Unmanned aerial vehicle fault important characteristic parameter optimal selection method and system based on causal reasoning and medium

Abstract

The invention provides an unmanned aerial vehicle fault important characteristic parameter optimal selection method and system based on causal reasoning and a medium, and the method comprises the steps: S1, collecting fault data which comprises flight fault data collected by a gyroscope, a GPS, an airspeed meter and the like; s2, performing data processing on the fault data; s3, causal structure learning: performing causal structure learning on the data after data processing through a VAE decoder and a VAE encoder in sequence to obtain a weighted causal DAG representing a causal relationship between the features; and S4, important fault feature selection: after the weighted causal DAG is obtained, node scores are calculated, then the node scores are sorted in a descending order, the top F fault feature variables are selected as feature subsets, and F is a positive integer. The method can effectively screen out the characteristic parameters of the fault of the unmanned aerial vehicle, can be used for fault diagnosis, facilitates the reduction of the fault frequency of the unmanned aerial vehicle, and improves the safety and reliability of the unmanned aerial vehicle.

Description

Technical Field

[0001] The present invention relates to the field of feature selection technology, and in particular to a method, system and medium for optimizing important feature parameters of unmanned aerial vehicle failure based on causal reasoning. Background Art

[0002] A drone is a high-end robot that integrates a mechanical structure and its various electronic systems. It is widely used in aerial reconnaissance, firefighting, and national defense because of its high efficiency, flexibility, low manufacturing cost, precision operation, and multi-tasking capabilities. However, compared with general manned aircraft, drones have more uncontrollable factors when performing various flight missions, face many challenges and potential dangers, and are prone to different failures. Before drone fault diagnosis (classification according to fault type labels), it is usually necessary to perform feature selection on the collected drone data to obtain the required feature subset; the purpose of this is to simplify the model structure, reduce computational complexity, and improve model performance.

[0003] The current situation is that either we rely on human judgment or prior knowledge to select characteristic parameters that can reflect UAV failures to form a feature subset, or we use a general feature selection method to obtain a feature subset. There is no feature parameter selection method specifically for UAV failures. Therefore, there is currently a lack of feature parameter selection methods specifically for UAV failure characteristics. This problem needs to be solved urgently to reduce the frequency of UAV failures, improve their safety and reliability, and ensure that UAVs can complete different tasks in various complex environments.

[0004] In addition, general feature selection methods are divided into unsupervised and supervised feature selection methods based on whether label information is used for feature selection. In practical applications, due to the large amount of data, many parameter variables, and many fault types, it is generally impossible to label in time, so unsupervised methods are often used for feature selection. Among them, the disadvantages of unsupervised feature selection methods are as follows: (1) Insufficient treatment of feature correlation: Unsupervised feature selection methods usually only consider the selection of a single feature, but do not fully consider the correlation between features. When dealing with highly correlated features, this method may not be able to effectively identify the most important feature combination, resulting in the selected feature subset being suboptimal. (2) Dependence on data distribution: Unsupervised feature selection methods usually need to select features based on the distribution of data. If the data distribution is uneven or there are heterogeneous samples, then this method may be greatly affected, resulting in the selected feature subset being inaccurate or unstable. (3) Poor interpretability: Since unsupervised feature selection methods do not rely on label information, the feature subsets they select may lack clear interpretability. This makes it difficult for users to understand why certain features are selected and others are excluded in practical applications. The disadvantages of supervised feature selection methods are as follows: (1) Supervised feature selection methods rely on labeled data, that is, each sample has a corresponding label. However, in practical applications, obtaining these labels often requires a lot of manpower and time, especially on large-scale data sets. (2) The quality of labeled data directly affects the accuracy of feature selection. If the labels are unreliable or erroneous, feature selection based on these labels may produce misleading results, thereby affecting the performance of subsequent models. (3) In some cases, supervised feature selection methods may focus too much on the relationship between a single feature and the target variable, while ignoring the correlation between features. This may result in the selected feature subset being suboptimal because some important feature combinations may be ignored. Summary of the invention

[0005] The present invention aims to at least solve the technical problems existing in the prior art, and in particular innovatively proposes a method, system and medium for optimizing important characteristic parameters of UAV failures based on causal reasoning.

[0006] In order to achieve the above-mentioned object of the present invention, the present invention provides a method for optimizing important characteristic parameters of UAV faults based on causal reasoning, comprising the following steps:

[0007] S1, collecting fault data, wherein the fault data includes flight fault data collected by a gyroscope, a GPS, an airspeed meter, etc.;

[0008] S2, data processing of the fault data: normalizing the data, and then using a filter to smooth and denoise the normalized data;

[0009] S3, causal structure learning: The processed data is input into the VAE model, and the causal structure learning is performed in sequence through the encoder and decoder of the VAE (variational autoencoder) to obtain a weighted causal DAG representing the causal relationship between features. The weighted causal DAG is in the form of a feature graph and is non-periodic.

[0010] Get the required weighted adjacency matrix through the VAE model , weighted adjacency matrix The element value in represents the weight w between nodes; the weighted adjacency matrix Converted into a weighted causal DAG feature graph.

[0011] For example: , the horizontal axis corresponds to the nodes from left to right , the vertical coordinates correspond to nodes from top to bottom , then it represents the node in the weighted causal DAG With Node Connected, with weight w=-1; node With Node connected, and weight w=1; nodes With Node connected, and weight w=1.

[0012] The encoder of the VAE is composed of a graph neural structure and a multi-layer perceptron, which is used to map the input data to the latent space to obtain the latent variables. The encoder of the VAE is expressed as follows:

[0013] ;

[0014] The decoder of the VAE is also composed of a graph neural structure and a multi-layer perceptron, which is used to reconstruct the latent variables into input data. The decoder of the VAE is expressed as follows:

[0015] ;

[0016] in, represents the identity matrix;

[0017] express The transpose of is the weighted adjacency matrix;

[0018] is the mean of the reconstructed data;

[0019] represents the standard deviation of the reconstructed data;

[0020] represents a multilayer perceptron;

[0021] represents the latent variables sampled in the latent space;

[0022] represents the parameter matrix of the MLP in the encoder of the VAE;

[0023] represents the parameter matrix of the MLP in the decoder of the VAE;

[0024] is the mean of the latent representation;

[0025] represents the standard deviation of the underlying representation;

[0026] Represents the input drone fault data;

[0027] represents the reconstructed drone failure data;

[0028] S4, selection of important fault features: After obtaining the weighted causal DAG, the node scores of the feature nodes are calculated based on the LeaderPageRank algorithm, and then the node scores are sorted in descending order. The top F fault feature variables are selected as the feature subset, where F is a positive integer.

[0029] Preferably, the normalization is min-max normalization, expressed as:

[0030] ;

[0031] In the formula, Represents the original UAV fault data in matrix form;

[0032] The normalized UAV fault data is also expressed in matrix form;

[0033] and Represent the minimum and maximum values ​​of each column in the fault data respectively.

[0034] Preferably, the filter is a Savitzky-Golay filter, which is expressed as:

[0035] ;

[0036] In the formula, Represents the value of the signal at position j after smoothing and denoising by the Savitzky-Golay filter;

[0037] Represents the value of the original signal at position j+v;

[0038] represents the noise reduction coefficient used in the vth operation;

[0039] The vth operation represents the polynomial fitting and signal replacement operation performed when the window moves to the vth position;

[0040] represents the half width of the smoothing window;

[0041] Represents an index variable;

[0042] Represents the total number of data points of the input data.

[0043] Preferably, the It is obtained by weighted averaging the original signal points in the window. The weight in the weighted averaging process is determined by the filter coefficient c; since the filter coefficient c is indirectly calculated through polynomial fitting. Therefore, It is obtained based on polynomial fitting. The expression of polynomial fitting is as follows:

[0044] ;

[0045] In the formula, Represents the coefficients of each polynomial;

[0046] Representation variables of Power;

[0047] represents the order of the polynomial;

[0048] Represents the index variable in the summation symbol;

[0049] Representing the polynomial The coefficient of

[0050] Representation variables of Power;

[0051] represents the independent variable of the polynomial;

[0052] Indicates the half-width of the smoothing window.

[0053] Preferably, since it is difficult to obtain the true posterior distribution p(Z│X) of the latent variable Z by maximizing the logarithmic evidence in actual operation,

[0054] The learning of the VAE decoder is based on the variational Bayesian technique: by introducing the variational posterior To approximate the true posterior distribution And the final ELBO is derived, and the final ELBO is expressed as:

[0055]

[0056] ;

[0057] In the formula, It represents the lower bound of evidence and is the objective function in the VAE model optimization process;

[0058] is the symbol of the expected operation, indicating that the latent variable Follow the distribution Under the condition, find the expectation of the expression in the brackets;

[0059] Indicates that given the latent variable Under the condition of probability;

[0060] represents the Kullback-Leibler divergence, which is used to measure the variational posterior With prior distribution The difference between

[0061] Represents the variational posterior distribution, which is used to approximate the true posterior distribution ;

[0062] Represents latent variables The prior distribution of

[0063] represents a constant term;

[0064] The operator symbol representing the norm;

[0065] Represents the input drone fault data;

[0066] express The reconstructed data.

[0067] In VAE (Variational Autoencoder), ELBO is a key concept, which is used to approximate the lower bound of the log-likelihood (log-evidence) function of the observed data. The goal of VAE is to train the model by maximizing ELBO, which involves optimizing the parameters of the encoder and decoder simultaneously. Specifically, the encoder is responsible for mapping the input data to a representation in the latent space, while the decoder reconstructs the input data from the representation in this latent space. During the training process, it is hoped that the ELBO can be maximized by adjusting the parameters of the encoder and decoder, thereby indirectly maximizing the log-likelihood function of the observed data.

[0068] Preferably, the learning problem for causal structure learning is defined as:

[0069] ;

[0070] ;

[0071] in, Indicates constraints;

[0072] Represents the constraint function, and the expression is:

[0073] ;

[0074] in, Represents the trace operation of a matrix;

[0075] represents the identity matrix;

[0076] represents a positive integer;

[0077] represents the real number parameter of the weight factor;

[0078] represents the original adjacency matrix.

[0079] The constraints can realize the non-periodicity of DAG, that is, directed acyclic graph.

[0080] Preferably, for the constraint , using the augmented Lagrangian method:

[0081] ;

[0082] in, is the Lagrange multiplier;

[0083] >0, is the penalty parameter;

[0084] In order to find the constraints The optimal solution of introduces updated λ and increased c1, and re-solves the minimization problem of the constrained augmented Lagrangian function after each increase, thus obtaining the final and The value of:

[0085] ;

[0086] ;

[0087] ;

[0088] in, Indicates that at the Kth iteration, the variable and The value of

[0089] represents the augmented Lagrangian function;

[0090] represents the weighted adjacency matrix;

[0091] Represents a variable in the optimization problem, which is the four important parameter matrices W of the VAE (Variational Autoencoder) model 1 ,W 2 ,W 3 ,W 4 A collection of;

[0092] represents the value of the Lagrange multiplier at the Kth iteration;

[0093] represents the penalty parameter at the K+1th iteration;

[0094] represents the constraint function;

[0095] Indicates that at the Kth iteration, the variable The value of

[0096] and To adjust the parameters, >1 to stimulate the fast convergence characteristics of the algorithm, and <1 is used to control the convergence process;

[0097] Represents the value of the penalty parameter at the Kth iteration.

[0098] Preferably, the calculating of the characteristic nodes based on LeaderPageRank and sorting the obtained node scores in descending order comprises:

[0099] Calculate the weighted causal DAG node x i to x j The probability of propagation The formula is as follows:

[0100] ;

[0101] in, Represents x i Node and x j The weight values ​​between nodes;

[0102] ;

[0103] in, It represents the score vector of all nodes after iterative update, which is composed of the scores of all nodes. The score of the node is the score of the fault feature. Generally, when calculating nodes through PageRank, the calculation is performed one by one. The present invention can calculate the values ​​of all nodes at one time through the vector.

[0104] represents the real number parameter of the weight factor;

[0105] Represents the transition probability Form a transition probability matrix P, P={ };

[0106] 1 represents the score vector of all nodes before iterative update;

[0107] m1 represents the total number of nodes in the obtained weighted causal DAG;

[0108] is the teleportation probability between 0 and 1, which determines The probability of continuing along the current path is 1- The probability of randomly jumping to any node in the graph is the random walk behavior mode. The value of The default value is set to 0.85.

[0109] The present invention also proposes a causal reasoning-based optimization system for important characteristic parameters of UAV faults, including a data processing module, a causal structure learning module, and a fault feature selection module.

[0110] Data processing module: normalizes and inputs the fault data into filters in turn to unify the data magnitude, smooth and denoise the data;

[0111] Causal structure learning module: The data output by the data processing module is sequentially passed through the encoder and decoder of the VAE (variational encoder) to perform causal structure learning; thus, a weighted causal DAG representing the causal relationship between features is obtained. The weighted causal DAG is non-periodic;

[0112] The encoder of the VAE is composed of a graph neural structure and a multi-layer perceptron, which is used to map the input data to the latent space to obtain the latent variables. The encoder of the VAE is expressed as follows:

[0113] ;

[0114] The decoder of the VAE is also composed of a graph neural structure and a multi-layer perceptron, which is used to reconstruct the latent variables into input data. The decoder of the VAE is expressed as follows:

[0115] ;

[0116] in, represents the identity matrix;

[0117] express The transpose of is the weighted adjacency matrix;

[0118] is the mean of the reconstructed data, represents the logarithmic standard deviation of the reconstructed data;

[0119] Represents the standard deviation of the reconstructed data

[0120] represents a multilayer perceptron;

[0121] represents the latent variables sampled in the latent space;

[0122] represents the parameter matrix of the MLP in the encoder of the VAE;

[0123] represents the parameter matrix of the MLP in the decoder of the VAE;

[0124] is the mean of the latent representation, is the log standard deviation of the latent representation;

[0125] represents the standard deviation of the underlying representation;

[0126] Represents the input drone fault data;

[0127] represents the reconstructed drone failure data;

[0128] Fault feature selection module: used to obtain feature subsets, calculate feature nodes through the LeaderPageRank algorithm and sort the obtained node scores in descending order, and select F fault feature variables as feature subsets by setting thresholds.

[0129] The present invention also provides a readable storage medium storing a program, wherein when the program is executed by a processor, the method described in claim 1 is implemented.

[0130] In summary, due to the adoption of the above-mentioned technical scheme, the method of the present invention can select from the UAV fault features and extract effective feature information; and based on these feature information, a corresponding causal network is constructed to represent the causal relationship, and the feature nodes and the weight values ​​between the two feature nodes are calculated, and then a suitable feature subset is selected. The present invention can effectively screen out characteristic parameters that reflect UAV failures, which can be used for fault diagnosis, thereby helping to reduce the frequency of UAV failures, improve their safety and reliability, and ensure that UAVs can complete different tasks in various complex environments. In addition, compared with the shortcomings of general feature selection methods, the present invention can also solve the shortcomings of unsupervised and supervised feature selection methods in the background technology: experiments have proved that the performance of the present invention is better than the unsupervised feature selection method, and is much better than the supervised feature selection method. Generally speaking, supervised methods are better than unsupervised methods. The specific advantages are as follows:

[0131] 1. The introduction of causal relationships enhances interpretability and can extract the potential influence relationship of variables in the data, thereby improving the performance of this method in feature selection.

[0132] 2. The method of the present invention is based on gradient causal structure learning, constructs the causal structure between feature variables, and calculates the importance of each feature through causal reasoning, so as to select the appropriate feature subset. It can also well mine the linear and nonlinear relationships of data without being restricted by data types.

[0133] 3. The method of the present invention adopts an unsupervised feature selection method, which does not require labeled data and is not affected by the quality of labeled data. It can be greatly applicable to data sets that are unlabeled or difficult to label. It has also been verified through experiments that although this method is unsupervised, its performance is not inferior to that of supervised methods, and is even better than some supervised methods.

[0134] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0135] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0136] Figure 1 This is the CausalAVFS: feature selection framework structure diagram of the present invention, Figure 1 (a) is the data processing part, Figure 1 (b) is the causal structure learning part, Figure 1 (c) is the important feature selection part.

[0137] Figure 2 It is a schematic diagram of the principle structure of the VAE model of the present invention.

[0138] Figure 3 This is the overall flow chart of the experiment.

[0139] Figure 4 This is the experimental result of four unsupervised feature parameter selection methods on the drone fault dataset. Figure 4 (a) is the experimental result on the fault dataset No. 12; Figure 4 (b) is the experimental result on the No. 13 fault dataset Figure 4 (c) is the experimental result on the No. 21 fault dataset; Figure 4 (d) is the experimental result on fault dataset No. 23.

[0140] Figure 5 This is the experimental result diagram of three supervised feature parameter selection methods and the CausalAVFS unsupervised feature parameter selection method on the drone fault dataset. Figure 5 (a) is the experimental result on the fault dataset No. 12; Figure 5 (b) is the experimental result on the No. 13 fault dataset (c) is the experimental result on the No. 21 fault dataset; Figure 5 (d) is the experimental results on the No. 23 fault dataset. DETAILED DESCRIPTION

[0141] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0142] The present invention uses a gradient-based causal structure learning method to study the causal relationship between drone fault feature variables, and proposes a method for selecting drone fault feature variables (CausalAVFS). In the study of feature selection, according to the different ways of combining feature selection with subsequent algorithms, feature selection methods can also be divided into four types: filtering, encapsulation, embedded and hybrid. As a causal reasoning method, causal structure learning discovers the causal relationship between variables in the observed data and constructs a directed acyclic graph (DAG). According to different structural learning principles, causal structure learning is divided into four types: constraint-based methods, score-based methods, function-based methods, and gradient-based methods. Among them, the gradient-based method can well mine the existing linear and nonlinear causal relationships in the data and construct a weighted causal DAG. CausalAVFS can automatically discover the causal relationship between drone fault features and output the corresponding weighted causal graph, and then calculate the score of each feature through causal reasoning and rank them in descending order. CausalAVFS mainly consists of three parts: data processing, causal structure learning and important fault feature selection, such as Figure 1 shown.

[0143] In the research on drone fault detection, the selected data include accelerometer (Ax, Ay, Az) and gyroscope (Gz, Gy, Gz) and other data. When a drone fails, the data generated by the gyroscope, accelerometer and other instruments will change accordingly. These data are collected to select the characteristic variables for drone fault detection. i Indicates that each feature variable is normalized and the data is smoothed and denoised using filters. Figure 1 Based on the data obtained by the data processing module, the gradient-based causal learning method is used to learn the causal structure and output the weighted causal DAG to represent the causal relationship between the features, as shown in (a). Figure 1 After obtaining the weighted causal DAG, the characteristic nodes are calculated based on LeaderPageRank and the obtained node scores are sorted in descending order. F fault characteristic variables are selected by setting the threshold, as shown in Figure 1 (c) shown.

[0144] S100, data processing

[0145] Since the fault data of the drone are collected by different sensors, these data have different magnitudes. Therefore, in order to eliminate the influence of data of different magnitudes on the model effect, the present invention uses the min-max normalization method to process the data. The principle of the normalization method is described as follows:

[0146] (1)

[0147] In the formula, Represents the original UAV fault data in matrix form;

[0148] Express the normalized UAV fault data in matrix form;

[0149] and Represent the minimum and maximum values ​​of each column of data in the fault data. This is because min-max normalization usually processes data column by column (or feature by feature). This means that for each feature in the data set (that is, each column of the data matrix), its minimum (min) and maximum value are calculated separately.

[0150] At the same time, after normalizing the data, the Savitzky-Golay filter is used to smooth and denoise the normalized data to reduce the impact of noise on the model effect. The output of the Savitzky-Golay filter is described as follows:

[0151] (2)

[0152] In the formula, Represents the value of the signal at position j after smoothing and denoising by the Savitzky-Golay filter;

[0153] Represents the value of the original signal at position j+v;

[0154] represents the noise reduction coefficient used in a particular operation v;

[0155] represents the half width of the smoothing window;

[0156] Represents an index variable;

[0157] Represents the total number of data points of the input data. One data point corresponds to the data of multiple variable fault parameters at a certain moment, corresponding to the row data of the drone fault data matrix; There is a close relationship, namely: .

[0158] The core principle of the Savitzky-Golay filter is to select the appropriate fitting order based on the least squares method to perform polynomial fitting on the continuous data points in the original data set, and output the obtained fitting smooth curve as the filtered data. The output value of the Savitzky-Golay filter obtained by formula (2) is , is obtained by weighted averaging the original signal points in the window. The weight in the weighted averaging process is determined by the filter coefficient c. Since the filter coefficient c is indirectly calculated by polynomial fitting, the output value of the filter is In fact, it is also based on the polynomial fitting results. The mathematical expression of polynomial fitting is formula (3), where: and are the polynomial coefficients and order that determine the complexity of the fitted curve.

[0159] (3)

[0160] In the formula, represents the coefficients of the polynomial;

[0161] Representing the polynomial The coefficient of the term;

[0162] Representation variables to the kth power;

[0163] represents the order of the polynomial;

[0164] Represents the index variable in the summation symbol;

[0165] Representing the polynomial The coefficient of

[0166] Representation variables τ to the power of

[0167] represents the independent variable of the polynomial;

[0168] Indicates the half-width of the smoothing window.

[0169] S200, Causal Structure Learning

[0170] The main function of causal structure learning is to construct a weighted adjacency matrix of fault feature variables to obtain a metric causal graph, in order to prepare for feature selection after calculating and sorting the causal strength of the nodes. In the present invention, DAG is used to represent the corresponding fault feature causal graph, in which each node represents the corresponding fault feature variable, and each edge represents the causal relationship between the two corresponding nodes. According to the research work on related causal structures, it is not difficult to find that some traditional causal structure learning methods will restrict the input data when performing structural learning, and cannot handle linear and nonlinear causal relationships in the data. Therefore, a gradient-based method is used to realize the causal structure learning between drone fault feature variables, which is to apply structural constraints to a variational autoencoder (VAE) that parameterizes a new graph neural network (GNN) to construct a DAG.

[0171] use ∈R m×n To indicate that the drone fault data needs to be input, m represents the number of feature variables, n is the data sample of each feature, The required causal structure is obtained in the linear SEM (linear structural equation model) as the data generation model to obtain the weighted adjacency matrix of DAG. The linear SEM model is:

[0172] (4)

[0173] In the formula, for The transpose of

[0174] is an m×m weighted adjacency matrix;

[0175] is an m×n noise matrix.

[0176] When the nodes of a directed acyclic graph (DAG) are arranged in their topological order, the matrix associated with the relationships between these nodes is It will show a strictly upper triangular characteristic. Based on this property, the sampling process of ancestor nodes (all nodes that can reach the node through a series of directed edges) of DAG can be equivalently regarded as first generating a random noise matrix , and then use the matrix The upper triangular characteristics of are used to perform efficient triangular solution to achieve target sampling, namely:

[0177] (5)

[0178] Formula (5) can be expressed as , this equation is a general form of parameterized graph neural network (GNN) with node features As input, As the output of the corresponding node, it is widely recognized in the deep learning community, and almost many graph neural networks can be expressed in this form. Therefore, in order to construct a VAE model that can generate a weighted adjacency matrix, a VAE encoder (6) based on the graph neural structure and multi-layer perceptron (MLP) of formula (5) is constructed to map the input data to the latent space and obtain the latent variables:

[0179] (6)

[0180] The corresponding VAE decoder is formula (7), which reconstructs the latent variables into input data:

[0181] (7)

[0182] in, represents the identity matrix;

[0183] express The transpose of is the weighted adjacency matrix;

[0184] is the mean of the reconstructed data;

[0185] represents the standard deviation of the reconstructed data;

[0186] represents a multilayer perceptron;

[0187] represents the latent variables sampled in the latent space;

[0188] represents the parameter matrix of the MLP in the encoder of the VAE;

[0189] represents the parameter matrix of the MLP in the decoder of the VAE;

[0190] is the mean of the latent representation;

[0191] represents the standard deviation of the underlying representation;

[0192] Represents the input drone fault data;

[0193] represents the reconstructed drone failure data;

[0194] Therefore, the VAE framework principle structure constructed by the present invention is as follows Figure 2 As shown in Figure 2. In the VAE framework, ELBO (Evidence Lower Bound) plays an important role. In practice, it is often difficult to learn the decoder model (6) by maximizing the logarithmic evidence. Therefore, based on the variational Bayesian technique, the variational posterior To approximate the true posterior distribution And the final ELBO is derived to solve this problem. The final ELBO can be expressed as:

[0195]

[0196]

[0197] In the formula, Stands for “Evidence Lower Bound”, which is the objective function in the variational autoencoder (VAE) optimization process;

[0198] is the symbol of the expected operation, indicating that the latent variable Follow the distribution Under the condition, find the expectation of the expression in the brackets;

[0199] Indicates that given the latent variable Under the condition of The probability of

[0200] represents the Kullback-Leibler divergence, which is used to measure the variational posterior With the prior distribution The difference between

[0201] Represents the variational posterior distribution, which is used to approximate the true posterior distribution ;

[0202] Represents latent variables The prior distribution of

[0203] represents a constant term;

[0204] The operator symbol representing the norm;

[0205] Represents the input drone fault data;

[0206] express The reconstruction data;

[0207] At the same time, in order to effectively improve the practicality and coding efficiency of the algorithm, maintain effective control over the non-periodicity of the graph, and reduce the complexity of implementation, the present invention uses a simpler and more practical constraint scheme (9).

[0208] (9)

[0209] in, represents the constraint function;

[0210] Represents the trace operation of a matrix;

[0211] represents the identity matrix;

[0212] represents a positive integer;

[0213] represents the real number parameter of the weight factor;

[0214] represents the original adjacency matrix;

[0215] Based on the above description, in order to better realize the weighted adjacency matrix learning of weighted causal DAG, the learning problem can be defined as:

[0216]

[0217] (10)

[0218] In the formula, is the unknown causal weighted adjacency matrix;

[0219] The four important parameter matrices W of the variational autoencoder (VAE) 1 ,W 2 ,W 3 ,W 4 In addition, the augmented Lagrangian method is used to deal with the existing nonlinear equality constraint problem. The principle is as follows:

[0220] (13)

[0221] In formula (13), λ is the Lagrange multiplier;

[0222] >0, is the penalty parameter. In order to find the optimal solution that meets the constraints, it is necessary to gradually increase The value of , and after each increase, the unconstrained augmented Lagrangian function minimization problem is re-solved. For this purpose, a method that can update λ and increase c1 is introduced:

[0223]

[0224] (14)

[0225]

[0226] in, Indicates that at the Kth iteration, the variable and The value of

[0227] represents the augmented Lagrangian function;

[0228] Represents a variable in the optimization problem, which is the four important parameter matrices W of the variational autoencoder (VAE) 1 ,W 2 ,W 3 ,W 4 A collection of;

[0229] represents the value of the Lagrange multiplier at the kth iteration;

[0230] represents the penalty parameter at the k+1th iteration;

[0231] represents the constraint function;

[0232] Indicates that at the kth iteration, the variable The value of

[0233] Indicates that this is a constant greater than 1, used to adjust the constraint;

[0234] represents the value of the penalty parameter at the kth iteration;

[0235] In this augmented Lagrangian function, the update strategy of c1 generally adopts an exponential increase based on the number of iterations, so as to promote the corresponding adjustment of λ and ensure that the algorithm gradually approaches the optimal solution. When dealing with optimization problems, reasonable adjustment of c1 is crucial to achieve a good balance between feasibility and optimality. At the same time, in order to accelerate the convergence process and optimize the solution quality, the strategy cleverly incorporates the adjustment parameters η and γ. Usually, η>1 is set to stimulate the fast convergence characteristics of the algorithm, while γ<1 is used to control the convergence process. In practice, η=10 and γ=0.25 have good results.

[0236] During the training process, the parameters and It will be updated once after each training cycle. After the training is completed, the weighted causal DAG can be obtained, such as Figure 1 As shown in (b), the graph depicts the causal relationship between different parameter variables. In the weighted causal DAG, each node x represents a fault feature parameter variable, and the directed edge between two nodes represents the causal relationship between the two nodes. The absolute value of the weight w on the edge reflects the impact of the causal relationship. The positive or negative value of the weight w indicates that the increase in the cause variable may lead to a decrease or increase in the result variable. Based on the weighted causal DAG, the present invention uses a causal reasoning filtering feature selection method to obtain the optimal feature subset that causes drone failures.

[0237] S300, selection of important characteristic parameters

[0238] Based on the obtained weighted causal DAG, the causal strength of each characteristic parameter variable in the graph is calculated and ranked by the LeaderRank algorithm, and a subset containing F fault characteristic variables is screened out by setting a threshold. The principle of the LeaderRank algorithm is to introduce a common node (Ground node) based on the PageRank algorithm to improve the accuracy of node importance assessment, and calculate the node importance based on the number of incoming edges and the probability of abnormal propagation in the graph. In the present invention, the node x is calculated. i to x j The probability of propagation The formula is as follows:

[0239] (15)

[0240] In formula 15, Represents x i Node and x j The weight value between nodes. In the present invention, the transfer probability Form a transition probability matrix P, P={ }, we can get the vector :

[0241] (16)

[0242] In formula 16, Represents the score vector of all nodes after iterative update;

[0243] Represents the transition probability Form a transition probability matrix P;

[0244] 1 represents the score vector of all nodes before iterative update;

[0245] m1 represents the total number of nodes in the obtained weighted causal DAG, is the teleportation probability between 0 and 1, which determines The probability of continuing along the current path is 1- The probability of randomly jumping to any node in the graph is the random walk behavior mode. The value of The default value of is set to 0.85. In the present invention, all nodes are sorted according to the obtained LeaderRank score to generate the following Figure 1 (c) shows a ranking list, where the first F parameter variables can be selected as a subset according to the requirements. In this list, the higher the ranking, the more likely it is that the node is the required parameter variable.

[0246] In order to verify the effectiveness of the method for selecting drone fault feature variables, the commonly used feature selection methods were used as benchmark algorithms to perform feature selection in public drone datasets, and experiments were conducted in multiple classification models. The selected benchmark algorithms are: unsupervised methods GURAG, MCFS, Laplacian-scores and supervised methods XGBoost feature selection (XGBFS), chi-square filter feature selection (CSFFS) and hybrid feature selection based on random forest (RFHFS).

[0247] After the CausalAVFS framework was designed, the corresponding experiments were conducted using the public UAV fault data set to evaluate the effectiveness and versatility of the method in selecting important feature parameters of UAV faults. The unsupervised methods GURAG, MCFS, Laplacian-scores and the supervised embedded feature selection XGBFS algorithm, hybrid feature selection RFHFS algorithm and filtered feature selection CSFFS algorithm were used to select the subset of important feature parameters of UAV faults from the data that had been normalized and processed by Savitzky-Golay filter. Then, 7 common classification algorithms such as KNN, SVM and AdaBoost were used for classification experiments to evaluate the advantages and disadvantages of using the CausalAVFS algorithm. Figure 3 This is the overall flow chart of the experiment.

[0248] (1) Experimental setup

[0249] The specific configuration of the computer used in the experiment is as follows: Windows 11 operating system, 12th Gen Intel(R)Core(TM) i7-12700H 2.30 GHz processor, 8GB RAM. The experiment was conducted in Python 3.7. The training set and validation set were randomly partitioned in a ratio of 8:2.

[0250] At present, fault diagnosis for UAV, an advanced machine, is a typical classification problem. Therefore, in order to evaluate the performance of the CausalAVFS method in selecting important characteristic parameters of UAV faults, this work uses the real UAV flight data set (https: / / github.com / mrtbrnz / fault_detection / tree / master / data) commonly used in the fault diagnosis method based on classification models by Bronz et al. for classification experiments. The data set contains more than 50 flight data variables, of which 16 data variables reflect the fault information of the UAV, that is, when the UAV fails, these 16 data variables will have clear changes. In this invention, these 16 variables are used to form the original feature set, which greatly reduces the model operation pressure. In the work, the data sets obtained by the sensors of the UAVs No. 12, No. 13, No. 21, and No. 23 were mainly intercepted for classification experiments. The basic information about these experimental data sets, including the name, sample size, number of features, and number of categories, is shown in Table 1:

[0251] Table 1 UAV fault experiment dataset

[0252] Dataset name Sample size Number of features Number of categories No. 12 UAV Fault Dataset 25400 16 2 No. 13 UAV Fault Dataset 20900 16 2 No. 21 UAV Fault Dataset 184600 16 9 No. 23 UAV Fault Dataset 205300 16 9

[0253] (2) Evaluation indicators

[0254] In order to rigorously verify the effectiveness of the CausalAVFS method, the core step is to use a variety of feature selection techniques to obtain and compare the impact of different input information on the prediction model. This verification process begins with a careful screening of key parameter features. Next, by monitoring the improvement of performance indicators in the prediction task, the effectiveness of this method in the selection of drone fault feature parameters can be directly demonstrated. For this purpose, seven classification models including KNN, SVM and AdaBoost were carefully selected for experiments. In order to comprehensively and accurately measure the classification effects of these models, accuracy was further used as an evaluation indicator.

[0255] Accuracy: The ratio of the number of samples correctly predicted by the classifier to the number of samples in the entire test data set.

[0256] (17)

[0257] (3) Comparison of unsupervised feature parameter selection methods and experimental results on drone datasets

[0258] In Figure 4 In this paper, we show the classification results of four unsupervised feature parameter selection algorithms, GURAG, MCFS, Laplacian-scores and CausalAVFS (ours), and seven classification algorithms, KNN, SVM, decision tree, Gaussian naive Bayes, neural network, logistic regression and AdaBoost, on the drone flight data set. Figure 4 In (a), (b) and (c), the x-axis represents the number of selected features (K), and the y-axis shows the average accuracy obtained by different classification algorithms after selecting important features using four unsupervised feature parameter selection algorithms.

[0259] Depend on Figure 4 From the comparative experimental results of the unsupervised feature parameter selection methods, it can be seen that CausalAVFS has the best accuracy performance except that the accuracy performance of the first two selected features of the 12th and 23rd data sets is similar to that of the other three unsupervised feature parameter selection algorithms. It can be seen that compared with other unsupervised feature parameter selection methods, CausalAVFS has a good performance in selecting and constituting feature parameter subsets of drone fault feature parameters.

[0260] (4) Comparison of supervised feature parameter selection methods on the experimental results of drone datasets

[0261] From (3), we can see that the CausalAVFS unsupervised feature parameter selection method has a better effect on the selection of UAV fault feature parameters than other unsupervised feature parameter selection methods. However, in order to better illustrate the practicality of the present invention, in this section, three supervised feature parameter selection methods, CSFFS, XGBFS, and RFHFS, and CausalAVFS (our) unsupervised feature parameter selection method, as well as seven classification algorithms, KNN, SVM, decision tree, Gaussian naive Bayes, neural network, logistic regression, and AdaBoost, are used to conduct classification experiments on the UAV flight data set. The results are shown in Figure 2. Figure 5 shown.

[0262] Here Figure 5 In (a), (b) and (c), the x-axis represents the number of selected features (F), while the y-axis shows the average accuracy obtained by different classification algorithms after using four feature parameter selection algorithms to select important features. This section also conducts an in-depth comparative analysis of the performance differences between the CausalAVFS algorithm and the other three feature selection algorithms when the number of features is between 7 and 14.

[0263] Figure 5 (a) shows the results of the binary classification experiment of the four feature parameter selection algorithms on the No. 12 UAV fault data set. When F≤10, the average accuracy of the CausalAVFS algorithm is weaker than that of the other three algorithms. When F=11, the four important feature selection methods show similar performance in the three evaluation criteria. Then, when F>11, the CausalAVFS algorithm is better than the other three algorithms in accuracy. With the further increase of the selected features, the results of the three evaluation criteria are basically stable, and the improvement of the index results is not obvious compared with before F=11.

[0264] Figure 5 (b) shows the binary classification experimental results of the 4 feature parameter selection algorithms on the No. 13 UAV fault data set. When F≤10, the CausalAVFS algorithm often outperforms the other three algorithms in terms of accuracy, and as the F value increases, that is, when F>10, the CausalAVFS algorithm outperforms the other three algorithms in this indicator and has an upward trend. However, it is worth noting that when k≥11, as the number of feature selections further increases, the corresponding evaluation indicators of the features selected by the other three feature selection methods (CSFFS, XGBFS and RFHFS) gradually stabilize after experiencing partial increases and decreases. In addition, during the experiment on this data set, it was found that the CSFFS algorithm had defects, that is, it could only select the first 13 feature parameters. Therefore, when the number of feature selections reached F=14, no experimental results of the CSFFS algorithm were obtained.

[0265] Figure 5 (c) shows the multi-classification experimental results of four feature parameter selection algorithms on the No. 21 UAV fault data set. When the feature selection number F value is between 7 and 10, the accuracy performance of the CausalAVFS algorithm is somewhat different from that of the CSFFS, XGBFS and RFHFS algorithms. However, when F≥11, CSFFS, XGBFS and RFHFS have similar performances on the three evaluation indicators and are accompanied by a downward trend, while the performance of the CausalAVFS algorithm is opposite to that of the other three algorithms and is higher than the other three methods.

[0266] Figure 5 (d) shows the multi-classification experimental results of four feature parameter selection algorithms on the No. 23 UAV fault data set. When F≤10, the accuracy of the CausalAVFS algorithm lags behind other algorithms. However, as the selected features increase, that is, when F≥11, although the performance of the CSFFS algorithm in the evaluation criteria has an upward trend, it is better than the CSFFS algorithm. At the same time, the evaluation index results of the CausalAVFS, XGBFS and RFHFS algorithms are relatively similar and tend to be stable.

[0267] In general, the classification experiment results using CSFFS, XGBFS, RFHFS and CausalAVFS algorithms to obtain feature selection results and using seven classification models such as KNN, SVM and AdaBoost show that each feature selection algorithm has its unique advantages in terms of accuracy evaluation indicators. Through comprehensive analysis, it can be seen that the CausalAVFS method can select feature subsets with better classification capabilities. In addition, to further illustrate the performance of the CausalAVFS method, we found that when F=11, the performance of the accuracy evaluation indicators of the four feature parameter selection methods remained roughly stable or showed a downward trend. This finding is consistent with the analysis conclusions in the literature. In the study of Zhang et al., about 11 variables were also selected for fault diagnosis, which further verified the experimental results. Therefore, the average accuracy of the seven classification methods when F=11 was selected as the benchmark for evaluating the performance of the four feature selection algorithms. The specific experimental results are shown in Table 2, where the highest accuracy values ​​of the four methods are bolded and the values ​​with the second highest accuracy are underlined.

[0268] Table 2 Accuracy results of four feature selection methods in four data sets

[0269] Table 2 shows the accuracy results of four datasets among seven classifiers: KNN, SVM, decision tree, Gaussian naive Bayes, neural network, logical algorithm and AdaBoost. The precision values ​​in the table are reserved to 4 decimal places.

[0270] For the No. 12 drone fault data set in Table 2, CausalAVFS achieved the highest accuracy value of 0.9993 in the neural network classification algorithm, which has a slight advantage over the other three feature selection algorithms. At the same time, when using the Gaussian Naive Bayes algorithm for binary classification, the CausalAVFS algorithm once again achieved the highest accuracy value of 0.8105, and its performance was 2.08% better than the CSFFS algorithm, 1.25% better than the XGBFS algorithm, and 0.71% better than the RFHFS algorithm. Unlike the first two cases, under the AdaBoost and SVM algorithms, the performance of the CausalAVFS algorithm lags behind the XGBFS algorithm, but is better than the RFHFS algorithm and the CSFFS algorithm. At the same time, the performance of the logical algorithm is slightly behind the XGBFS algorithm and RFHFS. In addition, the performance of the four feature selection methods is basically the same under the KNN algorithm and the decision tree algorithm. From the above, it can be seen that CausalAVFS and XGBFS have similar performances, and are better than the CSFFS and RFHFS methods. It can be seen that this method has a good advantage in this data set.

[0271] For the experimental results of the No. 13 UAV fault data set in Table 2, the CausalAVFS feature selection algorithm showed the best accuracy in five of the seven classification algorithms. First, under the AdaBoost algorithm, the accuracy of the CausalAVFS algorithm was significantly better than other feature selection methods, 3.72% higher than CSFFS, 2.19% higher than XGBFS, and 2.74% higher than RFHFS. Secondly, under the SVM algorithm, its accuracy also performed well, 4.99% higher than CSFFS, 2.62% higher than XGBFS, and although the gap with RFHFS was small, it was still 0.28% higher. Then, under the GNB algorithm, its advantage was also obvious, 5.47% higher than CSFFS, 2.79% higher than XGBFS, and 0.27% higher than RFHFS. Finally, for the neural network algorithm and decision tree algorithm, although the advantage of the CausalAVFS algorithm was not so significant, its accuracy was still slightly higher than CSFFS, XGBFS, and RFHFS. It is worth noting that although the performance of the CausalAVFS algorithm is not optimal under the KNN algorithm and the logical algorithm, it is similar to the performance of most feature selection methods. It can be seen that in this data set, the overall performance of the CausalAVFS algorithm is significantly better than that of the RFHFS, CSFFS and XGBFS algorithms.

[0272] For the multi-classification experimental results of the No. 21 UAV fault dataset in Table 2, the CausalAVFS method is the best only in the GNB classifier, with an accuracy of 0.3974, which is 1.89%, 5.72% and 4.33% higher than the CSFFS, XGBFS and RFHFS algorithms respectively. In addition, although the CausalAVFS feature selection algorithm lags behind other algorithms in the AdaBoost, KNN and logical algorithm classifiers, it performs better than CSFFS in the experimental results of SVM and neural network classifiers, and better than CSFFS and XGBFS in the decision tree classifier. From the above, it can be seen that although the CausalAVFS is not the best in the multi-classification experimental performance of the No. 21 UAV fault dataset, its performance is much better than the CSFFS algorithm.

[0273] For the multi-classification experimental results of the No. 23 UAV fault data set in Table 2, in the multi-classification experiment of the neural network classifier, the classification accuracy of the features selected by the CausalAVFS method is the highest, which is about 8.32% higher than the CSFFS method, 0.26% higher than the XGBFS method, and 2.19% higher than the RFHFS method. In the SVM and KNN classification experiments, although the CausalAVFS method lags behind the XGBFS and RFHFS methods, it is much better than the CSFFS method. In the logical algorithm and decision tree classification experiments, although the CausalAVFS method is not the best, the performance of the four feature selection methods is not much different. At the same time, in the AdaBoos classification experiment, the performance of the CausalAVFS method is not the best, but its performance is only behind the RFHFS method. In summary, although the performance of CausalAVFS in the multi-classification experiment of the No. 23 UAV fault data set is not the best, its performance is similar to that of RFHFS, and it is much better than the CSFFS method. It can be seen that it has a good performance in this data set.

[0274] In summary, the average performance of the proposed CausalAVFS algorithm is better than most of the compared algorithms. In the four datasets for classification experiments, it has obvious advantages in the first two datasets and has a good advantage in the fourth dataset. By comparing the four feature selection methods in the UAV fault dataset, it can be seen that the comprehensive performance of the CausalAVFS algorithm is significantly better than the CSFFS and RFHFS methods, and slightly weaker than the XGBFS method. At the same time, the experiment also shows that the performance of the CausalAVFS method in the No. 12 UAV fault dataset and the No. 13 UAV fault dataset is better than that in the No. 21 UAV fault dataset and the No. 23 UAV fault dataset. It can be seen that the CausalAVFS method is more suitable for the UAV fault dataset of binary classification supervised experiments.

[0275] In short, in order to improve the prediction performance of the model in the high-dimensional UAV fault data environment, it is an effective way to use feature selection methods to obtain a streamlined and effective feature subset as much as possible. However, traditional feature selection methods often focus on the correlation analysis between variables, lack of consideration of causal relationships, and may cause the selected features to contain insufficient information. Therefore, in this study, causal relationships are introduced into feature selection methods. By constructing a causal graph to reveal the causal relationship between variables and calculating and sorting the causal strength of variable nodes, a method for optimizing important feature parameters of UAV faults based on causal reasoning is proposed. In order to verify the effectiveness of feature selection of this method on UAV fault data sets, the present invention selects three unsupervised methods, GURAG, MCFS, and Laplacian-scores, and three supervised methods, CSFFS, XGBFS, and RFHFS, as comparisons. By using seven classification models, such as KNN, SVM, and decision tree, in four data sets, binary and multi-classification experiments are carried out to verify the effectiveness of the CausalAVFS method in feature selection.

[0276] The experimental results show that the CausalAVFS algorithm can effectively remove redundant features in four drone fault data sets, and shows excellent classification performance in comparison with the three unsupervised methods of GURAG, MCFS, and Laplacian-scores, and shows relatively excellent overall classification performance in comparison with the three supervised methods of CSFFS, XGBFS, and RFHFS. It can be seen that the performance of the CausalAVFS algorithm is better than all the compared unsupervised feature selection methods and most of the compared mainstream supervised feature selection algorithms. Therefore, its effectiveness in actual feature selection is verified through experiments. After using the four feature selection methods to extract effective features, unsupervised and supervised comparative classification experiments were carried out in 7 classification models. The experimental results show that the proposed method has good performance in the selection of drone fault feature parameters.

[0277] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A method for optimizing important characteristic parameters of UAV faults based on causal reasoning, characterized in that: The following steps are involved: S1, collect fault data; S2, data processing of the fault data: normalizing the data, and then using a filter to smooth and denoise the normalized data; S3, causal structure learning: The processed data is input into the VAE model, and the causal structure learning is performed through the VAE encoder and VAE decoder in turn to obtain the weighted causal DAG representing the causal relationship between features; The encoder of the VAE is composed of a graph neural structure and a multi-layer perceptron, which is used to map the input data to the latent space to obtain the latent variables. The encoder of the VAE is expressed as follows: ； The decoder of the VAE is also composed of a graph neural structure and a multi-layer perceptron, which is used to reconstruct the latent variables into input data. The decoder of the VAE is expressed as follows: ； in, represents the identity matrix; express The transpose of is the weighted adjacency matrix; is the mean of the reconstructed data; represents the standard deviation of the reconstructed data; represents a multilayer perceptron; represents the latent variables sampled in the latent space; represents the parameter matrix of the MLP in the encoder of the VAE; represents the parameter matrix of the MLP in the decoder of the VAE; is the mean of the latent representation; represents the standard deviation of the underlying representation; Represents the input drone fault data; represents the reconstructed drone failure data; S4, selection of important fault features: After obtaining the weighted causal DAG, the node scores of the feature nodes are calculated based on the LeaderPageRank algorithm, and then the node scores are sorted in descending order. The top F fault feature variables are selected as the feature subset, where F is a positive integer.

2. The method for optimizing important characteristic parameters of UAV faults based on causal reasoning according to claim 1 is characterized in that: The normalization is min-max normalization, which is expressed as: ； In the formula, Represents the original UAV fault data in matrix form; Describe the normalized drone failure data; and Represent the minimum and maximum values ​​of each column in the fault data respectively.

3. The method for optimizing important characteristic parameters of UAV failure based on causal reasoning according to claim 1 is characterized in that: The filter is a Savitzky-Golay filter, which is expressed as: ； In the formula, Represents the value of the signal at position j after smoothing and denoising by the Savitzky-Golay filter; Represents the value of the original signal at position j+v; represents the noise reduction coefficient used in the vth operation; represents the half width of the smoothing window; Represents an index variable; Represents the total number of data points of the input data.

4. The method for optimizing important characteristic parameters of UAV failure based on causal reasoning according to claim 3 is characterized in that: Said It is obtained based on polynomial fitting. The expression of polynomial fitting is as follows: ； In the formula, Represents the coefficients of each polynomial; Representation variables of Power; represents the order of the polynomial; Represents the index variable in the summation symbol; Representing the polynomial The coefficient of Representation variables of Power; represents the independent variable of the polynomial; Indicates the half-width of the smoothing window.

5. The method for selecting important characteristic parameters of UAV faults based on causal reasoning according to claim 1 is characterized in that: The learning of the VAE decoder is based on the variational Bayesian technique: by introducing the variational posterior To approximate the true posterior distribution And the final ELBO is derived, and the final ELBO is expressed as: ； ； In the formula, It represents the lower bound of evidence and is the objective function in the VAE model optimization process; is the symbol of the expected operation, indicating that the latent variable Follow the distribution Under the condition, find the expectation of the expression in the brackets; Indicates that given the latent variable Under the condition of probability; represents the Kullback-Leibler divergence; represents the variational posterior distribution; Represents latent variables The prior distribution of represents a constant term; The operator symbol representing the norm; Represents the input drone fault data; express The reconstructed data.

6. The method for selecting important characteristic parameters of UAV faults based on causal reasoning according to claim 5 is characterized in that: The learning problem for causal structure learning is defined as: ； ； in, Indicates constraints; Represents the constraint function, and the expression is: ； in, Represents the trace operation of a matrix; represents the identity matrix; represents a positive integer; represents the real number parameter of the weight factor; represents the original adjacency matrix.

7. The method for optimizing important characteristic parameters of UAV faults based on causal reasoning according to claim 6 is characterized in that: For constraints , using the augmented Lagrangian method: ； in, is the Lagrange multiplier; >0, is the penalty parameter; In order to find the constraints The optimal solution of introduces updated λ and increased c1, and re-solves the minimization problem of the constrained augmented Lagrangian function after each increase, thus obtaining the final The value of: ； ； ； in, Indicates that at the Kth iteration, the variable and The value of represents the augmented Lagrangian function; represents the weighted adjacency matrix; Represents a variable in the optimization problem, which is the four important parameter matrices W of the VAE model 1 ,W 2 ,W 3 ,W 4 A collection of; represents the value of the Lagrange multiplier at the Kth iteration; represents the value of the Lagrange multiplier at the Kth iteration; represents the penalty parameter at the K+1th iteration; represents the constraint function; Indicates that at the Kth iteration, the variable The value of and To adjust the parameters; Represents the value of the penalty parameter at the Kth iteration.

8. The method for selecting important characteristic parameters of UAV failure based on causal reasoning according to claim 1 is characterized in that: The method of calculating the characteristic nodes based on LeaderPageRank and sorting the obtained node scores in descending order includes: Calculate the weighted causal DAG node x i to x j The probability of propagation The formula is as follows: ； in, Represents x i Node and x j The weight values ​​between nodes; ； in, Represents the score vector of all nodes after iterative update; represents the real number parameter of the weight factor; Represents the transition probability Form a transition probability matrix P, P={ }; 1 represents the score vector of all nodes before iterative update; m1 represents the total number of nodes in the obtained weighted causal DAG; is the probability of teleportation between 0 and 1.

9. A system for optimizing important characteristic parameters of drone failures based on causal reasoning, characterized in that: It includes data processing module, causal structure learning module, and fault feature selection module. Data processing module: normalizes and inputs the fault data into filters in turn to unify the data magnitude, smooth and denoise the data; Causal structure learning module: The data output by the data processing module is sequentially passed through the VAE encoder and VAE decoder for causal structure learning; thus, a weighted causal DAG representing the causal relationship between features is obtained; The encoder of the VAE is composed of a graph neural structure and a multi-layer perceptron, which is used to map the input data to the latent space to obtain the latent variables. The encoder of the VAE is expressed as follows: ； The decoder of the VAE is also composed of a graph neural structure and a multi-layer perceptron, which is used to reconstruct the latent variables into input data. The decoder of the VAE is expressed as follows: ； in, represents the identity matrix; express The transpose of is the weighted adjacency matrix; is the mean of the reconstructed data, represents the logarithmic standard deviation of the reconstructed data; Represents the standard deviation of the reconstructed data represents a multilayer perceptron; represents the latent variables sampled in the latent space; represents the parameter matrix of the MLP in the encoder of the VAE; represents the parameter matrix of the MLP in the decoder of the VAE; is the mean of the latent representation, is the log standard deviation of the latent representation; represents the standard deviation of the underlying representation; Represents the input drone fault data; represents the reconstructed drone failure data; Fault feature selection module: used to obtain feature subsets, calculate feature nodes through the LeaderPageRank algorithm and sort the obtained node scores in descending order, and select F fault feature variables as feature subsets by setting thresholds.

10. A readable storage medium storing a program, characterized in that: When the program is executed by a processor, the method according to claim 1 is implemented.

Images