Radar ranging prediction method and device based on multivariate real-time combined time-series data

Abstract

To address the issue of conditional distribution offset in radar ranging data combined with experimental data, this invention proposes a radar ranging prediction method and apparatus based on multivariate combined digital and real-time data. This method combines sufficient digital radar ranging data generated from digital experiments such as radar ranging simulation systems with a small amount of physical radar ranging data generated from actual radar ranging experiments to construct a regression prediction model. This model not only fully considers the characteristics of multivariate time-time data but also aligns with the conditional distribution offset under the combined digital and real-time data, achieving more accurate radar ranging prediction with better generalization capabilities.

Description

Technical Field

[0001] This invention belongs to the field of radar ranging prediction technology, and in particular to a radar ranging prediction method and device based on multivariate real-time combined time series data. Background Technology

[0002] In actual radar ranging experiments, various factors affecting radar ranging are interdependent, with high-dimensional and complex influencing factors that may include static quantitative factors, static qualitative factors, and dynamic factors. The high-dimensional multivariate time-series dataset formed after labeling the physical test response is the foundation for radar ranging prediction model research. In recent years, data-driven prediction methods, represented by deep learning, have been extensively studied and proven to outperform traditional physics-based methods in many practical applications. Deep learning methods, by automatically extracting complex patterns and relationships from data, can effectively handle the complex characteristics of high data dimensionality, diverse factor types, and temporal correlations, demonstrating significant advantages in radar ranging prediction tasks. However, the scarcity of physical radar ranging test data poses a significant challenge to training a deep learning model with good prediction accuracy and generalization ability. Digital methods such as simulated radar ranging experiments provide a large amount of digital experimental data for combat effectiveness assessment. Combining digital radar ranging test data with physical radar ranging test data has become an effective way to solve this problem.

[0003] Digital radar ranging test data is usually readily available, compensating for the insufficient amount of physical radar ranging test data. However, digital radar ranging test data is generated by simulation systems based on the real world, resulting in distributional differences between it and physical radar ranging test data. Ignoring these distributional differences and directly training deep learning models will limit the predictive performance of the models. Therefore, there is an urgent need for a scheme that can eliminate the differences between digital radar ranging test data and physical radar ranging test data before predicting radar ranging response. Summary of the Invention

[0004] To address the aforementioned problems in existing technologies, this invention provides a radar ranging prediction method and apparatus based on multivariate real-time combined time-series data.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] On one hand, the present invention provides a radar ranging prediction method based on multivariate real-time combined time-series data, comprising:

[0007] Identify the mixed influence factors affecting radar ranging response, which include static factors that do not change with time and dynamic factors that change with time.

[0008] Set a time domain and obtain physical radar ranging data samples and digital analog radar ranging data samples with mixed influence factors based on the same time domain. The physical radar ranging data samples are divided into ranging response samples and non-ranging response samples. The ranging response in the physical radar ranging data samples and digital analog radar ranging data samples changes continuously with the time domain and is a time series dynamic response.

[0009] Data standardization processing was performed on physical radar ranging data samples and digital analog radar ranging data samples.

[0010] The standardized physical radar ranging data samples and the digital analog radar ranging data samples are processed by sliding time windows to obtain target domain window sequence samples and source domain window sequence samples.

[0011] Extract the latent space feature vector of the original ranging response from the target domain window sequence sample, and decode and reconstruct the ranging response based on the extracted latent space feature vector. Construct the target domain reconstruction loss function based on the original ranging response and the corresponding reconstructed ranging response from the target domain window sequence sample.

[0012] Extract the latent space feature vector of the original ranging response from the source domain window sequence samples, and decode and reconstruct the ranging response based on the extracted latent space feature vector. Construct the source domain reconstruction loss function based on the original ranging response and the corresponding reconstructed ranging response from the source domain window sequence samples.

[0013] The distribution difference of latent space feature vectors extracted by the target domain autoencoder and the source domain autoencoder is measured by multi-kernel maximum mean difference, and an alignment loss function is constructed.

[0014] The total loss function is constructed based on the target domain reconstruction loss function, the source domain reconstruction loss function, and the alignment loss function to complete the training of the target domain autoencoder;

[0015] The predictor model is trained to learn the mapping relationship between the mixed influence factors in the target domain window sequence samples and the latent space feature vectors extracted by the target domain autoencoder;

[0016] The measured mixed influence factor after standardizing the data to be predicted is input into the trained predictor model to obtain the predicted latent space feature vector. The predicted latent space feature vector is then decoded and reconstructed based on the decoder in the trained target domain autoencoder to obtain the final range response prediction value.

[0017] On the other hand, a radar ranging prediction device based on multivariate real-time combined time-series data is provided, comprising:

[0018] The first module is used to determine the mixed influence factors affecting the radar ranging response. The mixed influence factors include static factors that do not change with time and dynamic factors that change with time.

[0019] The second module is used to set the time domain and obtain physical radar ranging data samples and digital analog radar ranging data samples based on the same time domain regarding the mixed influence factors. The physical radar ranging data samples are divided into ranging response samples and non-ranging response samples. The ranging response in the physical radar ranging data samples and digital analog radar ranging data samples changes continuously with the time domain and is a time series dynamic response.

[0020] The third module is used to perform data standardization processing on physical radar ranging data samples and digital analog radar ranging data samples.

[0021] The fourth module is used to perform sliding time window processing on the standardized physical radar ranging data samples and the digital analog radar ranging data samples to obtain target domain window sequence samples and source domain window sequence samples.

[0022] The fifth module is used to extract the latent space feature vector of the original ranging response in the target domain window sequence sample, and decode and reconstruct the ranging response based on the extracted latent space feature vector. The target domain reconstruction loss function is constructed based on the original ranging response and the corresponding reconstructed ranging response in the target domain window sequence sample.

[0023] The sixth module is used to extract the latent space feature vector of the original ranging response in the source domain window sequence sample, and decode and reconstruct the ranging response based on the extracted latent space feature vector. The source domain reconstruction loss function is constructed based on the original ranging response and the corresponding reconstructed ranging response in the source domain window sequence sample.

[0024] The seventh module is used to measure the distribution difference of latent space feature vectors extracted by the target domain autoencoder and the source domain autoencoder using multi-kernel maximum mean difference, and to construct an alignment loss function.

[0025] The eighth module is used to construct the total loss function based on the target domain reconstruction loss function, the source domain reconstruction loss function, and the alignment loss function, thereby completing the training of the target domain autoencoder.

[0026] The ninth module is used to train the predictor model to learn the mapping relationship between the mixed influence factors in the target domain window sequence samples and the latent space feature vectors extracted by the target domain autoencoder.

[0027] The tenth module is used to input the measured mixed influence factor after the data to be predicted is standardized into the trained predictor model to obtain the predicted latent space feature vector. Based on the decoder in the trained target domain autoencoder, the predicted latent space feature vector is decoded and reconstructed to obtain the final range response prediction value.

[0028] On the other hand, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the radar ranging prediction method based on multivariate real-time data.

[0029] On the other hand, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the radar ranging prediction method based on multivariate real-time combined data.

[0030] On the other hand, the present invention provides a computer program product stored on a computer-readable storage medium and including computer instructions that, when executed by a processor, cause a computer device to implement the steps of the radar ranging prediction method based on multivariate real-time combined data.

[0031] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0032] Digital radar ranging data serves as the source domain, while physical radar ranging data serves as the target domain. Since the input to digital radar ranging experiments is typically controllable, the marginal distribution of digital radar ranging data can maintain a high degree of similarity to that of physical radar ranging data. The difference between the two stems from the mapping relationship between the mixed influence factor and the ranging response. To address the conditional distribution shift problem under combined digital and physical radar ranging experimental data, this invention proposes a radar ranging prediction method based on multivariate combined digital and physical time-series data. This method combines sufficient digital radar ranging data generated from digital experiments such as radar ranging simulation systems with a small amount of physical radar ranging data generated from physical radar ranging experiments to construct a regression prediction model. This model not only fully considers the characteristics of multivariate time-series data but also aligns with the conditional distribution shift under combined digital and physical data, achieving more accurate radar ranging prediction with better generalization capabilities. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0034] Figure 1 This is a flowchart of a radar ranging prediction method based on multivariate real-time combined time-series data;

[0035] Figure 2 This is a difference diagram between source domain ranging response data and target domain ranging response data in one embodiment;

[0036] Figure 3 This is a schematic diagram of an embodiment of an autoencoder network structure;

[0037] Figure 4 This is a diagram illustrating the effect of domain alignment in one embodiment;

[0038] Figure 5 This is a schematic diagram of the predictor network structure in one embodiment;

[0039] Figure 6 This is a line graph showing the relationship between the predicted ranging response value and the original ranging response value in one embodiment. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0041] Reference Figure 1 This invention provides a radar ranging prediction method based on multivariate real-time combined time-series data, comprising:

[0042] Identify the mixed influence factors affecting radar ranging response, which include static factors that do not change with time and dynamic factors that change with time.

[0043] Set a time domain and obtain physical radar ranging data samples and digital analog radar ranging data samples with mixed influence factors based on the same time domain. The physical radar ranging data samples are divided into ranging response samples and non-ranging response samples. The ranging response in the physical radar ranging data samples and digital analog radar ranging data samples changes continuously with the time domain and is a time series dynamic response.

[0044] Data standardization processing was performed on physical radar ranging data samples and digital analog radar ranging data samples.

[0045] The standardized physical radar ranging data samples and the digital analog radar ranging data samples are processed by sliding time windows to obtain target domain window sequence samples and source domain window sequence samples.

[0046] Extract the latent space feature vector of the original ranging response from the target domain window sequence sample, and decode and reconstruct the ranging response based on the extracted latent space feature vector. Construct the target domain reconstruction loss function based on the original ranging response and the corresponding reconstructed ranging response from the target domain window sequence sample.

[0047] Extract the latent space feature vector of the original ranging response from the source domain window sequence samples, and decode and reconstruct the ranging response based on the extracted latent space feature vector. Construct the source domain reconstruction loss function based on the original ranging response and the corresponding reconstructed ranging response from the source domain window sequence samples.

[0048] The distribution difference of latent space feature vectors extracted by the target domain autoencoder and the source domain autoencoder is measured by multi-kernel maximum mean difference, and an alignment loss function is constructed.

[0049] The total loss function is constructed based on the target domain reconstruction loss function, the source domain reconstruction loss function, and the alignment loss function to complete the training of the target domain autoencoder;

[0050] The predictor model is trained to learn the mapping relationship between the mixed influence factors in the target domain window sequence samples and the latent space feature vectors extracted by the target domain autoencoder;

[0051] The measured mixed influence factor after standardizing the data to be predicted is input into the trained predictor model to obtain the predicted latent space feature vector. The predicted latent space feature vector is then decoded and reconstructed based on the decoder in the trained target domain autoencoder to obtain the final range response prediction value.

[0052] Many factors are involved in radar ranging, including static factors and dynamic factors. Static factors include static quantitative factors and static qualitative factors. Static quantitative factors include filtering parameters and temperature parameters, while static qualitative factors include wind force level and navigation route. Dynamic factors include radar cross section and signal-to-noise ratio.

[0053] In one embodiment, [the following is a record] For the time domain of the experimental data, The total length of the time series must include the following time domains: Each time step involves acquiring physical radar ranging data samples and digital analog radar ranging data samples based on the mixed influence factors within this time domain. The physical radar ranging data samples are divided into samples with and without ranging responses. The ranging responses in both the physical and digital analog radar ranging data samples change continuously over time, exhibiting a time-series dynamic response. Specifically, the digital analog radar ranging data sample set... Digital analog radar ranging data sample set From a finite number of ranging response samples and distance-free response samples Composition. Among them... The mixing factor values ​​for digital analog radar ranging data samples include those that do not change with time. Changing static factors and with time Changing dynamic factors , The mixing factor values ​​for physical radar ranging data samples also include those that do not change with time. Changing static factors and with time Changing dynamic factors Static factor It also includes continuous static quantitative factors. and discrete static qualitative factors . and These are the ranging responses of digital analog radar ranging data samples and physical radar ranging data samples, respectively, over time. It is constantly changing, and is a time-series dynamic response. It is the sample size of digital analog radar ranging data. and These are the sample sizes of the physical radar ranging data, which include ranging response samples and those without ranging response samples.

[0054] Data standardization processing is performed on physical radar ranging data samples and digital analog radar ranging data samples, including: normal distribution standardization transformation is applied to the dynamic factors, static quantitative factors and ranging response of physical radar ranging data samples and digital analog radar ranging data samples, and one-hot encoding is applied to the static qualitative factors to enhance the robustness of the model to absolute values.

[0055] The standardized physical radar ranging data samples and the digital analog radar ranging data samples are processed using a sliding time window to obtain target domain window sequence samples and source domain window sequence samples. This invention generates window sequence samples for actual model training using the sliding time window method. The boundaries of each original sample point are clearly defined, representing a complete time series. The sliding window divides within each sample and cannot cross samples to avoid data confusion. The total time series length refers to the total sequence length of a single sample, i.e., ... Window length refers to the number of time steps the sliding window covers, denoted as ; Note that the window length must be less than the total timing length; the sliding step size refers to the number of time steps the window slides each time, denoted as... The default setting is generally 1 step to achieve maximum data utilization, generating highly overlapping but temporally continuous window sequence samples. When set to 1, the number of window sequence samples is... .

[0056] In one embodiment, the dataset is divided into digital analog radar ranging data samples and physical radar ranging data samples. To ensure the rigor of time-series data modeling and cross-domain independence, a block partitioning method is used to partition the digital analog radar ranging data sample set. Digital analog radar ranging data sample set Separate training, validation, and test sets are created, resulting in six data subsets to prevent data leakage. To preserve temporal integrity, the dataset is split into individual samples (e.g., ...). It contains a complete radar ranging test record, i.e., the time domain. middle Each time step of data is used. The training, validation, and test sets are allocated in a certain proportion by default to ensure strict isolation of each subset in the time dimension. Regarding the standardization strategy for each data subset, the validation and test sets must strictly adhere to the standardized parameters of the training set. This is to prevent implicit data leakage; if the standardized parameters of the validation or test sets are calculated based on their own data, the model has essentially "seen" the future data distribution during the evaluation phase, which may lead to overly optimistic evaluation results and fail to accurately reflect the model's generalization ability. Secondly, it meets the practical requirements of real-world application scenarios. New test data may be acquired in streaming form, making it impossible to obtain complete statistics; after model deployment, only the standardized parameters from training can be used to process new data. Furthermore, for the digital analog radar ranging data sample set... Digital analog radar ranging data sample set Standardized parameters need to be calculated separately to avoid systematic bias caused by cross-domain differences.

[0057] An autoencoder maps input data to a low-dimensional latent space using an encoder, obtaining the latent space feature representation of the input data. A decoder then restores this latent space feature representation to a reconstructed version of the input data, learning an effective representation of the data in the process. The network architecture in this invention for the cross-domain bias alignment stage consists of a source domain autoencoder, a target domain autoencoder, and a domain alignment module. Through iterative training, it performs feature decoupling and cross-domain alignment on dynamically responding data, obtaining an effective domain-invariant feature representation.

[0058] This invention constructs a target-domain autoencoder and a source-domain autoencoder. The target-domain autoencoder extracts the latent space feature vector of the original ranging response from the target-domain window sequence samples using an encoder, and then reconstructs the ranging response using a decoder based on the extracted latent space feature vector. The source-domain autoencoder extracts the latent space feature vector of the original ranging response from the source-domain window sequence samples using an encoder, and then reconstructs the ranging response using a decoder based on the extracted latent space feature vector.

[0059] In this invention, the target domain autoencoder and the source domain autoencoder employ the same autoencoder network structure. The encoder in this structure uses a single-layer LSTM network, and its input is the raw ranging response, with dimensions (batch_size, look_back, input_size=1). Here, batch_size represents the number of input window sequence samples during each training iteration, and look_back represents the window length. `input_size` represents the feature dimension at each time step, which is 1. The LSTM network processes the input raw ranging response and outputs the hidden state at each time step. Finally, a fully connected layer maps the hidden state of the last time step of the LSTM to the latent space, the dimension of which is represented as `latent_size`. The purpose of this process is to extract the latent space feature vector from the raw ranging response. The decoder in the autoencoder network structure also uses a single-layer LSTM network. First, the extracted latent space feature vector is repeatedly expanded so that its length is aligned with the length of the window sequence (i.e., `look_back`). Then, the expanded latent space feature vector is input into the LSTM network, and the output temporal data is reconstructed through the LSTM network. Finally, the decoder output is restored to the original feature dimension (i.e., `input_size`) through a fully connected layer. The role of the decoder is to reconstruct the ranging response from the latent space feature vector and evaluate the effectiveness of the latent space feature vector extracted by the encoder by comparing it with the original ranging response. During training, both the source domain autoencoder and the target domain autoencoder have their own reconstruction losses, which are used to measure the difference between the reconstructed ranging response and the original ranging response.

[0060] Specifically, the latent space feature vectors of the original ranging response in the target domain window sequence samples are extracted, and the ranging response is decoded and reconstructed based on the extracted latent space feature vectors. A target domain reconstruction loss function is constructed based on the original ranging response and the corresponding reconstructed ranging response in the target domain window sequence samples. Similarly, the latent space feature vectors of the original ranging response in the source domain window sequence samples are extracted, and the ranging response is decoded and reconstructed based on the extracted latent space feature vectors. A source domain reconstruction loss function is constructed based on the original ranging response and the corresponding reconstructed ranging response in the source domain window sequence samples. The target domain reconstruction loss function and the source domain reconstruction loss function are as follows:

[0061] ;

[0062] in Indicates the source domain reconstruction loss. For the target domain reconstruction loss, For the source domain autoencoder model parameters, For the parameters of the target domain autoencoder model, and These are the original source domain ranging response value and the original target domain ranging response value, respectively. and These are the source domain ranging response reconstruction values ​​and the target domain ranging response reconstruction values, respectively.

[0063] The domain alignment module uses multi-kernel maximum mean discrepancy (MMD) to measure the distribution difference of latent space feature vectors extracted by the source domain encoder and the target domain encoder. Using multiple kernels can measure the difference between the two distributions from multiple dimensions, improving the measurement effect. Through iterative training to minimize the multi-kernel MMD loss, the cross-domain feature distributions are gradually aligned, making the latent space distributions of the source domain ranging response and the target domain ranging response more similar.

[0064] In one embodiment, the alignment loss function is constructed as follows:

[0065] ;

[0066] in and The latent space feature vectors extracted by the source domain autoencoder and the target domain autoencoder are respectively, and the multi-kernel function is used. It is a linear combination of multiple Gaussian kernel functions;

[0067] ;

[0068] in It is the Gaussian kernel bandwidth. It is the first u One kernel function, These are the corresponding weights, and they satisfy... .

[0069] The total loss function is constructed based on the target domain reconstruction loss function, the source domain reconstruction loss function, and the alignment loss function, as follows:

[0070] ;

[0071] in The balancing factor is set to balance the relative importance of reconstruction loss and domain alignment loss. Reconstruction loss ensures feature representation capability, enabling the autoencoder to accurately reconstruct the input data; domain alignment loss enhances cross-domain generalization capability, making the feature distributions of the source and target domains closer.

[0072] After the target domain autoencoder is trained, model evaluation is required. Since the ranging response to be reconstructed has been standardized and contains zero values, the symmetric mean absolute percentage error (sMAPE) is chosen as the evaluation metric to effectively avoid the problem of a denominator of zero and to relatively fairly measure the deviation between the predicted and true values. The formula for calculating symmetric mean absolute percentage error is as follows:

[0073] ;

[0074] In the formula The sample size used in the calculation. To predict the ranging response value, This is the original ranging response value.

[0075] A small sMAPE value indicates that the target domain autoencoder can reconstruct the original ranging response data well, and the extracted latent space feature vectors retain the core features of the original ranging response data, indicating good model performance. Conversely, a large sMAPE value indicates that the reconstruction effect of the target domain autoencoder is poor, and it may be necessary to adjust the model parameters, increase training data, or improve the model structure.

[0076] After completing the training of the target domain autoencoder, its optimal parameters are obtained. The autoencoder parameters are frozen, and the training of the predictor model begins to obtain the mapping relationship between the mixed influence factor and the latent space feature vector extracted by the target domain autoencoder. Then, the ranging response is predicted by the decoder reconstruction in the target domain autoencoder.

[0077] This invention employs CNN + Bi-LSTM to construct a predictor. The predictor model includes a CNN module and a Bi-LSTM module. The CNN module receives static factors from the mixed influence factors and extracts local features of the static factors through a three-convolutional module structure. The local features of the static factors after convolution and pooling are flattened into static factor latent feature vectors. The Bi-LSTM module receives dynamic factors from the mixed influence factors, captures the long-term forward and backward dependencies in the dynamic factor sequence, and outputs the corresponding dynamic factor latent feature vectors. The static factor latent feature vectors and dynamic factor latent feature vectors are integrated through a fully connected layer to achieve prediction of the latent space features extracted by the target domain autoencoder.

[0078] During the predictor training process, the predictor parameters are optimized based on the mean squared error loss, so that the predicted latent space feature vector of the target domain is as close as possible to the latent space feature vector extracted by the autoencoder of the target domain. The calculation formula is as follows:

[0079] ;

[0080] in For predictor network parameters, The predictor outputs the latent space feature vector of the target domain. The latent space feature vectors extracted by the autoencoder in the target domain. The dimension of the latent space feature vector is [not specified]. The predictor model is still evaluated using the sMAPE metric to determine whether the pre-test has learned an effective mapping from the mixed influence factors to the latent space feature vector.

[0081] After the predictor is trained, its parameters are fixed. For the mixed influence factors of the target domain to be tested, the predicted latent space feature vector is first obtained through the predictor. Then, the decoder in the target domain autoencoder is used as input to decode and reconstruct the model, thus obtaining the final predicted ranging response value. Due to the sliding window setting, multiple response values ​​will be predicted for the same data point; the average value is taken to determine the final predicted ranging response value. The performance of the entire model is still evaluated using sMAPE. In the final evaluation, the predicted ranging response value needs to be de-standardized and compared with the original ranging response.

[0082] To demonstrate the effectiveness of the radar ranging prediction method based on multivariate real-time combined time-series data proposed in this invention, an example is provided below:

[0083] This example, based on a radar detection performance evaluation problem, identifies the aforementioned mixed influencing factors involved in radar ranging. Using the radar ranging prediction method based on multivariate numerical and real-time combined time-series data provided above, the radar measured distance is used as the response for prediction and evaluation. Samples of digital analog radar ranging data and physical radar ranging data are provided.

[0084] Specifically, 2000 digital simulated radar ranging data samples (i.e., source domain) are generated through function simulation, and 300 physical radar ranging data samples (i.e., target domain) are obtained through physical experiments. The length of each radar ranging data sample sequence in both the source domain and the target domain is 36. Figure 2 Using 10 samples as an example, the differences between source domain ranging response data and target domain ranging response data are shown.

[0085] Data preprocessing: For 2000 source domain radar ranging data samples and 300 target domain radar ranging data samples, the datasets were first divided for model training, creating training, validation, and test sets, forming six data subsets. Next, the training sets for both the source and target domains were standardized, while the validation and test sets were processed using the same standardization parameters. Then, a sliding time window was applied to the six data subsets, with a lookback value of 16 and a step size of 1. Finally, data loading was performed. Since the amount of source domain radar ranging data is much larger than that of target domain radar ranging data, the batch size was set to 64 for the source domain and 32 for the target domain.

[0086] Autoencoder Training: After inputting the dataset, the source domain autoencoder and the target domain autoencoder adopt the same autoencoder network structure, as shown in the figure below. Figure 3 As shown. After executing the program, the reconstruction loss of the training set decreased from 1.6773 to 0.0015, and the alignment loss decreased from 0.6151 to 0.0004. The latent space features extracted by the autoencoders in the source and target domains were visualized using t-SNE data dimensionality reduction. Figure 4 The alignment effect changes from iteration 0 to iteration 99. After training, reconstruction prediction is performed on the validation set data in the target domain. The loss value is 0.00084, and the calculated sMAPE index is 0.0854, indicating that the autoencoder model has excellent performance.

[0087] Predictor Training: Freeze the parameters of the trained target domain autoencoder and input the data into the predictor to begin training. In this example, the predictor uses three CNN modules and one Bi-LSTM module, with the network structure as follows: Figure 5 After executing the program, the prediction loss of the training set decreased from 0.0012 to 0.00000086. After training, prediction was performed on the validation set data in the target domain, and the loss was 0.000002. The calculated sMAPE index was 0.0396, indicating that the predictor model performed well.

[0088] Test sample prediction: After completing the training of the autoencoder and predictor, the optimal parameters of the model are frozen. First, the prepared target domain test samples are input into the predictor model to obtain the predicted latent space features. Then, the features are input into the target domain decoder to obtain the corresponding reconstructed prediction. The calculated sMAPE index is 0.0151. Figure 6 The graph shows the relationship between the predicted and original ranging response values. It can be seen that the predicted and original ranging response values ​​almost overlap, indicating that the model fits very well.

[0089] As can be seen from the above examples, the radar ranging prediction method based on multivariate digital-real combined time series data proposed in this invention can make full use of readily available radar ranging digital test data, combine it with scarce physical data to train the model, achieve accurate prediction of response, and the model has good generalization ability, which can effectively reduce test costs.

[0090] In one embodiment, a radar ranging prediction device based on multivariate real-time combined time-series data is provided, comprising:

[0091] The first module is used to determine the mixed influence factors affecting the radar ranging response. The mixed influence factors include static factors that do not change with time and dynamic factors that change with time.

[0092] The second module is used to set the time domain and obtain physical radar ranging data samples and digital analog radar ranging data samples based on the same time domain regarding the mixed influence factors. The physical radar ranging data samples are divided into ranging response samples and non-ranging response samples. The ranging response in the physical radar ranging data samples and digital analog radar ranging data samples changes continuously with the time domain and is a time series dynamic response.

[0093] The third module is used to perform data standardization processing on physical radar ranging data samples and digital analog radar ranging data samples.

[0094] The fourth module is used to perform sliding time window processing on the standardized physical radar ranging data samples and the digital analog radar ranging data samples to obtain target domain window sequence samples and source domain window sequence samples.

[0095] The fifth module is used to extract the latent space feature vector of the original ranging response in the target domain window sequence sample, and decode and reconstruct the ranging response based on the extracted latent space feature vector. The target domain reconstruction loss function is constructed based on the original ranging response and the corresponding reconstructed ranging response in the target domain window sequence sample.

[0096] The sixth module is used to extract the latent space feature vector of the original ranging response in the source domain window sequence sample, and decode and reconstruct the ranging response based on the extracted latent space feature vector. The source domain reconstruction loss function is constructed based on the original ranging response and the corresponding reconstructed ranging response in the source domain window sequence sample.

[0097] The seventh module is used to measure the distribution difference of latent space feature vectors extracted by the target domain autoencoder and the source domain autoencoder using multi-kernel maximum mean difference, and to construct an alignment loss function.

[0098] The eighth module is used to construct the total loss function based on the target domain reconstruction loss function, the source domain reconstruction loss function, and the alignment loss function, thereby completing the training of the target domain autoencoder.

[0099] The ninth module is used to train the predictor model to learn the mapping relationship between the mixed influence factors in the target domain window sequence samples and the latent space feature vectors extracted by the target domain autoencoder.

[0100] The tenth module is used to input the measured mixed influence factor after the data to be predicted is standardized into the trained predictor model to obtain the predicted latent space feature vector. Based on the decoder in the trained target domain autoencoder, the predicted latent space feature vector is decoded and reconstructed to obtain the final range response prediction value.

[0101] On the other hand, the present invention provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the radar ranging prediction method based on multivariate real-time combined time-series data provided in any of the above embodiments. The computer device can be a server. The computer device includes a processor, a memory, a network interface, and a database connected via a system bus. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer device stores sample data. The network interface of the computer device is used for communication with external terminals via a network connection.

[0102] On the other hand, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the steps of the radar ranging prediction method based on multivariate real-time combined time-series data provided in any of the above embodiments.

[0103] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0104] Matters not covered in this invention are common knowledge.

[0105] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0106] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application.

[0107] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A radar ranging and prediction method based on multivariate real-time combined time-series data, characterized in that, include: Identify the mixed influence factors affecting radar ranging response, which include static factors that do not change with time and dynamic factors that change with time. Set a time domain and obtain physical radar ranging data samples and digital analog radar ranging data samples with mixed influence factors based on the same time domain. The physical radar ranging data samples are divided into ranging response samples and non-ranging response samples. The ranging response in the physical radar ranging data samples and digital analog radar ranging data samples changes continuously with the time domain and is a time series dynamic response. Data standardization processing was performed on physical radar ranging data samples and digital analog radar ranging data samples. The standardized physical radar ranging data samples and the digital analog radar ranging data samples are processed by sliding time windows to obtain target domain window sequence samples and source domain window sequence samples. Extract the latent space feature vector of the original ranging response in the target domain window sequence sample, and decode and reconstruct the ranging response based on the extracted latent space feature vector. Construct the target domain reconstruction loss function based on the original ranging response and the corresponding reconstructed ranging response in the target domain window sequence sample. Extract the latent space feature vector of the original ranging response from the source domain window sequence samples, and decode and reconstruct the ranging response based on the extracted latent space feature vector. Construct the source domain reconstruction loss function based on the original ranging response and the corresponding reconstructed ranging response from the source domain window sequence samples. The distribution difference of latent space feature vectors extracted by the target domain autoencoder and the source domain autoencoder is measured by multi-kernel maximum mean difference, and an alignment loss function is constructed. The total loss function is constructed based on the target domain reconstruction loss function, the source domain reconstruction loss function, and the alignment loss function to complete the training of the target domain autoencoder; The predictor model is trained to learn the mapping relationship between the mixed influence factors in the target domain window sequence samples and the latent space feature vectors extracted by the target domain autoencoder; The measured mixed influence factor after standardizing the data to be predicted is input into the trained predictor model to obtain the predicted latent space feature vector. The predicted latent space feature vector is then decoded and reconstructed based on the decoder in the trained target domain autoencoder to obtain the final range response prediction value.

2. The radar ranging and prediction method based on multivariate real-time combined time-series data according to claim 1, characterized in that, The static factors include static quantitative factors and static qualitative factors. The static quantitative factors include filtering parameters and temperature parameters, while the static qualitative factors include wind force level and navigation route. The dynamic factors include radar cross section and signal-to-noise ratio.

3. The radar ranging and prediction method based on multivariate real-time combined time-series data according to claim 2, characterized in that, Data standardization processing is performed on physical radar ranging data samples and digital analog radar ranging data samples, including: normal distribution standardization transformation is applied to the dynamic factors, static quantitative factors and ranging response of physical radar ranging data samples and digital analog radar ranging data samples, and one-hot encoding is applied to the static qualitative factors.

4. The radar ranging prediction method based on multivariate real-time combined time-series data according to claim 1, 2, or 3, characterized in that, The target domain reconstruction loss function and the source domain reconstruction loss function are as follows: in Indicates the source domain reconstruction loss. For the target domain reconstruction loss, For the source domain autoencoder model parameters, For the parameters of the target domain autoencoder model, and These are the original source domain ranging response value and the original target domain ranging response value, respectively. and These are the source domain ranging response reconstruction values ​​and the target domain ranging response reconstruction values, respectively.

5. The radar ranging and prediction method based on multivariate real-time combined time-series data according to claim 4, characterized in that, The alignment loss function is: in and The latent space feature vectors extracted by the source domain autoencoder and the target domain autoencoder are respectively, and the multi-kernel function is used. It is a linear combination of multiple Gaussian kernel functions.

6. The radar ranging and prediction method based on multivariate real-time combined time-series data according to claim 5, characterized in that, The total loss function is constructed based on the target domain reconstruction loss function, the source domain reconstruction loss function, and the alignment loss function, as follows: in The balance factor is set.

7. The radar ranging prediction method based on multivariate real-time combined time-series data according to claim 5 or 6, characterized in that, The predictor model includes a CNN module and a Bi-LSTM module. The CNN module receives static factors from the mixed influence factors, extracts local features of the static factors through a three-convolutional module structure, and flattens the local features of the static factors after convolution and pooling into static factor latent feature vectors. The Bi-LSTM module receives dynamic factors from the mixed influence factors, captures the long-term forward and backward dependencies in the dynamic factor sequence, and outputs the corresponding dynamic factor latent feature vectors. The static factor latent feature vectors and dynamic factor latent feature vectors are integrated through a fully connected layer to achieve prediction of the latent space features extracted by the target domain autoencoder.

8. A radar ranging and prediction device based on multivariate real-time combined time-series data, characterized in that, include: The first module is used to determine the mixed influence factors affecting the radar ranging response. The mixed influence factors include static factors that do not change with time and dynamic factors that change with time. The second module is used to set the time domain and obtain physical radar ranging data samples and digital analog radar ranging data samples based on the same time domain regarding the mixed influence factors. The physical radar ranging data samples are divided into ranging response samples and non-ranging response samples. The ranging response in the physical radar ranging data samples and digital analog radar ranging data samples changes continuously with the time domain and is a time series dynamic response. The third module is used to perform data standardization processing on physical radar ranging data samples and digital analog radar ranging data samples. The fourth module is used to perform sliding time window processing on the standardized physical radar ranging data samples and the digital analog radar ranging data samples to obtain target domain window sequence samples and source domain window sequence samples. The fifth module is used to extract the latent space feature vector of the original ranging response in the target domain window sequence sample, and decode and reconstruct the ranging response based on the extracted latent space feature vector. The target domain reconstruction loss function is constructed based on the original ranging response and the corresponding reconstructed ranging response in the target domain window sequence sample. The sixth module is used to extract the latent space feature vector of the original ranging response in the source domain window sequence sample, and decode and reconstruct the ranging response based on the extracted latent space feature vector. The source domain reconstruction loss function is constructed based on the original ranging response and the corresponding reconstructed ranging response in the source domain window sequence sample. The seventh module is used to measure the distribution difference of latent space feature vectors extracted by the target domain autoencoder and the source domain autoencoder using multi-kernel maximum mean difference, and to construct an alignment loss function. The eighth module is used to construct the total loss function based on the target domain reconstruction loss function, the source domain reconstruction loss function, and the alignment loss function, thereby completing the training of the target domain autoencoder. The ninth module is used to train the predictor model to learn the mapping relationship between the mixed influence factors in the target domain window sequence samples and the latent space feature vectors extracted by the target domain autoencoder. The tenth module is used to input the measured mixed influence factor after the data to be predicted is standardized into the trained predictor model to obtain the predicted latent space feature vector. Based on the decoder in the trained target domain autoencoder, the predicted latent space feature vector is decoded and reconstructed to obtain the final range response prediction value.

9. The radar ranging and prediction device based on multivariate real-time combined time-series data according to claim 8, characterized in that, The static factors include static quantitative factors and static qualitative factors. The static quantitative factors include filtering parameters and temperature parameters, while the static qualitative factors include wind force level and navigation route. The dynamic factors include radar cross section and signal-to-noise ratio.

10. The radar ranging and prediction device based on multivariate real-time combined time-series data according to claim 8, characterized in that, The third module includes: applying a normal distribution normalization transformation to the dynamic factors, static quantitative factors, and ranging response of physical radar ranging data samples and digital analog radar ranging data samples, and using one-hot encoding for the static qualitative factors.

Images