A method for online anomaly detection of substation UAV flight data

Abstract

The invention relates to the technical field of power station monitoring, and discloses an online anomaly detection method for flight data of a substation unmanned aerial vehicle. The detection method comprises the following steps: S1, using a principal component analysis method to find a low-dimensional internal subspace matrix through data reconstruction in a multidimensional flight data matrix and calculate the main direction vector of the subspace; S2, using an oversampling principal component analysis method to amplify the influence of abnormal data on the main direction; S3, when the unmanned aerial vehicle switches the task mode, according to the characteristic that a data point can only be normal or abnormal at a certain moment, a type of support vector machine is selected to judge whether the data point is abnormal; the present invention adopts the OS-PCA abnormal data detection method to ensure that the unmanned aerial vehicle has the advantages of low delay, low complexity and low storage overhead when performing tasks; according to when the unmanned aerial vehicle switches the task mode, a type of support vector machine is selected to judge the abnormality score of the data point, thereby effectively reducing the false detection rate.

Description

Technical Field

[0001] The invention relates to the technical field of substation monitoring, and in particular to an online anomaly detection method for substation unmanned aerial vehicle flight data. Background Art

[0002] The society's demand for electric energy is increasing. As an important part of the power system, the construction scale of substations is constantly expanding, and the operation and maintenance workload is also increasing sharply. With the advantages of high maneuverability, compact and flexible fuselage, precise hovering, vertical take-off and landing, drones can observe equipment at close range and provide comprehensive coverage. They can make up for the shortcomings of blind spots in high-altitude equipment in stations during manual inspections and the limited access area for inspectors due to the safety distance of high-voltage live equipment. In addition, drones are equipped with a variety of sensors to perform refined visible light inspections and infrared inspections. Combined with image analysis technology and artificial intelligence cloud computing technology, the platform can automatically complete image analysis and diagnosis, quickly and accurately identify typical equipment defects, and monitor the operating status of substation equipment in real time, thereby improving the inspection efficiency of substations. However, in actual use, the deployment of drones needs to ensure their own safety as well as the safety of people and property around them.

[0003] During the inspection of the substation by the drone, the strong magnetic field interferes with the normal operation of the drone's electronic compass and magnetic compass, making it unable to accurately determine the heading, resulting in a shift in the hovering position. For example, the magnetic compass is interfered with, causing the GPS mode to fly in a circle, the GPS is interfered with, causing it to be unable to fly in a straight line, and the vibration is too large to exceed the IMU measurement value, making the flight unable to stabilize itself. These problems increase the risk of flight. By real-time monitoring and analysis of key data of the flight control system, the flight status and motion status of the drone can be monitored in real time, and the staff can be assisted to adjust the flight parameters in time. The reason why the failure rate of drones is relatively high when performing tasks is the lack of effective methods to detect their own faults. If the occurrence of abnormal values ​​in the system can be detected and early warning can be given, and the flight control can be effectively operated within a limited time, it will be of great significance to reduce the accident rate and improve flight safety.

[0004] Due to the complexity of drone systems, it is difficult to model each subsystem, and the adaptability and anti-interference ability of the model method are limited. Therefore, data-driven methods are widely used for data anomaly detection, with algorithm accuracy and real-time performance as the goal, and the temporal and heterogeneous nature of drone flight data as the focus. Drone flight data has the characteristics of high dimensionality, large data volume, few abnormal labels and strong real-time performance. Therefore, the selected anomaly detection algorithm needs to have: low latency and low complexity to meet the real-time requirements, unsupervised methods or semi-supervised methods to meet the requirements of few abnormal labels. In addition, the robustness of the algorithm should be high, that is, the algorithm false detection rate is low after the mode switch.

[0005] Therefore, the applicant proposed an online anomaly detection method for substation UAV flight data. Summary of the invention

[0006] In view of the shortcomings of the prior art, the present invention provides an online anomaly detection method for substation UAV flight data, which has low latency and low complexity to meet the real-time requirements, and an unsupervised method or a semi-supervised method to meet the requirement of fewer anomaly labels. In addition, the algorithm of the present invention has high robustness and low algorithm false detection rate after mode switching.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A method for detecting anomalies of substation UAV flight data online, the method comprising the following steps:

[0009] S1. Using principal component analysis (PCA) method on multidimensional flight data matrix Find a low-dimensional internal subspace matrix by data reconstruction And calculate the main direction vector of the subspace, which will contain the maximum amount of information in the original data space;

[0010] S2. Oversampling principal component analysis is used to amplify the influence of abnormal data on the main direction, while normal data has no influence on the main direction. The subspace vector is updated in a projection approximation way, and the main direction changes of the subspace vector are tracked and matched to realize instantaneous anomaly detection of flight data, and the anomaly score of each point is obtained;

[0011] S3. When the UAV switches the mission mode, in order to adaptively update the anomaly score threshold and further reduce the false detection rate of anomaly detection, based on the characteristic that a data point can only be either normal or abnormal at a certain moment, a class of support vector machines (OC-SVM) is selected to judge whether the data point is abnormal.

[0012] As a further solution of the present invention: Step S1 comprises the following steps:

[0013] S11. When the UAV performs operation and maintenance tasks in the substation, some key parameters are extremely important for the flight control of the UAV. The flight data matrix Including: longitude, latitude, altitude, pitch angle, roll angle, yaw angle, ground speed and air speed of the drone, etc. The flight parameters of the drone are monitored to detect potential anomalies. When the multidimensional flight data matrix as follows:

[0014]

[0015] In the above formula, Indicates when The flight data instance at is a multidimensional flight data matrix The row vector of ;

[0016] S12, PCA is mainly through the flight data matrix Reconstruction, in a multidimensional flight data matrix Find a low-dimensional interior subspace matrix in As a meaningful and lightweight data representation of the original flight data, the subspace matrix The row vector of is the vectorized representation of the corresponding flight parameters in the original input, where ;

[0017] S13. PCA essentially uses the square of the Euclidean norm as the distance metric and minimizes the reconstruction error of the projection data. The data reconstruction objective function is expressed as:

[0018] (1)

[0019] in, for , For the Raw flight data input, is the overall mean, To include The subspace matrix of the subspace vectors, For the The original data is projected into the data subspace, is the first estimate obtained from the projection subspace reconstruction indivual ;

[0020] S14. Solve the minimization reconstruction problem by deriving the eigenvalues ​​of the covariance data matrix as follows:

[0021] (2)

[0022] in, is the covariance matrix, is the mean, Each column of is a feature vector of The corresponding diagonal elements are the corresponding eigenvalues. The last few eigenvectors have little influence on the data distribution and will be discarded. Formula (1) is approximated in the least squares form as follows:

[0023] (3)

[0024] Among them, In dimensional space, for The approximate value of is approximately equal to , based on formula (3), the reconstruction error has a quadratic form and is The function can be calculated by solving the least squares problem. In the online environment, it is assumed that the adjacent subspaces of the flight data in the time series direction have similarity, that is, .

[0025] As a further solution of the present invention: the specific steps of step S2 include:

[0026] The amount of normal data and abnormal data of drones is unbalanced. The OS-PCA anomaly detection method is used because the proportion of abnormal data is very small. By oversampling the abnormal data points, the main direction of the data is far away from the main direction of the original data, while the oversampling of normal data points has little effect on the change of the main direction. When drones perform inspection tasks in substations, they will contain a large amount of data, and resampling the data points This can amplify the influence of the current data point on the main direction. The data instance at time is , the data input is expressed as , then the objective function is:

[0027] (4)

[0028] In addition, if The impact before the moment can be reduced to amplify the current moment The effect on the main direction, that is, applying the forgetting factor ,at this time , is the number of oversampling The original dataset size The objective function is shown in formula (5), where the current moment The input is , and Use and approximate,

[0029] (5)

[0030] The optimization problem of solving equation (5) is the problem of data point incremental strategy anomaly detection. The input of the oversampling PCA anomaly detection algorithm is the flight data matrix X and the parameter , the algorithm output is the anomaly score, which is obtained by taking formula (5) The derivative of The solution is as follows:

[0031] (6)

[0032] For equation (6), if we only care about the main direction, that is, the current The subspace vector corresponding to the maximum eigenvalue at time is , Time reference subspace vector Zhang Cheng's subspace is , using the estimated most significant subspace vector and the most significant subspace vector of the reference , respectively, to the reference subspace and the estimated subspace Projection, the average of the angles calculated by formula (7) is The anomaly score of the input vector at time instant,

[0033] (7)

[0034] Among them, the anomaly score for The impact of the input vector on the data subspace vector at the moment. The larger the value, the greater the abnormality of the input vector. Through a large number of experiments, it is verified that when the abnormal score When it is greater than 0.5, it can be determined that an abnormality has occurred in a certain flight state of the drone.

[0035] As a further solution of the present invention: the specific steps of step S3 include:

[0036] OS-PCA makes use of the similarity of data subspace in the time series direction to provide the possibility for anomaly detection of time series. It has the characteristics of lightweight calculation and storage and adaptability to the dynamic changes of flight data. However, when the UAV switches mission modes, such as takeoff, climb, inspection, landing, etc., the collected flight data may be different from historical data in normal performance. In order to adaptively update the anomaly score threshold and further reduce the false detection rate of anomaly detection, OC-SVM is selected to perform anomaly discrimination on the anomaly score based on the characteristic that the data point at a certain moment can only be either normal or abnormal.

[0037] A type of support vector machine algorithm does not need to obtain the classification label of the data. It aims to construct a hyperplane or hypersphere to contain the distribution area of ​​most data points and isolate a small number of data points. For a given training data set, OC-SVM uses a mapping function The training data points Project the original feature space to a higher-dimensional space and establish a hyperplane in the mapped space ,in and Assume that the origin is an outlier in the OC-SVM model, and then you need to obtain a hyperplane to separate all training sample points from the point, and ensure that the distance between the hyperplane and the origin is as far as possible, that is, you need to optimize the distance from the origin to the hyperplane. is the maximum value. The hyperplane decision equation of a type of support vector machine is shown in formula (8):

[0038] (8)

[0039] in, is the Lagrange multiplier, and By any satisfaction A sample of support vectors Calculate and obtain, is a symbolic function, if If it is greater than 0, the output is 1, otherwise the output is 0.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] 1. The present invention adopts the OS-PCA abnormal data detection method, which only needs to update the main direction of the subspace matrix through formula (7). The calculation and storage complexity of the algorithm are low, which ensures the requirements of low latency, low complexity and low storage overhead for drones in performing tasks.

[0042] 2. The present invention monitors the flight data of the UAV in an online monitoring manner, and can detect abnormal data of the UAV in a timely manner. These data usually indicate an existing fault or potential fault, effectively ensuring the safe operation and good condition of the UAV, and reducing the possibility of risks and accidents.

[0043] 3. When the UAV switches the mission mode, the present invention selects a type of support vector machine to perform abnormality discrimination on the abnormal scores of data points in order to adaptively update the abnormal score threshold and further reduce the false detection rate of abnormal detection, because the collected flight data is different from the historical data in normal performance, thereby effectively reducing the false detection rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without creative work. Among them:

[0045] Figure 1 is a data processing flow chart of the present invention;

[0046] Figure 2 is a flow chart of the data preprocessing method of the present invention;

[0047] Figure 3 Heat map of the subset extracted by the analysis method of the present invention. DETAILED DESCRIPTION

[0048] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0049] Example

[0050] See also Figure 1-Figure 3 As shown, the present invention provides a method for online abnormality detection of substation UAV flight data, and the detection method comprises the following steps:

[0051] S1. Using principal component analysis (PCA) method on multidimensional flight data matrix Find a low-dimensional internal subspace matrix by data reconstruction And calculate the main direction vector of the subspace, which will contain the maximum amount of information in the original data space;

[0052] S2. Oversampling principal component analysis is used to amplify the influence of abnormal data on the main direction, while normal data has no influence on the main direction. The subspace vector is updated in a projection approximation way, and the main direction changes of the subspace vector are tracked and matched to realize instantaneous anomaly detection of flight data, and the anomaly score of each point is obtained;

[0053] S3. When the UAV switches the mission mode, in order to adaptively update the anomaly score threshold and further reduce the false detection rate of anomaly detection, based on the characteristic that a data point can only be either normal or abnormal at a certain moment, a class of support vector machines (OC-SVM) is selected to judge whether the data point is abnormal.

[0054] Preferably, step S1 comprises the following steps:

[0055] (1) When a drone performs maintenance tasks at a substation, some key parameters are extremely important for the flight control of the drone. The flight data matrix Including: longitude, latitude, altitude, pitch angle, roll angle, yaw angle, ground speed and air speed of the drone, etc. The flight parameters of the drone are monitored to detect potential anomalies. When the multidimensional flight data matrix as follows:

[0056]

[0057] In the above formula, Indicates when The flight data instance at is a multidimensional flight data matrix The row vector of ;

[0058] (2) PCA is mainly based on the flight data matrix Reconstruction, in a multidimensional flight data matrix Find a low-dimensional interior subspace matrix in As a meaningful and lightweight data representation of the original flight data, the subspace matrix The row vector of is the vectorized representation of the corresponding flight parameters in the original input, where ;

[0059] (3) Due to the electromagnetic interference environment of the substation and poor contact of hardware equipment, the UAV flight data matrix It is often noisy, so it is necessary to clean the data before model training. The box plot can identify several values ​​that are significantly higher than the normal range. In addition, the normalization method can scale the data of different features to the same range to eliminate the dimensional differences between different features, thereby avoiding some features dominating the model training process due to excessive numerical range. Therefore, the normalization method of formula (9) is used to process the data.

[0060] (9)

[0061] in, is the mean of the data, is the standard deviation of the data, and the data processing method flow is as follows Figure 2 As shown;

[0062] (4) For multi-dimensional UAV flight data, there is often a certain correlation between different dimensions, and they are not completely independent of each other. Therefore, from the flight data matrix Extracting highly correlated subsets for anomaly detection can not only reduce computational complexity but also improve detection accuracy to a certain extent. Extracting highly correlated subsets of drone flight data can be done through heat map analysis, such as Figure 3As shown in the figure, the higher the score of each block, the greater the correlation. Generally speaking, the absolute value of the correlation score is above 0.8, and the data is considered to be strongly correlated. Taking the altitude and speed subsets in the flight data as an example, GPS altitude (alt), navigation altitude (navalt), barometric altimeter altitude (h), GPS north speed (vn), navigation north speed (navvn), navigation vertical speed (navvd) and GPS vertical speed (vd) are used for analysis. The highly correlated data subsets are {alt, h, navalt}, {navvd, vd}, {navve, ve}, {navvn, vn};

[0063] (5) PCA essentially uses the square of the Euclidean norm as the distance metric and minimizes the reconstruction error of the projected data. The data reconstruction objective function can be expressed as:

[0064] (1)

[0065] in, for , For the Raw flight data input, is the overall mean, To include The subspace matrix of the subspace vectors, For the The original data is projected into the data subspace, is the first estimate obtained from the projection subspace reconstruction indivual ;

[0066] (6) The minimization reconstruction problem is solved by deriving the eigenvalues ​​of the covariance data matrix as follows:

[0067] (2)

[0068] in, is the covariance matrix, is the mean, Each column of is a feature vector of The corresponding diagonal elements are the corresponding eigenvalues. The last few eigenvectors have little influence on the data distribution and will be discarded. Formula (1) is approximated by the least squares form as follows:

[0069] (3)

[0070] Among them, In dimensional space, for The approximate value of is approximately equal to , based on formula (3), the reconstruction error has a quadratic form and is The function can be calculated by solving the least squares problem. In the online environment, it is assumed that the adjacent subspaces of the flight data in the time series direction have similarity, that is, .

[0071] Preferably, the specific steps of step S2 include:

[0072] The amount of normal data and abnormal data of drones is unbalanced. The OS-PCA anomaly detection method is used because the proportion of abnormal data is very small. By oversampling the abnormal data points, the main direction of the data is far away from the main direction of the original data, while the oversampling of normal data points has little effect on the change of the main direction. When drones perform inspection tasks in substations, they will contain a large amount of data, and resampling the data points This can amplify the influence of the current data point on the main direction. The data instance at time is , the data input is expressed as , then the objective function is:

[0073] (4)

[0074] In addition, if The impact before the moment can be reduced to amplify the current moment The effect on the main direction, that is, applying the forgetting factor ,at this time , is the number of oversampling The original dataset size The objective function is shown in formula (5), where the current moment The input is , and Use and approximate,

[0075] (5)

[0076] The optimization problem of solving equation (5) is the problem of data point incremental strategy anomaly detection. The input of the oversampling PCA anomaly detection algorithm is the flight data matrix X and the parameter , the algorithm output is the anomaly score, which is obtained by taking formula (5) The derivative of The solution is as follows:

[0077] (6)

[0078] For equation (6), if we only care about the main direction, that is, the current The subspace vector corresponding to the maximum eigenvalue at time is , Time reference subspace vector Zhang Cheng's subspace is , using the estimated most significant subspace vector and the most significant subspace vector of the reference , respectively, to the reference subspace and the estimated subspace Projection, the average of the angles calculated by formula (7) is The anomaly score of the input vector at time instant,

[0079] (7)

[0080] Among them, the anomaly score for The impact of the input vector on the data subspace vector at the moment. The larger the value, the greater the abnormality of the input vector. Through a large number of experiments, it is verified that when the abnormal score When it is greater than 0.5, it can be determined that an abnormality has occurred in a certain flight state of the drone.

[0081] Prioritize, OS-PCA uses the similarity characteristics of data subspace in the time series direction to provide a basis for anomaly detection of time series. It has the characteristics of lightweight calculation and storage and adaptability to dynamic changes in flight data. However, when the UAV switches mission modes, such as takeoff, climb, inspection, landing, etc., the collected flight data may be different from historical data in normal performance. In order to adaptively update the anomaly score threshold and further reduce the false detection rate of anomaly detection, according to the characteristic that a data point at a certain moment can only be either normal or abnormal, OC-SVM is selected to perform anomaly discrimination on the anomaly score;

[0082] A type of support vector machine algorithm does not need to obtain the classification label of the data. It aims to construct a hyperplane or hypersphere to contain the distribution area of ​​most data points and isolate a small number of data points. For a given training data set, OC-SVM uses a mapping function The training data points Project the original feature space to a higher-dimensional space and establish a hyperplane in the mapped space ,in and Assume that the origin is an outlier in the OC-SVM model, and then you need to obtain a hyperplane to separate all training sample points from the point, and ensure that the distance between the hyperplane and the origin is as far as possible, that is, you need to optimize the distance from the origin to the hyperplane. is the maximum value. The hyperplane decision equation of a type of support vector machine is shown in formula (8):

[0083] (8)

[0084] in, is the Lagrange multiplier, and By any satisfaction A sample of support vectors Calculate and obtain, is a symbolic function, if If it is greater than 0, the output is 1, otherwise the output is 0;

[0085] OS-PCA is used as the measurement model of drone anomaly score, and OC-SVM is used as the judgment model of whether the data is abnormal. The main process is as follows Figure 1 As shown: First, use the normal data in the historical data set The OC-SVM model is trained by scoring. Since the training data provided are all normal samples, the decision boundary can be trained by a class of support vector machine algorithms. The anomaly score calculated by OS-PCA is used as the input feature of OC-SVM. If the anomaly score of the data at a certain moment exceeds the normal boundary, it is judged to be abnormal, otherwise it is judged to be normal.

[0086] The present embodiment records a method for online abnormality detection of substation UAV flight data. By detecting whether the flight data of the UAV is abnormal during the operation and maintenance of the substation, it is judged whether the flight state of the UAV is safe. The abnormal data of the UAV may reflect potential faults, system errors or unexpected situations. Therefore, it is necessary to judge the abnormal state of the current data input as quickly and accurately as possible. The present invention first extracts a subset of high correlation from the flight data after data cleaning and normalization, and reconstructs the data of the subset of high correlation through OS-PCA. The reconstruction process is also a process of finding a low-dimensional subspace of the flight data. The subspace after dimensionality reduction is not only highly representative but also reduces the amount of online calculation of the flight data. If only the main direction of the subspace is considered, the online detection of flight data point abnormalities is realized by measuring the change of the main direction vector of the subspace; secondly, in order to adaptively update the abnormal score threshold and further reduce the false detection rate of abnormal detection, according to the characteristics that the data point at a certain moment can only be normal or abnormal, a class of support vector machines is selected to perform abnormal discrimination on the abnormal score.

[0087] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for online anomaly detection of substation UAV flight data, characterized by: The detection method comprises the following steps: S1. Using principal component analysis (PCA) method on multidimensional flight data matrix Find a low-dimensional internal subspace matrix by data reconstruction And calculate the main direction vector of the subspace, which will contain the maximum amount of information in the original data space; S2. Oversampling principal component analysis is used to amplify the influence of abnormal data on the main direction, while normal data has no influence on the main direction. The subspace vector is updated in a projection approximation way, and the main direction changes of the subspace vector are tracked and matched to realize instantaneous anomaly detection of flight data, and the anomaly score of each point is obtained; S3. When the UAV switches the mission mode, in order to adaptively update the anomaly score threshold and further reduce the false detection rate of anomaly detection, a class of support vector machines (OC-SVM) is selected to determine whether the data points are abnormal; Step S1 includes the following steps: S11. When the UAV performs operation and maintenance tasks in the substation, some key parameters are extremely important for the flight control of the UAV. The flight data matrix Including: longitude, latitude, altitude, pitch angle, roll angle, yaw angle, ground speed and air speed parameters of the drone. The flight parameters of the drone are monitored to detect potential anomalies. , multidimensional flight data matrix as follows: In the above formula, each row represents an n-dimensional flight data instance, and t is the number of instances; S12, PCA is mainly through the flight data matrix Reconstruction, in a multidimensional flight data matrix Find a low-dimensional interior subspace matrix in As a meaningful and lightweight data representation of the raw flight data, the subspace matrix The row vector of is the vectorized representation of the corresponding flight parameters in the original input, where ; S13. PCA essentially uses the square of the Euclidean norm as the distance metric and minimizes the reconstruction error of the projection data. The data reconstruction objective function is expressed as: （1） in, for , For the Raw flight data input, is the overall mean, To include The subspace matrix of the subspace vectors, For the The original data is projected into the data subspace, is the first estimate obtained from the projection subspace reconstruction indivual ; S14. Solve the above-mentioned minimization reconstruction problem by deriving the eigenvalues ​​of the covariance data matrix, as follows: （2） in, is the covariance matrix, is the mean, Each column of is a feature vector of The corresponding diagonal elements are the corresponding eigenvalues. The last few eigenvectors have little influence on the data distribution and will be discarded. Formula (1) is approximated by the least squares form as follows: （3） in, for The approximate value of exist In the dimensional space, based on formula (3), the reconstruction error has a quadratic form and is The function can be calculated by solving the least squares problem. In the online environment, it is assumed that the adjacent subspaces of the flight data in the time series direction have similarity, that is, .

2. The method for online anomaly detection of substation UAV flight data according to claim 1 is characterized by: The specific steps of step S2 include: The OS-PCA method is used for anomaly detection of drone flight data because the amount of normal data and abnormal data is unbalanced. The main idea of ​​this method is that oversampling of outliers will make the main direction of the data far away from the main direction of the original data, while oversampling of normal data points will have little effect on the change of the main direction. When drones perform inspection tasks in substations, they will contain a large amount of data, and resampling of data points is necessary. This can amplify the influence of the current data point on the main direction. The data input is represented as , then the objective function is: （4） In addition, if The impact before the moment can be reduced to amplify the current moment The effect on the main direction, that is, applying the forgetting factor ,at this time , is the number of oversampling The original dataset size The objective function is shown in formula (5), where the current moment The input is , and Use and approximate, （5） The optimization problem of solving equation (5) is the problem of data point incremental strategy anomaly detection. The input of the oversampling PCA anomaly detection algorithm is the flight data matrix X and the parameter , the algorithm output is the anomaly score, which is obtained by taking formula (5) The derivative of The solution is as follows: （6） For equation (6), if we only care about the main direction, that is, the subspace vector corresponding to the maximum eigenvalue is , Time reference subspace vector Zhang Cheng's subspace is , using the estimated most significant subspace vector and the most significant subspace vector of the reference , respectively, to the reference subspace and the estimated subspace Projection, the average of the angles calculated by formula (7) is The anomaly score of the input vector at time instant, （7） Among them, the anomaly score for The impact of the input vector on the data subspace vector at the moment. The larger the value, the greater the abnormality of the input vector. Through a large number of experiments, it is verified that when the abnormal score When it is greater than 0.5, it can be determined that an abnormality has occurred in a certain flight state of the drone.

3. The method for online anomaly detection of substation UAV flight data according to claim 2 is characterized by: The specific steps of step S3 include: OS-PCA uses the similarity characteristics of data subspace in the time series direction to provide a basis for anomaly detection in time series. It has the characteristics of lightweight calculation and storage and adaptability to dynamic changes in flight data. However, when the UAV switches mission modes, such as takeoff, climb, inspection, and landing, the collected flight data is different from historical data in normal performance. In order to adaptively update the anomaly score threshold and further reduce the false detection rate of anomaly detection, OC-SVM is selected to perform anomaly discrimination on the anomaly score based on the characteristic that a data point at a certain moment can only be either normal or abnormal. A type of support vector machine algorithm does not need to obtain the classification label of the data. It aims to construct a hyperplane or hypersphere to contain the distribution area of ​​most data points and isolate a small number of data points. For a given training data set, OC-SVM uses a mapping function The training data points Project the original feature space to a higher-dimensional space and establish a hyperplane in the mapped space ,in and Assume that the origin is an outlier in the OC-SVM model, and then you need to obtain a hyperplane to separate all training sample points from the point, and ensure that the distance between the hyperplane and the origin is as far as possible, that is, you need to optimize the distance from the origin to the hyperplane. is the maximum value, the hyperplane decision equation of a class of support vector machines is shown in formula (8): （8） in, and By any satisfaction A sample of support vectors Calculate and obtain, is a sign function that outputs 1 if x is greater than 0, otherwise it outputs 0.

Images