Space-time security protection situation prediction method based on machine learning

Abstract

The invention discloses a machine learning-based space-time security protection situation prediction method, which comprises the following steps of S1, performing feature extraction on a real navigation signal and an interference signal captured by a navigation receiver, and constructing a navigation quality feature index system; s2, acquiring navigation quality characteristics in the collected sample, and preprocessing the navigation quality characteristics of the sample data; s3, dividing the preprocessed sample data into a long-term sample set and a recent sample set; s4, for the long-term sample set, enabling the model to learn universal features by adopting a self-supervised pre-training mode; step S5, for the recent sample set, adjusting the model by using a fine adjustment method of a low-rank decomposition idea, and performing situation prediction training of key facility space-time safety protection by using a prediction model based on Transform; s6, predicting navigation quality characteristics capable of reflecting the space-time security protection situation by using the prediction model; and S7, evaluating the space-time safety protection prediction situation according to the predicted navigation quality characteristics.

Description

Technical Field

[0001] The present invention relates to the field of satellite navigation and communication technology, and in particular to a method for predicting spatiotemporal security protection situation based on machine learning. Background Art

[0002] The Global Navigation Satellite System (GNSS), with its unique all-weather, high-precision and high-efficiency characteristics, has become an indispensable technological pillar of modern society, playing a vital role in both civilian and military applications. In particular, my country's Beidou satellite navigation system, with the successful realization of its global networking, is not only widely used in China, but its international influence is also expanding, and its future application potential is unlimited. However, the satellite navigation system is not impeccable. Its navigation signal has relatively low power when it reaches the ground and has limited penetration. These inherent limitations make the system particularly vulnerable to electromagnetic interference, which in turn affects its stability and reliability.

[0003] In this context, ensuring the space-time security of satellite navigation systems has become an urgent issue to be resolved. Therefore, the space-time security protection situation prediction based on navigation quality is particularly important, as it can provide early warning and guarantee for the normal operation of the system.

[0004] Although a variety of algorithms have been developed in the field of time series prediction and have been applied in various scenarios, the existing methods are still insufficient in the specific practice of spatiotemporal security protection situation prediction, which is specifically manifested in the following aspects:

[0005] (1) Lack of a systematic prediction parameter system. Existing research has not yet clearly proposed a complete prediction parameter framework to guide the prediction of spatiotemporal security protection situation.

[0006] (2) Current time series forecasting methods still face challenges in dealing with the randomness, uncertainty, and nonlinear relationships of spatiotemporal data. Summary of the invention

[0007] In response to the above-mentioned technical problems to be solved, the present invention provides a method for predicting the spatiotemporal security protection situation based on machine learning, which takes into account the dynamic characteristics of spatiotemporal data and the complexity of security protection, and provides a reliable basis for the analysis and prediction of the spatiotemporal security situation of Beidou navigation and other global satellite navigation systems.

[0008] In order to solve the above technical problems, the technical solution proposed by the present invention is:

[0009] A method for predicting spatiotemporal security protection situation based on machine learning, specifically comprising the following steps:

[0010] Step S1, performing feature extraction on the real navigation signal and the interference signal captured by the navigation receiver to construct a navigation quality feature index system;

[0011] Step S2, obtaining navigation quality features in the collected samples, and preprocessing the navigation quality features of the sample data;

[0012] Step S3, dividing the preprocessed sample data into a long-term sample set and a recent sample set;

[0013] Step S4, for the long-term sample set, a self-supervised pre-training method is used to enable the model to learn common features;

[0014] Step S5: For the recent sample set, the model is adjusted by using the low-rank decomposition fine-tuning method, and the Transformer-based prediction model is used to perform situation prediction training for the spatiotemporal security protection of key facilities;

[0015] Step S6, using the prediction model to predict navigation quality characteristics that can reflect the spatiotemporal security protection situation;

[0016] Step S7, evaluating the spatiotemporal security protection prediction situation based on the predicted navigation quality characteristics.

[0017] As a further improvement of the above technical solution:

[0018] Preferably, in step S1, the navigation quality characteristics include navigation signal integrity analysis, navigation signal availability analysis, navigation signal continuity analysis, and navigation positioning error.

[0019] Preferably, the navigation quality characteristics and influencing factors are dedimensionalized:

[0020]

[0021] Among them, x ij is the jth dimensionless influencing factor of the i-th evaluation index, min i (factor ij ) is the minimum value of a known single influencing factor, min i (factor ij ) is the maximum value of a known single influencing factor.

[0022] Preferably, in step S2, the preprocessing process is specifically as follows:

[0023] S2-1, fill in the missing values ​​of the navigation quality characteristics in the collected samples; fill in the missing values ​​according to the following formula:

[0024]

[0025] in, is the observed value of the jth feature of the ith sample, x kj This is the actual observed value of the jth feature of the kth sample. K is a constant indicating how many nearest neighbors or observations are considered during the calculation.

[0026] S2-2, perform data centering and standardization on the navigation quality features in the collected samples; fill in the missing values ​​according to the following formula:

[0027] x'=x-μ

[0028]

[0029] In the formula, x', x, u, and σ represent the processed data, sample data, sample mean, and sample standard deviation, respectively;

[0030] S2-3, performing data equalization processing on the navigation quality features in the collected samples; performing data equalization processing according to the following formula:

[0031] x new =x i +λ(x j -x i )

[0032] Among them, x new is a newly generated synthetic sample, x i is the original minority class sample, λ is a random number, uniformly distributed between 0 and 1, used to control the distance between the newly generated sample and the original sample, x j It is a sample randomly selected from the K nearest neighbors of the minority class sample.

[0033] Preferably, the step S4 specifically includes the following steps:

[0034] S4-1, initialize data segmentation: for the initialized data, split each time series into non-overlapping subsequence blocks;

[0035] S4-2, self-supervised pre-training:

[0036] Build a Transformer-based pre-trained model: Use a standard Transformer encoder, including self-attention layers and feed-forward network layers to map the observed signal to a latent representation;

[0037] Masking strategy: randomly select some subsequence blocks in the input signal and set their values ​​to zero to construct mask data;

[0038] Pre-training prediction: The masked fragment sequence is input into the Transformer-based pre-training model, and the model outputs the pre-training prediction result, which is the reconstructed value of the masked part;

[0039] Loss calculation: The MSE loss function is used to calculate the difference between the reconstructed value and the true value to evaluate the reconstruction ability of the model;

[0040] Optimization process: Update the model parameters through the gradient descent algorithm to minimize the loss function.

[0041] Preferably, in step S5, the fine-tuning method of low-rank decomposition is used to adjust the model specifically as follows:

[0042] S5-1-1, select the weight matrix:

[0043] The fine-tuning phase focuses on the query W in the self-attention layer q , key W k Sum value W v Adjustment of projection matrix;

[0044] W q and W k Considered as a matrix W qk , dimension d model ×d model , where d model is the model input / output dimension; W v The dimension is d model ×d model ;

[0045] S5-1-2, low-rank decomposition:

[0046] Low-rank decomposition decomposes the weight matrix into the product of two matrices;

[0047] For W qk , use random Gaussian distribution to initialize the low-rank decomposition into matrix A qk and B qk ;

[0048] For W v , use random Gaussian distribution to initialize the low-rank decomposition into matrix A v and B v ;

[0049] Use low-rank decomposition to represent the update of the weight matrix:

[0050] ΔW qk =A qk ·B qk

[0051] ΔW v =A v ·Bv

[0052] S5-1-3, Model Update:

[0053] Freeze the weight w0 of the pre-trained model and decompose the low-rank matrix A qk , B qk , A v and B v Perform training; use the back-propagation algorithm and optimizer to calculate A qk , B qk , A v and B v The gradient of , and update its parameters;

[0054] S5-1-4, Model deployment:

[0055] At deployment time, calculate the new weight matrix W = W0 + ΔW qk +ΔW v ; Use the updated weight matrix for inference to get the prediction result.

[0056] Preferably, in step S5, the situation prediction training process is:

[0057] S5-2-1, model initialization: first initialize the Transformer encoder, linear projection layer, and position encoding parameters;

[0058] S5-2-2, forward propagation: divide the time series data into multiple blocks and map these blocks one by one to the latent space of Transformer;

[0059] S5-2-3, multi-head attention mechanism: focus on different parts of the input sequence at the same time to capture the complex relationships in time series data;

[0060] S5-2-4, feature extraction: extract the features processed by the attention mechanism through the feedforward network;

[0061] S5-2-5, feature concatenation and prediction: concatenate the extracted features, integrate them through the linear layer and linear head, and finally output the prediction results in training;

[0062] S5-2-6, loss calculation: use the mean square error loss function to calculate the difference between the predicted value and the true value;

[0063] S5-2-7, Backpropagation: Use the gradient descent algorithm to update the model parameters and minimize the loss function.

[0064] Preferably, in step S7, the entropy method is used in combination with expert advice to confirm the weights of various navigation quality features to calculate a comprehensive navigation interference degree index that can evaluate the spatiotemporal security protection prediction situation.

[0065] Compared with the prior art, the spatiotemporal security protection situation prediction method based on machine learning provided by the present invention has the following advantages:

[0066] (1) The spatiotemporal security protection situation prediction method based on machine learning of the present invention has established a systematic prediction parameter system, which comprehensively considers the integrity, availability, continuity and positioning error of the navigation signal, ensuring the comprehensiveness of the prediction. In addition, various methods in the data preprocessing stage, such as filling missing values, centralization, standardization and equalization processing, have greatly improved the data quality and laid a solid foundation for accurate prediction.

[0067] (2) The spatiotemporal security protection situation prediction method based on machine learning of the present invention significantly improves the generalization ability of the model through self-supervised pre-training and fine-tuning strategies, making the prediction results more reliable. At the same time, the optimized data segmentation and model training process improves the prediction efficiency, can quickly respond to the needs of spatiotemporal security protection, and enhances the timeliness in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 This is an algorithm flow chart of the spatiotemporal security protection situation prediction method based on machine learning provided by the present invention.

[0069] Figure 2 The method for predicting spatiotemporal security protection situation based on machine learning provided by the present invention is in an embodiment.

[0070] Figure 3 This is a graph showing the change in the loss function when the prediction model is used to train and predict the navigation signal integrity index in the present invention.

[0071] Figure 4 The present invention uses a prediction model to train and predict the comparison between the predicted value and the true value of the navigation quality feature. DETAILED DESCRIPTION

[0072] The specific embodiments of the present invention are described in detail below. It should be understood that the specific embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0073] like Figures 1 to 4 As shown in the figure, the spatiotemporal security protection situation prediction method based on machine learning of the present invention comprises the following steps:

[0074] Step S1, performing feature extraction on the real navigation signal and the interference signal captured by the navigation receiver to construct a navigation quality feature index system.

[0075] Taking the spatiotemporal security protection of key infrastructure of mobile communications but not limited to mobile communications as the application scenario, feature extraction operations are performed on the real navigation signals and interference signals captured by the navigation receiver to identify the typical features of the navigation quality, which are recorded as "navigation quality features".

[0076] In the process of constructing the navigation quality characteristic index system, it is ensured that the selected indexes can comprehensively and accurately map the comprehensive situation of the navigation system and reflect the intuitiveness of its prediction results. The index selection follows the following principles:

[0077] (1) Principle of practicality: The selected indicators have practical application value and can intuitively and truly reflect the performance of navigation signals in actual operating environments. Key indicators such as navigation signal integrity, navigation signal availability, navigation signal continuity, and navigation positioning error not only reveal the essential characteristics of navigation signals, but also provide a direct reference for the prediction results of the situation.

[0078] (2) Systematic principle: When selecting typical characteristics of navigation quality, the entire life cycle of the navigation signal is considered comprehensively, covering the generation, transmission, reception and processing links. This ensures the integrity of the evaluation index system and enables the assessment of the navigation situation from multiple dimensions.

[0079] (3) Operability: The selected indicators are easy to obtain and quantify through existing technical means, which is convenient for technical personnel to operate and evaluate, while improving the efficiency and accuracy of the prediction process.

[0080] Based on the above principles, practicality and systematicity are taken into consideration when selecting navigation quality characteristic indicators, while operability is fully considered.

[0081] The navigation quality characteristics include the following:

[0082] (1) Navigation signal integrity analysis:

[0083] Integrity=N / M

[0084] In the formula, Integrity represents the completeness of the data, M represents the number of complete observations, and N is defined as the number of observed satellites that meet the satellite elevation angle exceeding the predetermined cutoff elevation angle, where the satellite elevation angle is calculated based on the broadcast ephemeris data. The broadcast ephemeris provides satellite orbit parameters and allows the calculation of the position of the satellite relative to the observation point at any point in time.

[0085] (2) Navigation signal availability analysis:

[0086] According to the constellation value a of the satellite navigation system k,n and Service AvailabilityA service , M is the total number of satellites in the constellation, is the number of all possible combinations of K satellite failures among N satellites, a k,n is the constellation value in the nth combination, and the constellation state probability P in the nth combination is obtained using the following service availability formula: k,n :

[0087]

[0088] Using the constellation state probability calculation formula, the navigation signal availability p is obtained:

[0089]

[0090] In the formula, p n,m is the signal availability of the mth working satellite under the nth combination, m ranges from 1 to (Nk), (1-p n,i ) is the failure probability of a specific i-th satellite under the n-th combination, the value of i is N-k+1, and there are k cases in total.

[0091] (3) Navigation signal continuity analysis:

[0092] 1) Within the set evaluation time range, continuously monitor the operating status of each satellite and record the frequency of its transition from "normal" to "abnormal" status, including the specific time points when these transitions begin and end;

[0093] 2) When analyzing the period of time when the health status changes, the expected data outage events should be excluded and only the unexpected data outage events should be counted. The frequency calculation is based on the time unit of each hour to quantify the number of unexpected outages for each satellite;

[0094] 3) Accumulate the total working time of each satellite since it joined the system;

[0095] 4) Using a specific continuity evaluation formula, quantify and evaluate the ability of each satellite to provide uninterrupted services within a given period.

[0096]

[0097] Where, t start ,t top ,t end They are the recording start time, selection time, and recording end time respectively. If the satellite health status is healthy at a certain moment, bool(Status) takes 1, otherwise it takes 0.

[0098] (4) Navigation positioning error:

[0099] The pseudorange mean absolute deviation index is used to average the pseudorange absolute deviations to obtain the pseudorange mean absolute deviation value MAD(ρ):

[0100]

[0101] In the formula, ρ ri is the pseudorange of the measured position of the interfered target machine r in the i-th experiment, and mean(ρ0i) is the average of the pseudoranges of the interference position initially set in the i-th experiment.

[0102] Step S2, obtaining navigation quality features in the collected samples, and preprocessing the navigation quality features of the sample data.

[0103] Specifically, the collected samples refer to the raw data collected by the satellite receiver, including pseudorange, Doppler shift, signal-to-noise ratio and other parameters related to navigation quality. By processing these satellite data, the navigation quality characteristics reflecting the navigation quality are extracted according to the indicator system and calculation formula constructed in step S1. Then, these data are preprocessed, which includes filling the missing values ​​in the data to make it complete; centering the data so that the data is centered on the origin; standardizing the data so that the data has a uniform scale; and finally balancing the data to ensure that the data is evenly distributed among different categories. These preprocessing steps help improve the quality of the data and lay a good foundation for subsequent analysis and modeling.

[0104] The specific preprocessing operations are:

[0105] S2-1, fill in the missing values ​​of the navigation quality characteristics in the collected samples. Fill in the missing values ​​according to the following formula:

[0106]

[0107] in, is the observed value of the jth feature of the ith sample, x kj This is the actual observed value of the jth feature of the kth sample; K is a constant that indicates how many nearest neighbors or observations are considered in the calculation process. The steps for determining the nearest neighbor are as follows:

[0108] 1) For each observation i, calculate its Euclidean distance to all other observations j:

[0109]

[0110] where x ik is the value of observation i on the kth feature, x jk is the value of observation j on the kth feature;

[0111] 2) For each observation i, sort all other observations according to the calculated Euclidean distance;

[0112] 3) Select the first K observations after sorting as the nearest neighbors of observation i.

[0113] S2-2, perform data centering and standardization on the navigation quality features processed in the previous step. Fill in the missing values ​​according to the following formula:

[0114] x'=x-μ

[0115]

[0116] In the formula, x', x, u, and σ represent the processed data, sample data, sample mean, and sample standard deviation, respectively.

[0117] S2-3, perform data equalization on the navigation quality features processed in the previous step. Perform data equalization according to the following formula:

[0118] x new =x i +λ(x j -x i )

[0119] Among them, x new is a newly generated synthetic sample, x i is the original minority class sample, λ is a random number, uniformly distributed between 0 and 1, used to control the distance between the newly generated sample and the original sample, x j It is a sample randomly selected from the K nearest neighbors of the minority class sample. The specific steps are as follows:

[0120] 1) Randomly select a sample from the minority class;

[0121] 2) Find its K nearest neighbors;

[0122] 3) Interpolate between the sample and the randomly selected nearest neighbor to generate a new synthetic sample;

[0123] 4) Repeat this process until the desired number of synthetic samples is reached.

[0124] Step S3, dividing the preprocessed sample data into a long-term sample set and a recent sample set.

[0125] Long-term sample sets usually contain more data points, which helps the model learn more general features and patterns during the pre-training stage; and they contain more historical information and periodic changes, which helps the model understand the long-term trends and patterns of the data; by pre-training on a large amount of data, the model can learn more general feature representations, which helps improve the model's generalization ability on new data; in addition, long-term sample sets require more computing resources for pre-training. Fine-tuning of short-term sample sets is usually faster and can be performed more frequently so that the model can be updated and adjusted in a timely manner. Therefore, long-term sample sets collect situation analysis data for one year.

[0126] The short-term sample set has less data and may contain more recent trends and changes, which is suitable for fine-tuning and final training of the model. It can also reflect more recent dynamics and emergencies, which is more important for real-time prediction and rapid adaptation of the model. Therefore, the short-term sample set uses situation analysis data from the past month.

[0127] Step S4: for the long-term sample set, a self-supervised pre-training method is used to enable the model to learn common features.

[0128] When using large-scale sample sets for self-supervised pre-training, for datasets of tens of thousands, 60% of the data is allocated as a training set for the model to perform preliminary learning and feature extraction. At the same time, 20% of the data is reserved as a validation set to monitor model performance during training and adjust hyperparameters to prevent overfitting. The remaining 20% ​​of the data should be used as a test set to ultimately evaluate the generalization ability of the model and ensure that it remains stable when facing new data.

[0129] When the size of the data set processed reaches millions or more, due to the extremely large amount of data, even a smaller proportion can provide sufficient data points. In this case, 98% of the data should be allocated to the training set to ensure that the model can learn complex features and patterns from the rich data. At the same time, 1% of the data is divided as a validation set to monitor model performance and adjust hyperparameters during training. Finally, 1% of the data is reserved as a test set for a comprehensive evaluation of its performance after the model training is completed.

[0130] Pre-train the long-term sample set, and the pre-trained model will be saved for subsequent use.

[0131] The specific steps include:

[0132] S4-1, initialize data segmentation: For the preprocessed data, the data set arranged in chronological order can be regarded as a continuous set of time series, and each time series is divided into non-overlapping subsequence blocks (patches). Set the length of the time series to T. Set the subsequence block length P and the step size S. Starting from the beginning of the time series, create a subsequence block of length P. According to the step size S, move to the next position and create a new subsequence block. Repeat this process until the end of the time series;

[0133] S4-2, self-supervised pre-training:

[0134] Build a pre-trained Transformer-based model: Use a standard Transformer encoder, including self-attention layers and feed-forward network layers to map the observed signal to a latent representation.

[0135] Masking strategy: Randomly select a portion of subsequence blocks in the input signal and set their values ​​to zero to construct masked data.

[0136] Pre-trained prediction: The masked fragment sequence is input into the Transformer-based pre-trained model, and the model outputs the pre-trained prediction result, that is, the reconstructed value of the masked part.

[0137] Loss calculation: The difference between the reconstructed value and the true value is calculated using the MSE loss function. The MSE loss function is calculated as follows:

[0138]

[0139] in, is the true value - predicted value on the test set.

[0140] Optimization: Use the gradient descent algorithm to update the model parameters and minimize the loss function. The update formula is as follows:

[0141]

[0142] Among them, θ old is the vector of current model parameters, θ new is the updated parameter vector, α is the learning rate, which controls the step size of parameter update. is the gradient of the loss function loss with respect to the parameter θ, which is a vector whose components are the partial derivatives of the loss function corresponding to each parameter in θ. i , the update of gradient descent can be expressed as:

[0143]

[0144] in, is the value at iteration t, is the loss function L over θ i The partial derivative of .

[0145] Step S5, based on the recent sample set after feature extraction, the model is adjusted by using the fine-tuning method of low-rank decomposition idea, and the situation prediction training of spatiotemporal security protection of key facilities is carried out using the Transformer-based prediction model.

[0146] The model is adjusted using the fine-tuning method of low-rank decomposition. The self-attention module has four weight matrices: query matrix (W q ), key matrix (W k ), value matrix (W v ). Fine-tuning mainly focuses on the query in the self-attention layer (W q ), key(W k ) and value (W v )Adjustment of the projection matrix.

[0147] W q and W k Considered as a matrix W qk , whose dimension is d model ×d model , where d model is the model input / output dimension. v The dimension is d model ×d model .

[0148] Low rank decomposition: For W qk , use random Gaussian distribution to initialize the low-rank decomposition into matrix A qk and B qk For W v , use random Gaussian distribution to initialize the low-rank decomposition into matrix A v and B v .

[0149] Use low-rank decomposition to represent the update of the weight matrix:

[0150] ΔW qk =A qk ·B qk

[0151] ΔW v =A v ·B v

[0152] Model update: Freeze the weight w0 of the pre-trained model and only update the low-rank decomposition matrix A qk , B qk , A v and B vPerform training. Use the back-propagation algorithm and optimizer (such as Adam) to calculate A qk , B qk , A v and B v The gradient of θ is calculated and its parameters are updated. Repeat the above steps until the model converges or reaches the training target.

[0153] Use the Transformer-based prediction model to conduct situation prediction training for the spatiotemporal security protection of key facilities.

[0154] The Transformer prediction model consists of an embedding layer, position encoding, MSA (multi-head attention mechanism), FFN (feedforward neural network) residual and normalization, and an output layer:

[0155] Embedding layer: Converts the time series of recent sample sets into continuous vectors to better represent their features and functions.

[0156] Multi-head attention layer: The multi-head attention layer divides the continuous vector obtained in the previous step into three parts: Q (query), K (key), and V (value). Then, Q, K, and V are projected through h different linear transformations, where h is the number of heads in the multi-head attention mechanism. The corresponding outputs are all vectors. The output is the weighted sum of V (Value), where the weight is the value calculated by the corresponding combination of Q (Query) and K (key). Finally, the different attention results are spliced ​​together.

[0157] The calculation of Attention is shown as follows, where d k The dimension of the key:

[0158]

[0159] head i =Attention(QW i Q ,KW i K ,VW i V )

[0160] The final concatenation operation of the multi-head attention mechanism can be expressed as:

[0161] Multihead(Q,K,V)=Concar(head1,head2,...,head i ...,head h )W o

[0162] Positional encoding: Since the Attention mechanism cannot effectively focus on the position information of the input sequence when performing global attention, Transformer designs a positional encoding module (PositionalEncoding), which is usually calculated by sine and cosine functions, so that the encoding vectors at different positions have obvious differences in space, as shown in the following formula, where pos refers to the position of the data in the sequence, d model is the dimension of the model:

[0163]

[0164] Feedforward neural network: The feedforward neural network is used to process the output of multi-head attention. It consists of two linear transformation layers with a ReLU activation function in the middle. The calculation formula is as follows:

[0165] m i =MLP(output i )=W2*RELU(W1×output i +b1)+b2

[0166] Residual connections and normalization: The outputs of MSA and FFN are combined through residual connections to increase the depth of the model. Layer normalization is then performed, which helps speed up the training process and makes the model easier to optimize.

[0167] Step S6, using the prediction model to predict navigation quality characteristics that can reflect the spatiotemporal security protection situation.

[0168] The prediction model is used to predict the navigation quality characteristics that can reflect the time and space security protection situation, including navigation signal integrity analysis, navigation signal availability analysis, navigation signal continuity analysis, and navigation positioning error.

[0169] Step S7, evaluating the spatiotemporal security protection prediction situation based on the predicted navigation quality characteristics.

[0170] The entropy method is used in combination with expert advice to confirm the weights of each navigation quality feature to calculate the comprehensive navigation interference index that can evaluate the prediction situation of spatiotemporal security protection.

[0171] Construct navigation quality feature matrix R x ,Navigation quality characteristics include four evaluation indicators: navigation signal integrity analysis, navigation signal availability analysis, navigation signal continuity analysis, and navigation positioning error.,X ij is the jth indicator of the i-th evaluation object (where i=1, 2, ..., m, j=1, 2, 3, 4):

[0172]

[0173] Construct navigation quality feature matrix R y ,The extreme value method is used to standardize the navigation quality characteristic data, and the following formula is used for the positive indicator navigation signal integrity analysis, navigation signal availability analysis, and navigation signal continuity analysis:

[0174]

[0175] The following formula is used to process the negative indicator navigation signal integrity analysis, navigation signal availability analysis, and navigation signal continuity analysis:

[0176]

[0177] Calculate each y ij Proportion:

[0178]

[0179] Calculate the entropy value of each indicator:

[0180]

[0181] Calculate the weight of each indicator:

[0182]

[0183] Through the above calculation, the weight of each indicator is recorded as W1=(w1,w 2, w3,w4), combined with the navigation quality feature weight W2 given by the expert, the final weight W of the navigation quality feature and the comprehensive interference degree index F of navigation can be:

[0184] W=α*W1+(1-α)W2

[0185] F=W*X ij

[0186] The space-time security protection situation is evaluated through the comprehensive navigation interference level index. When the comprehensive navigation interference level index drops to the set threshold, interference is considered to exist.

[0187] Experimental verification

[0188] With the protection of key infrastructure of base stations as the background, the data of the isolation protection device equipped by the base station is recorded and monitored. The isolation protection device equipped by the base station can be regarded as a navigation receiver to capture the real navigation signal and interference signal data. By performing feature extraction on the real navigation signal and interference signal captured by the navigation receiver, a navigation quality feature indicator system is constructed and the quality feature is predicted, and finally the prediction of the spatiotemporal security situation of the base station is realized. In this scenario, the number of satellites that can be observed by the isolation protection device equipped on each base station is 32, and the step length is set to 1 second. The prediction target is to predict the interference situation within 192 steps, that is, 192 seconds.

[0189] Data collection: Figure 2 As shown, in a certain area centered on the geographic coordinates of 112.877932 East longitude and 27.880724 North latitude, the isolation protection devices carried by ten base stations were monitored and data was collected. 12,960,000 real navigation signal and interference signal sample data were collected at a frequency of 1 Hz for five months as collection samples. The collection frequency is sufficient to capture subtle changes in the navigation signal.

[0190] Data recording: The collected data is stored in the database to ensure the integrity and security of the data. Each data record includes a timestamp and the operating information of the relevant equipment.

[0191] The locations of the ten basic protection facilities in this embodiment are shown in Table 1.

[0192] Table 1 Basic protection facilities data table

[0193]

[0194] Step S1, performing feature extraction on the real navigation signal and the interference signal captured by the navigation receiver to construct a navigation quality feature index system.

[0195] The navigation quality characteristic index system is constructed by the following formula:

[0196] (1) Navigation signal integrity analysis:

[0197] Integrity=N / M

[0198] M represents the number of complete observation values. The total number of satellites that can be observed by the isolation protection device carried by the base station is 32, and M is set to 32. N is the number of observed satellites that meet the satellite altitude angle exceeding the predetermined cut-off altitude angle, and the cut-off altitude angle is set to 15 degrees.

[0199] (2) Navigation signal availability analysis:

[0200]

[0201] In the formula, p n,mis the signal availability of the mth working satellite under the nth combination, m ranges from 1 to (Nk), (1-p n,i ) is the failure probability of a specific i-th satellite under the n-th combination, the value of i is N-k+1, and there are k cases in total.

[0202] (3) Navigation signal continuity analysis:

[0203]

[0204] t start ,t top ,t end They are the recording start time, selection time, and recording end time respectively. If the satellite health status is healthy at a certain moment, bool(Status) takes 1, otherwise it takes 0.

[0205] (4) Navigation positioning error:

[0206]

[0207] ρ ri is the pseudo-range of the measured position of the target machine r interfered in the i-th experiment, mean(ρ 0i ) is the average of the pseudoranges of the interference positions initially set for the i-th trial.

[0208] Step S2, obtaining navigation quality features in the collected samples, and preprocessing the navigation quality features of the sample data.

[0209] For the 12,960,000 real navigation signal and interference signal sample data stored in the database, the navigation quality characteristics are calculated according to the typical characteristic index calculation formula in S1, including: navigation signal integrity, availability, continuity, interference degree, navigation positioning error. Check the calculated navigation quality characteristic data and delete missing values ​​and outliers. Use interpolation to fill missing values ​​to ensure data continuity. Remove values ​​that exceed 3 times the standard deviation.

[0210] The checked navigation quality characteristic data is standardized, the mean u and standard deviation σ of the data are calculated, and each data point is standardized. The formula is as follows;

[0211]

[0212] Step S3, dividing the preprocessed sample data into a long-term sample set and a recent sample set.

[0213] The amount of all navigation quality feature data collected and preprocessed in this experiment reached tens of millions. All the navigation quality feature data collected and preprocessed constituted a long-term sample set, and all the navigation quality feature data collected and preprocessed in nearly 15 days constituted a recent sample set.

[0214] Step S4: For the long-term sample set, self-supervised pre-training is used to enable the model to learn common features.

[0215] Build a pre-trained Transformer-based model: Use a standard Transformer encoder, including self-attention layers and feed-forward network layers to map the observed signal to a latent representation.

[0216] Data segmentation: Treat every 1024 navigation quality feature data as a time series. Set the subsequence block length P to 64 and the step length S to 16. Starting from the first time step, extract a time series block of length 64. Then skip 16 time steps and extract a time series block of length 64 again. Repeat the step of extracting time series blocks until the entire time series is covered. This will result in 1024 / 64=64 subsequence blocks.

[0217] Mask strategy: Set the mask ratio to 0.2, that is, 20% of the subsequence blocks are masked. Set all data values ​​in the randomly selected subsequence blocks to 0 to construct the mask data.

[0218] The masked fragment sequences are input into the pre-trained model, and the model will output the reconstructed values ​​of these fragment sequences, that is, the predicted values.

[0219] Loss calculation: The difference between the reconstructed value and the true value is calculated using the MSE loss function.

[0220] Optimization process: Use Adam optimizer to update model parameters, set learning rate to 1e-4, batch size to 32, and training rounds to 100. In each training round, the training data is randomly shuffled and then divided into multiple batches for training. Use gradient descent algorithm to update model parameters:

[0221]

[0222] Among them, θ old is the vector of current model parameters, θ new is the updated parameter vector, α is the learning rate, which controls the step size of parameter update. is the gradient of the loss function loss with respect to the parameter θ, which is a vector whose components are the partial derivatives of the loss function corresponding to each parameter in θ. i , the update of gradient descent can be expressed as:

[0223]

[0224] in, is the value at iteration t, is the loss function L over θ i The partial derivative of .

[0225] Repeat the above optimization steps until the model converges or reaches the preset 100 rounds.

[0226] Step S5: For the recent sample set, the fine-tuning method based on the low-rank decomposition idea is used to adjust the model, and the Transformer-based prediction model is used to conduct situation prediction training for the spatiotemporal security protection of key facilities.

[0227] Fine-tune the model: Use low-rank decomposition to represent the weight matrix W qk and W v renew.

[0228] ΔW qk =A qk ·B qk

[0229] ΔW v =A v ·B v

[0230] Freeze the weights w0 of the pre-trained model and only decompose the low-rank matrix A qk , B qk , A v and B v Use the Adam optimizer to calculate A qk , B qk , A v and B v The gradient of θ is calculated and its parameters are updated. Repeat the above steps until the model converges or reaches the training target.

[0231] Transformer-based prediction model:

[0232] Position encoding: Add position encoding to each time series data point to introduce time dimension information. Use sine and cosine functions for encoding, where pos refers to the position of the data in the sequence, d model is the dimension of the model:

[0233]

[0234] Multi-head self-attention mechanism: Initialize the weight matrix of query vector Q, key vector K and value vector V. Calculate the attention weights and perform weighted summation, concatenate the outputs of multiple heads, and obtain the final output after linear transformation:

[0235] Multihead(Q,K,V)=Concar(head1,head2,...,head i ...,head h )W o

[0236] Feedforward neural network: consists of two linear layers and a ReLU activation function:

[0237] m i =MLP(output i )=W2*RELU(W1×output i +b1)+b2

[0238] The output of the multi-head attention mechanism is input into the feedforward neural network for nonlinear transformation.

[0239] The depth of the Transformer network is set to d = 1, and the hidden layer size is set to h = 32. The number of training rounds is 100.

[0240] Model training: Use recent sample sets and the Adam optimization algorithm to train the Transformer model, with a learning rate of τ = 0.0001, a learning rate of 1e-4, and a loss function of mean square error (MSE). Monitor the loss function value during training until the model converges or reaches the preset 100 rounds.

[0241] Step S6, using the prediction model to predict navigation quality characteristics that can reflect the spatiotemporal security protection situation.

[0242] like Figure 3 As shown in the figure, the prediction model is used to predict the navigation quality characteristics that can reflect the spatiotemporal security protection situation, namely the navigation signal integrity, availability, continuity, interference level, and navigation positioning error. Experiments have shown that when the algorithm is used to predict 192-step navigation quality feature data, the average mean square error between the predicted value and the true value is 0.32.

[0243] Step S7, evaluating the spatiotemporal security protection prediction situation based on the predicted navigation quality characteristics.

[0244] like Figure 4 As shown, the blue line is the true value and the yellow line is the predicted value. The weight of the navigation quality feature of expert opinion is set to 0.5, and the total weight calculated by the entropy method is W1 = (0.2, 0.2, 0.2, 0.4) T , combined with the navigation quality feature weights given by experts W2 = (0.3, 0.2, 0.2, 0.1) T , then the final weight of the navigation quality feature can be

[0245] W=α*W1+(1-α)W2=(0.25,0.2,0.2,0.25). The comprehensive interference level index of navigation is used to evaluate the spatiotemporal security protection situation, and the threshold of the comprehensive interference level of navigation with interference is set to 0.7. The situation prediction of the ten basic protection facilities obtained in the experiment can be used to determine whether interference will occur in the next three minutes and predict the trend of the spatiotemporal security protection situation.

[0246] The above implementation cases are only preferred embodiments of the present invention and are not intended to limit the present invention in any form. Although the present invention has been disclosed as above with preferred embodiments, they are not intended to limit the present invention. Therefore, any simple modification, equivalent changes and modifications made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the scope of protection of the technical solution of the present invention.

Claims

1. A method for predicting spatiotemporal security protection situation based on machine learning, characterized in that: The following steps are involved: Step S1, performing feature extraction on the real navigation signal and the interference signal captured by the navigation receiver to construct a navigation quality feature index system; Step S2, obtaining navigation quality features in the collected samples, and preprocessing the navigation quality features of the sample data; Step S3, dividing the preprocessed sample data into a long-term sample set and a recent sample set; Step S4, for the long-term sample set, a self-supervised pre-training method is used to enable the model to learn common features; Step S5: For the recent sample set, the model is adjusted using the low-rank decomposition fine-tuning method, and the Transformer-based prediction model is used to perform situation prediction training for the spatiotemporal security protection of key facilities; Step S6, using the prediction model to predict navigation quality characteristics that reflect the spatiotemporal security protection situation; Step S7, evaluating the spatiotemporal security protection prediction situation based on the predicted navigation quality characteristics.

2. The method for predicting spatiotemporal security protection situation based on machine learning according to claim 1 is characterized in that: In step S1, the navigation quality characteristics include navigation signal integrity analysis, navigation signal availability analysis, navigation signal continuity analysis, and navigation positioning error.

3. The method for predicting spatiotemporal security protection situation based on machine learning according to claim 2 is characterized in that: In step S2, the preprocessing is specifically: S2-1, fill in the missing values ​​of the navigation quality characteristics in the collected samples; fill in the missing values ​​according to the following formula: in, is the observed value of the jth feature of the ith sample, x kj This is the actual observed value of the jth feature of the kth sample. K is a constant indicating how many nearest neighbors or observations are considered during the calculation. S2-2, perform data centering and standardization on the navigation quality features in the collected samples; fill in the missing values ​​according to the following formula: x'=x-μ In the formula, x', x, u, and σ represent the processed data, sample data, sample mean, and sample standard deviation, respectively; S2-3, performing data equalization processing on the navigation quality features in the collected samples; performing data equalization processing according to the following formula: x new =x i +λ(x j -x i ) Among them, x new is a newly generated synthetic sample, x i is the original minority class sample, λ is a random number, uniformly distributed between 0 and 1, used to control the distance between the newly generated sample and the original sample, x j It is a sample randomly selected from the K nearest neighbors of the minority class sample.

4. The method for predicting spatiotemporal security protection situation based on machine learning according to claim 1, characterized in that: The step S4 specifically includes the following steps: S4-1, initialize data segmentation: for the initialized data, split each time series into non-overlapping subsequence blocks; S4-2, self-supervised pre-training: Build a Transformer-based pre-trained model: Use a standard Transformer encoder, including self-attention layers and feed-forward network layers to map the observed signal to a latent representation; Masking strategy: randomly select some subsequence blocks in the input signal and set their values ​​to zero to construct mask data; Pre-training prediction: The masked fragment sequence is input into the Transformer-based pre-training model, and the model outputs the pre-training prediction result, which is the reconstructed value of the masked part; Loss calculation: The MSE loss function is used to calculate the difference between the reconstructed value and the true value to evaluate the reconstruction ability of the model; Optimization process: Update the model parameters through the gradient descent algorithm to minimize the loss function.

5. The method for predicting spatiotemporal security protection situation based on machine learning according to claim 1, characterized in that: In step S5, the model is adjusted by using the low-rank decomposition fine-tuning method as follows: S5-1-1, select the weight matrix: The fine-tuning phase focuses on the query W in the self-attention layer q , key W k Sum value W v Adjustment of projection matrix; W q and W k Considered as a matrix W qk , dimension d model ×d model , where d model is the model input / output dimension; W v The dimension is d model ×d model ; S5-1-2, low-rank decomposition: Low-rank decomposition decomposes the weight matrix into the product of two matrices; For W qk , use random Gaussian distribution to initialize the low-rank decomposition into matrix A qk and B qk ; For W v , use random Gaussian distribution to initialize the low-rank decomposition into matrix A v and B v ; Use low-rank decomposition to represent the update of the weight matrix: ΔW qk =A qk ·B qk ΔW v =A v ·B v S5-1-3, Model Update: Freeze the weight w0 of the pre-trained model and decompose the low-rank matrix A qk , B qk , A v and B v Perform training; use the back-propagation algorithm and optimizer to calculate A qk , B qk , A v and B v The gradient of , and update its parameters; S5-1-4, Model deployment: At deployment time, calculate the new weight matrix W = W0 + ΔW qk +ΔW v ; Use the updated weight matrix for inference to get the prediction result.

6. The method for predicting spatiotemporal security protection situation based on machine learning according to claim 5 is characterized in that: In step S5, the situation prediction training process is: S5-2-1, model initialization: first initialize the Transformer encoder, linear projection layer, and position encoding parameters; S5-2-2, forward propagation: divide the time series data into multiple blocks and map these blocks one by one to the latent space of Transformer; S5-2-3, multi-head attention mechanism: focus on different parts of the input sequence at the same time to capture the complex relationships in time series data; S5-2-4, feature extraction: extract the features processed by the attention mechanism through the feedforward network; S5-2-5, feature concatenation and prediction: concatenate the extracted features, integrate them through the linear layer and linear head, and finally output the prediction results in training; S5-2-6, loss calculation: use the mean square error loss function to calculate the difference between the predicted value and the true value; S5-2-7, Backpropagation: Use the gradient descent algorithm to update the model parameters and minimize the loss function.

7. The method for predicting spatiotemporal security protection situation based on machine learning according to claim 1, characterized in that: In step S7, the entropy method is used in combination with expert advice to confirm the weights of various navigation quality features to calculate a comprehensive navigation interference degree index that can evaluate the spatiotemporal security protection prediction situation.

Images