and the normalized value is calculated for each variable;

[0068] Different variables in drone flight data, such as velocity, acceleration, and attitude angle, often have different dimensions and orders of magnitude. Separate normalization can eliminate these differences, allowing different variables to be compared and analyzed on the same scale. For example, velocity might be expressed in meters per second, while acceleration might be expressed in meters per second squared. Without separate normalization, small changes in the acceleration data can be masked by the larger values ​​in the velocity data, making it difficult for the model to accurately capture the characteristics and patterns of the different variables during learning. Separate normalization allows the model to treat each variable equally, improving its ability to handle multivariate data. Calculating normalized values ​​separately can better preserve these characteristics, allowing the model to more accurately capture normal and abnormal patterns in each variable during learning. For example, if one variable has a small normal range, abnormal changes may cause large fluctuations in its value; while for another variable with a large normal range, smaller changes in its value may indicate an anomaly. Separate normalization allows the model to learn based on the characteristics of each variable, more accurately detecting and localizing anomalies, and thus improving the model's accuracy in anomaly detection and localization tasks. Drone flight environments are complex and changeable, and flight data is susceptible to noise. Separately normalizing each variable can mitigate 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.

[0069] 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 the abnormal injection method. The injected abnormality types include deviation abnormality, drift abnormality and stuck abnormality. The abnormal injection formula is as follows:

[0070] Deviation anomaly:

[0071] in is a constant, is the original flight data, t is the time period;

[0072] Drift anomaly: ,in is the drift rate;

[0073] Stuck exception: ,in is a constant.

[0074] Drones may encounter various abnormal conditions during actual flight. Deviation anomalies simulate situations where data deviates from the normal range due to momentary sensor deviation or sudden system failure, such as sensor deviation caused by electromagnetic interference. Drift anomalies correspond to data trend anomalies caused by gradual system aging and slow changes in component performance, such as the gradual change of certain parameters due to a declining battery charge. Stuck anomalies reflect sudden component failure that causes data to become fixed, such as a sensor that stops updating after a malfunction. These three anomaly types can fully simulate problems that may occur in real flight, providing realistic anomaly data for model training, making the trained model more practical in real-world applications. Furthermore, different types of anomalies have significant differences in data characteristics and patterns of change. Deviation anomalies are discrete, sudden changes, drift anomalies are continuous, gradual changes, and stuck anomalies are fixed. By learning these different anomaly data characteristics, the model can enhance its adaptability and recognition capabilities for various anomaly situations, enabling more accurate detection and location of anomalies even in the face of complex and diverse real-world anomalies.

[0075] In the S02, in the local feature extraction module based on one-dimensional convolution, for the input data , output after Conv1D operation , No. The elements are , is the convolution kernel and , K is the length of the convolution kernel, yes No. elements, is the convolution kernel UAV flight data is multidimensional data with time series characteristics, containing a wealth of local correlation information. Through this convolution operation, Conv1D can slide a fixed-size convolution kernel over the data sequence, focusing each operation on a local window of the data. For example, when processing time-varying data such as flight speed and acceleration, the convolution kernel can capture patterns in these parameters over a short period of time, such as sharp changes in speed or fluctuations in acceleration. This effective capture of local patterns helps extract key information reflecting the UAV's flight status, providing important evidence for subsequent anomaly detection and localization. Furthermore, its weight sharing and fixed parameter count make the model highly adaptable to input data of varying lengths, reducing the risk of overfitting. Regardless of the length of the input data, Conv1D consistently extracts local features, thereby improving the model's generalization. For example, in real-world applications, the model can effectively process and accurately extract features from data generated by different flight missions, even though the length may vary. This ensures stable performance across multivariate flight data in various flight scenarios, enabling more reliable anomaly detection and localization.

[0076] In S02, in the key feature learning module based on the attention mechanism, given the input data and query vector q , The weight is

[0077]

[0078] In the formula is the scoring function, soft is used to Converted to a probability distribution softmax function, 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:

[0079]

[0080] UAV flight data contains multiple parameters, and in different flight states, the importance of each parameter in judging the flight situation is different. and softmaxThe attention mechanism automatically weights each input data point. For example, during a drone's takeoff and landing phases, attitude angle data is crucial for determining flight safety. The attention mechanism assigns higher weights to attitude angle-related data, allowing the model to focus on this critical information, thereby accurately capturing key features related to flight anomalies and improving the accuracy of anomaly detection and localization. The drone flight environment is complex and ever-changing, and the important features of flight data change dynamically at different times. The attention mechanism can adjust weights in real time, flexibly allocating attention based on the current flight conditions. When encountering airflow disturbances, the importance of data such as speed and acceleration changes. The attention mechanism can promptly capture these changes and reallocate attention, ensuring that the model always focuses on the most representative features. This enhances the model's adaptability to complex and dynamic environments and effectively handles anomaly detection tasks in various flight scenarios.

[0081] In the S02, in the data representation module based on the autoencoder, the encoder of the autoencoder AE compresses the input data D into latent variables z ,in are the encoder weights, is the encoder offset. is the activation function, the decoder will z Decoding to reconstructed data ,in are the decoder weights, 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:

[0082]

[0083] Since the training goal of the autoencoder is to make the original data D The reconstruction error between the original data and the reconstructed data is as small as possible. 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 collection 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 better reflects the real flight status of the drone, improve the data quality, and provide a more reliable data basis for anomaly detection and positioning. By calculating the original data DThe loss function is constructed based on the reconstruction error between the original and reconstructed data, providing an effective basis for anomaly detection. Under normal flight conditions, the autoencoder accurately reconstructs the data with a small reconstruction error. However, when the data is abnormal, the reconstruction error increases significantly. Based on this characteristic, by setting an appropriate threshold, the data is considered abnormal when the reconstruction error exceeds the threshold.

[0084] In S03, the residual is reconstructed using the training set As anomaly score S Initialization threshold th , the part exceeding the threshold can be written as S-th Satisfies the generalized Pareto distribution, and its cumulative function

[0085]

[0086] in and σ They are the shape parameter and scale parameter of GPD, which can be obtained by maximizing the likelihood function

[0087] estimate and σ , is greater than th The total number of peaks, the final threshold ,

[0088]

[0089] in q represents the predefined extreme point quartiles, n Indicates the total number of samples. If v The variable i Anomaly score for each data point Greater than the threshold value of this point , then the point is considered abnormal, otherwise it is normal.

[0090] UAV flight data changes dynamically with the flight environment and mission. Traditional fixed thresholds cannot adapt to such changes, which can easily lead to high false alarm rates. This method uses the training set to reconstruct the residual initialization threshold th , and estimate the shape parameters based on GPD γ and scale parameters σ To dynamically adjust the final threshold When flight data fluctuates or changes in complexity, γ and σ change accordingly, allowing the threshold to reflect the real-time characteristics of the data. For example, in a complex electromagnetic environment, flight data fluctuations increase, and the threshold will be adjusted accordingly to ensure the accuracy of anomaly detection and reduce false positives and missed reports.

[0091] In S04, for the test set middle v The variable c The residual is defined as

[0092]

[0093] In the formula 、 and They are Middle v The variable c test residuals, original values, and reconstructed values, v The residuals of the variables and

[0094] As the abnormality score of this variable, The higher the variable v The more likely it is an abnormal parameter.

[0095] By calculating the residual sum as an anomaly score, this method simply and intuitively quantifies the degree of deviation of each variable from its normal state. When a variable is normal, its reconstructed value is very close to the original value, resulting in a small residual and a correspondingly small residual sum. Conversely, if a variable is abnormal, the difference between the reconstructed value and the original value increases, and the residual sum also increases. Taking the acceleration variable in drone flight data as an example, during normal flight, the reconstructed value of the acceleration data differs very little from the original value, and the residual sum is close to zero. However, when the accelerometer malfunctions or the aircraft is subjected to abnormal external forces, the deviation between the reconstructed value and the original value increases, and the residual sum rises rapidly, making the degree of anomaly clear at a glance and facilitating the rapid identification of abnormal variables. The residual sum, as an anomaly score, can accurately indicate which variable or variables are abnormal. The residual sums of different variables are calculated independently. Even in complex situations where multiple variables are abnormal simultaneously, the variable with the greatest likelihood of anomaly can be quickly identified by comparing the residual sums 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.

[0096] The experimental data used in the above method includes simulated data and real drone flight data. The simulated data comes from the XTDrone simulation platform based on PX4, ROS and Gazebo, and the real data uses the ThorFlight98 dataset 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 to obtain multiple sets of data for model training and evaluation, ensuring the reliability and repeatability of the experimental results.

[0097] In S03, accuracy (Acc), true positive rate (TPR), recall, precision, false positive rate (FPR), and F1 score are used as evaluation metrics. In the anomaly localization task, ALC and NLC are used as evaluation metrics. Accuracy (Acc) intuitively reflects the overall proportion of normal and anomalous samples correctly identified by the model, providing an overall overview of the model's performance. However, relying solely on accuracy can be misleading due to issues such as data imbalance. The true positive rate (TPR, equivalent to recall) focuses on the proportion of anomalous samples that are correctly detected, reflecting the model's ability to capture anomalous samples. Precision measures the proportion of truly anomalous samples identified as anomalous by the model, preventing the model from misclassifying too many normal samples as anomalous. The false positive rate (FPR) indicates the proportion of normal samples incorrectly labeled as anomalous. It complements precision and evaluates the model from the perspective of misclassification. The F1 score combines precision and recall, balancing their impact and providing a more comprehensive and objective evaluation of the model's performance in anomaly detection tasks. For example, in drone flight data anomaly detection, if a model has a high Acc but also a high FPR, this indicates that while the overall correct judgment ratio is high, there are many misjudgments. Combining these metrics can clearly identify model issues. In anomaly localization tasks, ALC directly measures the model's ability to accurately identify anomalous parameters, while NLC reflects the model's accuracy in distinguishing between normal and abnormal parameters. These two metrics can accurately assess the model's effectiveness in locating anomalies in multivariate flight data. In complex flight data, multiple parameters may be anomalies. ALC can intuitively reflect the model's probability of accurately identifying anomalous parameters, while NLC can reflect the model's misjudgment of normal parameters, helping to determine the reliability of the model's localization.

[0098] Example 2:

[0099] 1 Methodological Framework

[0100] like Figure 1 To achieve adaptive anomaly detection and localization in multivariate UAV flight data, we propose the DCANNs-ADL framework (see figure). This framework primarily consists of three steps: data preparation, model training and testing, and anomaly detection and localization. In the data preparation step, UAV flight data is collected and preprocessed. In the model training and testing step, the DCANNs autoencoder model is constructed by leveraging the advantages of one-dimensional convolutional neural networks and attention mechanisms in extracting local patterns and global features, as well as the advantages of autoencoders in data representation. The model can fully extract and learn key data features, thereby achieving accurate data reconstruction. In the anomaly detection and localization step, a dynamic threshold strategy is adopted to reduce the false alarm rate and improve the model's anomaly detection performance in complex dynamic environments. Furthermore, residual statistical features are used to characterize the degree of flight parameter anomalies, enabling rapid and accurate localization of abnormal parameters.

[0101] 1.1 Neural Network Reconstruction Model Architecture

[0102] Figure 2 The designed neural network model is presented. The model first utilizes a one-dimensional convolutional layer to capture the local features of drone flight data. These local features reflect key flight and environmental information collected by the drone during flight, such as velocity, acceleration, and attitude angle. Through the convolutional layer, the model can effectively extract local patterns in time series data, revealing potential correlations and short-term trends in the data. At the same time, the model also introduces an embedded attention mechanism to capture global dependencies in the data. By assigning dynamic weights to each sampling point, the attention layer enables the model to focus on the most representative features in the flight data, thereby better capturing global patterns in long time series. This allows the model to not only process local features but also consider longer-range dependencies in the time series, improving its understanding of complex drone flight data.

[0103] The model also utilizes autoencoders to extract and learn key features of the data. This process not only compresses the high-dimensional representation of the data but also effectively filters noise from the input data, minimizing its impact on the reconstruction results. This noise reduction process enables the model to more accurately reconstruct flight data, enhancing fidelity to the original signal. Ultimately, by combining one-dimensional convolution, an attention mechanism, and autoencoder technology, the model demonstrates superior capabilities in flight data reconstruction, achieving high-precision data reconstruction.

[0104] Local feature extraction module based on one-dimensional convolution

[0105] Figure 3As shown in Figure 1, one-dimensional convolution (Conv1D) is used to capture local patterns in time series data, effectively extracting locally relevant features using a sliding window. Its weight-sharing mechanism between convolution kernels ensures computational efficiency, resulting in a fixed parameter count independent of input length. This lightweight structure minimizes the risk of over-coupling and ensures scalability across varying input lengths.

[0106] Given input data , after Conv1D operation, the output can be obtained , as shown below:

[0107]

[0108] In the formula yes No. elements, is the convolution kernel and , K is the length of the convolution kernel, yes No. elements, is the convolution kernel k elements.

[0109] (2) Key feature learning module based on attention mechanism

[0110] The attention mechanism extracts information from the input data, enabling the model to assign attention weights at different time steps or spatial positions, thereby processing the data more efficiently. and query vector q , The weight can be expressed as follows:

[0111]

[0112] In the formula S（∙） is the scoring function, soft is used to S（∙） Converted to a probability distribution softmax function, 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:

[0113]

[0114] (3) Data representation module based on autoencoder

[0115] like Figure 4As 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.

[0116] Given input data D , which is first compressed by the encoder into a latent variable z, which can be expressed as follows:

[0117]

[0118] In the formula are the encoder weights, is the encoder offset. f（∙） is the activation function. Then, the decoder will z Decode back to the original data space to obtain the reconstructed data , which can be expressed as follows:

[0119]

[0120] In the formula are the decoder weights, 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:

[0121]

[0122] In the formula loss is the loss function, Represents the square of the Euclidean norm of a vector.

[0123] 1.2 Adaptive Anomaly Detection

[0124] During model training and testing, the reconstruction residuals of the training set and test set can be calculated by the following formula:

[0125]

[0126]

[0127] 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 streampeak-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. The determination of 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

[0128]

[0129] In the formula th is the initial threshold for retrieving the peak anomaly score, which is used As anomaly score Initialize. and σ are the shape parameter and scale parameter of GPD, respectively, and ε represents the amplitude exceeding the threshold th, which can be estimated by maximizing the likelihood function as follows:

[0130]

[0131] In the formula is greater than th Then, we can get the following formula: and the estimated value of σ and estimated value.

[0132]

[0133] in q represents the predefined extreme point quartiles, n Represents the total number of samples. ∆ is the final threshold With the initial threshold th The difference between Rewriting the formula, we can get The expression is defined as follows:

[0134]

[0135] Finally, by and anomaly score S Specifically, if the first v The variable i Anomaly score for each data point Greater than the threshold value of this point , then the point is marked as 1, that is, abnormal; vice versa, it is marked as 0, as shown in the following formula:

[0136]

[0137] 1.3 Abnormal Location

[0138] In practical applications, flight data is generated in the form of data streams, which makes it 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. middle v The variable c The residuals are defined as follows:

[0139] In the formula 、 and They are Middle v The variable c test residuals, original values, and reconstructed values. , No. v The residual sum of the variables can be calculated according to the following formula:

[0140]

[0141] In the formula It is v The residual sum of the variables is used as the abnormality score of the variable. The higher the variable v The more likely it is an abnormal parameter, and vice versa.

[0142] 2 Experimental data description

[0143] Figure 5 As shown in Figure 1, the experimental data in this paper includes simulated data and real drone 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 simulated data has 22,340 sampling points for each parameter, with a sampling frequency of 10 Hz. The real data has 21,015 sampling points for each parameter, with a sampling frequency of 50 Hz. 12,000 sampling points were selected as experimental data, which include 9 flight parameters from magnetometers, accelerometers and gyroscopes. Table 1 and Figure 5 Descriptions and original curves of these parameters are shown separately.

[0144] Table 1 Description of experimental parameters

[0145]

[0146] 3 Data partitioning and exception injection

[0147] Figure 6 As shown, before model training, the original input needs to be x Divide it into training set and test set to train and evaluate the model. x Divided into 7:3 ratio and ,in Contains only normal data, Contains abnormal data. Considering the difficulty and risk of obtaining real abnormal data, anomaly injection is used to generate abnormal data. To simulate the abnormal conditions in drone flight data as much as possible, different types of anomalies are injected, including deviation anomalies, drift anomalies, and stuck anomalies, as shown in the following formula:

[0148]

[0149]

[0150]

[0151] Where α and β are constants, is the original flight data, φ is the drift rate, t For 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.

[0152] 4 Data Preprocessing

[0153] For the original data x , use the following formula to preprocess it:

[0154]

[0155] in is the normalized data, and They are x The maximum and minimum values ​​of v Normalized data of variables It can be calculated according to the following formula:

[0156]

[0157] In the formula and They are The maximum and minimum values ​​of .

[0158] 5 evaluation indicators

[0159] For anomaly detection tasks, accuracy (Acc), true positive rate (TPR), recall rate (recall), precision rate (precision), false positive rate (FPR), and F1 score are used as evaluation metrics, as shown in the following formula. The larger the values ​​of Acc, TPR (recall rate), precision, and F1 score, and the smaller the value of FPR, the better the model performance.

[0160]

[0161]

[0162]

[0163]

[0164]

[0165] Where 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:

[0166]

[0167]

[0168] In the formula ALC and NLC They represent the correct positioning accuracy of abnormal parameters and normal parameters respectively.

[0169] 6 Experimental results

[0170] 6.1 Comparison with Baseline Methods

[0171] To comprehensively evaluate the effectiveness of the proposed method, we used leading multi-source anomaly detection models in the same and related but different domains, including LSTM-AE, TranAD, DAGMM, and USAD, as baselines for comparison. 1) LSTM-AE is a hybrid neural network method based on LSTM and AE. This method leverages LSTM and AE to capture the spatiotemporal characteristics of multivariate UAV flight data and uses the reconstruction error as an anomaly score. 2) TranAD is a Transformer-based multivariate anomaly detection method. It uses the Transformer's multi-head attention mechanism to capture the global characteristics of multivariate data, enabling fast inference. 3) DAGMM combines a Gaussian mixture model (GMM) with a deep autoencoder to generate a low-dimensional representation. This method performs joint optimization in an end-to-end manner, while using a separate estimation network to facilitate learning of the hybrid model. 4) USAD is an anomaly detection method that utilizes two autoencoders and adversarial training to reconstruct the data. This method uses the reconstruction error of the two autoencoders to calculate the anomaly score.

[0172] 6.2 Anomaly Detection

[0173] Figure 7 Figure 2 shows the visual anomaly detection results achieved by DCANNs-ADL on simulated and real UAV flight data. DCANNs-ADL can clearly distinguish between normal and abnormal patterns and effectively identify anomalous 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 metrics are marked in bold. DCANNs-ADL performs best on most metrics for both simulated and real UAV flight data. In particular, it accurately detects all anomalous data, as shown by an FPR of 0.00000 and a Precision of 1.00000. Furthermore, DCANNs-ADL achieves the highest Acc and F1 scores of 0.99481 and 0.99728, and 0.99623 and 0.99803, respectively, on simulated and real UAV flight data. Taking all metrics into account, 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.

[0174] LSTM-AE achieved average TPR (Recall), Accuracy, Precision, and F1 values ​​exceeding 0.97000 on both simulated and real drone flight data. However, its average FPR was higher, at 0.19607. This is primarily because LSTM-AE utilizes LSTM to capture features, neglecting the learning of global information. Furthermore, while selectively retaining historical information, LSTM-AE may lose some useful information. TranAD uses an attention mechanism to learn dependencies between different time series, thereby more comprehensively capturing the temporal trends in drone time series data. Although TranAD's average FPR of 0.27857 is relatively high, its overall performance is superior to LSTM-AE. In particular, TranAD achieved the highest average TPR (Recall) of 0.99613. However, TranAD may amplify noise during inference. Therefore, TranAD's overall performance is inferior to DCANNs-ADL. The average TPR (Recall), FPR, Accuracy, Precision, and F1 scores of DAGMM on simulated and real UAV flight data were 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 DAGMM's input data consists of only single observations, not historical time series windows. However, temporal dependencies are crucial for anomaly detection in multivariate time series. Therefore, DAGMM is not suitable for modeling temporal dependencies. With appropriate regularization, USAD can fully leverage the advantages of AE and achieve good overall performance. However, like TranAD, USAD also amplifies noise during inference. Therefore, its overall performance is inferior to that of DCANNs-ADL.

[0175] Table 2. Anomaly detection results of DCANNs-ADL and baseline methods.

[0176]

[0177] Table 3 lists the time spent per epoch during model training for these methods. DCANNs-ADL achieved the shortest single-shot training time on real and simulated UAV flight data, at 0.18413 seconds and 0.18490 seconds, respectively. This is due to the use of one-dimensional convolutional layers and attention layers in DCANNs-ADL. The one-dimensional convolutional layers effectively reduce the number of parameter layers through weight sharing and local connections, thereby improving the model's computational efficiency. The attention layers selectively focus on learning key features, further reducing unnecessary computational overhead. Specifically, compared to LSTM-AE, TranAD, DAGMM, and USAD, DCANNs-ADL achieved an average training time of only 0.18451 seconds per epoch, representing reductions of 92.99%, 97.31%, 99.12%, and 99.86%, respectively. These results further validate the proposed method's superior computational efficiency.

[0178] Table 3. Training time (in seconds) per epoch for the above methods.

[0179]

[0180] 6.3 Abnormal Location

[0181] 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 of the test set, where x The axis represents the order of multivariable flight parameters, y The axis represents the reconstruction residual of the test set. The right sub-figure shows the histogram of the anomaly localization results, where x The axes represent multiple variable flight parameters. y The axis represents the sum of residuals. As mentioned above, the larger the v The higher the probability that the variable corresponds to the abnormal parameter, the higher the probability that the abnormal parameter is located incorrectly. "×" means that the abnormal parameter is located correctly, and "√" means that the abnormal parameter is located correctly. 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.

[0182] Table 4 lists the anomaly localization results of the aforementioned methods. DCANNs-ADL achieved the best anomaly localization results, with ALC and NLC values ​​of 1.000 and 1.000, and 0.800 and 0.750, respectively, for real and simulated UAV flight data. TranAD achieved the second best results, with ALC and NLC values ​​of 0.750 and 0.800, and 0.800 and 0.750, respectively, for real and simulated UAV flight data. This is primarily because DCANNs-ADL and TranAD employ an attention mechanism to capture the global dependencies of the input data, thereby better distinguishing and localizing anomalous and normal parameters. In comparison, LSTM-AE, DAGMM, and USAD performed relatively poorly overall. Compared with LSTM-AE, TranAD, DAGMM, and USAD, DCANNs-ADL improves the average ALC and NLC by 33.33% and 34.62%, 16.13% and 12.90%, 33.33% and 34.62%, and 33.33% and 34.62%, respectively. Experimental results show that DCANNs-ADL has excellent consistency in anomaly detection and localization.

[0183] Table 4. Anomaly localization results of DCANNs-ADL and baseline methods.

[0184]

[0185] The above are only embodiments of the present invention. Common knowledge such as the known specific structures and characteristics in the scheme are not described in detail here. Ordinary technicians in the relevant field are aware of all the common technical knowledge in the technical field of the invention 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 inspiration 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 variations and improvements can be made without departing from the structure of the present invention. These should also be regarded as the scope of protection of the present invention. 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 and test sets 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 the normal data. S02 Model Construction and Training Phase: Construct a 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 an autoencoder. The model is trained using the training set data to minimize the reconstruction error of the autoencoder. In the S02, in the local feature extraction module based on one-dimensional convolution, for the input data , output after Conv1D operation , No. The elements are , is the convolution kernel and , K is the length of the convolution kernel, yes No. elements, is the convolution kernel elements; In S02, in the key feature learning module based on the attention mechanism, given the input data and query vector q , The weight is In the formula is a scoring function, soft is used to Converted to the softmax function of the probability distribution, yes The weight of and Perform weighted summation to obtain the information that the model focuses on, expressed as Atten , as shown below: ； In the said S02, in the said data representation module based on the autoencoder, the encoder of the autoencoder AE compresses the input data D into the latent variable z ,in are the encoder weights, is the encoder bias, Is the activation function, the decoder decodes z into reconstructed data ,in are the decoder weights, 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 is defined as follows: ； 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 residuals, 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; In S04, for the test set The first v variables c The residual is defined as In the formula 、 and They are Middle v The variable c test residuals, original values, and reconstructed values, v The residuals of the variables and As the abnormality score of this variable, The higher the variable v The more likely it is an abnormal parameter.

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 the abnormal injection method. The injected abnormality 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, t is the time period; Drift anomaly: ,in is the drift rate; 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 S03, the residual is reconstructed using the training set As anomaly score S Initialization threshold th , the part exceeding the threshold is written as S-th Satisfies the generalized Pareto distribution, and its cumulative function in and σ They are the shape parameter and scale parameter of GPD, which are obtained by maximizing the likelihood function estimate and σ , is greater than The total number of peaks, the final threshold , in q represents the predefined extreme point quartiles, n Indicates the total number of samples. If v The variable i Anomaly score for each data point Greater than the threshold value of this point , then the point is considered abnormal, otherwise it is normal.

Images