Multi-source navigation abnormal data detection method

Abstract

The application discloses a kind of multi-source navigation anomaly data detection methods, including collection summary navigation data, and navigation data is preprocessed to;Navigation data is fused after being preprocessed;Navigation anomaly data detection model is constructed, and the data after fusion is handled, and the navigation anomaly data detection model constructed is trained and optimized;Navigation anomaly data detection model after training, optimization is used, and the detection of abnormal data is completed to actual navigation data;The robustness of the method of the application is enhanced, accuracy is improved, and safety is increased.

Description

Technical Field

[0001] This invention belongs to the field of advanced data processing technology for navigation systems, and specifically relates to a method for detecting multi-source navigation anomaly data. Background Technology

[0002] The environment in which navigation systems operate is relatively complex, making navigation data susceptible to various factors, including sensor malfunctions or inaccuracies, signal interference, environmental factors, human error, and even malicious attacks. These factors may act individually or in combination, leading to data anomalies in the navigation system. Effectively identifying and processing navigation data from multiple sources, as well as detecting anomalies that may affect the performance of the navigation system, is of great significance for improving navigation safety and efficiency.

[0003] Currently, methods for detecting anomaly data in multi-source navigation mainly employ Kalman filtering, particle filtering, weight fusion, feature fusion, and deep learning to detect anomalies. However, these methods suffer from issues such as sensor inconsistency, varying noise levels, data latency, and nonlinearity and non-Gaussianity. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-source navigation anomaly data detection method with enhanced robustness, improved accuracy, and increased security.

[0005] The multi-source navigation anomaly data detection method provided by this invention includes the following steps:

[0006] S1. Collect and summarize navigation data, and preprocess the navigation data;

[0007] S2. Perform fusion processing on the navigation data preprocessed in step S1;

[0008] S3. Construct a navigation anomaly data detection model, and use the data fused and processed in step S2 to train and optimize the constructed navigation anomaly data detection model;

[0009] S4. Using the navigation anomaly data detection model trained and optimized in step S3, anomaly data detection is performed on the actual navigation data;

[0010] Step S1, which involves collecting and summarizing navigation data and preprocessing it, specifically includes:

[0011] (1-1) Navigation data:

[0012] Collect and summarize navigation data, including satellite navigation data (GPS), lidar (LiDAR) data, and gyroscope data for the same detected target;

[0013] (1-2) Navigation data classification:

[0014] The navigation data quality is checked to verify its completeness and accuracy; the data is checked for missing or outlier values. If there are no missing or outlier values, the data is classified as "normal data"; if there are missing or outlier values, the data is classified as "abnormal data".

[0015] (1-3) Navigation data preprocessing:

[0016] Perform data time and sampling rate preprocessing on the classified data;

[0017] Time preprocessing includes: standardizing timestamps and checking the timestamps to ensure data time synchronization;

[0018] Sampling rate preprocessing includes: determining the least common multiple of the sampling rates of all sensors as the target sampling rate, resampling the data, and verifying the sampled data;

[0019] The resampled data is standardized in terms of unit format;

[0020] Step S2, which involves fusing the preprocessed navigation data from step S1, specifically includes:

[0021] The three types of navigation data preprocessed in step S1 are fused using an unscented Kalman filter (UKF), specifically including:

[0022] Set the state vector x0 = [p x ,p y ,p z ,v x ,v y ,v z State covariance matrix Where, p x Indicates the predicted x-axis position; p y Indicates the predicted y-axis position; p z Indicates the predicted z-axis position; v x Indicates the predicted x-axis velocity; v y This indicates the predicted y-axis velocity; v z This represents the predicted z-axis velocity; diag(·) represents the diagonal matrix; σ represents the initial uncertainty of each state variable; Indicates position p x The variance; Indicates position p y The variance; Indicates position p z The variance; Indicates velocity v x The variance; Indicates velocity vy The variance; Indicates velocity v z The variance;

[0023] The sigma point generated based on the current state and covariance is represented by the following formula:

[0024] x k =UT(x k ,P k ,λ)

[0025] Where, x k denoted as sigma point; UT(·) represents a function that generates sigma points using an unscented transformation; x k P represents the current state vector; k Let λ represent the covariance matrix of the current state; λ represents the parameter that adjusts the sigma distribution, and its calculation formula is shown below:

[0026] λ=α 2 (L+κ)-L

[0027] Where α represents the parameter controlling the density of the sigma point distribution; L represents the dimension of the state vector; and κ represents the scale parameter.

[0028] The state transition function for propagating the sigma point is expressed by the following formula:

[0029] x k|k-1 =f(x) k ,u k )

[0030] Where, x k|k-1 Let f(·) represent the predicted value at time k-1, i.e., the prediction of the state before time k; f(·) represents the state transition function; u k Indicates control input;

[0031] The predicted state and prediction covariance are calculated using the propagated sigma point set.

[0032] The formula for calculating the predicted state is as follows:

[0033]

[0034] in, This represents the predicted state of the system at time k-1; denoted by sigma point weight coefficients, f() represents the state transition function, which describes the transition process of the system state from time k to time k-1; U represents the state of the i-th sigma point. k This represents the control input at time k;

[0035] The formula for calculating the predicted covariance is shown below:

[0036]

[0037] Among them, P k|k-1 The predictive covariance matrix represents the state at time k-1; Q represents the weighting coefficient; k The process noise covariance matrix is ​​represented by the following formula:

[0038] Q k =αQ k-1 +(1-α)Cov(w k )

[0039] Where α is a parameter used to control the update rate of the process noise covariance matrix; Cov(·) represents the process noise covariance matrix; w k Indicates process noise;

[0040] Collect and organize observation data from GPS and LiDAR to generate the observation vector shown in the following formula:

[0041]

[0042] in, Indicates the GPS x-axis observation position; Indicates the GPS y-axis observation position; Indicates the GPS observation position along the z-axis; Indicates the LiDAR x-axis observation position; Indicates the GPS y-axis observation position; Indicates the GPS z-axis observation position;

[0043] Set the sigma point set Z k|k-1 =h(x k|k-1 The predicted observations are obtained by propagating through the observation model; where h(·) represents the observation model that maps the state space to the observation space.

[0044] The predicted values ​​and prediction covariance are calculated using the results of the sigma point set in the observation space.

[0045] The formula for calculating the predicted value is as follows:

[0046]

[0047] in, This represents the predicted value of the state at time k-1. denoted by sigma point weight coefficients, h() is the observation function used to describe the mapping of the system state in the observation space; This represents the predicted state of the i-th sigma point;

[0048] The formula for calculating the predicted covariance is shown below:

[0049]

[0050] in, R represents the covariance matrix of the predicted values ​​at time k-1; k The observation noise covariance matrix is ​​represented by the following formula:

[0051] R k =βR k-1 +(1-β)Cov(v k )

[0052] Where β represents the update rate adjustment factor for the observation noise covariance matrix; v k This represents the system's observation noise;

[0053] The Kalman gain is calculated using the prediction covariance and the observation residual covariance, as shown in the following formula:

[0054]

[0055] Among them, K k Indicates Kalman gain; The inverse of the covariance matrix of the observation residuals at time k;

[0056] Update the state estimate using Kalman gain;

[0057] The covariance matrix of the updated state estimate is calculated using the following formula:

[0058]

[0059] Based on actual observations and predicted residuals, dynamically adjust the process noise covariance Q. k and the observation noise covariance matrix R k ;

[0060] Final prediction vector x k =[p x ,p y ,p z ,v x ,v y ,v z ]; where p x Indicates the predicted x-axis position; p y Indicates the predicted y-axis position; p z Indicates the predicted z-axis position; v x Indicates the predicted x-axis velocity; v yThis indicates the predicted y-axis velocity; v z This indicates the predicted z-axis velocity.

[0061] Step S3, which involves constructing a navigation anomaly data detection model and training and optimizing the model using the fused data from step S2, specifically includes:

[0062] (3-1) Construction of a navigation anomaly data detection model:

[0063] The navigation anomaly data detection model includes a coder network, a decoder network, a discriminator network, and a classifier network;

[0064] (1) Encoder network structure:

[0065] The encoder network receives an input sample x and a label c. By capturing the key features of the input sample, it maps the input sample to a latent vector z, which represents the position of the input sample in the latent space.

[0066] An encoder network consists of an input layer, a hidden layer, and an output layer;

[0067] The input layer consists of 8 input neurons;

[0068] The hidden layer consists of three layers: the first layer is a fully connected layer with 512 neurons and the activation function is ReLU; the second layer is a fully connected layer with 256 neurons and the activation function is ReLU; and the third layer is a fully connected layer with 64 neurons and the activation function is ReLU.

[0069] The output layer consists of two parts: the output mean μ and the logarithm of the output contrast logσ. 2 Each part consists of M output neurons; where M represents the latent spatial dimension.

[0070] (2) Decoder network structure:

[0071] The decoder network receives the latent vector z and the true class label c, and uses the latent vector to reconstruct or generate data x′ such that x′ is as close as possible to the original input sample x.

[0072] The decoder network consists of an input layer, a hidden layer, and an output layer;

[0073] The input layer consists of M input neurons;

[0074] The hidden layer consists of three layers: the first layer is a fully connected layer with 64 neurons and the activation function is ReLU; the second layer is a fully connected layer with 256 neurons and the activation function is ReLU; and the third layer is a fully connected layer with 512 neurons and the activation function is ReLU.

[0075] The output layer consists of an 8-dimensional column vector;

[0076] (3) Discriminator network structure:

[0077] The discriminator network is used to learn to distinguish between real samples and fake samples generated by the decoder;

[0078] The discriminator network consists of an input layer, a hidden layer, and an output layer;

[0079] The input layer consists of 8 input neurons;

[0080] The hidden layer consists of three layers: the first layer is a fully connected layer with 128 neurons and the activation function is LeakyReLU; the second layer is a fully connected layer with 32 neurons and the activation function is LeakyReLU; the third layer is a fully connected layer with 32 neurons and the activation function is LeakyReLU. The fully connected layer is represented by the following formula:

[0081] F(X) = g(WX + b)

[0082] Where F(·) represents the output of the fully connected layer; X represents the input vector; g(·) represents the activation function; W represents the weight matrix connecting the input X and the output F; and b represents the bias vector.

[0083] The output layer consists of one neuron; the activation function is sigmoid; there is one output value, ranging from 0 to 1, representing the probability that a sample is real or fake; the output layer is represented by the following formula:

[0084] O(X) = σ(WX + b)

[0085] Where O(·) represents the output of the output layer; σ(·) represents the activation function;

[0086] (4) Classifier network structure:

[0087] The classifier network predicts the probability of the true class label c from the input sample x;

[0088] A classifier network consists of an input layer, a hidden layer, and an output layer.

[0089] The input layer consists of 8 input neurons;

[0090] The hidden layer consists of three layers: the first layer is a convolutional layer with a kernel size of 3, 64 filters, and ReLU activation function; the second layer is a convolutional layer with a kernel size of 3, 128 filters, and ReLU activation function; the third layer is a convolutional layer with a kernel size of 3, 256 filters, and ReLU activation function; and the fourth layer is a fully connected layer with 128 neurons and ReLU activation function.

[0091] The output layer consists of one neuron; the activation function is sigmoid; there is one output value, which ranges from 0 to 1, and the output value represents the probability that the sample is real or fake.

[0092] The output of the encoder network serves as the input to the decoder network; the output of the decoder network serves as the input to both the discriminator network and the classifier network; the output of the discriminator network is the discrimination result; and the output of the decoder network is the classification result.

[0093] (3-2) Training and optimization of the navigation anomaly data detection model:

[0094] Using the data fused in step S2, the encoder network and decoder network in the model are trained, and the output data of the encoder network and the decoder network are obtained.

[0095] The discriminator network and classifier network in the model are trained using the output data of the encoder network and the decoder network.

[0096] Repeatedly iterate the training, update the parameters of the navigation anomaly data detection model, until the set conditions are met, and complete the training of the navigation anomaly data detection model.

[0097] The navigation anomaly data detection model needs to meet certain conditions to complete training, including: the outputs of the discriminator network and the classifier network both meet the set thresholds; and the training parameters of the decoder network converge.

[0098] Optimize the navigation anomaly data detection model by using its loss function;

[0099] The loss functions of the navigation anomaly data detection model include KL divergence loss function, reconstruction loss function, discriminator loss function, classifier loss function, and generator loss function;

[0100] 1) KL divergence loss function:

[0101] The encoder network uses the KL divergence loss function to ensure that the distribution of the latent space is as close as possible to the prior distribution.

[0102] The KL divergence loss function is calculated using the following formula:

[0103]

[0104] Among them, L KL σ represents the KL divergence loss function value; μ represents the mean of the latent vector output by the encoder; σ represents the covariance of the latent vector output by the encoder network.

[0105] 2) Reconstruct the loss function:

[0106] The reconstruction loss function is used to measure the difference between the decoder network output and the original input data;

[0107] The reconstruction loss function is calculated using Euclidean distance, and the formula is shown below:

[0108]

[0109] Among them, L rec represents the reconstruction loss function value; x represents the input sample; D(·) represents the decoder network output; E(·) represents the encoder network output; This represents the square of the vector norm;

[0110] 3) Discriminator network loss function:

[0111] The discriminator network loss function is calculated using the following formula:

[0112]

[0113] Among them, L D This represents the value of the discriminator network loss function; p represents the expectation of the actual data sample; data Represents the distribution of real data; Let represent the expected value of the noise distribution; z represents the latent vector; p z G(·) represents the latent spatial distribution; G(·) represents the samples generated by the generator network based on the input noise.

[0114] 4) Loss function of classifier network:

[0115] The loss function of the classifier network is calculated using the following formula:

[0116]

[0117] Among them, L C denoted by , representing the loss function value of the classifier network; P(·) represents the classifier function, whose output is the probability of whether the given input data is abnormal; c represents the true class label.

[0118] 5) Generator loss function:

[0119] The generator loss function is calculated using the following formula:

[0120]

[0121] Among them, L G This represents the generator loss function value;

[0122] 6) Loss function of the navigation anomaly data detection model:

[0123] The loss function of the navigation anomaly data detection model is calculated using the following formula:

[0124] L = L D +α1L KL +α2L rec +α3L C +α4L G

[0125] Where L represents the loss function value of the navigation anomaly data detection model; α i The hyperparameters, i = 1, 2, 3, 4, are used to balance the weights of each component in the loss function of the navigation anomaly data detection model.

[0126] Step S4 describes using the navigation anomaly data detection model trained and optimized in step S3 to detect anomalies in the actual navigation data. Specifically, this includes:

[0127] The navigation anomaly data detection model trained and optimized in step S3 is used to detect anomalies in the actual navigation data.

[0128] The multi-source navigation anomaly data detection method provided by this invention uses UKF (Universal Keyless Filtering) to fuse multi-source navigation data. UKF has the characteristics of high accuracy, high robustness and nonlinearity, providing better filtering performance and more accurate estimation. By constructing a navigation data anomaly detection model, it can better capture complex patterns and relationships in the data. The robustness, accuracy and security of the method of this invention are enhanced. Attached Figure Description

[0129] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0130] Figure 2 This is a schematic diagram of the navigation data anomaly detection model of the method of the present invention. Detailed Implementation

[0131] like Figure 1 The diagram shown is a flowchart of the method of the present invention: The multi-source navigation anomaly data detection method provided by the present invention includes the following steps:

[0132] S1. Collect and summarize navigation data, and preprocess the navigation data; specifically including:

[0133] (1-1) Navigation data:

[0134] Collect and summarize navigation data, including satellite navigation data (GPS), lidar (LiDAR) data, and gyroscope data for the same detected target;

[0135] In the method of this invention, the collected and summarized gyroscope data is ten-axis gyroscope data;

[0136] (1-2) Navigation data classification:

[0137] The navigation data quality is checked to verify its completeness and accuracy; the data is checked for missing or outlier values. If there are no missing or outlier values, the data is classified as "normal data"; if there are missing or outlier values, the data is classified as "abnormal data".

[0138] (1-3) Navigation data preprocessing:

[0139] Perform data time and sampling rate preprocessing on the classified data;

[0140] Time preprocessing includes: standardizing timestamps and checking the timestamps to ensure data time synchronization;

[0141] Sampling rate preprocessing includes: determining the least common multiple of the sampling rates of all sensors as the target sampling rate, resampling the data, and verifying the sampled data;

[0142] The resampled data is standardized in terms of unit format;

[0143] In the method of this invention, the unit of velocity is set to m / s and the unit of angular velocity is set to rad / s; the formatted data is converted into column vector form.

[0144] S2. Perform fusion processing on the preprocessed navigation data from step S1; specifically including:

[0145] The three types of navigation data preprocessed in step S1 are fused using an unscented Kalman filter (UKF), specifically including:

[0146] Set the state vector x0 = [p x ,p y ,p z ,v x ,v y ,v z State covariance matrix Where, p x Indicates the predicted x-axis position; p y Indicates the predicted y-axis position; p z Indicates the predicted z-axis position; v x Indicates the predicted x-axis velocity; v y This indicates the predicted y-axis velocity; v z This represents the predicted z-axis velocity; diag(·) represents the diagonal matrix; σ represents the initial uncertainty of each state variable; Indicates position p x The variance; Indicates position p y The variance; Indicates position p z The variance; Indicates velocity v x The variance; Indicates velocity v y The variance; Indicates velocity v z The variance;

[0147] The sigma point generated based on the current state and covariance is represented by the following formula:

[0148] x k =UT(x k ,P k ,λ)

[0149] Where, x k denoted as sigma point; UT(·) represents a function that generates sigma points using an unscented transformation; x k P represents the current state vector; k Let λ represent the covariance matrix of the current state; λ represents the parameter that adjusts the sigma distribution, and its calculation formula is shown below:

[0150] λ=α 2 (L+κ)-L

[0151] Where α represents the parameter controlling the density of the sigma point distribution; L represents the dimension of the state vector; and κ represents the scale parameter.

[0152] The state transition function for propagating the sigma point is expressed by the following formula:

[0153] x k|k-1 =f(x) k ,u k )

[0154] Where, x k|k-1 Let f(·) represent the predicted value at time k-1, i.e., the prediction of the state before time k; f(·) represents the state transition function; u k Indicates control input;

[0155] The predicted state and prediction covariance are calculated using the propagated sigma point set.

[0156] The formula for calculating the predicted state is as follows:

[0157]

[0158] in, This represents the predicted state of the system at time k-1; denoted by sigma point weight coefficients, f() represents the state transition function, which describes the transition process of the system state from time k to time k-1; U represents the state of the i-th sigma point. k This represents the control input at time k;

[0159] The formula for calculating the predicted covariance is shown below:

[0160]

[0161] Among them, P k|k-1 The predictive covariance matrix represents the state at time k-1; Q represents the weighting coefficient; k The process noise covariance matrix is ​​represented by the following formula:

[0162] Q k =αQ k-1 +(1-α)Cov(w k )

[0163] Where α controls the update rate of the process noise covariance matrix; Cov(·) represents the process noise covariance matrix, and w k Indicates process noise;

[0164] Collect and organize observation data from GPS and LiDAR to generate the observation vector shown in the following formula:

[0165]

[0166] in, Indicates the GPS x-axis observation position; Indicates the GPS y-axis observation position; Indicates the GPS observation position along the z-axis; Indicates the LiDAR x-axis observation position; Indicates the GPS y-axis observation position; Indicates the GPS z-axis observation position;

[0167] Set the sigma point set Z k|k-1 =h(x kk-1 The predicted observations are obtained by propagating through the observation model; where h(·) represents the observation model that maps the state space to the observation space.

[0168] The predicted values ​​and prediction covariance are calculated using the results of the sigma point set in the observation space.

[0169] The formula for calculating the predicted value is as follows:

[0170]

[0171] in, This represents the predicted value of the state at time k-1. denoted by sigma point weight coefficients, h() is the observation function used to describe the mapping of the system state in the observation space; This represents the predicted state of the i-th sigma point;

[0172] The formula for calculating the predicted covariance is shown below:

[0173]

[0174] in, R represents the covariance matrix of the predicted values ​​at time k-1; k The observation noise covariance matrix is ​​represented by the following formula:

[0175] R k =βR k-1 +(1-β)Cov(v k )

[0176] Where β represents the update rate adjustment factor for the observation noise covariance matrix; v k This represents the system's observation noise;

[0177] The Kalman gain is calculated using the prediction covariance and the observation residual covariance, as shown in the following formula:

[0178]

[0179] Among them, K k Indicates Kalman gain; The inverse of the covariance matrix of the observation residuals at time k;

[0180] Update the state estimate using Kalman gain;

[0181] The covariance matrix of the updated state estimate is calculated using the following formula:

[0182]

[0183] Based on actual observations and predicted residuals, dynamically adjust the process noise covariance Q. k and the observation noise covariance matrix R k ;

[0184] Final prediction vector x k =[p x ,p y ,p z ,vx ,v y ,v z ]; where p x Indicates the predicted x-axis position; p y Indicates the predicted y-axis position; p z Indicates the predicted z-axis position; v x Indicates the predicted x-axis velocity; v y This indicates the predicted y-axis velocity; v z This indicates the predicted z-axis velocity;

[0185] S3. Construct a navigation anomaly data detection model, and train and optimize the constructed model using the fused data from step S2; specifically including:

[0186] (3-1) Construction of a navigation anomaly data detection model:

[0187] The navigation anomaly data detection model includes a coder network, a decoder network, a discriminator network, and a classifier network;

[0188] (1) Encoder network structure:

[0189] The encoder network receives an input sample x and a label c. By capturing the key features of the input sample, it maps the input sample to a latent vector z, which represents the position of the input sample in the latent space.

[0190] An encoder network consists of an input layer, a hidden layer, and an output layer;

[0191] The input layer consists of 8 input neurons;

[0192] The hidden layer consists of three layers: the first layer is a fully connected layer with 512 neurons and the activation function is ReLU; the second layer is a fully connected layer with 256 neurons and the activation function is ReLU; and the third layer is a fully connected layer with 64 neurons and the activation function is ReLU.

[0193] The output layer consists of two parts: the output mean μ and the logarithm of the output contrast logσ. 2 Each part consists of M output neurons; where M represents the latent spatial dimension.

[0194] (2) Decoder network structure:

[0195] The decoder network receives the latent vector z and the true class label c, and uses the latent vector to reconstruct or generate data x′, making x′ close to the original input sample x;

[0196] The decoder network consists of an input layer, a hidden layer, and an output layer;

[0197] The input layer consists of M input neurons;

[0198] The hidden layer consists of three layers: the first layer is a fully connected layer with 64 neurons and the activation function is ReLU; the second layer is a fully connected layer with 256 neurons and the activation function is ReLU; and the third layer is a fully connected layer with 512 neurons and the activation function is ReLU.

[0199] The output layer consists of an 8-dimensional column vector;

[0200] (3) Discriminator network structure:

[0201] The discriminator network is used to learn to distinguish between real samples and fake samples generated by the decoder;

[0202] The discriminator network consists of an input layer, a hidden layer, and an output layer;

[0203] The input layer consists of 8 input neurons;

[0204] The hidden layer consists of three layers: the first layer is a fully connected layer with 128 neurons and the activation function is LeakyReLU; the second layer is a fully connected layer with 32 neurons and the activation function is LeakyReLU; the third layer is a fully connected layer with 32 neurons and the activation function is LeakyReLU. The fully connected layer is represented by the following formula:

[0205] F(X) = g(WX + b)

[0206] Where F(·) represents the output of the fully connected layer; X represents the input vector; g(·) represents the activation function; W represents the weight matrix connecting the input X and the output F; and b represents the bias vector.

[0207] The output layer consists of one neuron; the activation function is sigmoid; there is one output value, ranging from 0 to 1, representing the probability that a sample is real or fake; the output layer is represented by the following formula:

[0208] O(X) = σ(WX + b)

[0209] Where O(·) represents the output of the output layer; σ(·) represents the activation function;

[0210] (4) Classifier network structure:

[0211] The classifier network predicts the probability of the true class label c from the input sample x;

[0212] A classifier network consists of an input layer, a hidden layer, and an output layer.

[0213] The input layer consists of 8 input neurons;

[0214] The hidden layer consists of three layers: the first layer is a convolutional layer with a kernel size of 3, 64 filters, and ReLU activation function; the second layer is a convolutional layer with a kernel size of 3, 128 filters, and ReLU activation function; the third layer is a convolutional layer with a kernel size of 3, 256 filters, and ReLU activation function; and the fourth layer is a fully connected layer with 128 neurons and ReLU activation function.

[0215] The output layer consists of one neuron; the activation function is sigmoid; there is one output value, which ranges from 0 to 1, and the output value represents the probability that the sample is real or fake.

[0216] The output of the encoder network serves as the input to the decoder network; the output of the decoder network serves as the input to both the discriminator network and the classifier network; the output of the discriminator network is the discrimination result; and the output of the decoder network is the classification result.

[0217] like Figure 2 The diagram shown is a schematic representation of the navigation data anomaly detection model of the present invention:

[0218] (3-2) Training and optimization of the navigation anomaly data detection model:

[0219] Using the data fused in step S2, train the encoder network and decoder network in the model, and obtain the output data of the encoder network and the output data of the decoder network;

[0220] The discriminator and classifier networks in the model are trained using the output data of the encoder network and the decoder network.

[0221] Repeatedly iterate the training, update the parameters of the navigation anomaly data detection model, until the set conditions are met, and complete the training of the navigation anomaly data detection model.

[0222] The navigation anomaly data detection model needs to meet certain conditions to complete training, including: the outputs of the discriminator network and the classifier network both meet the set thresholds; and the training parameters of the decoder network converge.

[0223] The navigation anomaly data detection model is optimized by using its loss function.

[0224] The loss functions of the navigation anomaly data detection model include KL divergence loss function, reconstruction loss function, discriminator loss function, classifier loss function, and generator loss function;

[0225] 1) KL divergence loss function:

[0226] The encoder network uses the KL divergence loss function to ensure that the distribution of the latent space is as close as possible to the prior distribution;

[0227] The KL divergence loss function is calculated using the following formula:

[0228]

[0229] Among them, L KL σ represents the KL divergence loss function value; μ represents the mean of the latent vectors output by the encoder; σ represents the covariance of the latent vectors output by the encoder network.

[0230] 2) Reconstruct the loss function:

[0231] The reconstruction loss function is used to measure the difference between the decoder network output and the original input data;

[0232] The reconstruction loss function is calculated using Euclidean distance, and the formula is shown below:

[0233]

[0234] Among them, L rec represents the reconstruction loss function value; x represents the input sample; D(·) represents the encoder output; E(·) represents the decoder output; L represents the vector 2 The square of the norm;

[0235] 3) Discriminator network loss function:

[0236] The discriminator network loss function is calculated using the following formula:

[0237]

[0238] Among them, L D This represents the value of the discriminator network loss function; p represents the expected value of the real data sample x; data Represents the distribution of real data; Let represent the expected value of the noise distribution; z represents the latent vector; p z G(·) represents the latent spatial distribution; G(·) represents the samples generated by the generator network based on the input noise.

[0239] 4) Loss function of classifier network:

[0240] The loss function of the classifier network is calculated using the following formula:

[0241]

[0242] Among them, L Crepresents the loss function value of the classifier network; P(·) represents the classifier function, the output of which is the probability that the given input data is abnormal; c represents the true class label;

[0243] 5) Generator loss function:

[0244] The generator loss function is calculated using the following formula:

[0245]

[0246] Among them, L G This represents the generator loss function value;

[0247] 6) Loss function of the navigation anomaly data detection model:

[0248] The loss function of the navigation anomaly data detection model is calculated using the following formula:

[0249] L = L D +α1L KL +α2L rec +α3L C +α4L G

[0250] Where L represents the loss function value of the navigation anomaly data detection model; α i The hyperparameters, i = 1, 2, 3, 4, are used to balance the weights of each component in the loss function of the navigation anomaly data detection model.

[0251] S4. Using the navigation anomaly data detection model trained and optimized in step S3, perform anomaly data detection on the actual navigation data; specifically including:

[0252] The navigation anomaly data detection model trained and optimized in step S3 is used to detect anomalies in the actual navigation data.

Claims

1. A method for detecting multi-source navigation anomaly data, comprising the following steps: S1. Collect and summarize navigation data, and preprocess the navigation data; specifically including: (1-1) Navigation data: Collect and summarize navigation data, including satellite navigation data, lidar data, and gyroscope data for the same target; (1-2) Navigation data classification: The navigation data quality is checked to verify its completeness and accuracy; the data is checked for missing or outlier values. If there are no missing or outlier values, the data is classified as "normal data"; if there are missing or outlier values, the data is classified as "abnormal data". (1-3) Navigation data preprocessing: Preprocess the categorized data by adjusting the data time and sampling rate. Time preprocessing includes: standardizing timestamps and checking the timestamps to ensure data time synchronization; Sampling rate preprocessing includes: determining the least common multiple of the sampling rates of all sensors as the target sampling rate, resampling the data, and verifying the sampled data; The resampled data is standardized in terms of unit format; S2. Perform fusion processing on the preprocessed navigation data from step S1; specifically including: The three types of navigation data preprocessed in step S1 are fused using an unscented Kalman filter, specifically including: Set the state vector State covariance matrix ;in, Indicates the predicted x-axis position; Indicates the predicted y-axis position; Indicates the predicted z-axis position; This indicates the predicted x-axis velocity; This indicates the predicted y-axis velocity; This indicates the predicted z-axis velocity; Represents a diagonal matrix; This represents the initial uncertainty of each state variable; Indicates position The variance; Indicates position The variance; Indicates position The variance; Indicates speed The variance; Indicates speed The variance; Indicates speed The variance; The sigma point generated based on the current state and covariance is represented by the following formula: in, Represents a sigma point; This represents a function that generates sigma points using an unscented transformation. Represents the current state vector; The covariance matrix representing the current state; The parameter representing the adjustment of the sigma distribution is calculated using the following formula: in, This represents a parameter that controls the density of the sigma point distribution; The dimension of the state vector; Indicates the scale parameter; The state transition function for propagating the sigma point is expressed by the following formula: in, This represents the predicted value at time k-1, that is, the prediction of the state before time k. Represents the state transition function; Indicates control input; The predicted state and prediction covariance are calculated using the propagated sigma point set. The formula for calculating the predicted state is as follows: in, This represents the predicted state of the system at time k-1; This represents the weight coefficient of the sigma point. This represents the state transition function, used to describe the transition process of the system state from time k to time k-1; This represents the state of the i-th sigma point. This represents the control input at time k; The formula for calculating the predicted covariance is shown below: in, The predictive covariance matrix represents the state at time k-1; Indicates the weighting coefficient; The process noise covariance matrix is ​​represented by the following formula: in, These are variables used to control the update rate of the process noise covariance matrix; The covariance matrix represents the process noise; Indicates process noise; Collect and organize observation data from satellite navigation and lidar to generate the observation vector shown in the following formula: in, Indicates the GPS x-axis observation position; Indicates the GPS y-axis observation position; Indicates the GPS observation position along the z-axis; Indicates the LiDAR x-axis observation position; Indicates the GPS y-axis observation position; Indicates the GPS z-axis observation position; Set of sigma points Predicted observations are obtained through propagation via an observation model; among which, This represents an observation model that maps the state space to the observation space; The predicted values ​​and prediction covariance are calculated using the results of the sigma point set in the observation space. The formula for calculating the predicted value is as follows: in, This represents the predicted value of the state at time k-1. This represents the weight coefficient of the sigma point. The observation function describes the mapping of the system state in the observation space; This represents the predicted state of the i-th sigma point; The formula for calculating the predicted covariance is shown below: in, The covariance matrix represents the predicted values ​​at time k-1; The observation noise covariance matrix is ​​represented by the following formula: in, This represents the adjustment factor for the update rate of the observation noise covariance matrix; This represents the system's observation noise; The Kalman gain is calculated using the prediction covariance and the observation residual covariance, as shown in the following formula: in, Indicates Kalman gain; The inverse of the covariance matrix of the observation residuals at time k; Update the state estimate using Kalman gain; The covariance matrix of the updated state estimate is calculated using the following formula: Based on actual observations and predicted residuals, dynamically adjust the process noise covariance. and observation noise covariance matrix ; Final prediction vector ;in, Indicates the predicted x-axis position; Indicates the predicted y-axis position; Indicates the predicted z-axis position; This indicates the predicted x-axis velocity; This indicates the predicted y-axis velocity; This indicates the predicted z-axis velocity; S3. Construct a navigation anomaly data detection model, and use the data fused and processed in step S2 to train and optimize the constructed navigation anomaly data detection model; S4. Using the navigation anomaly data detection model trained and optimized in step S3, anomaly data detection is performed on the actual navigation data.

2. The method for detecting multi-source navigation anomaly data according to claim 1, characterized in that... Step S3 involves constructing a navigation anomaly data detection model and training and optimizing the model using the fused data from step S2. Specifically, this includes: (3-1) Construction of the navigation anomaly data detection model: The navigation anomaly data detection model includes an encoder network, a decoder network, a discriminator network, and a classifier network; The output of the encoder network serves as the input to the decoder network; the output of the decoder network serves as the input to both the discriminator network and the classifier network; the output of the discriminator network is the discrimination result; and the output of the decoder network is the classification result. (3-2) Training and optimization of the navigation anomaly data detection model: Using the data fused in step S2, train the encoder network and decoder network in the model, and obtain the output data of the encoder network and the output data of the decoder network; The discriminator and classifier networks in the model are trained using the output data of the encoder network and the decoder network. Repeatedly iterate the training, update the parameters of the navigation anomaly data detection model, until the set conditions are met, and complete the training of the navigation anomaly data detection model. The navigation anomaly data detection model needs to meet certain conditions to complete training, including: the outputs of the discriminator network and the classifier network both meet the set thresholds; and the training parameters of the decoder network converge. The navigation anomaly data detection model is optimized by using the loss function of the navigation anomaly data detection model.

3. The method for detecting multi-source navigation anomaly data according to claim 2, characterized in that... The construction of the navigation anomaly data detection model specifically includes: The navigation anomaly data detection model includes an encoder network, a decoder network, a discriminator network, and a classifier network; (1) Encoder network structure: Encoder network receives input samples and tags By capturing the key features of the input samples, the input samples are mapped to latent vectors. latent vector This represents the position of the input sample in the latent space; An encoder network consists of an input layer, a hidden layer, and an output layer; The input layer consists of 8 input neurons; The hidden layer consists of three layers: the first layer is a fully connected layer with 512 neurons and the activation function is ReLU; the second layer is a fully connected layer with 256 neurons and the activation function is ReLU; and the third layer is a fully connected layer with 64 neurons and the activation function is ReLU. The output layer consists of two parts: the output mean. Output the logarithm of the contrast Each part includes There are 1 output neuron; among them, Represents the potential spatial dimension; (2) Decoder network structure: The decoder network receives the latent vector. and real category labels Data reconstruction or generation using latent vectors , making Closest to the original input sample ; The decoder network consists of an input layer, a hidden layer, and an output layer; The input layer includes One input neuron; The hidden layer consists of three layers: the first layer is a fully connected layer with 64 neurons and the activation function is ReLU; the second layer is a fully connected layer with 256 neurons and the activation function is ReLU; and the third layer is a fully connected layer with 512 neurons and the activation function is ReLU. The output layer consists of an 8-dimensional column vector; (3) Discriminator network structure: The discriminator network is used to learn to distinguish between real samples and fake samples generated by the decoder; The discriminator network consists of an input layer, a hidden layer, and an output layer; The input layer consists of 8 input neurons; The hidden layer consists of three layers: the first layer is a fully connected layer with 128 neurons and the activation function is LeakyReLU; the second layer is a fully connected layer with 32 neurons and the activation function is LeakyReLU; the third layer is a fully connected layer with 32 neurons and the activation function is LeakyReLU. The fully connected layer is represented by the following formula: in, Indicates the output of the fully connected layer; Represents the input vector; Indicates the activation function; This represents the weight matrix connecting the input X and the output F; Represents the bias vector; The output layer consists of one neuron; the activation function is sigmoid; there is one output value, ranging from 0 to 1, representing the probability that a sample is real or fake; the output layer is represented by the following formula: in, Indicates the output of the output layer; Indicates the activation function; (4) Classifier network structure: The classifier network uses input samples Predict the true category label The probability of; A classifier network consists of an input layer, a hidden layer, and an output layer. The input layer consists of 8 input neurons; The hidden layer consists of three layers: the first layer is a convolutional layer with a kernel size of 3, 64 filters, and ReLU activation function; the second layer is a convolutional layer with a kernel size of 3, 128 filters, and ReLU activation function; the third layer is a convolutional layer with a kernel size of 3, 256 filters, and ReLU activation function; and the fourth layer is a fully connected layer with 128 neurons and ReLU activation function. The output layer consists of one neuron; the activation function is sigmoid; there is one output value, which ranges from 0 to 1, and the output value represents the probability that the sample is real or fake.

4. The method for detecting multi-source navigation anomaly data according to claim 3, characterized in that... The loss function of the navigation anomaly data detection model specifically includes: The loss functions of the navigation anomaly data detection model include KL divergence loss function, reconstruction loss function, discriminator loss function, classifier loss function, and generator loss function; 1) KL divergence loss function: The encoder network uses the KL divergence loss function to ensure that the distribution of the latent space is as close as possible to the prior distribution. The KL divergence loss function is calculated using the following formula: in, This represents the KL divergence loss function value; This represents the mean of the latent vectors output by the encoder. This represents the covariance of the latent vectors output by the encoder network; 2) Reconstruct the loss function: The reconstruction loss function is used to measure the difference between the decoder network output and the original input data; The reconstruction loss function is calculated using Euclidean distance, and the formula is shown below: in, This represents the value of the reconstruction loss function; Indicates the input sample; Indicates encoder output; Indicates the decoder output; Representing vectors The square of the norm; 3) Discriminator network loss function: The discriminator network loss function is calculated using the following formula: in, This represents the value of the discriminator network loss function; This represents the expected value of the real data sample x; Represents the distribution of real data; This represents the expected value of the noise distribution; Represents the latent vector; Indicates the potential spatial distribution; This represents the samples generated by the generator network based on the input noise; 4) Loss function of classifier network: The loss function of the classifier network is calculated using the following formula: in, This represents the loss function value of the classifier network; This represents a classifier function, whose output is the probability of whether a given input data is abnormal. Indicates the true category label; 5) Generator loss function: The generator loss function is calculated using the following formula: in, This represents the generator loss function value; 6) Loss function of the navigation anomaly data detection model: The loss function of the navigation anomaly data detection model is calculated using the following formula: in, This represents the loss function value of the navigation anomaly data detection model; Indicates hyperparameters, The weights of each component in the loss function of the navigation anomaly data detection model are balanced by hyperparameters.

5. The method for detecting multi-source navigation anomaly data according to claim 4, characterized in that... Step S4 describes using the navigation anomaly data detection model trained and optimized in step S3 to detect anomalies in the actual navigation data. Specifically, this includes: The navigation anomaly data detection model trained and optimized in step S3 is used to detect anomalies in the actual navigation data.

Images