Abnormal signal detection method and system for navigation and positioning based on knowledge distillation

Abstract

The present invention discloses a navigation and positioning anomaly signal detection method and system based on knowledge distillation, which relates to the field of satellite navigation. The method comprises: constructing an untrained new-scenario navigation anomaly signal detection model and a trained original-scenario navigation anomaly signal detection model; adjusting the parameters of the untrained new-scenario navigation anomaly signal detection model using a knowledge distillation loss function based on input features in training sample data and the trained original-scenario navigation anomaly signal detection model to obtain a first-trained new-scenario navigation anomaly signal detection model; iteratively updating the parameters of the first-trained new-scenario navigation anomaly signal detection model using a multi-task loss function based on the training sample data to obtain a trained new-scenario navigation anomaly signal detection model; and inputting the feature quantity to be detected into the trained new-scenario navigation anomaly signal detection model to obtain a detection result. The present invention improves the model's navigation and positioning anomaly signal detection performance in different scenarios.

Description

Technical Field

[0001] The present invention relates to the field of satellite navigation, and in particular to a method and system for detecting abnormal navigation and positioning signals based on knowledge distillation. Background Art

[0002] In recent years, with the rapid development of wireless location-based service technologies such as the Global Navigation Satellite System (GNSS), mobile cellular communication positioning, and Ultra Wide Band (UWB), navigation and positioning services have gained widespread application in transportation, surveying and mapping, urban management, the smart Internet of Things, and other fields. However, because wireless positioning technology transmits signals from the positioning source device to the receiver device via radio transmission, the positioning signal is easily affected by the transmission environment during the propagation process. For example, in complex environments such as cities, tunnels, valleys, and military theaters, positioning signals are easily blocked, multipath, and interference, resulting in degraded navigation and positioning accuracy, or even inability to use for positioning services.

[0003] To further improve navigation positioning accuracy, it is usually necessary to detect abnormal navigation signals and then exclude or suppress their weight in the positioning solution. Abnormal navigation signals here mainly include low-quality signals (for example, low signal-to-noise ratio), non-line-of-sight signals (for example, signals received after being reflected by an object), and fault signals (for example, the pseudorange of the observed signal deviates from the true range). The observed values ​​of these signals differ significantly from the actual results, so it is necessary to detect them before positioning calculation to ensure the accuracy and reliability of the positioning results.

[0004] To effectively detect abnormal navigation and positioning signals, existing methods typically require a linear Gaussian assumption about the signal's error characteristics. These assumptions are then used, such as hypothesis testing, to determine if the signal is faulty. However, in complex signal propagation environments, the linear Gaussian assumption is often overly idealistic, resulting in poor performance in practical applications.

[0005] In recent years, with the rapid development of deep learning technology, the use of deep neural networks to model navigation signal characteristics has become an important development trend. However, the performance of such methods often depends on the training data of the model. Generally speaking, the model trained with the training data of a certain scene is often difficult to apply to new scenes. Therefore, in order to improve the abnormal signal detection performance of the model in new scenes, the existing technology usually uses the data of the new scene to fine-tune an existing model (Fine-Tune, FT). For example, the model trained with street navigation data often has poor performance in tunnel scenarios; by using tunnel scene data to fine-tune the existing street scene model, the model's detection performance of navigation abnormal signals in tunnel environments can be improved.

[0006] However, fine-tuning an existing model often results in catastrophic loss of its recognition performance for the original scenario. In other words, in the example above, while the model significantly improves its detection rate for anomaly signals in tunnel scenarios, its anomaly detection performance for the original street navigation signals degrades significantly. Summary of the Invention

[0007] The purpose of the present invention is to provide a navigation and positioning anomaly signal detection method and system based on knowledge distillation, so as to improve the navigation and positioning anomaly signal detection effect of the model in different scenarios.

[0008] To achieve the above object, the present invention provides the following solutions:

[0009] A method for detecting abnormal navigation and positioning signals based on knowledge distillation, the method comprising:

[0010] Construct an untrained new scene navigation anomaly signal detection model and a trained original scene navigation anomaly signal detection model;

[0011] Acquire training sample data; the input features in the training sample data are sample feature quantities obtained based on the new scene navigation signal observation sample processing, and the output features in the training sample data are the labeling results of whether the new scene navigation signal observation is abnormal;

[0012] According to the input features in the training sample data and the trained original scene navigation anomaly signal detection model, the knowledge distillation loss function is used to adjust the parameters of the untrained new scene navigation anomaly signal detection model to obtain the first trained new scene navigation anomaly signal detection model;

[0013] Iteratively updating the parameters of the new scene navigation abnormal signal detection model trained for the first time using a multi-task loss function according to the training sample data to obtain a trained new scene navigation abnormal signal detection model;

[0014] Obtaining the observation quantity of the navigation signal of the new scene to be detected, and calculating and obtaining the feature quantity to be detected;

[0015] The feature quantity to be detected is input into the trained new scene navigation abnormal signal detection model to obtain the detection result.

[0016] Optionally, obtaining training sample data specifically includes:

[0017] Obtaining new scene navigation signal observation samples;

[0018] Extracting features from the new scene navigation signal observation sample to obtain sample feature quantities;

[0019] Normalizing the sample feature quantity to obtain a normalized sample feature quantity as an input feature;

[0020] Obtain the labeling result of whether the navigation signal observation quantity of the new scene is abnormal as the output feature.

[0021] Optionally, a knowledge distillation loss function is used to adjust parameters of an untrained new scene navigation anomaly signal detection model based on input features in the training sample data and the trained original scene navigation anomaly signal detection model to obtain a first trained new scene navigation anomaly signal detection model, specifically including:

[0022] Inputting the input features in the training sample data into the trained original scene navigation abnormal signal detection model to obtain a first output result;

[0023] Inputting the input features in the training sample data into an untrained new scene navigation abnormal signal detection model to obtain a second output result;

[0024] Calculating the similarity between the first output result and the second output result using a knowledge distillation loss function;

[0025] Determine whether the similarity is greater than a similarity threshold, and obtain a determination result;

[0026] If the judgment result is yes, outputting the untrained new scene navigation abnormal signal detection model as the first trained new scene navigation abnormal signal detection model;

[0027] If the judgment result is no, the parameters of the untrained new scene navigation abnormal signal detection model are adjusted, and the step of inputting the input features in the training sample data into the untrained new scene navigation abnormal signal detection model is returned to obtain the second output result.

[0028] Optionally, the knowledge distillation loss function is formulated as follows:

[0029]

[0030] Among them, g(f0,h i ,θ0) represents the first output result; h i represents the i-th sample feature obtained based on the i-th new scene navigation signal observation sample processing; θ0 represents the model parameters of the trained original scene navigation anomaly signal detection model; θ1 represents the model parameters of the new scene navigation anomaly signal detection model trained for the first time; g(f1,h i ,θ1) represents the second output result; l KL () represents KL divergence; Represents the similarity; f0 represents the nonlinear mapping function corresponding to the trained original scene navigation anomaly signal detection model; f1 represents the nonlinear mapping function corresponding to the new scene navigation anomaly signal detection model trained for the first time.

[0031] Optionally, iteratively updating the parameters of the new scene navigation abnormal signal detection model trained for the first time using a multi-task loss function according to the training sample data to obtain a trained new scene navigation abnormal signal detection model specifically includes:

[0032] Based on the training sample data, the parameters of the new scene navigation abnormal signal detection model trained for the first time are adjusted with the purpose of minimizing the multi-task loss function to obtain a trained new scene navigation abnormal signal detection model.

[0033] Optionally, the multi-task loss function is,

[0034]

[0035]

[0036] Among them, l CE () represents the cross entropy loss function; y i The labeling result indicating whether the new scene navigation signal observation corresponding to the i-th sample feature is abnormal; Represents the cross entropy classification loss value of the new scene navigation anomaly signal detection model trained for the first time; represents the multi-task loss function; α represents the weight coefficient.

[0037] A navigation and positioning abnormal signal detection system based on knowledge distillation is applied to the above-mentioned navigation and positioning abnormal signal detection method based on knowledge distillation. The system includes:

[0038] A construction module is used to construct an untrained new scene navigation anomaly signal detection model and a trained original scene navigation anomaly signal detection model;

[0039] A first acquisition module is configured to acquire training data for a new scene navigation signal; the input of the training data is a feature set for training the new scene navigation signal, and the output of the training data is a detection result for training the new scene navigation signal; the feature set includes: observed pseudorange, signal-to-noise ratio, three-dimensional coordinate position of the satellite in the ECEF coordinate system, pitch angle, pseudorange residual, and pseudorange change rate;

[0040] The first training module is used to adjust the parameters of the untrained new scene navigation anomaly signal detection model using a knowledge distillation loss function based on the trained feature set of the new scene navigation signal, the untrained new scene navigation anomaly signal detection model, and the trained original scene navigation anomaly signal detection model to obtain the first trained new scene navigation anomaly signal detection model;

[0041] The second training module is used to iteratively update the parameters of the new scene navigation anomaly signal detection model trained for the first time using a multi-task loss function according to the training data of the new scene navigation signal to obtain a trained new scene navigation anomaly signal detection model;

[0042] The second acquisition module is used to obtain the observation value of the navigation signal of the new scene;

[0043] The prediction module is used to input the observation amount into the trained new scene navigation abnormal signal detection model to obtain the detection result of the new scene navigation signal.

[0044] Optionally, the first acquisition module specifically includes:

[0045] An acquisition unit, used to acquire observation samples of navigation signals of a new scene;

[0046] A feature extraction unit, configured to extract features from the new scene navigation signal observation sample to obtain sample feature quantities;

[0047] A normalization unit, configured to perform normalization processing on the sample feature quantity to obtain the normalized sample feature quantity as an input feature;

[0048] The output feature unit is used to obtain the labeling result of whether the navigation signal observation quantity of the new scene is abnormal as the output feature.

[0049] An electronic device includes a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the computer program, it implements the navigation positioning abnormal signal detection method based on knowledge distillation as described above.

[0050] A computer-readable storage medium stores a computer program, which, when executed, implements the navigation and positioning abnormal signal detection method based on knowledge distillation as described above.

[0051] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0052] The embodiment of the present invention discloses a navigation and positioning anomaly signal detection method and system based on knowledge distillation, which constructs an untrained new scene navigation anomaly signal detection model and a trained original scene navigation anomaly signal detection model. The knowledge distillation loss function is used to adjust the parameters of the untrained new scene navigation anomaly signal detection model according to the input features in the training sample data and the trained original scene navigation anomaly signal detection model to obtain a first trained new scene navigation anomaly signal detection model. The multi-task loss function is then used to iteratively update the parameters of the first trained new scene navigation anomaly signal detection model according to the training sample data to obtain a trained new scene navigation anomaly signal detection model. Because the present invention introduces the original model's knowledge about navigation anomaly signal detection as an auxiliary learning task on the basis of existing fine-tuning training technology, it guides the new model to learn new scene signal anomaly detection through incremental learning during the fine-tuning process while retaining the model's recognition ability in the original scene as much as possible, thereby improving the model's navigation and positioning anomaly signal detection effect in different scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0054] Figure 1 This is a flowchart of a method for detecting abnormal navigation and positioning signals based on knowledge distillation in an embodiment of the present invention. DETAILED DESCRIPTION

[0055] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0056] To address the problem of catastrophic forgetting of the model's recognition performance for the original scene after fine-tuning an existing model, this paper proposes a navigation and positioning anomaly signal detection method and system based on knowledge distillation. Building on existing fine-tuning training techniques, this method introduces the original model's knowledge of navigation anomaly signal detection as an auxiliary learning task. This guides the new model during the fine-tuning process, using the concept of incremental learning to learn signal anomaly detection for new scenes while preserving the model's recognition capabilities for the original scene as much as possible. Ultimately, this method improves the model's navigation and positioning anomaly signal detection performance across different scenes.

[0057] The purpose of the present invention is to provide a navigation and positioning anomaly signal detection method and system based on knowledge distillation, so as to improve the navigation and positioning anomaly signal detection effect of the model in different scenarios through knowledge distillation.

[0058] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0059] like Figure 1 As shown, the present invention provides a navigation positioning abnormal signal detection method based on knowledge distillation, the method comprising:

[0060] Step 101: Construct an untrained new scene navigation anomaly signal detection model and a trained original scene navigation anomaly signal detection model.

[0061] In the specific implementation, firstly, the trained original scene navigation abnormal signal detection model f0(h i ,θ0), where f0(·) is the nonlinear mapping function corresponding to the trained original scene navigation abnormal signal detection model, h i represents the i-th sample feature obtained based on the i-th new scene navigation signal observation sample processing, and θ0 is the model parameter of the trained original scene navigation anomaly signal detection model obtained by training with the original scene training data.

[0062] Step 102: Acquire training sample data; the input features in the training sample data are sample feature quantities obtained based on the new scene navigation signal observation sample processing, and the output features in the training sample data are the labeling results of whether the new scene navigation signal observation is abnormal.

[0063] Step 103: Based on the input features in the training sample data and the trained original scene navigation anomaly signal detection model, the knowledge distillation loss function is used to adjust the parameters of the untrained new scene navigation anomaly signal detection model to obtain the first trained new scene navigation anomaly signal detection model.

[0064] Step 104: Iteratively update the parameters of the new scene navigation abnormal signal detection model trained for the first time using a multi-task loss function according to the training sample data to obtain a trained new scene navigation abnormal signal detection model.

[0065] Step 105: Obtain observation quantities of navigation signals of a new scene to be detected, and perform calculations to obtain feature quantities to be detected.

[0066] Step 106: Input the feature quantity to be detected into the trained new scene navigation abnormal signal detection model to obtain a detection result.

[0067] In the specific implementation, the navigation and positioning anomaly signal detection method based on knowledge distillation mainly includes five specific steps, as follows: 1) Model and data preparation, including preparing the trained original scene navigation anomaly signal detection model and the training sample data of the new scene navigation signal. 2) Feature engineering: Use the same feature extraction method as the trained original scene navigation anomaly signal detection model to extract features from the training sample data. 3) Knowledge distillation loss function construction: Construct a model training loss function based on Kullback-Leible (KL) divergence to retain the original recognition performance of the model. 4) Multi-task loss function construction: Use the knowledge distillation loss function and the cross entropy loss function to construct a multi-task learning function. 5) Model training: Using the training sample data, on the basis of the trained original scene navigation anomaly signal detection model, the untrained new scene navigation anomaly signal detection model is incrementally trained until convergence to obtain a trained new scene navigation anomaly signal detection model.

[0068] As a specific embodiment, obtaining training sample data specifically includes:

[0069] Get new scene navigation signal observation samples.

[0070] The navigation signal scenes are divided into two categories: original scenes and new scenes. In the training process of the present invention, only the new scene navigation signal observation sample set d is used. i ={x i ,y i} to train the model, where i∈[1,N], N is the number of training samples in the new scene navigation signal observation sample set, x i is the navigation signal observation sample of the i-th new scene, y i is the labeling result of whether the navigation signal observation of the i-th new scene is abnormal, where 1 represents an abnormal signal and 0 represents a normal signal.

[0071] The new scene navigation signal observation sample is subjected to feature extraction to obtain sample feature quantities. In a specific implementation, the feature extraction method for the new scene navigation signal observation sample must be consistent with the feature extraction method of the trained original scene navigation abnormal signal detection model.

[0072] The sample feature quantity is normalized to obtain the normalized sample feature quantity as the input feature.

[0073] Obtain the labeling result of whether the navigation signal observation quantity of the new scene is abnormal as the output feature.

[0074] In a specific application, taking GNSS satellite navigation signals as an example, an exemplary description is given of how to obtain training sample data.

[0075] New scene navigation signal observation sample x i Including observation pseudorange, signal-to-noise ratio, and the three-dimensional coordinate position of the satellite in the ECEF (Earth Centered Earth Fixed) coordinate system. The input sample feature h of the model used is i It is calculated from the new scene navigation signal observation sample, including the following 8 features:

[0076] (1) Observation pseudorange: the original observation value of the new scene navigation signal observation value sample.

[0077] (2) Signal-to-noise ratio: the original observation value of the new scene navigation signal observation sample.

[0078] (3) Satellite ECEF three-dimensional coordinates: the original observation x of the new scene navigation signal observation sample.

[0079] (4) Satellite ECEF three-dimensional coordinates: the original observation y of the new scene navigation signal observation sample.

[0080] (5) Satellite ECEF three-dimensional coordinates: the original observation z of the new scene navigation signal observation sample. Using the existing least squares iterative algorithm, the receiver position is estimated and the following features are calculated:

[0081] (6) Pitch angle: Calculate the pitch angle using the receiver's estimated position and the satellite's ECEF three-dimensional coordinates.

[0082] (7) Pseudorange residual: The estimated pseudorange is calculated using the receiver's estimated position and the three-dimensional coordinates of the satellite ECEF. The estimated pseudorange is then subtracted from the observed pseudorange to obtain the pseudorange residual.

[0083] (8) Pseudorange change rate: The pseudorange change rate is obtained by calculating the difference between the observed pseudoranges of the same observation satellite at adjacent times.

[0084] The Min-Max normalization method is used to normalize each feature so that all features are between 0 and 1. For example, for a certain dimension feature β, the maximum value of this dimension feature in all data samples is β max , the minimum characteristic value β min , then the Min-Max normalized feature is λ=(β-β min ) / (β max -β min ).

[0085] Using the above steps, the new scene navigation signal observation samples are extracted. Note that no matter what feature engineering method is used, the method for extracting the new scene navigation signal observation samples requires the same feature extraction method as the original scene model.

[0086] As a specific embodiment, according to the input features in the training sample data and the trained original scene navigation anomaly signal detection model, the knowledge distillation loss function is used to adjust the parameters of the untrained new scene navigation anomaly signal detection model to obtain the first trained new scene navigation anomaly signal detection model, specifically including:

[0087] The input features in the training sample data are input into the trained original scene navigation abnormal signal detection model to obtain a first output result.

[0088] The input features in the training sample data are input into an untrained new scene navigation abnormal signal detection model to obtain a second output result.

[0089] A knowledge distillation loss function is used to calculate the similarity between the first output result and the second output result.

[0090] It is determined whether the similarity is greater than a similarity threshold, and a determination result is obtained.

[0091] If the judgment result is yes, the untrained new scene navigation abnormal signal detection model is output as the first trained new scene navigation abnormal signal detection model.

[0092] If the judgment result is no, the parameters of the untrained new scene navigation abnormal signal detection model are adjusted, and the step of inputting the input features in the training sample data into the untrained new scene navigation abnormal signal detection model is returned to obtain the second output result.

[0093] The knowledge distillation loss function is formulated as follows:

[0094]

[0095] Among them, g(f0,h i ,θ0) represents the first output result; h irepresents the i-th sample feature obtained based on the i-th new scene navigation signal observation sample processing; θ0 represents the model parameters of the trained original scene navigation anomaly signal detection model; θ1 represents the model parameters of the new scene navigation anomaly signal detection model trained for the first time; g(f1,h i ,θ1) represents the second output result; l KL () represents KL divergence; Represents the similarity; f0 represents the nonlinear mapping function corresponding to the trained original scene navigation anomaly signal detection model; f1 represents the nonlinear mapping function corresponding to the new scene navigation anomaly signal detection model trained for the first time.

[0096] In specific applications, the i-th sample feature h obtained by processing the i-th new scene navigation signal observation sample i , the trained original scene navigation abnormal signal detection model f0(h i ,θ0) and the untrained new scene navigation abnormal signal detection model f1(h i ,θ1), where the untrained new scene navigation abnormal signal detection model f1(h i ,θ1) and the structure of the trained original scene navigation abnormal signal detection model remain the same, and the outputs of the two models are recorded as g(f0,h i ,θ0) and g(f1,h i ,θ1), the standard KL divergence is used to calculate the knowledge distillation loss function to measure the similarity of the outputs of the two models, so that the untrained new scene navigation anomaly signal detection model and the trained original scene navigation anomaly signal detection model can obtain similar outputs under the same input conditions.

[0097] As a specific embodiment, the parameters of the new scene navigation abnormal signal detection model trained for the first time are iteratively updated using a multi-task loss function according to the training sample data to obtain a trained new scene navigation abnormal signal detection model, specifically including:

[0098] Based on the training sample data, the parameters of the new scene navigation abnormal signal detection model trained for the first time are adjusted with the purpose of minimizing the multi-task loss function to obtain a trained new scene navigation abnormal signal detection model.

[0099] For h i and y i , by using the output g(f1,h i ,θ1) is the labeling result y of whether the new scene navigation signal observation quantity corresponding to the i-th sample feature quantity is abnormal iPerform cross entropy calculation to obtain the classification task loss function, so that during the training process, it is possible to learn the abnormal signal detection task in new scenarios.

[0100] As a specific embodiment, the multi-task loss function is:

[0101]

[0102]

[0103] Among them, l CE () represents the cross entropy loss function; y i The labeling result indicating whether the new scene navigation signal observation corresponding to the i-th sample feature is abnormal; Represents the cross entropy classification loss value of the new scene navigation anomaly signal detection model trained for the first time; represents the multi-task loss function; α represents the weight coefficient.

[0104] The weight coefficient α is used to control the weights of the trained original scene navigation anomaly signal detection model and the trained new scene navigation anomaly signal detection model. When α = 0, the method of the present invention degenerates into an ordinary fine-tuning model. However, there may be a problem of performance degradation of the trained new scene navigation anomaly signal detection model in the original scene. The larger the α value, the stronger the memory of the trained new scene navigation anomaly signal detection model for the trained original scene navigation anomaly signal detection model. When the α value is too large, it may affect the learning ability of the trained new scene navigation anomaly signal detection model for the new scene, and unilaterally retain the original knowledge. In practical applications, by adjusting the appropriate α value, the learning effect of the model in the new scene task and the retention of the original scene knowledge can be balanced.

[0105] The training of an untrained new scene navigation anomaly signal detection model consists of two parts, including model initialization and model parameter update.

[0106] In terms of model initialization, g(f1,h i ,θ1) is initialized using g(f0,h i The advantage of this sampling method is that it can improve the convergence efficiency of the model without retraining the model. At the same time, the new model inherits all the parameters of the original model, which can better preserve the detection performance of the model in the original scene.

[0107] In terms of model parameter updating, the present invention adopts a multi-task loss function To iteratively update the model parameters, the model can improve the performance of the model in detecting abnormal navigation and positioning signals in different scenarios. Specifically, the existing ADAM optimization algorithm is used with gradient backpropagation. ADAM adopts a warm-up strategy with a learning rate peak of 0.001 and a warm-up strategy step size of 1000 steps.

[0108] Example 2

[0109] A navigation and positioning abnormal signal detection system based on knowledge distillation is applied to the navigation and positioning abnormal signal detection method based on knowledge distillation described in Example 1. The system includes:

[0110] The construction module is used to construct an untrained new scene navigation anomaly signal detection model and a trained original scene navigation anomaly signal detection model.

[0111] The first acquisition module is used to obtain training data of the new scene navigation signal; the input of the training data is a feature set of the training of the new scene navigation signal, and the output of the training data is the detection result of the training of the new scene navigation signal; the feature set includes: observed pseudorange, signal-to-noise ratio, three-dimensional coordinate position of the satellite in the ECEF coordinate system, pitch angle, pseudorange residual and pseudorange change rate.

[0112] The first training module is used to adjust the parameters of the untrained new scene navigation anomaly signal detection model using the knowledge distillation loss function based on the trained feature set of the new scene navigation signal, the untrained new scene navigation anomaly signal detection model and the trained original scene navigation anomaly signal detection model to obtain the new scene navigation anomaly signal detection model trained for the first time.

[0113] The second training module is used to iteratively update the parameters of the new scene navigation anomaly signal detection model trained for the first time using a multi-task loss function according to the training data of the new scene navigation signal to obtain a trained new scene navigation anomaly signal detection model.

[0114] The second acquisition module is used to acquire the observation value of the new scene navigation signal.

[0115] The prediction module is used to input the observation amount into the trained new scene navigation abnormal signal detection model to obtain the detection result of the new scene navigation signal.

[0116] As a specific embodiment, the first acquisition module specifically includes:

[0117] The acquisition unit is used to obtain the new scene navigation signal observation sample.

[0118] The feature extraction unit is used to extract features from the new scene navigation signal observation sample to obtain sample feature quantities.

[0119] The normalization unit is used to perform normalization processing on the sample feature quantity to obtain the normalized sample feature quantity as the input feature.

[0120] The output feature unit is used to obtain the labeling result of whether the navigation signal observation quantity of the new scene is abnormal as the output feature.

[0121] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method for detecting abnormal navigation and positioning signals based on knowledge distillation as described in Example 1 is implemented.

[0122] A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed, the method for detecting abnormal navigation and positioning signals based on knowledge distillation as described in Example 1 is implemented.

[0123] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0124] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A navigation positioning abnormal signal detection method based on knowledge distillation, characterized in that: The method comprises: Construct an untrained new scene navigation anomaly signal detection model and a trained original scene navigation anomaly signal detection model; Acquire training sample data; the input features of the training sample data are sample feature quantities obtained based on processing of new scene navigation signal observation samples; the output features of the training sample data are labeling results of whether the new scene navigation signal observation is abnormal; specifically, the features include: observed pseudorange, signal-to-noise ratio, three-dimensional coordinate position of the satellite in the ECEF coordinate system, pitch angle, pseudorange residual, and pseudorange change rate; According to the input features in the training sample data and the trained original scene navigation anomaly signal detection model, the knowledge distillation loss function is used to adjust the parameters of the untrained new scene navigation anomaly signal detection model to obtain the first trained new scene navigation anomaly signal detection model, specifically including: Inputting the input features in the training sample data into the trained original scene navigation abnormal signal detection model to obtain a first output result; Inputting the input features in the training sample data into an untrained new scene navigation abnormal signal detection model to obtain a second output result; Calculating the similarity between the first output result and the second output result using a knowledge distillation loss function; Determine whether the similarity is greater than a similarity threshold, and obtain a determination result; If the judgment result is yes, outputting the untrained new scene navigation abnormal signal detection model as the first trained new scene navigation abnormal signal detection model; If the judgment result is no, adjusting the parameters of the untrained new scene navigation abnormal signal detection model, returning to the step of inputting the input features in the training sample data into the untrained new scene navigation abnormal signal detection model to obtain a second output result; Iteratively updating the parameters of the new scene navigation abnormal signal detection model trained for the first time using a multi-task loss function according to the training sample data to obtain a trained new scene navigation abnormal signal detection model; Obtaining the observation quantity of the navigation signal of the new scene to be detected, and calculating and obtaining the feature quantity to be detected; The feature quantity to be detected is input into the trained new scene navigation abnormal signal detection model to obtain the detection result.

2. The navigation positioning abnormal signal detection method based on knowledge distillation according to claim 1 is characterized in that: Obtain training sample data, including: Obtaining new scene navigation signal observation samples; Extracting features from the new scene navigation signal observation sample to obtain sample feature quantities; Normalizing the sample feature quantity to obtain a normalized sample feature quantity as an input feature; Obtain the labeling result of whether the navigation signal observation quantity of the new scene is abnormal as the output feature.

3. The navigation positioning abnormal signal detection method based on knowledge distillation according to claim 1 is characterized in that: The knowledge distillation loss function is formulated as follows: ； in, Represents the first output result; Indicates that based on i The first sample obtained by processing the new scene navigation signal observation i Sample characteristics; Represents the model parameters of the trained original scene navigation anomaly signal detection model; Represents the model parameters of the new scene navigation anomaly signal detection model trained for the first time; Represents the second output result; represents KL divergence; Indicates similarity; Represents the nonlinear mapping function corresponding to the trained original scene navigation anomaly signal detection model; Represents the nonlinear mapping function corresponding to the new scene navigation anomaly signal detection model trained for the first time.

4. The navigation positioning abnormal signal detection method based on knowledge distillation according to claim 3 is characterized in that: Iteratively updating the parameters of the new scene navigation abnormal signal detection model trained for the first time using a multi-task loss function according to the training sample data to obtain a trained new scene navigation abnormal signal detection model, specifically including: Based on the training sample data, the parameters of the new scene navigation abnormal signal detection model trained for the first time are adjusted with the purpose of minimizing the multi-task loss function to obtain a trained new scene navigation abnormal signal detection model.

5. The navigation positioning abnormal signal detection method based on knowledge distillation according to claim 4 is characterized in that: The multi-task loss function is, ； ； in, represents the cross entropy loss function; Indicates the i The labeling result of whether the new scene navigation signal observation quantity corresponding to the sample feature quantity is abnormal; Represents the cross entropy classification loss value of the new scene navigation anomaly signal detection model trained for the first time; represents the multi-task loss function; Represents the weight coefficient.

6. The navigation and positioning abnormal signal detection system based on knowledge distillation is characterized by: The navigation and positioning abnormal signal detection system based on knowledge distillation is applied to the navigation and positioning abnormal signal detection method based on knowledge distillation according to any one of claims 1 to 5, and the system includes: A construction module is used to construct an untrained new scene navigation anomaly signal detection model and a trained original scene navigation anomaly signal detection model; A first acquisition module is configured to acquire training data of a new scene navigation signal; the input of the training data is a feature set of the training of the new scene navigation signal, and the output of the training data is a detection result of the training of the new scene navigation signal; The first training module is used to adjust the parameters of the untrained new scene navigation anomaly signal detection model using a knowledge distillation loss function based on the trained feature set of the new scene navigation signal, the untrained new scene navigation anomaly signal detection model, and the trained original scene navigation anomaly signal detection model to obtain the first trained new scene navigation anomaly signal detection model; The second training module is used to iteratively update the parameters of the new scene navigation anomaly signal detection model trained for the first time using a multi-task loss function according to the training data of the new scene navigation signal to obtain a trained new scene navigation anomaly signal detection model; The second acquisition module is used to obtain the observation value of the navigation signal of the new scene; The prediction module is used to input the observation amount into the trained new scene navigation abnormal signal detection model to obtain the detection result of the new scene navigation signal.

7. The navigation and positioning abnormal signal detection system based on knowledge distillation according to claim 6 is characterized in that: The first acquisition module specifically includes: An acquisition unit, used to acquire observation samples of navigation signals of a new scene; A feature extraction unit, configured to extract features from the new scene navigation signal observation sample to obtain sample feature quantities; A normalization unit, configured to perform normalization processing on the sample feature quantity to obtain the normalized sample feature quantity as an input feature; The output feature unit is used to obtain the labeling result of whether the navigation signal observation quantity of the new scene is abnormal as the output feature.

8. An electronic device, characterized in that: The method comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method according to any one of claims 1 to 5 when executing the computer program.

9. A computer-readable storage medium, characterized in that The storage medium stores a computer program, which implements the method according to any one of claims 1 to 8 when executed.

Images