Unmanned aerial vehicle flight trajectory anomaly tracing method based on fusion clustering algorithm

Abstract

Disclosed in the present invention is an unmanned aerial vehicle flight trajectory anomaly tracing method based on a fusion clustering algorithm. A hierarchical clustering algorithm is applied to classify and organize complex and sensitive spatio-temporal data features to identify a potential anomaly point; then, weights of various data features obtained by using a Lasso regression algorithm are used to calculate the sum of products of the potential anomaly point and the weights of the various data features; a system detects an anomaly point in a flight trajectory of an unmanned aerial vehicle to ensure flight safety and stability; the anomaly point is confirmed; and the system instantly calls an improved KD tree algorithm to quickly notify nearest on-duty personnel to go to the site, thereby comprehensively inspecting and thoroughly verifying factors interfering with the normal flight of the unmanned aerial vehicle, quickly and accurately processing anomaly conditions of the flight of the unmanned aerial vehicle, ensuring the flight safety and data integrity of the unmanned aerial vehicle, and providing reliable guarantee and support for a flight system of the unmanned aerial vehicle.

Description

An unmanned aerial vehicle flight trajectory anomaly tracing method based on a fusion clustering algorithm TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned aerial vehicle anomaly detection, and particularly relates to an unmanned aerial vehicle flight trajectory anomaly tracing method based on a fusion clustering algorithm. BACKGROUND

[0002] With the rapid evolution and continuous development of new power systems, data service applications based on unmanned aerial vehicles have ushered in a thriving development momentum in power systems. This trend brings many opportunities, but also poses serious challenges. In particular, since the data transmission between unmanned terminals and power systems involves highly sensitive key data information, security issues in this field have become particularly prominent and important. At the same time, there are potential security risks in the uncontrollable behavior of unmanned aerial vehicles during use, which increases the challenge to the security of data transmission and processing. In order to ensure that unmanned aerial vehicles are controllable, auditable and traceable throughout the process under the driving of data services, it is urgent to conduct in-depth research on security monitoring and tracing technology for unmanned aerial vehicles. This research not only concerns the development of the technical level, but also involves the healthy development of the entire power system and unmanned aerial vehicle industry. Effective security monitoring and tracing technology will provide reliable protection for the sustainable development of unmanned aerial vehicle data applications in power systems, promoting the safe, stable and efficient operation of services. Therefore, research and technological innovation in this regard are of great significance to ensure the security and controllability of unmanned aerial vehicle data service applications.

[0003] According to data, fragmentary change trends are common abnormal phenomena in industrial systems. Trends are roughly divided into four forms: slow to fast, fast to slow, constant to change, and change to constant. Unmanned aerial vehicle trajectory change is a typical fragmentary change trend anomaly, and this fragmentary anomaly from constant to change usually shows a slow rise, slow decline or slow fluctuation in parameters. Limited by the control mechanism of unmanned aerial vehicles and actual application scenarios, when unmanned aerial vehicles are disturbed, their flight will slowly deviate from the established trajectory. In the case of abnormal changes in the segment, the data changes very slowly at the beginning, and the result is that the residual error between the predicted point of the single-point prediction model and the true point is much smaller than the threshold, thereby reducing the anomaly detection effect.

[0004] Chinese patent document CN112100137A discloses a "drone anomaly detection method based on multi-log collaborative analysis". The following steps are adopted: template extraction is performed in the drone flight log, the features of different events are extracted, and the text information of the drone flight log is represented in the form of a feature vector; PCA algorithm is used to perform dimensionality reduction operation on the obtained feature vector; k-medoids algorithm is used to classify events into a known cluster according to the shortest Euclidean metric; events are identified, and anomaly detection is performed on the events to determine whether they are abnormal events, if not, they are put back into the training data set for training in the learning stage, and if so, the attack model is determined, if not, the abnormal event is defined as a new attack model, and if so, the abnormal event is classified as a known attack model. The above technical solution cannot quickly and accurately handle the abnormal situation of the flight of the drone, and ensure the stability and safety of the drone inspection.

[0005] SUMMARY

[0006] The present application mainly solves the technical problem that the original technical solution cannot quickly and accurately handle the abnormal situation of the flight of the drone, and ensure the stability and safety of the drone inspection, and provides a drone flight trajectory anomaly tracing method based on a fusion clustering algorithm. The hierarchical clustering algorithm is used to classify and organize complex and sensitive spatio-temporal data features to identify potential abnormal points. Then, the weight of each data feature obtained by the Lasso regression algorithm is used to calculate the sum of the product of the potential abnormal points and the weight of each data feature. The system detects the abnormal points in the flight trajectory of the drone, ensures the flight safety and stability, and confirms the abnormal points. The system immediately calls the improved KD tree algorithm to quickly notify the nearest on-duty personnel to the scene, so as to fully patrol and thoroughly verify the factors interfering with the normal flight of the drone, quickly and accurately handle the abnormal situation of the flight of the drone, ensure the flight safety and data integrity of the drone, and provide reliable protection and support for the flight system of the drone.

[0007] The above technical problems of the present application are mainly solved by the following technical solution: the present application comprises the following steps:

[0008] S1 extracts the data of the drone and performs decryption and preprocessing;

[0009] S2 finds out the potential abnormal points of the flight trajectory of the drone by using the hierarchical clustering algorithm;

[0010] S3 determines the influence of each data feature on the normal flight trajectory of the drone, obtains the weight of each data feature by using the Lasso regression algorithm, calculates the sum of the product of the potential abnormal points and the weight of each data feature, and traces to obtain the exact trajectory abnormal point;

[0011] S4 arranges personnel to verify and investigate the abnormal points.

[0012] The hierarchical clustering algorithm is applied to classify and organize the complex sensitive spatio-temporal data features to identify potential abnormal points. Then, the sum of the product of the potential abnormal points and the weight of each data feature obtained by the Lasso regression algorithm is calculated. The system can accurately and effectively detect the abnormal points in the UAV flight trajectory, ensure flight safety and stability, and maintain power grid safety. Once the abnormal points are confirmed, the improved KD-tree algorithm is called to quickly notify the nearest on-duty personnel to the scene for comprehensive inspection and thorough verification of the factors interfering with the normal flight of the UAV.

[0013] As preferred, the step S1 extracts the data of the UAV, including the data features of the UAV flight trajectory. The trajectory data of the UAV mainly involves complex sensitive spatio-temporal data, including three-dimensional position information, roll angle, pitch angle, yaw angle, speed, and acceleration based on time series. The data processor is used to decrypt and preprocess the data from the UAV.

[0014] As preferred, the step S1 preprocessing includes removing irrelevant data, and then assuming that the three-dimensional space coordinates of the kth UAV at t and t+Δt in the ith flight are: k，i (x k，i (t),y k，i (t),z k，i (t)),P k，i (x k，i (t+Δt),y k，i (t+Δt),z k，i (t+Δt))

[0015] The Euclidean distance between them is calculated,

[0016] For time series data, the instantaneous speed and instantaneous acceleration are estimated by using difference, i.e. the instantaneous speed is

[0017] The acceleration is

[0018] The roll angle The pitch angle θki(t) and the yaw angle ψki(t) come from the decrypted data. As the input of the algorithm, the sample will inevitably affect the analysis results of the algorithm output, so good data cleaning and preprocessing work is crucial to improve the accuracy of the algorithm.

[0019] As preferred, the above multiple data features are standardized, and the normalization formula is:

[0020] The data feature of the kth UAV at the tth time of the ith flight is denoted as:

[0021] In order to scale the value range of different data features to similar scales, avoid the problem of unstable model training or slow convergence speed caused by too large value difference of different features, and cope with data of different dimensions, normalization processing should be performed on the data.

[0022] Preferably, the step S2 specifically comprises:

[0023] S2.1, for each single data feature of the multiple data sets of each UAV, a similarity measure matrix is calculated and denoted as:

[0024] G k,m , which is a symmetric matrix, i.e. The matrix element is defined as:

[0025] d in , represents the Euclidean distance between C k,m,i (t) and C k，m,n (t);

[0026] S2.2, initialization, each d in is processed as a separate cluster;

[0027] S2.3, single linkage clustering is adopted, the minimum distance is taken as the distance measure of the cluster, and the trajectory points with the closest distance are gradually merged to form a hierarchical clustering tree;

[0028] S2.4, the normal trajectories are finally merged into one cluster, and the remaining few isolated clusters are potential abnormal trajectories, a segmentation threshold is set to divide the clusters into normal clusters and potential abnormal clusters.

[0029] The UAV flight trajectory anomaly finder finds the potential abnormal points of the UAV flight trajectory, adopts a hierarchical clustering algorithm to calculate the similarity measure matrix between data points, finds the two most similar clusters according to the similarity measure matrix, merges the two most similar clusters into a new cluster, updates the similarity measure matrix to reflect the similarity between the new merged cluster and other clusters, repeatedly merges the most similar clusters until all data points reach a pre-set threshold, determines the final clustering result, and divides the clusters into normal clusters and potential abnormal clusters.

[0030] Preferably, the step S3 specifically comprises:

[0031] S3.1 For the kth unmanned aerial vehicle, if its data feature m is determined as potentially abnormal by the unmanned aerial vehicle trajectory anomaly finder in the ith flight, mark it as p(k, m, i) = 1, otherwise, mark it as p(k, m, i) = 0;

[0032] S3.2 Mark the weight of the data feature m as ω m , and obtain the weight of each data feature by using the Lasso regression algorithm;

[0033] S3.3 For the same time t, the different data features m of the same trajectory of the kth unmanned aerial vehicle, first perform an or operation on p(k, m, i) obtained by the different data features, if the or operation result is not 0, it means that there may be an abnormality in the flight trajectory of the kth unmanned aerial vehicle at time t;

[0034] S3.4 Comprehensive determination is made on the potential abnormal points at this time t

[0035] In the formula, κ o represents the determination threshold.

[0036] By using the Lasso regression algorithm by the abnormal behavior determinator, the influence of each data feature on the normal flight trajectory of the unmanned aerial vehicle is determined according to the data set of the normal flight trajectory of the unmanned aerial vehicle. According to the abnormality of the data features calculated by the unmanned aerial vehicle trajectory anomaly finder, the weight of each data feature is obtained by using the Lasso regression algorithm, and the sum of the product of the potential abnormal points and the weight of each data feature is calculated, so as to make a comprehensive determination and trace back to obtain the exact trajectory abnormal point.

[0037] As preferred, the step S3.2 specifically comprises: the normal patrol trajectory data of the unmanned aerial vehicle is divided into a training set and a test set in advance by using the Lasso regression model, and then the training set is used to train the Lasso regression model; the influence of each data feature on the normal flight trajectory of the unmanned aerial vehicle is determined by using the Lasso regression model, and the coefficients of the Lasso regression model reflect the influence degree of each data feature on the normal flight trajectory of the unmanned aerial vehicle. By observing the coefficients obtained by the model, the data feature with the greatest influence on the flight of the unmanned aerial vehicle and the data feature with small or even negligible influence are determined.

[0038] As preferred, the step S4 specifically comprises:

[0039] S4.1 Obtain the longitude and latitude (u x , u y ) of the kth unmanned aerial vehicle at time t;

[0040] S4.2 For the position data set of the on-duty personnel on the same day, the KD tree is processed according to the construction method of the KD tree, and the KD tree is recursively constructed;

[0041] S4.3 Starting from the root node of the KD tree, recursively traverse the KD tree;

[0042] S4.4 When backtracking to the root node, find the nearest neighbor point;

[0043] S4.5 Notify the on-duty personnel of the nearest neighbor point through real-time communication to verify and investigate the abnormal point.

[0044] The abnormal point given by the abnormal behavior determinator is issued to the system, and the system notifies the on-duty personnel closest to the abnormal point position to the scene to verify the abnormal reason and investigate whether there is a violation of deliberately interfering with the normal inspection of the unmanned aerial vehicle.

[0045] As preferred, the step S4.3 specifically comprises comparing the target point by using Manhattan distance, calculating the distance between the current on-duty personnel position and the unmanned aerial vehicle coordinates, recording the current nearest point and distance; determining the next subtree to traverse according to the position of the unmanned aerial vehicle coordinates; continuing to recursively traverse the subtree until the leaf node, and using the pruning strategy to stop searching the child nodes of the node when a node close enough to the target point is found, thereby improving the search efficiency.

[0046] As preferred, the step S4.3 further comprises backtracking, and in the backtracking process, checking whether the parent node and other subtree nodes have closer on-duty personnel positions, and updating the nearest neighbor point and distance.

[0047] The present application has the beneficial effect of mainly solving the traceability problem of abnormal points in the flight trajectory of the unmanned aerial vehicle during the inspection of the power grid line. By using the method proposed in the present application, the sensitive space-time information of the unmanned aerial vehicle can be obtained according to the decrypted unmanned aerial vehicle inspection data set, the abnormal points of the unmanned aerial vehicle trajectory can be found by fusing the multi-weight hierarchical clustering algorithm, and the on-duty personnel closest to the abnormal point can be notified by the system to go to the scene for processing by improving the KD tree algorithm. BRIEF DESCRIPTION OF DRAWINGS

[0048] Fig. 1 is a flowchart of identifying abnormal points according to the present application.

[0049] Fig. 2 is a flowchart of screening processing nodes according to the present application. DETAILED DESCRIPTION

[0050] The technical solutions of the present application will be further specifically described below by examples in combination with the drawings.

[0051] Embodiment: A kind of unmanned aerial vehicle flight trajectory abnormal traceability method based on fusion clustering algorithm in the present embodiment, as shown in Figure 1,

[0052] Comprising the following steps:

[0053] S1 extracts data of the UAV and performs decryption and preprocessing. The data includes data features of the flight trajectory of the UAV. The trajectory data related to the UAV is mainly complex sensitive spatio-temporal data, including three-dimensional position information based on time series, roll angle, pitch angle, yaw angle, speed, and acceleration. The data processor is used to decrypt and preprocess the data from the UAV.

[0054] The preprocessing includes removing irrelevant data, and assuming that the three-dimensional space coordinates of the kth UAV at the t time and t+Δt of the ith flight are respectively: k，i (x k，i (t),y k，i (t),z k，i (t)) and P k，i (x k，i (t+Δt),y k，i (t+Δt),z k，i (t+Δt))

[0055] The Euclidean distance between them is calculated,

[0056] For time series data, the instantaneous speed and the instantaneous acceleration are estimated by using the difference, i.e. the instantaneous speed is

[0057] The acceleration is

[0058] The roll angle pitch angle θ k，i (t) and yaw angle ψ k,i (t) of the kth UAV at the t time of the ith flight are from the decrypted data. The sample as the input of the algorithm will affect the analysis result of the algorithm output, so it is very important to do the data cleaning and preprocessing work to improve the accuracy of the algorithm.

[0059] As preferred, the above multiple data features are standardized, and the normalization formula is:

[0060] Therefore, the data features of the kth UAV at the t time of the ith flight are recorded as:

[0061] In order to scale the value range of different data features to a similar scale, avoid the problem that the value difference of different features is too large, and cause the model training to be unstable or the convergence speed to be slow, the data of different dimensions should be normalized.

[0062] Step S2 specifically includes:

[0063] S2.1 Compute the similarity measure matrix for each data feature of the multi-group dataset of each UAV, denoted as:

[0064] G k,m : represents the similarity measure matrix of data feature m of the kth UAV, which is a symmetric matrix, i.e. The matrix element is defined as:

[0065] d in represents the Euclidean distance between C k，m,i (t) and C k，m,n (t);

[0066] S2.2 Initialization, each d in is processed as a separate cluster;

[0067] S2.3 Single-link clustering is used, taking the minimum distance as the distance measure of the cluster, gradually merging the nearest trajectory points to form a hierarchical clustering tree;

[0068] S2.4 The normal trajectory is finally merged into a cluster, and the remaining few isolated clusters are potential abnormal trajectories. Set the segmentation threshold to divide the cluster into normal clusters and potential abnormal clusters.

[0069] Through the UAV trajectory anomaly finder, the potential abnormal points of the UAV flight trajectory are found. The hierarchical clustering algorithm is used to calculate the similarity measure matrix between data points. According to the similarity measure matrix, the two most similar clusters are found. The two most similar clusters are merged into a new cluster. Update the similarity measure matrix to reflect the similarity of the new merged cluster with other clusters. Repeat the merging of the most similar clusters until all data points reach the pre-set threshold. Determine the final clustering result and divide the cluster into normal clusters and potential abnormal clusters.

[0070] Through the UAV trajectory anomaly finder, the potential abnormal points of the UAV flight trajectory are found. The hierarchical clustering algorithm is used to calculate the similarity measure matrix between data points. According to the similarity measure matrix, the two most similar clusters are found. The two most similar clusters are merged into a new cluster. Update the similarity measure matrix to reflect the similarity of the new merged cluster with other clusters. Repeat the merging of the most similar clusters until all data points reach the pre-set threshold. Determine the final clustering result and divide the cluster into normal clusters and potential abnormal clusters.

[0071] S3 Determine the influence of each data feature on the normal flight trajectory of the UAV, use Lasso regression algorithm to obtain the weight of each data feature, and calculate the sum of the product of the potential abnormal points and the weight of each data feature. Trace back to get the exact trajectory abnormal point, which includes:

[0072] S3.1 For the k-th UAV, if its data feature m is identified as a potential anomaly by the UAV trajectory anomaly detector during the i-th flight, then it is marked as ρ(k, m, i) = 1; otherwise, it is marked as ρ(k, m, i) = 0.

[0073] S3.2 Label the weights of data feature m as ω m The Lasso regression algorithm is used to derive the weights of each data feature. Specifically, this involves: first, using the Lasso regression model to divide the normal inspection trajectory data of the UAV into a training set and a test set, and then using the training set to train the Lasso regression model; using the Lasso regression model to determine the impact of each data feature on the normal flight trajectory of the UAV. The coefficients of the Lasso regression model reflect the degree of influence of each data feature on the normal flight trajectory of the UAV. By observing the coefficients obtained from the model, the data features with the greatest impact on the flight of the UAV and the data features with little or no impact can be identified.

[0074] S3.3 For the same time t, the different data features m of the same trajectory of the kth UAV are obtained by ρ(k, m, i). First, an OR operation is performed. If the result of the OR operation is not 0, it means that there may be an anomaly in a certain data feature of the flight trajectory of the kth UAV at time t.

[0075] S3.4 makes a comprehensive judgment on possible outliers at time t.

[0076] In the formula, κ o This indicates the threshold for judgment.

[0077] The abnormal behavior detector employs the Lasso regression algorithm to determine the impact of various data features on the normal flight trajectory of the UAV using a dataset of normal flight trajectories. Based on the anomalies in the data features calculated by the UAV trajectory anomaly detector, the Lasso regression algorithm is used to derive the weights of each data feature. The sum of the products of potential anomalies and the weights of each data feature is then calculated to make a comprehensive judgment and trace the exact trajectory anomalies.

[0078] S4 assigned personnel to verify and investigate the anomalies, specifically including:

[0079] S4.1 Obtain the latitude and longitude (u) of the k-th UAV at time t. x ,u y );

[0080] S4.2 For the location dataset of the staff on duty that day, process it according to the construction method of KD tree, and recursively construct KD tree;

[0081] S4.3 Starting from the root node of the KD tree, recursively traverse the KD tree. Specifically, compare the target point with the current node's value guard location and the UAV coordinate using Manhattan distance, record the current nearest point and distance; according to the position of the UAV coordinate, determine the next subtree to traverse; continue to recursively traverse the subtree until the leaf node, use the pruning strategy, when a node close enough to the target point is found, stop searching its child nodes, improve the search efficiency. Also includes backtracking, during the backtracking process, check if the parent node and other subtree nodes have closer value guard locations, update the nearest neighbor point and distance.

[0082] S4.4 When backtracking to the root node, find the nearest neighbor point;

[0083] S4.5 Inform the value guard of the nearest neighbor point through real-time communication to verify and investigate abnormal points.

[0084] The abnormal behavior determinator gives the abnormal point to the system, which informs the value guard closest to the abnormal point location to the scene to verify the abnormal reason and investigate whether there is a deliberate interference with the normal inspection of the UAV.

[0085] The hierarchical clustering algorithm is applied to classify and organize complex sensitive spatio-temporal data features to identify potential abnormal points. Then, using the weights of each data feature obtained by the Lasso regression algorithm, the sum of the product of the potential abnormal points and the weights of each data feature is calculated, the system can accurately and effectively detect the abnormal points in the UAV flight trajectory, ensure flight safety and stability, and maintain power grid safety. Once the abnormal point is confirmed, the system immediately calls the improved KD tree algorithm to quickly inform the nearest value guard to the scene for comprehensive inspection and thorough verification of the factors interfering with the normal flight of the UAV.

[0086] First, construct the data trajectory anomaly effective identification structure, as shown in FIG. 2:

[0087] Data processor: The data processor mainly decrypts and preprocesses the data from the UAV. The trajectory data of the UAV mainly involves complex sensitive spatio-temporal data, i.e. three-dimensional position information, roll angle, pitch angle, yaw angle, speed, acceleration, etc. based on time series. In order to scale the value range of different data features to a similar scale, avoid the problem of unstable model training or slow convergence speed caused by too large value difference between different features, different dimensional data should be normalized.

[0088] UAV trajectory anomaly finder: the UAV trajectory anomaly finder is mainly to find out the potential abnormal points of the UAV flight trajectory. This patent adopts hierarchical clustering algorithm to calculate the similarity measurement matrix between data points, and finds the two most similar clusters according to the similarity measurement matrix. The two most similar clusters are merged into a new cluster. The similarity measurement matrix is updated to reflect the similarity of the new merged cluster with other clusters. Repeat the merging of the most similar clusters until all data points reach the pre-set threshold. Determine the final clustering result, and divide the cluster into normal cluster and potential abnormal cluster.

[0089] Abnormal behavior determinator: this patent adopts Lasso regression algorithm to determine the influence of each data feature on the normal flight trajectory of the UAV through the data set of the normal flight trajectory of the UAV. According to the abnormal situation of the data features calculated by the UAV trajectory anomaly finder, the weight of each data feature is obtained by using the Lasso regression algorithm, and the sum of the product of the potential abnormal points and the weight of each data feature is calculated, so as to make comprehensive judgment and trace the exact trajectory abnormal point.

[0090] Adaptive alarm mechanism: the abnormal points given by the abnormal behavior determinator are issued to the system, and the system notifies the on-duty personnel closest to the abnormal point position to the scene to verify the abnormal reason and investigate whether there is a deliberate interference with the normal inspection of the UAV.

[0091] The working method of the data trajectory anomaly effective recognition component structure is as follows

[0092] 1. Data processor

[0093] The sample as the input of the algorithm will inevitably affect the analysis result of the algorithm output, so it is very important to do a good job of data cleaning and preprocessing to improve the accuracy of the algorithm. The rules of data cleaning mainly include:

[0094] (1) Eliminate irrelevant data and retain data of interest.

[0095] (2) Assuming that the three-dimensional space coordinates of the kth UAV at the tth time and t+Δt of the ith flight are: k，i (x k，i (t), y k，i (t), z k，i (t)), P k，i (x k，i (t+Δt), y k，i (t+Δt), z k，i (t+Δt))

[0096] Calculate the Euclidean distance between them,

[0097] For time series data, the instantaneous velocity and the instantaneous acceleration are estimated using the difference, i.e., the instantaneous velocity is

[0098] The acceleration is

[0099] The roll angle of the kth UAV at the tth time of the ith flight The pitch angle θ k,i The yaw angle ψ k,i (t) come from the decrypted data. The samples, as the input of the algorithm, must affect the analysis results of the algorithm output, so it is very important to do data cleaning and preprocessing work to improve the accuracy of the algorithm.

[0100] As preferred, the above multiple data features are standardized, and the normalization formula is:

[0101] Therefore, the data features of the kth UAV at the tth time of the ith flight are denoted as:

[0102] In order to scale the value range of different data features to a similar scale, avoid the problem that the value difference of different features is too large, and cause the model training to be unstable or the convergence speed to be slow, the data of different dimensions should be normalized.

[0103] 2. UAV trajectory anomaly finder

[0104] Usually, the inspection route of each UAV to the power grid line is fixed, and the spatial coordinate information of each inspection should show a similar trajectory curve. On the contrary, the abnormal trajectory curve does not satisfy the rule.

[0105] The steps of detecting trajectory abnormal data based on the fusion multi-weight hierarchical clustering algorithm are as follows:

[0106] (1) For each single data feature of the multiple data sets of each UAV, the similarity measure matrix is calculated, denoted as:

[0107] G k,m : represents the similarity measure matrix of the data feature m of the kth UAV, which is a symmetric matrix, i.e. The matrix elements are defined as:

[0108] d in represents the Euclidean distance of C k,m,i (t) and C k，m，n (t);

[0109] S2.2 initialization, each d inTreat it as a separate cluster;

[0110] (3) Single-link clustering is adopted, with the minimum distance as the distance metric for clusters, and the closest trajectory points are gradually merged to form a hierarchical clustering tree;

[0111] (4) Normal trajectories are eventually merged into one cluster, and the remaining few isolated clusters are potential abnormal trajectories. Set a segmentation threshold to divide the clusters into normal clusters and potential abnormal clusters.

[0112] 3. Abnormal Behavior Detector

[0113] Based on the anomalies in the data features calculated by the drone trajectory anomaly detector, the weights of each data feature are derived using the Lasso regression algorithm, and a comprehensive judgment is then made. The specific steps are as follows:

[0114] (1) For the kth UAV, if its data feature m is identified as a potential anomaly by the UAV trajectory anomaly detector during the i-th flight, then it is marked as ρ(k, m, i) = 1; otherwise, it is marked as ρ(k, m, i) = 0.

[0115] (2) Mark the weight of data feature m as ωm, and use the Lasso regression algorithm to obtain the weight of each data feature; specifically, it includes: using the Lasso regression model to divide the normal inspection trajectory data of the UAV into training set and test set in advance, and then using the training set to train the Lasso regression model; using the Lasso regression model to determine the influence of each data feature on the normal flight trajectory of the UAV, the coefficient of the Lasso regression model reflects the degree of influence of each data feature on the normal flight trajectory of the UAV, and by observing the coefficients obtained by the model, determine the data features that have the greatest impact on the flight of the UAV and the data features that have a small impact or can even be ignored.

[0116] (3) For the same time t, the different data features m of the same trajectory of the kth UAV are obtained by ρ(k, m, i). First, perform an OR operation. If the result of the OR operation is not 0, it means that there may be an anomaly in a certain data feature of the flight trajectory of the kth UAV at time t.

[0117] (4) Make a comprehensive judgment on the possible outliers at time t.

[0118] In the formula, κ o This indicates the threshold for judgment.

[0119] The Lasso regression algorithm is adopted by the abnormal behavior determinator, and the influence of each data feature on the normal flight trajectory of the UAV is determined through the data set of the normal flight trajectory of the UAV. According to the abnormal situation of the data features calculated by the UAV trajectory anomaly finder, the weight of each data feature is obtained by using the Lasso regression algorithm, and the sum of the product of the potential abnormal point and each data feature weight is calculated, so as to make a comprehensive judgment and trace back to obtain the exact trajectory abnormal point.

[0120] S4 arranges personnel to verify and investigate abnormal points, specifically including:

[0121] S4.1 obtains the longitude and latitude (u x ,u y ) of the kth UAV at time t.

[0122] (2) Construct a KD tree. For the position data set of the on-duty personnel on the same day, process it according to the construction method of the KD tree, and recursively construct the KD tree.

[0123] (3) Traverse the KD tree. Starting from the root node of the KD tree, recursively traverse:

[0124] a) Compare the target point: use Manhattan distance to calculate the distance between the current on-duty personnel position and the UAV coordinate, and record the current nearest point and distance.

[0125] b) Select the subtree: according to the position of the UAV coordinate, determine the next subtree (left subtree or right subtree) for traversal.

[0126] c) Recursively traverse the subtree: continue to recursively traverse the subtree downward until the leaf node. Use the pruning strategy, when a node close enough to the target point has been found, the search for its child nodes can be stopped to improve the search efficiency.

[0127] d) Backtracking to update the nearest neighbor point: in the process of backtracking, check whether the parent node and other subtree nodes have closer on-duty personnel positions, and update the nearest neighbor point and distance.

[0128] (4) Find the nearest neighbor point. When backtracking to the root node, the nearest on-duty personnel position is determined.

[0129] (5) The system notifies the on-duty personnel through a mobile application, a short message, a call, or other real-time communication methods.

[0130] As shown in FIG. 1 and FIG. 2, the main workflow is:

[0131] Step 1: First, the data set from the UAV external terminal is decrypted using the decryption algorithm, which contains multiple data features of the UAV, and the data required for trajectory anomaly analysis is extracted, i.e. three-dimensional position information based on time series, roll angle, pitch angle, yaw angle, and irrelevant data is removed. Go to Step 2.

[0132] Step 2: Use the difference to get the instantaneous speed and instantaneous acceleration of the UAV. Go to Step 3.

[0133] Step 3: Due to the difference in dimensions, the 8 data features obtained in Step 1 and Step 2 need to be standardized. The normalization formula is:

[0134] Go to Step 4.

[0135] Step 4: For each single data feature of the multiple data sets of each UAV, calculate the similarity measure matrix. The calculation formula is

[0136] Go to Step 5.

[0137] Step 5: Initialization. Each d in is processed as a separate cluster. Go to Step 6.

[0138] Step 6: Single-link clustering is used, with the minimum distance as the distance measure of the cluster, and the nearest trajectory points are gradually merged to form a hierarchical clustering tree. Go to Step 7.

[0139] Step 7: Normal trajectories will eventually merge into one cluster, and the remaining few clusters will exist in isolation. These clusters may be abnormal trajectories. Based on human experience, set the segmentation threshold to divide the clusters into normal clusters and potential abnormal clusters. Go to Step 8.

[0140] Step 8: For the kth UAV, if its data feature m is determined by the UAV trajectory anomaly detector as a potential abnormal point in the ith flight, it is marked as ρ(k,m,i) = 1, otherwise, it is marked as ρ(k,m,i) = 0. Go to Step 9.

[0141] Step 9: The weight of data feature m is marked as ω m , and the normal patrol trajectory data of the UAV is divided into training set and test set by Lasso regression in advance, and then the training set is used to train the Lasso regression model. In this process, the model will determine the influence of each data feature m on the normal flight trajectory of the UAV.

[0142] Step 10: For the same t time, the different data characteristics m of the kth UAV in the same trajectory, first, the or operation is performed, if the or operation result is 0, it means that all data characteristics of the flight trajectory of the kth UAV at t time do not exist abnormality. If the or operation result is not 0, it means that some data characteristics of the flight trajectory of the kth UAV at t time may exist abnormality, and step 11 is entered.

[0143] Step 11: Comprehensive judgment is made on the possible abnormal points at t time.

[0144] In the formula, κ o represents the judgment threshold. If ∑ω m < κ o , then k = k + 1, and return to step 1; otherwise, step 12 is entered.

[0145] Step 12: When the abnormal behavior judge determines that there is abnormality at a certain t time of the kth UAV, the adaptive alarm mechanism immediately alarms. The system obtains the abnormal point of the flight trajectory of the UAV, and notifies the nearest on-duty personnel to the scene to verify the abnormal reason through the improved KD tree algorithm, and investigates whether there is a lawless behavior of deliberately interfering with the normal inspection of the UAV.

[0146] Embodiment

[0147] Suppose that a power grid company in a certain city uses UAVs to conduct normal inspection of power grid lines, and the UAVs are dispersed in UAV warehouses near each inspection target, and each UAV is equipped with an external terminal module. First, the data from the UAVs is decrypted, and then the data is preprocessed. The hierarchical clustering algorithm is used to find the potential abnormal points of the UAV flight trajectory. Then, the weights of each data characteristic are obtained by the Lasso regression algorithm during the normal inspection of the UAV, and the potential abnormal points are comprehensively judged to trace the exact abnormal point. Finally, the nearest on-duty personnel to the abnormal point is notified by the system through the improved KD tree algorithm.

[0148] The specific implementation scheme is as follows:

[0149] (1) First, the trajectory data from the UAV is decrypted, the data characteristics are extracted, and the data is preprocessed by using the standardization formula to normalize the data.

[0150] (2) The hierarchical clustering algorithm is used to cluster the UAV flight trajectory data to find the potential abnormal points.

[0151] (3) In the abnormal behavior judge, the Lasso regression algorithm is used to obtain the weights of each data characteristic, and the sum of the product of the potential abnormal point and each data characteristic weight is calculated, and compared with the threshold κ oComparison is made to determine whether the t moment of the k unmanned aerial vehicle is abnormal.

[0152] (4) The system obtains the abnormal point position of the unmanned aerial vehicle flight trajectory, improves the KD tree algorithm, informs the nearest on-duty personnel of the abnormal point to verify the abnormal reason on site, and investigates whether there is a lawless behavior of deliberately interfering with the normal inspection of the unmanned aerial vehicle.

[0153] Improvement and optimization:

[0154] The unmanned aerial vehicle trajectory anomaly finder finds the possible abnormal points of the unmanned aerial vehicle flight trajectory through the hierarchical clustering method. Compared with the traditional K-Means clustering algorithm, the hierarchical clustering method is not sensitive to the selection of the initial clustering center. The hierarchical clustering method does not need to assume that the data is convex, and therefore has better effect when processing non-convex clusters. The hierarchical clustering needs to set a segmentation threshold based on artificial experience to divide the cluster into normal clusters and possible abnormal clusters.

[0155] The abnormal behavior determinator first performs an or operation on p(k,m,i). This step can save system consumption. If the or operation result is not 0, the possible abnormal point at the t moment is further comprehensively determined. The Lasso regression algorithm is used to determine the influence of each data feature on the normal flight trajectory of the unmanned aerial vehicle through the data set of the normal flight trajectory of the unmanned aerial vehicle. According to the abnormality of the data features calculated by the unmanned aerial vehicle trajectory anomaly finder, the weight of each data feature is obtained by using the Lasso regression algorithm, and the sum of the product of the potential abnormal point and each data feature weight is calculated. The threshold κ o Comparison is made to determine whether the t moment of the k unmanned aerial vehicle is abnormal.

[0156] The adaptive alarm mechanism improves the KD algorithm and introduces a pruning strategy to reduce unnecessary search. When a node that is close enough to the target point is found, the search of its child nodes can be stopped, thereby improving the search efficiency and saving system calculation overhead.

[0157] The preferred embodiments of the embodiments of the present application are described above with reference to the accompanying drawings, but this does not limit the scope of the rights of the embodiments of the present application. Any modification, equivalent replacement and improvement made by those skilled in the art without departing from the scope and essence of the embodiments of the present application shall be within the scope of the rights of the embodiments of the present application.

Claims

1. A method for tracing abnormal flight trajectory of a UAV based on a fusion clustering algorithm, characterized in that, The method comprises the following steps: S1 extracts data of the unmanned aerial vehicle and performs decryption and preprocessing, and extracts data features of the flight trajectory of the unmanned aerial vehicle; S2 adopts a hierarchical clustering algorithm to find potential abnormal points of the flight trajectory of the unmanned aerial vehicle, and specifically comprises: S2.1, for each single data feature of the multiple data sets of each unmanned aerial vehicle, a similarity measure matrix is calculated; G k,m a similarity measure matrix of data features m of the kth UAV, the matrix elements are defined as: d in represents C k,m,i (t) and C k,m,n (t) are Euclidean distances; S2.2, each Euclidean distance is processed as a separate cluster; S2.3, single-link clustering is adopted, the minimum distance is taken as the distance measure of the cluster, the trajectory points with the closest distance are gradually merged, and a hierarchical clustering tree is formed; S2.4, the normal trajectory is finally merged into one cluster, and the remaining few isolated clusters are potential abnormal trajectories, a segmentation threshold is set, and the clusters are divided into normal clusters and potential abnormal clusters; S3 determines the influence of each data feature on the normal flight trajectory of the unmanned aerial vehicle, uses a Lasso regression algorithm to obtain the weight of each data feature, and obtains the sum of the product of the potential abnormal points and the weight of each data feature, and traces to obtain the exact trajectory abnormal point, and specifically comprises: S3.1, for the k unmanned aerial vehicle, if the data m of the unmanned aerial vehicle is determined as potential abnormal by the unmanned aerial vehicle trajectory anomaly finder in the i flight, the unmanned aerial vehicle is marked as p(k,m,i)=1, otherwise, p(k,m,i)=0 is marked; S3.2 mark the weight of data feature m as ω m , and the weight of each data feature is obtained by using Lasso regression algorithm. S3.3, for the same t time, the p(k,m,i) obtained by the same trajectory of the k unmanned aerial vehicle for different data features m is first operated, if the operation result is not 0, it indicates that there may be an abnormality in the flight trajectory of the k unmanned aerial vehicle at t time; S3.4 make a comprehensive decision on possible outliers at this time t In the formula, κ o represents a determination threshold value; S4 arranges personnel to verify and investigate abnormal points.

2. The unmanned aerial vehicle flight trajectory abnormality backtracking method based on the fusion clustering algorithm according to claim 1, characterized in that, The step S1 extracts data of the unmanned aerial vehicle, which comprises extracting data features of the flight trajectory of the unmanned aerial vehicle. The trajectory data of the unmanned aerial vehicle mainly involves complex sensitive spatio-temporal data, including three-dimensional position information and speed information based on time series.

3. The unmanned aerial vehicle flight trajectory abnormality backtracking method based on the fusion clustering algorithm according to claim 1 or 2, characterized in that, The step S1 preprocessing comprises excluding irrelevant data, and then assuming that the three-dimensional space coordinates of the k unmanned aerial vehicle at t time and t+Δt in the i flight are: P k,i (x k,i (t),y k,i (t),z k,i (t)),P k,i (x k,i (t+Δt),y k,i (t+Δt),z k,i (t+Δt)) calculate the Euclidean distance between them, for time series data, utilize the difference to estimate the instantaneous velocity and instantaneous acceleration.

4. The unmanned aerial vehicle flight trajectory abnormality backtracking method based on the fusion clustering algorithm according to claim 3, characterized in that, The above plurality of data features are standardized, and the data feature of the kth UAV at the tth moment of the ith flight is denoted as:

5. The unmanned aerial vehicle flight trajectory abnormality backtracking method based on the fusion clustering algorithm according to claim 1, characterized in that, The step S3.2 specifically comprises: the normal patrol trajectory data of the unmanned aerial vehicle is divided into a training set and a test set in advance by using a Lasso regression model, and then the training set is used to train the Lasso regression model; the influence of each data feature on the normal flight trajectory of the unmanned aerial vehicle is determined by using the Lasso regression model. The coefficients of the Lasso regression model reflect the influence degree of each data feature on the normal flight trajectory of the unmanned aerial vehicle. By observing the coefficients obtained by the model, the data features with the greatest influence on the flight of the unmanned aerial vehicle and the data features with small or negligible influence are determined.

6. The unmanned aerial vehicle flight trajectory abnormality backtracking method based on a fusion clustering algorithm according to claim 1, characterized in that, The step S4 specifically comprises: S4.1 Obtain the longitude and latitude (u x , y ) of the kth UAV at time t; S4.2, for the position data set of the on-duty personnel of the day, the KD tree is processed according to the construction mode of the KD tree, and the KD tree is recursively constructed; S4.3, starting from the root node of the KD tree, the KD tree is recursively traversed; S4.4, when backtracking to the root node, the nearest neighbor point is found; S4.5, the on-duty personnel of the nearest neighbor point is notified by a real-time communication mode to verify and investigate abnormal points.

7. The unmanned aerial vehicle flight trajectory abnormality backtracking method based on the fusion clustering algorithm according to claim 6, characterized in that, The step S4.3 specifically comprises comparing the target point by using Manhattan distance, calculating the distance between the current on-duty position of the node and the UAV coordinate, recording the current nearest point and distance, recursively traversing the sub-tree downward according to the position of the UAV coordinate until the leaf node, and improving the search efficiency by using the pruning strategy.

8. The unmanned aerial vehicle flight trajectory abnormality backtracking method based on the fusion clustering algorithm according to claim 7 or 6, characterized in that, The step S4.3 further comprises backtracking, in which process, whether the parent node and other sub-tree nodes have closer on-duty positions is checked, and the nearest neighbor point and distance are updated.

Images