Segment-level unmanned aerial vehicle track anomaly detection method and system
By using trajectory prediction based on multiple coordinate coupling relationships and segment-level residual analysis, the problem of detecting slow drift anomalies in UAV tracks was solved, achieving high-precision identification of slow drift anomalies and improving the stability and reliability of the system.
Patent Information
- Application Number
- CN202610650858.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-12
- Publication Date
- 2026-08-25
AI Technical Summary
Existing methods for detecting drone flight anomalies are insufficient to effectively detect slow drift anomalies during low-altitude flight in urban areas, especially in areas with dense high-rise buildings, narrow streets and valleys, or severe obstruction leading to satellite signal interference. Existing methods struggle to capture the continuity and cumulative nature of slow drift anomalies.
Trajectory prediction is performed using a multi-coordinate coupling relationship. A reference trajectory is formed by a Transformer prediction network and an adaptive robust unscented Kalman filter correction module. Fragment-level residual analysis is then performed to explicitly model the error level, error energy, and growth trend, thereby improving the accuracy of anomaly detection.
It effectively improves the accuracy of UAV trajectory anomaly detection, can accurately identify slow drift anomalies, avoid false alarms and missed detections, and improve the stability and reliability of the system.
Smart Images
Figure CN122634427A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of unmanned aerial vehicles (UAVs), and more particularly to a method and system for detecting anomalies in UAV flight paths at the fragment level. Background Technology
[0002] In urban low-altitude flight missions, UAVs typically rely on GNSS, RTK, or integrated navigation systems to obtain location information. However, in areas with dense high-rise buildings, narrow streets and valleys, or severe obstruction, satellite signals are susceptible to multipath reflections, reduced visible satellite numbers, non-line-of-sight propagation, and environmental disturbances, leading to navigation observation deviations, drift, and instability.
[0003] Unlike sudden spikes, the aforementioned deviations often do not manifest as a sudden, large-scale shift, but rather accumulate gradually over a continuous period in a weaker form, eventually forming a smoother but continuously expanding track deviation. This type of anomaly is commonly referred to as a slow drift anomaly.
[0004] Existing technologies for detecting UAV trajectory anomalies mainly include threshold statistics, traditional machine learning, and deep learning. Threshold statistics typically set thresholds for position errors, velocity errors, filtering innovation, or prediction residuals; when the detected quantity exceeds the threshold at a certain moment, that moment is considered an anomaly. While simple to implement, this method usually detects only a single moment, making it difficult to characterize the continuity and cumulative nature of slow drift anomalies. Because slow drift amplitudes are weak in the early stages, it often mixes with fluctuations caused by normal maneuvers, leading to point-level thresholding methods easily missing early anomalies or generating false alarms due to short-term fluctuations. Traditional machine learning or deep learning algorithms generally use time-series prediction models such as LSTM, GRU, and Transformer to predict the future trajectory of the UAV, and then determine anomalies based on the deviation between the observed and predicted trajectories. These methods can incorporate historical context information, offering some improvement over static thresholding methods; however, many existing solutions still rely on single-point residuals for anomaly determination, lacking modeling of the overall rise, continuous growth, and structural stability of residuals over continuous time segments. For urban low-altitude slow drift anomalies, what is truly meaningful is not whether a certain point is too large, but whether a certain local time segment shows a continuous deviation and a gradual accumulation development pattern.
[0005] Therefore, in order to solve the above-mentioned technical problems, it is urgent to propose a new technical approach. Summary of the Invention
[0006] In view of this, the purpose of this invention is to provide a fragment-level UAV trajectory anomaly detection method and system. By using multi-coordinate coupling relationships for trajectory prediction, a reference trajectory is formed, which can fully consider the coupling relationship of each coordinate dimension. Furthermore, the anomaly detection is upgraded from single-point residual judgment to fragment-level residual analysis, explicitly modeling the error level, error energy, growth trend and persistence, thereby effectively improving the accuracy of anomaly detection and effectively avoiding the impact of slow drift.
[0007] This invention provides a fragment-level UAV trajectory anomaly detection method, comprising the following steps:
[0008] The position sequence of the UAV is obtained as the observation trajectory, and the UAV position sequence is sampled by a sliding window of length L to form an L×d-dimensional trajectory window matrix;
[0009] Transpose the L×d-dimensional trajectory window matrix to form a d×L-dimensional transpose matrix;
[0010] The d×L dimensional transpose matrix is input into the neural network prediction model to obtain the predicted trajectory of the UAV;
[0011] The trajectory residual sequence is formed by subtracting the predicted trajectory from the observed trajectory;
[0012] The trajectory residual sequence is fragmented, with a fragment length of W and a step size of S. The starting time of the m-th fragment is... m = 1, 2, ..., N;
[0013] Feature extraction is performed on each trajectory residual segment, and the extracted features are input into a shallow multilayer perceptron to generate an anomaly score.
[0014] The anomaly score is compared with the score threshold. If the anomaly score is greater than the score threshold, the observed trajectory corresponding to the current trajectory residual is anomaly.
[0015] Furthermore, in the feature extraction step for each trajectory residual segment, the extracted features include mean features, root mean square features, trend slope features, and duration features.
[0016] Wherein: the mean feature is to calculate the average value of the data within each trajectory residual segment;
[0017] The root mean square feature is to take the square root of the average of the squared values of the data within each trajectory residual segment;
[0018] The trend slope value is obtained by least-squares linear fitting of the data within each trajectory residual segment according to the time index.
[0019] The duration feature is the longest duration within each trajectory residual segment where the data is above the baseline level.
[0020] Furthermore, according to claim 1, the fragment-level UAV trajectory anomaly detection method is characterized in that: the neural network prediction model includes a Transformer prediction network and an adaptive robust unscented Kalman filter correction module;
[0021] The Transformer prediction network takes a d×L dimensional transpose matrix as input and outputs a prediction mean vector and a diagonal covariance matrix, where the prediction mean vector represents the center estimate of the location and the covariance matrix represents the prediction uncertainty.
[0022] The output features of the Transformer prediction network are input into the adaptive robust unscented Kalman filter correction module, which outputs the predicted trajectory.
[0023] Furthermore, the loss function of the Transformer prediction network is the Gaussian negative log-likelihood loss function.
[0024] Furthermore, after feature extraction for each trajectory residual segment, each extracted feature is normalized and then input into a shallow multilayer perceptron to generate anomaly scores.
[0025] Accordingly, the present invention provides a fragment-level UAV trajectory anomaly detection system, including a data input module, a window construction module, a variable token module, a trajectory prediction module, a residual construction module, a fragment segmentation module, a feature extraction module, a scoring module, and an anomaly judgment module;
[0026] The data input module is used to acquire the UAV position sequence as the observation trajectory;
[0027] The window construction module, whose input end is connected to the data input module, is used to sample the UAV position sequence using a sliding window of length L to form an L×d-dimensional trajectory window matrix;
[0028] The variable token module is used to transpose an L×d-dimensional trajectory window matrix into a d×L-dimensional transpose matrix;
[0029] The trajectory prediction module is used to receive the output features of the variable token module and output the predicted trajectory.
[0030] The residual construction module is used to receive the observed trajectory output by the data input module and the predicted trajectory output by the trajectory prediction module, and output the trajectory residual sequence.
[0031] The segmentation module is used to receive the trajectory residual sequence and form several trajectory residual segments according to a sliding window of length W and step size S.
[0032] The feature extraction module is used to perform feature extraction on each trajectory residual segment and then perform normalization processing;
[0033] The scoring module receives the normalized features from the feature extraction module and generates anomaly scores.
[0034] The anomaly detection module receives an anomaly score and compares the anomaly score with a scoring threshold. If the anomaly score is greater than the scoring threshold, then the observed trajectory corresponding to the current trajectory residual is anomaly.
[0035] Furthermore, the trajectory prediction module includes a Transformer prediction network and an adaptive robust unscented Kalman filter correction module;
[0036] The Transformer prediction network takes a d×L dimensional transpose matrix as input and outputs a prediction mean vector and a diagonal covariance matrix, where the prediction mean vector represents the center estimate of the location and the covariance matrix represents the prediction uncertainty.
[0037] The output features of the Transformer prediction network are input into the adaptive robust unscented Kalman filter correction module, which outputs the predicted trajectory.
[0038] Furthermore, the features extracted by the feature extraction module include:
[0039] Mean characteristics, root mean square characteristics, trend slope characteristics, and duration characteristics;
[0040] Wherein: the mean feature is to calculate the average value of the data within each trajectory residual segment;
[0041] The root mean square feature is to take the square root of the average of the squared values of the data within each trajectory residual segment;
[0042] The trend slope value is obtained by least-squares linear fitting of the data within each trajectory residual segment according to the time index.
[0043] The duration feature is the longest duration within each trajectory residual segment where the data is above the baseline level.
[0044] Furthermore, the loss function of the Transformer prediction network is the Gaussian negative log-likelihood loss function.
[0045] The beneficial effects of this invention are as follows: By using multiple coordinate coupling relationships for trajectory prediction, a reference trajectory is formed, which can fully consider the coupling relationship of each coordinate dimension. Furthermore, the anomaly detection is improved from single-point residual judgment to segment-level residual analysis, and the error level, error energy, growth trend and persistence are explicitly modeled, thereby effectively improving the accuracy of anomaly detection and effectively avoiding the impact of slow drift. Attached Figure Description
[0046] The present invention will be further described below with reference to the accompanying drawings and embodiments:
[0047] Figure 1 This is a schematic diagram of the process of the present invention.
[0048] Figure 2 This is a schematic diagram of the system structure of the present invention. Detailed Implementation
[0049] The present invention will be further described in detail below:
[0050] This invention provides a fragment-level UAV trajectory anomaly detection method, characterized by the following steps:
[0051] The UAV position sequence is obtained as the observation trajectory, and a sliding window of length L is used to sample the UAV position sequence to form an L×d-dimensional trajectory window matrix. The UAV position sequence is represented as P_t, where t = 1, 2, ..., T, and P_t is a d-dimensional position vector, where d equals 3, corresponding to the position components along the x-axis (East direction), y-axis (North direction), and z-axis, respectively. In this process, continuous UAV position sequences are converted into time window samples with a certain degree of overlap (i.e., the window length is greater than the sliding step size), and each sample retains the motion context information of the most recent L time moments.
[0052] The L×d-dimensional trajectory window matrix is transposed to form a d×L-dimensional transpose matrix. Through this process, the sequence of each coordinate dimension at L historical moments is used to form an independent token. That is to say, the East sequence, North sequence, and Up sequence are used as three variable tokens to input the model. Each token completely preserves the temporal evolution pattern of a certain coordinate dimension. On the other hand, the self-attention in the subsequent prediction model can directly establish a connection between dimensions, which is more suitable for modeling the dynamic coupling relationship between the coordinate components of the UAV.
[0053] The d×L dimensional transpose matrix is input into the neural network prediction model to obtain the predicted trajectory of the UAV;
[0054] The residual sequence is formed by subtracting the predicted trajectory from the observed trajectory. Under normal flight conditions or with good measurement quality, the residual usually fluctuates around a small range. However, in the case of slow drift anomalies, although the residual may not be significant at the beginning, it will show a gradual upward trend and continue to rise over a continuous period of time. Therefore, the residual sequence can clearly characterize the slow drift anomaly, thus ensuring the accuracy of UAV trajectory anomaly detection.
[0055] The trajectory residual sequence is fragmented, with a fragment length of W and a step size of S. The starting time of the m-th fragment is... , m=1,2,...,N; The fragmentation process is a sliding window of length W with step size S for fragmented detection. Through fragmentation, each detection unit is no longer the instantaneous error at a certain moment, but the error evolution mode in a certain continuous local interval. This makes the anomaly judgment based on the overall structure inside the local interval rather than a single peak value, which is more suitable for detecting smooth cumulative slow drift anomalies.
[0056] Feature extraction is performed on each trajectory residual segment, and the extracted features are input into a shallow multilayer perceptron to generate an anomaly score; the shallow multilayer perceptron is an existing network, which will not be described in detail here;
[0057] Anomaly scores are compared with scoring thresholds. If the anomaly score exceeds the threshold, the observed trajectory corresponding to the current trajectory residual is considered to have an anomaly. The scoring threshold for each trajectory residual segment can be determined based on actual working conditions. This method, employing multi-coordinate coupling relationships for trajectory prediction to form a reference trajectory, fully considers the coupling relationships across coordinate dimensions and elevates anomaly detection from single-point residual judgment to segment-level residual analysis. It explicitly models error levels, error energy, growth trends, and persistence, thereby effectively improving the accuracy of anomaly detection and mitigating the impact of slow drift.
[0058] In this embodiment, the features extracted in the feature extraction step for each trajectory residual segment include mean features, root mean square features, trend slope features, and duration features.
[0059] Among them: the mean feature is the average value of the data within each trajectory residual segment; this feature reflects the overall deviation level of the segment;
[0060] The root mean square feature is the square root of the average of the squared values of the data within each trajectory residual segment; this feature reflects the energy and intensity of the error and is more sensitive to larger deviations.
[0061] The trend slope value is obtained by least squares linear fitting of the data in each trajectory residual segment according to the time index; where least squares linear fitting is an existing algorithm, which will not be described in detail here. This feature describes whether the error has a tendency to continue to grow and is a key feature for identifying early drift anomalies.
[0062] The duration feature is the longest duration by which data within each trajectory residual segment exceeds a baseline level. This baseline level is a set value determined based on actual operating conditions. This feature reflects whether the abnormal deviation has a persistent characteristic. Through the above, the accuracy of the scoring can be effectively ensured.
[0063] Of course, to further improve accuracy, the following features can also be used as inputs to the shallow multilayer perceptron, such as the proportion of duration and the mean difference between the first and second halves of a segment.
[0064] In this embodiment, the fragment-level UAV trajectory anomaly detection method according to claim 1 is characterized in that: the neural network prediction model includes a Transformer prediction network and an adaptive robust unscented Kalman filter correction module;
[0065] The Transformer prediction network takes a d×L dimensional transpose matrix as input and outputs a predicted mean vector and a diagonal covariance matrix. The predicted mean vector represents the center estimate of the location, and the covariance matrix represents the prediction uncertainty. These two outputs provide richer prediction information and offer directly usable prior parameters for the subsequent adaptive robust unscented Kalman filter correction module. Furthermore, the Transformer prediction network provides confidence constraints for the filtering recursion of the adaptive robust unscented Kalman filter correction module, enhancing the coupling between prediction and filtering. The Transformer prediction network is a prior art technology, consisting of multiple stacked encoding modules. Each encoding module includes a multi-head self-attention layer, a feedforward network layer, and a normalization layer.
[0066] The multi-head self-attention sublayer is used to capture the interaction relationship between tokens with different coordinates, the feedforward network sublayer is used to further enhance the non-linear representation ability of each token, and the normalization sublayer is used to improve training stability.
[0067] The loss function of the Transformer prediction network is the Gaussian negative log-likelihood loss function. This loss function comprehensively considers the deviation between the predicted mean and the true position, as well as the fitting of the prediction covariance to the error distribution. It can simultaneously constrain the "accuracy of the predicted value" and the "reasonableness of the uncertainty estimation", making the diagonal covariance matrix output by the Transformer prediction network more physically meaningful and directly serving as the confidence basis for subsequent filtering priors, thereby enhancing the effect of integrated prediction and filtering modeling.
[0068] The output features of the Transformer prediction network are input into the adaptive robust unscented Kalman filter correction module, which outputs the predicted trajectory.
[0069] After filtering by the adaptive robust unscented Kalman filter correction module, a predicted trajectory that is aligned with the observations but is smoother, more stable, and more resistant to anomalies can be obtained. This predicted trajectory provides a basic reference for subsequent residual detection. The adaptive robust unscented Kalman filter correction module uses an existing adaptive robust unscented Kalman filter algorithm to process the data and generate the predicted trajectory; its specific details are not elaborated here.
[0070] In this embodiment, after feature extraction for each trajectory residual segment, each extracted feature is normalized and then input into a shallow multilayer perceptron to generate an anomaly score. The normalization process can be carried out using Z-SCORE standard normalization, scaling normalization based on maximum and minimum values, etc. After normalization, the large differences in the dimensions of different features are avoided, which could lead to a dominant role of a certain feature in the score.
[0071] Accordingly, the present invention provides a fragment-level UAV trajectory anomaly detection system, including a data input module, a window construction module, a variable token module, a trajectory prediction module, a residual construction module, a fragment segmentation module, a feature extraction module, a scoring module, and an anomaly judgment module;
[0072] The data input module is used to acquire the UAV position sequence as the observation trajectory;
[0073] The window construction module, whose input end is connected to the data input module, is used to sample the UAV position sequence using a sliding window of length L to form an L×d-dimensional trajectory window matrix;
[0074] The variable token module is used to transpose an L×d-dimensional trajectory window matrix into a d×L-dimensional transpose matrix;
[0075] The trajectory prediction module is used to receive the output features of the variable token module and output the predicted trajectory.
[0076] The residual construction module is used to receive the observed trajectory output by the data input module and the predicted trajectory output by the trajectory prediction module, and output the trajectory residual sequence.
[0077] The segmentation module is used to receive the trajectory residual sequence and form several trajectory residual segments according to a sliding window of length W and step size S.
[0078] The feature extraction module is used to perform feature extraction on each trajectory residual segment and then perform normalization processing;
[0079] The scoring module receives the normalized features from the feature extraction module and generates anomaly scores.
[0080] The anomaly detection module receives an anomaly score and compares the anomaly score with a scoring threshold. If the anomaly score is greater than the scoring threshold, then the observed trajectory corresponding to the current trajectory residual is anomaly.
[0081] Specifically: the trajectory prediction module includes a Transformer prediction network and an adaptive robust unscented Kalman filter correction module;
[0082] The Transformer prediction network takes a d×L dimensional transpose matrix as input and outputs a prediction mean vector and a diagonal covariance matrix, where the prediction mean vector represents the center estimate of the location and the covariance matrix represents the prediction uncertainty.
[0083] The output features of the Transformer prediction network are input into the adaptive robust unscented Kalman filter correction module, which outputs the predicted trajectory.
[0084] The features extracted by the feature extraction module include:
[0085] Mean characteristics, root mean square characteristics, trend slope characteristics, and duration characteristics;
[0086] Wherein: the mean feature is to calculate the average value of the data within each trajectory residual segment;
[0087] The root mean square feature is to take the square root of the average of the squared values of the data within each trajectory residual segment;
[0088] The trend slope value is obtained by least-squares linear fitting of the data within each trajectory residual segment according to the time index.
[0089] The duration feature is the longest duration within each trajectory residual segment where the data is above the baseline level.
[0090] The loss function of the Transformer prediction network is the Gaussian negative log-likelihood loss function.
[0091] The system described above uses multi-coordinate coupling relationships for trajectory prediction to form a reference trajectory. This fully considers the coupling relationships of each coordinate dimension and elevates anomaly detection from single-point residual judgment to segment-level residual analysis. It explicitly models the error level, error energy, growth trend, and persistence, thereby effectively improving the accuracy of anomaly detection and effectively avoiding the impact of slow drift.
[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A method for detecting anomalies in UAV flight paths at the fragment level, characterized in that: Includes the following steps: The position sequence of the UAV is obtained as the observation trajectory, and the UAV position sequence is sampled by a sliding window of length L to form an L×d-dimensional trajectory window matrix; Transpose the L×d-dimensional trajectory window matrix to form a d×L-dimensional transpose matrix; The d×L dimensional transpose matrix is input into the neural network prediction model to obtain the predicted trajectory of the UAV; The trajectory residual sequence is formed by subtracting the predicted trajectory from the observed trajectory; The trajectory residual sequence is fragmented, with a fragment length of W and a step size of S. The starting time of the m-th fragment is... m = 1, 2, ..., N; Feature extraction is performed on each trajectory residual segment, and the extracted features are input into a shallow multilayer perceptron to generate an anomaly score. The anomaly score is compared with the score threshold. If the anomaly score is greater than the score threshold, the observed trajectory corresponding to the current trajectory residual is anomaly.
2. The fragment-level UAV trajectory anomaly detection method according to claim 1, characterized in that: In the feature extraction step for each trajectory residual segment, the extracted features include mean features, root mean square features, trend slope features, and duration features. Wherein: the mean feature is to calculate the average value of the data within each trajectory residual segment; The root mean square feature is to take the square root of the average of the squared values of the data within each trajectory residual segment; The trend slope value is obtained by least-squares linear fitting of the data within each trajectory residual segment according to the time index. The duration feature is the longest duration within each trajectory residual segment where the data is above the baseline level.
3. The fragment-level UAV trajectory anomaly detection method according to claim 1, characterized in that: The fragment-level UAV trajectory anomaly detection method according to claim 1 is characterized in that: the neural network prediction model includes a Transformer prediction network and an adaptive robust unscented Kalman filter correction module; The Transformer prediction network takes a d×L dimensional transpose matrix as input and outputs a prediction mean vector and a diagonal covariance matrix, where the prediction mean vector represents the center estimate of the location and the covariance matrix represents the prediction uncertainty. The output features of the Transformer prediction network are input into the adaptive robust unscented Kalman filter correction module, which outputs the predicted trajectory.
4. The fragment-level UAV trajectory anomaly detection method according to claim 3, characterized in that: The loss function of the Transformer prediction network is the Gaussian negative log-likelihood loss function.
5. The fragment-level UAV trajectory anomaly detection method according to claim 1, characterized in that: After feature extraction for each trajectory residual segment, each extracted feature is normalized and then input into a shallow multilayer perceptron to generate an anomaly score.
6. A fragment-level UAV trajectory anomaly detection system, characterized in that: It includes a data input module, a window construction module, a variable token module, a trajectory prediction module, a residual construction module, a segmentation module, a feature extraction module, a scoring module, and an anomaly detection module; The data input module is used to acquire the UAV position sequence as the observation trajectory; The window construction module, whose input end is connected to the data input module, is used to sample the UAV position sequence using a sliding window of length L to form an L×d-dimensional trajectory window matrix; The variable token module is used to transpose an L×d-dimensional trajectory window matrix into a d×L-dimensional transpose matrix; The trajectory prediction module is used to receive the output features of the variable token module and output the predicted trajectory. The residual construction module is used to receive the observed trajectory output by the data input module and the predicted trajectory output by the trajectory prediction module, and output the trajectory residual sequence. The segmentation module is used to receive the trajectory residual sequence and form several trajectory residual segments according to a sliding window of length W and step size S. The feature extraction module is used to perform feature extraction on each trajectory residual segment and then perform normalization processing; The scoring module receives the normalized features from the feature extraction module and generates anomaly scores. The anomaly detection module receives an anomaly score and compares the anomaly score with a scoring threshold. If the anomaly score is greater than the scoring threshold, then the observed trajectory corresponding to the current trajectory residual is anomaly.
7. The fragment-level UAV trajectory anomaly detection system according to claim 6, characterized in that: The trajectory prediction module includes a Transformer prediction network and an adaptive robust unscented Kalman filter correction module; The Transformer prediction network takes a d×L dimensional transpose matrix as input and outputs a prediction mean vector and a diagonal covariance matrix, where the prediction mean vector represents the center estimate of the location and the covariance matrix represents the prediction uncertainty. The output features of the Transformer prediction network are input into the adaptive robust unscented Kalman filter correction module, which outputs the predicted trajectory.
8. The fragment-level UAV trajectory anomaly detection system according to claim 6, characterized in that: The features extracted by the feature extraction module include Mean characteristics, root mean square characteristics, trend slope characteristics, and duration characteristics; Wherein: the mean feature is to calculate the average value of the data within each trajectory residual segment; The root mean square feature is to take the square root of the average of the squared values of the data within each trajectory residual segment; The trend slope value is obtained by least-squares linear fitting of the data within each trajectory residual segment according to the time index. The duration feature is the longest duration within each trajectory residual segment where the data is above the baseline level.
9. The fragment-level UAV trajectory anomaly detection system according to claim 7, characterized in that: The loss function of the Transformer prediction network is the Gaussian negative log-likelihood loss function.