Method, device and equipment for abnormal detection and root cause analysis of unmanned aerial vehicle flight parameter data

Abstract

The present invention proposes a method, device and equipment for detecting and analyzing the abnormality of UAV flight parameter data, which overcomes the defects of high computational complexity, large generalization error and low algorithm accuracy of abnormality detection and root cause analysis in the prior art. The method includes: obtaining the causal relationship between parameters in the UAV flight parameter data; obtaining a data set; training the graph neural network by combining the graph structure with the causal relationship and optimizing the loss function of the graph neural network; establishing a causal root cause analysis optimization model by combining the binary classification model with the causal relationship between the parameters, and solving it, locating the UAV abnormality to the parameter, and finding the parameter causing the abnormality. The method steps of the present invention improve the accuracy of the abnormality detection of flight parameter data and the interpretation of the root cause analysis model, and reduce the computational complexity.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence and aviation, and in particular relates to a method, device and equipment for detecting and analyzing abnormalities in flight parameter data of unmanned aerial vehicles (UAVs). Background Art

[0002] Drone technology is becoming the leader of a new wave of technological innovation, and it has shown broad market potential and application value in many civil fields such as intelligent agriculture, logistics and transportation, emergency rescue and environmental protection. Due to operational errors and system failures, drone safety accidents occur frequently. The reason is that drones lack effective abnormality detection methods when performing tasks. Flight parameter data records the state variables of each system during the flight. It is a typical multi-dimensional time series data, including sensor data and flight control quantities of actuators. It is the key basis for abnormality detection, maintenance and accident analysis.

[0003] At present, with the rapid development of technologies such as the computer revolution and artificial intelligence, data-driven methods provide cases and extract rules for drone anomaly detection, including methods based on multivariate statistics, machine learning methods, and deep learning methods. However, most of these methods only deal with correlations and lack in-depth exploration of causal relationships, resulting in high computational complexity and insufficient accuracy of anomaly detection models. This is mainly because drone flight parameter data comes from various sensors of the subsystem and is composed of multiple indicator data closely related to the state of the subsystem. With the advancement of drone technology and the increase in the number of sensors, the collected drone non-parameters present high-dimensional characteristics. If the original high-dimensional data is directly used as the training set of the anomaly detection algorithm, on the one hand, the high dimensionality may cause "dimensionality disaster" and affect the efficiency of algorithm detection. On the other hand, too much irrelevant data leads to sparseness of training data in terms of dimension, which brings about large generalization errors and affects the accuracy of the algorithm.

[0004] At the same time, the cost of collecting abnormal flight parameter data of drones is also high, and it is difficult to label the data, which makes it difficult to obtain abnormal data and brings challenges to model building. As the complexity of drone missions and the harshness of the environment increase, the accuracy of state information acquisition may decrease, making it more difficult to obtain a large amount of labeled, high-quality abnormal data, which is difficult to meet the needs of traditional model building. Therefore, the existing technologies for drone anomaly detection are more focused on using unsupervised learning methods, which rely on the inherent information of the flight parameter data itself, but the high dimensionality and high coupling characteristics of the flight parameter data make the analysis work complicated, and the large amount of data also brings high computational complexity.

[0005] In addition, after detecting anomalies, no root cause analysis is provided for the anomalies, which cannot provide sufficient decision-making information for the subsequent handling of the anomalies. In addition, the mainstream of existing root cause analysis methods is also based on correlation, including association analysis, knowledge graphs, Bayesian networks, fuzzy cognition, etc. These analysis methods can identify factors related to anomalies, but these factors are not necessarily the root causes of anomalies, which can easily lead to misjudgment of the root cause, thereby affecting the accuracy of the root cause analysis results and the effectiveness of subsequent handling measures. In order to improve the accuracy of root cause analysis, it is necessary to propose a root cause analysis method specifically suitable for anomaly detection of UAV flight parameter data. Summary of the invention

[0006] The purpose of the present invention is to overcome the above-mentioned problems existing in the prior art, and to provide a method, device and equipment for anomaly detection and root cause analysis of UAV flight parameter data, so as to solve the problem of anomaly detection and root cause analysis of UAV flight parameter time series data in high-dimensional and massive measured data, and to improve the accuracy of anomaly detection and root cause analysis of flight parameter data while balancing the computational complexity.

[0007] The present invention proposes a method for detecting and analyzing abnormalities in UAV flight parameter data, including:

[0008] Step 110: Obtain the causal relationship between parameters in the high-dimensional and highly coupled flight parameter data of the UAV;

[0009] Step 120: Obtain a data set of flight parameter data, and sample a training set and a test set from the data set;

[0010] Step 130: construct a graph neural network based on a graph structure, perform a graph aggregation operation based on the causal relationship between the parameters, optimize the loss function of the graph neural network, train the graph neural network using the training set, and obtain a trained graph neural network;

[0011] Step 140: Input the test set into the trained graph neural network and perform anomaly detection on the flight parameter data in the test set;

[0012] Step 150: Establishing a causal root cause analysis optimization model using the binary classification model and the causal relationship between the parameters;

[0013] Step 160: By solving the causal root cause analysis optimization model, the flight parameter data determined to be abnormal data is located to the parameter, the root cause analysis result is generated, and the parameter causing the abnormal data is found.

[0014] Specifically, in step 110, the process of obtaining the causal relationship between the parameters includes:

[0015] Step 111: using the maximum mutual information coefficient to detect the correlation between the time series trends of the parameters in the nonlinear flight parameter data, and eliminating weakly correlated parameters according to the screening factor;

[0016] Step 112: For the parameters with weak correlation removed in step 111, the causal relationship is determined by using the transfer entropy method to obtain a causal relationship matrix.

[0017] Preferably, the step 130 includes:

[0018] Step 131: Construct a graph neural network by building a graph structure of flight parameter data, including:

[0019] The intrinsic attribute characteristics and temporal characteristics of the parameters are represented as nodes;

[0020] By introducing an embedding vector for each parameter to represent the intrinsic attribute characteristics of the parameter;

[0021] A time series decomposition method is used to capture the time series characteristics of parameters, so as to capture the dynamic information of time in the time dimension; the time series decomposition method uses two different linear layer strategies to capture the time series characteristics, and the linear layer strategies include: all parameters share the same linear layer, and each parameter has its own linear layer;

[0022] The intrinsic attribute characteristics and time series characteristics of the parameters are superimposed and represented as node characteristics;

[0023] After the causal relationship between the parameters is superimposed on the attention coefficient, a causally enhanced adjacency relationship is obtained as an edge of the graph structure;

[0024] Step 132: Based on the causal relationship between the parameters, a graph aggregation operation is performed using a graph neural network to integrate the temporal and spatial associations within the subsequence data in the training set or the test set, including:

[0025] The time series feature is used as the input of the graph neural network, and the graph neural network is used to aggregate and update each node and its neighboring nodes in the graph structure to obtain an output result generated by graph aggregation;

[0026] The output result of the graph aggregation is multiplied by the corresponding embedding vector representing the intrinsic attribute feature to obtain a product result; the product results of all nodes are input into a fully connected layer to predict flight parameter data;

[0027] Step 133: Optimizing the loss function of the graph neural network, including:

[0028] Two different loss functions are used according to the parameter type. For continuous parameters, the mean square error is used as the first loss function, which is given by the following formula:

[0029] ;

[0030] in, is the dimension of the continuous parameter, It is A continuous parameter at time The predicted value of It is A continuous parameter at time The observed value of yes The square of the norm;

[0031] For 0-1 type parameters, binary cross entropy is used as the second loss function, which is given by the following formula:

[0032] ;

[0033] in, is the dimension of the 0-1 type parameter, It is A 0-1 parameter at time The predicted value of It is A 0-1 parameter at time Observed value of

[0034] A constraint term is introduced to alleviate overfitting. The constraint term is given by the following formula:

[0035] ;

[0036] in, is the adjacency matrix of the graph structure, Represents the matrix singular values, and achieves low-rank constraints by suppressing high-order singular values;

[0037] The final total loss function of the graph neural network is obtained by taking the linear sum of the first loss function, the second loss function and the constraint term. , is given by:

[0038] ;

[0039] in, represents the Lagrange multiplier, which is used to balance the influence of the constraint terms; is the penalty parameter used to control the strength of the constraint;

[0040] Minimize the total loss function , build a loss function optimization model;

[0041] Step 134: Use an unsupervised anomaly detection method to train the graph neural network, including: superimposing the causal relationship between parameters with the attention coefficient to weaken the non-causal components in the time series features; using the augmented Lagrangian method to solve the loss function optimization model, updating the parameters through the AdamW optimizer, and obtaining accurate parameter predictions and a more reasonable graph structure through a learning method.

[0042] Furthermore, in step 150, the process of establishing a causal root cause analysis optimization model using the causal relationship between the binary classification model and the parameters includes:

[0043] Step 151: using a support vector machine to obtain a binary classification model, for finding the nearest counterfactual explanation that can make the classification result of the binary classification model less than 0.5;

[0044] Step 152: Construct a structural causal model to take action to transform the actual value of the parameter into the nearest counterfactual explanation through structural intervention;

[0045] Step 153: Define the intervention and obtain the structural causal model after the intervention;

[0046] Step 154: Using the structural causal model after the intervention, the optimization problem in the binary classification model is reformulated to obtain a causal root cause analysis optimization model.

[0047] On the other hand, the present invention provides a device for detecting and analyzing the abnormality of UAV flight parameter data and the root cause thereof, and the device is used to implement the steps of the aforementioned method for detecting and analyzing the abnormality of UAV flight parameter data and the root cause thereof, and the device comprises:

[0048] The first module is used to obtain the causal relationship between parameters in the high-dimensional and highly coupled flight parameter data of UAVs;

[0049] The second module is used to obtain a data set of flight parameter data, and sample a training set and a test set from the data set;

[0050] The third module is used to construct a graph neural network based on a graph structure, perform graph aggregation operations based on the causal relationship between the parameters, optimize the loss function of the graph neural network, train the graph neural network using the training set, and obtain a trained graph neural network;

[0051] The fourth module is used to input the test set into the trained graph neural network and perform anomaly detection on the flight parameter data in the test set;

[0052] The fifth module is used to establish a causal root cause analysis optimization model by using the binary classification model and the causal relationship between the parameters;

[0053] The sixth module is used to locate the flight parameter data determined as abnormal data to the parameters by solving the causal root cause analysis optimization model, generate the root cause analysis results, and find the parameters that cause the abnormal data.

[0054] In addition, the present invention also protects a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the aforementioned UAV flight parameter data anomaly detection and root cause analysis method are implemented.

[0055] In summary, the present invention provides a method, device and equipment for detecting and analyzing abnormalities in flight parameter data of unmanned aerial vehicles. Compared with the prior art, the technical problems solved by the technical solution of the present invention and the beneficial effects produced include:

[0056] 1) Comprehensively considering the relationship between the intrinsic attribute characteristics and time series characteristics of flight parameter data, the graph neural network based on graph structure is used to combine the time series feature extraction of time series with the causal relationship between parameters, so as to more comprehensively capture the intrinsic connection and time series characteristics of UAV flight parameters, and build an integrated framework to combine anomaly detection and root cause analysis, which not only detects anomalies but also provides explanations for their occurrence. This process not only saves computing costs by reducing the amount of calculation of the information transfer entropy part, but also improves the detection and analysis accuracy of the method.

[0057] 2) By adopting a causal analysis method based on maximum information transfer entropy, the maximum mutual information coefficient is used to establish the correlation of parameter scaffolds, and weak correlations are filtered through a screening mechanism to reduce the impact of irrelevant variables on modeling. The information transfer direction between strongly correlated variables is reflected through transfer entropy to form a one-way loop-free network. The screening of weak correlations also reduces the amount of calculation of the transfer entropy part.

[0058] 3) In the process of flight parameter data anomaly detection, by combining the graph structure with the causal relationship, causal enhancement in the anomaly detection process is achieved, thereby improving the accuracy of flight parameter data anomaly detection; in the process of abnormal root cause analysis of flight parameter data, through the backtracking and counterfactual interpretation of the structural causal model, and by combining it with the binary classification model of the support vector machine, a causal root cause analysis optimization model is constructed. Through model solving, the drone anomaly is located to the parameter, and the cause parameter of the anomaly is found, which improves the accuracy of the counterfactual interpretation of the root cause analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 A schematic diagram of a flow chart of a method for abnormal detection and root cause analysis of flight parameter data in a first embodiment of the present invention;

[0060] Figure 2This is a diagram of the graph structure and operation flow framework of the graph neural network for flight parameter data anomaly detection in the first embodiment of the present invention, wherein: , , , , , is an embedding vector, and Dlinear is a time series decomposition method;

[0061] Figure 3 Schematic diagram of super-mirror separation of the support vector machine model on the coordinate plane in the first embodiment of the present invention, wherein: is the separating hyperplane, and are two boundary hyperplanes, is the weight coefficient, is the input vector, is the bias term, is the distance between the two boundary hyperplanes, is the distance from the separating hyperplane to the origin. DETAILED DESCRIPTION

[0062] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0063] The present invention provides a method, device and equipment for detecting anomalies and performing root cause analysis on unmanned aerial vehicle (UAV) flight parameter data, which solves the problem of performing anomaly detection and root cause analysis on unmanned aerial vehicle (UAV) flight parameter time series data in high-dimensional and massive measured data.

[0064] In the first embodiment of the present invention, referring to Figure 1 , provides a method for detecting and analyzing abnormalities in UAV flight parameter data, including:

[0065] Step 110: Obtain the causal relationship between parameters in the high-dimensional and highly coupled flight parameter data of the UAV;

[0066] Step 120: Obtain a data set of flight parameter data, and sample a training set and a test set from the data set;

[0067] Step 130: construct a graph neural network based on a graph structure, perform a graph aggregation operation based on the causal relationship between the parameters, optimize the loss function of the graph neural network, train the graph neural network using the training set, and obtain a trained graph neural network;

[0068] Step 140: Input the test set into the trained graph neural network and perform anomaly detection on the flight parameter data in the test set;

[0069] Step 150: Establishing a causal root cause analysis optimization model using the binary classification model and the causal relationship between the parameters;

[0070] Step 160: By solving the causal root cause analysis optimization model, the flight parameter data determined to be abnormal data is located to the parameter, the root cause analysis result is generated, and the parameter causing the abnormal data is found.

[0071] Specifically, in step 110, the process of obtaining the causal relationship between the parameters includes two stages:

[0072] In the first stage, the correlation between the time series trends of the parameters in the nonlinear flight parameter data is detected using the maximum mutual information coefficient, and the weakly correlated parameters are eliminated according to the screening factors, including:

[0073] Given the time series data as the flight parameter data input ,in, is the length of the time series, is the number of parameters; let , For two time series of length Time series data and The mutual information (MI) value between can be given by:

[0074] (1)

[0075] in, and The parameters are and The variables that take values ​​in the probability density function, is a parameter and The joint probability density of and The parameters are and The marginal probability density of

[0076] The maximum mutual information coefficient is the highest normalized MI value obtained. Let , then the maximum mutual information coefficient (MIC) value is given by:

[0077] (2)

[0078] in, Represents the length of the time series Related parameters, and meet , The purpose is to balance computational efficiency and result accuracy. The growth rate of is faster than any constant function, but faster than The growth rate is slower; statistical results show that when When , the MIC value effect is better;

[0079] After further calculation, we can get MIC values ​​between parameters to obtain the correlation matrix ,in, Indicates Parameters and The correlation between the parameters is between 0 and 1.

[0080] Before calculating the transfer entropy, the screening factor Calculate the threshold Filter out weakly correlated variables, set the filter factor to the unit of the second significant digit, and filter the threshold The calculation is:

[0081]

[0082] in, is the screening factor, For rounding operations, Non-zero The mean of .

[0083] The second stage, The parameters of the process of determining causal relationships using the transfer entropy method include:

[0084] right , The lengths of the two time series are time series data, and Is has The Markov process of order Is has For a Markov process of order, we can get the following formula:

[0085] (3)

[0086] The above formula shows The next state With its own predecessor The same is true for The next state is the same as its previous state. state, , The lengths are and time series.

[0087] consider Towards Information conveyed:

[0088] (4)

[0089] It can be considered Status and The transition probability on The next state depending on Before oneself Status and Before states, then the formula for transfer entropy is defined as:

[0090] (5)

[0091] like ,but yes cause, causality ,on the contrary, yes The reason, , Respectively and The position in the parameter set after weak correlation screening, the causal relationship values ​​constitute the causal relationship matrix of all parameters (after screening) It can be clearly seen that the purpose of anomaly detection is to detect flight parameter data that deviates from the normal state, that is, abnormal data in the flight parameter data.

[0092] Specifically, in step 120, since each flight parameter data covers multiple key stages such as take-off, cruising, mission execution, return and landing, these stages together constitute the flight parameter data of the drone, which fully reflects a complete flight process. Therefore, all flight parameter data can be represented in the form of time series data. A dataset representing flight parameter data, ,in, is the length of the time series corresponding to multiple key stages, is the number of parameters.

[0093] First, in involving A dataset of all flight parameters On, sampling time samples, and obtain multivariate time series data as the training set of the graph neural network , the flight parameter data of the training set is recorded as:

[0094] (6)

[0095] Among them, each moment , flight data , forming a dimensional vector, indicating that the records are Moment The flight parameter data of each parameter; the flight parameter data used as a training set is composed only of normal data in the data set.

[0096] The goal of graph neural network anomaly detection is to detect anomalies in the flight parameter data that constitute the test set. These anomalies come from the Specifically, the test set is recorded as:

[0097] (7)

[0098] in, represents the sampling time of the test set.

[0099] After the anomaly detection process of the method described in this embodiment, the output data is a set of length The binary label ( , used to represent each test moment Is it an exception? ,in, , indicating that at time The flight parameter data is normal; , indicating that at time The flight parameter data is abnormal.

[0100] Specifically, Figure 2 As shown, in step 130, a graph neural network based on a graph structure is constructed, and the graph neural network is trained using the training set to obtain a trained graph neural network, including:

[0101] Step 131: Build the graph structure of flight parameter data, such as Figure 2 As shown, the parameters are taken as nodes of the graph structure, the intrinsic attribute characteristics and time series characteristics of the parameters are represented as node characteristics, and after the causal relationship between the parameters in the time window is superimposed with the attention coefficient, the causal enhanced adjacency relationship is obtained as the edge of the graph structure;

[0102] Specifically, the main means of anomaly detection in this embodiment includes learning the relationship between parameters in the form of a graph structure, that is, the structural learning process in the graph neural network. To achieve this, a directed graph is used to construct the graph structure, and its nodes represent parameters. In order to achieve subsequent feedback, performance differentiation and updating, each parameter exhibits both intrinsic attribute characteristics and timing characteristics within a time window, and the intrinsic attribute characteristics and timing characteristics of the parameter are represented as node features. After extracting the node features, they are recorded in a node feature matrix, and all the node features are normalized for better information transmission.

[0103] First, an embedding vector is introduced for each parameter to represent its intrinsic attribute characteristics. The embedding vector is given by the following formula:

[0104] (8)

[0105] Among them, the embedding vector is randomly initialized and then participates in subsequent training with the rest of the graph neural network. The similarity of the embedding vectors is used to represent the similarity of the parameters corresponding to the UAV flight behavior, that is, parameters with similar embedding vectors should have a high tendency to be correlated with each other. Based on this, ideally, this embodiment adopts a flexible way to represent each parameter to capture the intrinsic attribute characteristics of the UAV flight behavior in a multi-dimensional way.

[0106] In graph neural networks, these embedding vectors are used in two ways:

[0107] (i) For structural learning, by learning node embedding representations to better capture the topological structure of the graph and the relationships between nodes, including graph classification and clustering;

[0108] (ii) In the attention mechanism, attention is paid to neighboring nodes in a way that allows different types of parameters to have heterogeneous influences.

[0109] The capture of time series features is to capture the dynamic information of time in the time dimension. Flight parameter data has unique time series features. The data features and abnormal representations at different stages are very different. Therefore, a simple time series decomposition method (Dlinear) is used to capture time series features. The Dlinear decomposes the time series of the original flight parameter data into Decomposed into moving average classification quantity and stage trend component, that is, ,in, is the phase trend component, is the moving average classification quantity, " represents the addition operation of the vector; then two single-layer linear layers are applied to each component respectively, and the two component features are added together to obtain the time series features:

[0110] (9)

[0111] (10)

[0112] (11)

[0113] in, and Represents the moving average classification quantity and phase trend component The component features obtained after the single linear layer is applied respectively, and Respectively for and The weighting coefficient of .

[0114] Since each key stage of the flight parameter data has different characteristic manifestations, that is, different stage characteristics and trends, sharing weights among different parameters may not perform well. Therefore, a different parameter-independent linear layer strategy is proposed, that is, each parameter has its own linear layer.

[0115] The edges of the graph structure represent the dependencies between parameters. This dependency is related to the causal relationship between parameters and the influence weight of the causal relationship, that is, the causal attention coefficient. The reason for using a directed graph is that the dependency between parameters does not need to be symmetric. Use an adjacency matrix To represent this directed graph, is a matrix The elements in represent the slave nodes To Node The directed edge of Parameter modeling The behavior of a parameter.

[0116] In order to facilitate the unified discussion of the mechanism of training set and test set in graph neural network, we use Indicates the sampling time, for node With Node The intrinsic attribute characteristics and time series characteristics of and Respectively represent nodes With Node The above eigenvectors of . or The subsequence data set of the network is defined as time and length Historical subsequence data of the sliding time window:

[0117] (12)

[0118] for time and length The sliding time window, the attention coefficient is Indicates that the calculation process is as follows:

[0119] (13)

[0120] (14)

[0121] in, Is a node The input features of Is a node The reason parameter set, represents a trainable weight matrix, the ReLU is an activation function, , used to introduce nonlinearity, is the initial attention value, which is normalized by formula (14) to obtain the final attention coefficient .

[0122] Step 132: Use the graph neural network to perform graph aggregation operations, such as Figure 2 As shown, the graph aggregation operation is used to integrate the temporal and spatial associations within the subsequence data.

[0123] Specifically, the timing characteristics As the input of the graph neural network, each node and its neighbor nodes are aggregated and updated, such as The final output is expressed as :

[0124] (15)

[0125] in, Is a node The time series characteristics of the input, Is a node With Node The adjacency relationship, Is a node A collection of reason parameters.

[0126] Through the above graph aggregation process, we get all The output of a node is expressed as . For each node, the graph aggregation output results , multiply it by the corresponding embedding vector representing the intrinsic attribute features , get the product result, and then input the product result of all nodes into a fully connected layer to predict the time Flight parameter data:

[0127] (16)

[0128] in," " represents the Hadamard product of vector element-wise multiplication, represents the fully connected layer as a prediction function, The parameterization of the fully connected layer includes the weight and bias of the fully connected layer. Indicates time The predicted flight parameters data.

[0129] Step 133: Optimizing the loss function of the graph neural network, specifically including: using two different loss functions according to the parameter type, using the mean square error as the first loss function for continuous parameters, and setting For the A continuous parameter at time The predicted value of For the A continuous parameter at time The first loss function is given by:

[0130] (17)

[0131] in, is the dimension of the continuous parameter, yes The square of the norm.

[0132] For 0-1 type parameters, binary cross entropy is used as the second loss function, assuming For the A 0-1 parameter at time The predicted value of For the A 0-1 parameter at time The second loss function is given by:

[0133] (18)

[0134] in, is the dimension of the 0-1 type parameter.

[0135] Considering the sparsity of the graph structure based on causality, a constraint is introduced to reduce overfitting:

[0136] (19)

[0137] in, is the adjacency matrix of the graph structure, Represents the matrix singular values, and low rank constraints are achieved by suppressing high-order singular values.

[0138] Take the linear sum of the first loss function, the second loss function and the constraint term to obtain the final total loss function of the graph neural network , which is given by the following formula:

[0139] (20)

[0140] in, represents the Lagrange multiplier, which is used to balance the influence of the constraint term. is a penalty parameter used to control the strength of the constraint term by minimizing the total loss function A loss function optimization model is obtained, and the loss function optimization model is solved by the augmented Lagrangian method.

[0141] Step 134: Use an unsupervised anomaly detection method to train the graph neural network, including: superimposing the causal relationship between parameters and the attention coefficient to weaken the non-causal components in the features, minimizing the loss function as the goal, using the AdamW optimizer to update the parameters, and obtaining accurate parameter predictions and a reasonable graph structure through learning.

[0142] Furthermore, in step 140, the test set is input into the trained graph neural network, and the process of performing anomaly detection on the flight parameter data in the test set is achieved by calculating a separate outlier value for each parameter and then merging them into isolated outliers at each moment. The graph neural network needs to design outlier determination rules specifically according to the characteristics of the network itself. The anomaly detection process specifically includes:

[0143] Step 141: Calculate the time The difference between the predicted value and the observed value :

[0144] (twenty one)

[0145] Step 142: To prevent any parameter from producing a deviation that exceeds that of the other parameters, the error value of each parameter is After normalization, we get :

[0146] (twenty two)

[0147] in, for The median value, yes The median and IQR are used instead of the mean and standard deviation because they are more robust.

[0148] Step 143: During the flight of the drone, the value of the parameter may suddenly change, which may lead to a sharp increase in erroneous data even when the flight behavior is normal. In order to suppress such sudden changes, the statistical tool SMA (Simple Moving Average) is used to generate a smooth score , if the smoothing score If a given fixed threshold is exceeded, then a moment The flight parameter data will be marked as abnormal data.

[0149] Specifically, in step 150, the process of establishing a causal root cause analysis optimization model using the causal relationship between the binary classification model and the parameters includes:

[0150] Step 151: Use support vector machine (SVM) to obtain a binary classification model , find the nearest counterfactual explanation that can make the classification result of the binary classification model less than 0.5, including:

[0151] The set of parameters corresponding to the flight parameter data is used Indicates that the corresponding abnormal data is expressed as By constructing a binary classification model , using the causal tracing method to study How to select values ​​so that anomalies do not occur, that is, abnormal data does not appear. The causal tracing method uses support vector machine (SVM) so that The value of is less than 0.5.

[0152] Support vector machine is a typical binary classification model. The basic model is a classifier with the largest interval defined in the feature space. The purpose is to find a hyperplane to divide the samples so that the interval is maximized. The basic idea of ​​SVM is to solve the separation hyperplane that can correctly divide the training data set of the binary classification model and has the largest geometric interval. Figure 3 As shown, , is the separating hyperplane, where is the weight coefficient, is the input vector, For a linearly separable data set, there are infinitely many such hyperplanes, but the separating hyperplane with the largest geometric interval is unique.

[0153] The problem of constructing a binary classification model is described as finding the closest counterfactual explanation, which is defined as: "How different the results would be in order to achieve the desired results". The purpose is to reveal the changes that need to be made to achieve the desired results by comparing the current results with the desired results. The goal of the counterfactual explanation is to find a set of input features that are as close as possible to the actual observations, but will cause the model to predict a different category than the actual results. This method can help understand the decision boundary of the model and reveal which changes in features may lead to changes in the predicted results. Therefore, finding the closest counterfactual explanation in the root cause analysis of flight parameter data anomalies can be expressed as the following optimization problem:

[0154] (twenty three)

[0155] in, Represents the actual value of the parameter, which is the original input data received by the binary classification model; represents the most recent counterfactual explanation; It is a binary classification model; represents the value space of the parameter; the dist() function represents the distance between variable data; the argmin function is used to obtain the value of the parameter that minimizes the objective function. By solving this optimization problem, we can find the nearest counterfactual explanation that can make the classification result of the binary classification model less than 0.5.

[0156] Step 152: Build a structural causal model to take action through structural interventions. become , including:

[0157] From a causal perspective, action can be carried out through structural interventions, through structural causal models represents the causal relationship between parameters, where Represents a collection of parameters, represents an exogenous variable, represents the structural equation and is given by:

[0158] (twenty four)

[0159] In the formula, is the function in the structural equation, Indicates the number of parameters. is the parameter in the causal relationship The parent node of is the component of the exogenous variable; the structural equation quantitatively describes the causal relationship between the indicators. The structural causal model is usually represented by its related causal diagram, which is constructed by combining the results of the causal analysis in step 110. The parent node is the node that directly affects Those parameters.

[0160] Step 153: Define the intervention and obtain the structural causal model after the intervention, including:

[0161] The interventions corresponding to the actions can simulate the changes in the intervention parameters and the impact of the intervention on its downstream (non-intervention) parameters. Interventions can be considered as transformations between structural causal models. The set of interventions can be constructed as:

[0162] (25)

[0163] in, Contains the index of the subset of parameters that are intervened, for each , Operation, representing the execution of intervention operations: replace .

[0164] Intervening in a parameter is equivalent to cutting off all edges pointing to it. Implementation of actions Produces a structure equation , and obtain the structural causal model after intervention , whose structural equation is given by:

[0165] (26)

[0166] Step 154: Using the structural causal model after intervention, the optimization problem in the binary classification model is reformulated to obtain a causal root cause analysis optimization model, including:

[0167] Intervention is used to predict the impact of actions on results, and to determine whether the expected results are achieved after the actions are implemented. Therefore, the optimization problem in the binary classification model can be restated as the following causal root cause analysis optimization model:

[0168] (27)

[0169] (28)

[0170] (29)

[0171] In the formula, represents the optimal intervention, which achieves the desired effect at the lowest cost; Indicates the cost of taking action; represents all possible intervention spaces; The value of the exogenous variable can be calculated; Representing the obtained structural counterfactual, we can know the values ​​of each parameter after the action is taken.

[0172] Specifically, in step 160, by solving the causal root cause analysis optimization model, locating the flight parameter data determined as abnormal data to the parameter, generating the root cause analysis result, and finding the parameter causing the abnormal data, the process includes:

[0173] Step 161: Obtain the values ​​of each parameter corresponding to when the action occurs. The parameter value corresponding to the abnormal state is 1, that is, the abnormality occurs; obtain the time before the abnormality occurs, and obtain the parameter value corresponding to the time. The parameter value corresponding to the normal state is 0, that is, the abnormality does not occur;

[0174] Step 162: Based on the causal graph, the aforementioned structural causal model is reconstructed using Gaussian process to obtain a Gaussian process structural causal model (GP-SCM), and the parameters therein are intervened in step 152, and the GP-SCM after intervention is used to calculate the values ​​of the parameters after the intervention, that is, the counterfactual distribution.

[0175] Specifically, the GP-SCM Defined by:

[0176] (30)

[0177] in, is the covariance function, such as continuous The kernel of the radial basis function (RBF) is:

[0178] (31)

[0179] also, is a normal variance, exogenous variable Follows normal distribution, the function in the structural equation They are all Gaussian processes.

[0180] For a single , calculated using the GP-SCM after intervention The counterfactual distribution of :

[0181] (32)

[0182] in:

[0183] (33)

[0184] represents the Gram matrix, is the posterior mean, is the variance.

[0185] Step 163: Solve the causal root cause analysis optimization model through an optimization algorithm to obtain the cause parameters that lead to the abnormality and what kind of intervention can prevent the event from occurring (i.e., whether the abnormality can be effectively prevented by intervening in the parameter).

[0186] There is a wide range of intervention options available, and the best intervention must meet the following three conditions:

[0187] 1) The best intervention needs to be able to successfully prevent the causal event from occurring;

[0188] 2) When selecting intervention parameters, it is necessary to ensure that the selected parameters are easy to intervene;

[0189] 3) The value of each intervention parameter must be within the normal value range of the parameter, and the closer to the original value of the parameter, the better.

[0190] use Indicates whether to intervene in the parameter index. A value of 0 means intervention, and a value of 1 means no intervention. It is expressed in an array. Indicates the difficulty of intervening in the parameter indicators. The smaller the value, the simpler and easier it is to operate. represents the numerical value of each intervention parameter, The value of the parameter needs to be within the normal range.

[0191] Assume that the objective function of the optimization algorithm is , Specifically by , and It consists of three sub-objective functions, corresponding to the three conditions for optimal intervention.

[0192] 1) After intervening in the parameters, the GP-SCM model is used to obtain the counterfactual distribution values ​​of each parameter, assuming that it is recorded as ,Will Input the SVM model to obtain the probability of abnormal occurrence. The smaller the better. Given by:

[0193] (34)

[0194] 2) The intervention parameters should be easy to operate, so Given by:

[0195] (35)

[0196] 3) Set represents the original value of the intervened parameter, let , represents the intervention value. The intervention value should be as close to the original value of the parameter as possible. As shown below:

[0197] (36)

[0198] Furthermore, the objective function of the optimization algorithm is:

[0199] (37)

[0200] in, Represents the weight of each sub-objective function.

[0201] Generate a large number of and The initial value of The constraints are as follows:

[0202] 1) It is a binary array with values ​​of 0 and 1;

[0203] 2) It is a real number array, and the value must be within the normal range of the parameter.

[0204] In the second embodiment of the present invention, in step 163, the optimization algorithm used to solve the causal root cause analysis optimization model is the particle swarm optimization algorithm, which is a random search algorithm based on group assistance. It randomly initializes a group of random particles, finds the optimal solution through iteration, and updates itself by tracking two extreme values ​​in each iteration. By proposing a search-adaptive particle swarm algorithm optimization function, the initial population (initial value) of the optimization algorithm is the candidate intervention, and the candidate intervention includes intervention parameters and intervention values. By randomly generating intervention parameters and intervention values ​​according to the above rules, the structural causal model after intervention can be calculated. By judging whether the convergence conditions are met, the optimal intervention , the optimal intervention parameter at this time This is the reason that triggered the exception.

[0205] The above two embodiments comprehensively consider the relationship between the intrinsic attribute characteristics and time series characteristics of the flight parameter data, and use the graph neural network based on the graph structure to combine the time series feature extraction of the time series with the causal relationship between the parameters, so as to more comprehensively capture the intrinsic connection and time series characteristics of the UAV flight parameters. By building an integrated framework, anomaly detection and root cause analysis are combined to not only detect anomalies, but also provide explanations for the occurrence of anomalies. This process not only saves computing costs by reducing the amount of calculation of the information transfer entropy part, but also improves the detection and analysis accuracy of the method.

[0206] Specifically, a causal analysis method based on maximum information transfer entropy is adopted, the maximum mutual information coefficient is used to establish the correlation of parameter scaffolds, and weak correlations are filtered through a screening mechanism to reduce the impact of irrelevant variables on modeling. The information transfer direction between strongly correlated variables is reflected through transfer entropy to form a one-way loop-free network. The screening of weak correlations also reduces the amount of calculation of the transfer entropy part.

[0207] Secondly, in the process of flight parameter data anomaly detection, by combining the graph structure with the causal relationship, causal enhancement in the anomaly detection process is achieved, thereby improving the accuracy of flight parameter data anomaly detection; in the process of abnormal root cause analysis of flight parameter data, through the backtracking and counterfactual interpretation of the structural causal model and combining it with the binary classification model of the support vector machine, a causal root cause analysis optimization model is constructed. Through model solving, the drone anomaly is located to the parameter, and the cause parameter of the anomaly is found, which improves the accuracy of the counterfactual interpretation of the root cause analysis.

[0208] In a third embodiment, the present invention provides a device for detecting and analyzing abnormalities in flight parameter data of a UAV, and the steps of the method described in the first embodiment are implemented by using the device, and the device includes:

[0209] The first module is used to obtain the causal relationship between parameters in the high-dimensional and highly coupled flight parameter data of UAVs;

[0210] The second module is used to obtain a data set of flight parameter data, and sample a training set and a test set from the data set;

[0211] The third module is used to construct a graph neural network based on a graph structure, perform graph aggregation operations based on the causal relationship between the parameters, optimize the loss function of the graph neural network, train the graph neural network using the training set, and obtain a trained graph neural network;

[0212] The fourth module is used to input the test set into the trained graph neural network and perform anomaly detection on the flight parameter data in the test set;

[0213] The fifth module is used to establish a causal root cause analysis optimization model by using the binary classification model and the causal relationship between the parameters;

[0214] The sixth module is used to locate the flight parameter data determined as abnormal data to the parameters by solving the causal root cause analysis optimization model, generate the root cause analysis results, and find the parameters that cause the abnormal data.

[0215] The present invention also provides, in one embodiment, a computer device, including a memory and a processor, the memory storing a computer program, and the processor implementing the steps of the flight parameter data anomaly detection and root cause analysis method described in any of the first embodiments when executing the computer program. The computer device may be a server. The computer device includes a processor, a memory, a network interface, and a database connected via a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store sample data. The network interface of the computer device is used to communicate with an external terminal via a network connection.

[0216] In addition, in another embodiment of the present invention, a storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the flight parameter data anomaly detection and root cause analysis method described in the first embodiment are implemented.

[0217] In a typical configuration, the computer device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0218] The memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.

[0219] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, disk storage or other magnetic storage devices or any other non-transmission media for storing information accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0220] A person of ordinary skill in the art can understand that all or part of the process of implementing the aforementioned embodiment method can be completed by instructing related hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the embodiment process of the flight parameter data anomaly detection and root cause analysis method.

[0221] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0222] The above-mentioned embodiments only express several implementation modes of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the invention. It should be pointed out that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the attached claims.

Claims

1. A method for detecting and analyzing abnormalities in UAV flight parameter data, characterized in that: include: Step 110: Obtain the causal relationship between parameters in the high-dimensional and highly coupled flight parameter data of the UAV; Step 120: Obtain a data set of flight parameter data, and sample a training set and a test set from the data set; Step 130: construct a graph neural network based on a graph structure, perform a graph aggregation operation based on the causal relationship between the parameters, optimize the loss function of the graph neural network, train the graph neural network using the training set, and obtain a trained graph neural network; Step 140: Input the test set into the trained graph neural network and perform anomaly detection on the flight parameter data in the test set; Step 150: Establishing a causal root cause analysis optimization model using the binary classification model and the causal relationship between the parameters; Step 160: By solving the causal root cause analysis optimization model, the flight parameter data determined to be abnormal data is located to the parameter, the root cause analysis result is generated, and the parameter causing the abnormal data is found.

2. The method for detecting and analyzing abnormalities in UAV flight parameter data according to claim 1, characterized in that: In step 110, the process of obtaining the causal relationship between the parameters includes: Step 111: using the maximum mutual information coefficient to detect the correlation between the time series trends of the parameters in the nonlinear flight parameter data, and eliminating weakly correlated parameters according to the screening factor; Step 112: For the parameters with weak correlation removed in step 111, the causal relationship is determined by using the transfer entropy method to obtain a causal relationship matrix.

3. The method for detecting and analyzing abnormalities in UAV flight parameter data according to claim 2, characterized in that: In step 120, a data set of flight parameter data is obtained, and a process of sampling a training set and a test set from the data set includes: Step 121: A data set of UAV flight parameter data is formed by collecting data of multi-stage flight parameters of the UAV take-off, cruising, mission execution, return and landing; Step 122: sampling time samples in the data set of flight parameter data to obtain multivariate time series data as a training set; the training set is composed only of normal data in the data set of flight parameter data; Step 123: Sampling a test set from the data set of flight parameter data, wherein the test set includes abnormal data in the data set; using a set of binary labels with the same length as the test set to indicate whether each test moment is abnormal, if the label of a test moment is 0, it indicates that the flight parameter data at the test moment is normal, and the label is 1, it indicates that the flight parameter data at the test moment is abnormal.

4. The method for detecting and analyzing abnormalities in UAV flight parameter data according to claim 3 is characterized in that: The step 130 includes: Step 131: Construct a graph neural network by building a graph structure of flight parameter data, including: The intrinsic attribute characteristics and temporal characteristics of the parameters are represented as nodes; By introducing an embedding vector for each parameter to represent the intrinsic attribute characteristics of the parameter; A time series decomposition method is used to capture the time series characteristics of parameters, so as to capture the dynamic information of time in the time dimension; the time series decomposition method uses two different linear layer strategies to capture the time series characteristics, and the linear layer strategies include: all parameters share the same linear layer, and each parameter has its own linear layer; The intrinsic attribute characteristics and time series characteristics of the parameters are superimposed and represented as node characteristics; After the causal relationship between the parameters is superimposed on the attention coefficient, a causally enhanced adjacency relationship is obtained as an edge of the graph structure; Step 132: Based on the causal relationship between the parameters, a graph aggregation operation is performed using a graph neural network to integrate the temporal and spatial associations within the subsequence data in the training set or the test set, including: The time series feature is used as the input of the graph neural network, and the graph neural network is used to aggregate and update each node and its neighboring nodes in the graph structure to obtain an output result generated by graph aggregation; The output result of the graph aggregation is multiplied by the corresponding embedding vector representing the intrinsic attribute feature to obtain a product result; the product results of all nodes are input into a fully connected layer to predict flight parameter data; Step 133: Optimizing the loss function of the graph neural network, including: Two different loss functions are used according to the parameter type. For continuous parameters, the mean square error is used as the first loss function, which is given by the following formula: ； in, is the dimension of the continuous parameter, It is A continuous parameter at time The predicted value of It is A continuous parameter at time The observed value of yes The square of the norm; For 0-1 type parameters, binary cross entropy is used as the second loss function, which is given by the following formula: ； in, is the dimension of the 0-1 type parameter, It is A 0-1 parameter at time The predicted value of It is A 0-1 parameter at time Observed value of A constraint term is introduced to alleviate overfitting. The constraint term is given by the following formula: ； in, is the adjacency matrix of the graph structure, Represents the matrix singular values, and achieves low-rank constraints by suppressing high-order singular values; The final total loss function of the graph neural network is obtained by taking the linear sum of the first loss function, the second loss function and the constraint term. , is given by: ； in, represents the Lagrange multiplier, which is used to balance the influence of the constraint terms; is the penalty parameter used to control the strength of the constraint; Minimize the total loss function , build a loss function optimization model; Step 134: Use an unsupervised anomaly detection method to train the graph neural network, including: superimposing the causal relationship between parameters with the attention coefficient to weaken the non-causal components in the time series features; using the augmented Lagrangian method to solve the loss function optimization model, updating the parameters through the AdamW optimizer, and obtaining accurate parameter predictions and a more reasonable graph structure through a learning method.

5. The method for detecting and analyzing abnormalities in UAV flight parameter data according to claim 4, characterized in that: In step 140, the test set is input into the trained graph neural network, and the process of performing anomaly detection on the flight parameter data in the test set includes: Step 141: Calculate the time The error between the predicted value and the observed value : ； Step 142: Error value for each parameter Normalization is performed to prevent the deviation of any parameter from exceeding that of other parameters. The normalized error of the parameter error value is Given by the following formula: ； in, for The median value, yes The interquartile range of the values, using the median and interquartile range instead of the mean and standard deviation, makes the normalization error more robust; Step 143: Generate smoothed scores using the statistical tool SMA , used to suppress data mutation; if the smoothing score If a fixed threshold is exceeded, then The flight parameter data will be marked as abnormal data.

6. The method for detecting and analyzing abnormalities in UAV flight parameter data according to claim 5, characterized in that: In step 150, the process of establishing a causal root cause analysis optimization model using the causal relationship between the binary classification model and the parameters includes: Step 151: Using a support vector machine to obtain a binary classification model, for finding the nearest counterfactual explanation that can make the classification result of the binary classification model less than 0.5, including: By constructing a binary classification model , select parameters and filter abnormal events by causal tracing method; the causal tracing method adopts support vector machine to find a hyperplane to divide the samples so as to maximize the interval and ensure The value of is less than 0.5; The problem of constructing a binary classification model is described as finding the closest counterfactual explanation, which is used to reveal the intervention actions needed to achieve the expected results by comparing the current results with the expected results. Finding the closest counterfactual explanation in the abnormal root cause analysis of flight parameter data is expressed as the following optimization problem: ； in, The actual value of the representative parameter is the original input data received by the binary classification model; represents the most recent counterfactual explanation; It is a binary classification model; Indicates the value space of the parameter; the dist() function indicates the distance between variable data; the argmin function is used to obtain the value of the parameter that minimizes the objective function; By solving the above optimization problem, we can find the nearest counterfactual explanation that can make the classification result of the binary classification model less than 0.

5. Step 152: Construct a structural causal model to take action to transform the actual value of the parameter into the nearest counterfactual explanation through structural intervention, including: Building a structural causal model , used to express the causal relationship between parameters, where Represents a collection of parameters, represents an exogenous variable, represents the structural equation and is given by: ； In the above formula, is the function in the structural equation, Indicates the number of parameters. is the parameter in the causal relationship The parent node of is the component of the exogenous variable; the structural equation quantitatively describes the causal relationship between the indicators; The structural causal model is represented by a corresponding causal graph, which is constructed by combining the causal relationships obtained in step 110. The parent node is the node that directly affects Those parameters; Step 153: Define the intervention and obtain the structural causal model after the intervention, including: Construct a collection of interventions: ； in, Contains the index of the subset of parameters that are intervened, for each , Operation, representing the execution of intervention operations: replace ; Intervening in a parameter is equivalent to cutting off all edges pointing to it; In structural causal models Implementing intervention , which produces a structure equation , and obtain the structural causal model after intervention , Given by: Step 154: Using the structural causal model after the intervention, the optimization problem in the binary classification model is reformulated to obtain a causal root cause analysis optimization model: in, represents the best intervention for achieving the desired effect at the lowest cost; Indicates the cost of taking action; represents all possible intervention spaces; Calculate the value of the exogenous variable; Represents the obtained structural counterfactual, obtaining the values ​​of each parameter after the action is taken.

7. The method for detecting and analyzing abnormalities in UAV flight parameter data according to claim 6, characterized in that: In step 160, by solving the causal root cause analysis optimization model, locating the flight parameter data determined as abnormal data to the parameter, generating the root cause analysis result, and finding the parameter causing the abnormal data, the process includes: Step 161: Obtain the values ​​of each parameter corresponding to when the action occurs. The parameter value is 1, indicating that an exception occurs; obtain the time before the exception occurs, and obtain the value of the parameter corresponding to the time; the parameter value is 0, which corresponds to the normal state, indicating that the exception does not occur; Step 162: Based on the causal graph, the aforementioned structural causal model is reconstructed using a Gaussian process to obtain a Gaussian process structural causal model, and the parameters therein are intervened as described in step 152. The values ​​of the parameters after the intervention are calculated using the Gaussian process structural causal model after the intervention to obtain the counterfactual distribution: ； in, is the value of the parent node under intervention, express It follows a normal distribution. is the component of the exogenous variable The variance of the normal distribution; represents the Gram matrix, is the identity matrix, is the posterior mean, is the variance, is the autocorrelation matrix, ； is the covariance function, including continuous parent nodes The kernel of the radial basis function is: ； Step 163: Solving the causal root cause analysis optimization model through an optimization algorithm to obtain the parameters that cause the abnormality to occur and what kind of intervention can be used to prevent the event from occurring, including: Establishing the objective function of the optimization algorithm , is given by: ； in, , and There are three sub-objective functions, Represents the weight of each sub-objective function; , and They correspond to the three constraints of optimal intervention: (i) The optimal intervention needs to be able to successfully prevent the occurrence of the cause event, and the sub-objective function Given by: ； in, It represents the counterfactual distribution value of each parameter obtained by using the Gaussian process structural causal model after the parameter is intervened. Indicates that Input the binary classification model and obtain the probability of abnormal occurrence. The smaller the probability, the better. (ii) When selecting intervention parameters, it is necessary to ensure that the selected parameters are easy to intervene. Given by: ； in, Indicates whether to intervene in the parameter index. The component value of 0 indicates intervention, and the value of 1 indicates no intervention; (iii) Constrain the values ​​of each intervention parameter to be within the normal range of the parameter, and the closer to the original value of the parameter, the better; sub-objective function Given by: ； in, , Indicates the original value of the intervened parameter. represents the numerical value of each intervention parameter, The value of is a real number and needs to be within the normal range of the parameter; Generate according to the constraints and The initial value of Iterative optimization is performed under constraints for the best intervention.

8. The method for detecting and analyzing abnormalities in UAV flight parameter data according to claim 7, characterized in that: The optimization algorithm adopts the particle swarm optimization algorithm. By proposing a search-adaptive particle swarm optimization function, the initial population of the optimization algorithm is used as a candidate intervention. The candidate intervention includes intervention parameters and intervention values. According to the above constraints, the intervention parameters and intervention values ​​are randomly generated, and the structural causal model after intervention is used to calculate ; By judging whether the convergence conditions of the causal root cause analysis optimization model are met, the optimal intervention is obtained , the corresponding optimal intervention parameter This is the reason that triggered the exception.

9. A device for detecting abnormalities in flight parameter data of unmanned aerial vehicles and analyzing root causes, characterized in that: The device is used to implement the steps of the method for abnormal detection and root cause analysis of UAV flight parameter data as claimed in claim 1, and the device includes: The first module is used to obtain the causal relationship between parameters in the high-dimensional and highly coupled flight parameter data of UAVs; The second module is used to obtain a data set of flight parameter data, and sample a training set and a test set from the data set; The third module is used to construct a graph neural network based on a graph structure, perform graph aggregation operations based on the causal relationship between the parameters, optimize the loss function of the graph neural network, train the graph neural network using the training set, and obtain a trained graph neural network; The fourth module is used to input the test set into the trained graph neural network and perform anomaly detection on the flight parameter data in the test set; The fifth module is used to establish a causal root cause analysis optimization model by using the binary classification model and the causal relationship between the parameters; The sixth module is used to locate the flight parameter data determined as abnormal data to the parameters by solving the causal root cause analysis optimization model, generate the root cause analysis results, and find the parameters that cause the abnormal data.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

Images