Radar measurement distance fitting method and device based on deep learning

Abstract

The present invention proposes a radar distance fitting method and device based on deep learning, including inputting radar distance measurement data samples, the radar distance measurement data samples including the true value of radar distance measurement data and the corresponding dynamic factor data affecting radar distance measurement and static factor data affecting radar distance measurement; extracting static data features of static factor data and dynamic data features of dynamic factor data; splicing the static data features and dynamic data features corresponding to the radar distance measurement data samples to obtain a high-dimensional target feature vector; predicting the radar distance measurement data based on the high-dimensional target feature vector, and constructing a loss function between the predicted value of the radar distance measurement data and the true value of the radar distance measurement data to realize model training and obtain an optimal training model. The present invention takes into account the trend and periodicity of factors and the dependency between factors, realizes feature extraction of mixed factors, and makes the prediction result more accurate.

Description

Technical Field

[0001] The present invention mainly relates to the field of radar ranging technology, and in particular to a radar measurement distance fitting method and device based on deep learning. Background Art

[0002] Radar is the main equipment for target detection, and radar detection performance is an important performance indicator. Evaluating the detection capability of radar through experiments is of great significance for ensuring the performance of radar. In the detection performance test of radar, there are often a large number of influencing factors of nonlinear relationships, and due to the interference of uncontrollable factors such as the environment in actual tests, the data is usually nonlinear, non-stationary, and contains different types of high-dimensional complex data. To evaluate the detection performance of radar, it is necessary to conduct a flight test, that is, the tested radar measures the target set on the aircraft and calculates indicators such as distance, pitch angle, and azimuth. At the same time, in order to ensure the detection accuracy, the same target will be used under the same test conditions and repeated measurements will be made over multiple flights.

[0003] Traditional radar ranging accuracy evaluation methods are mostly based on radar equation analysis, but there are many influencing factors in the equation, and some environmental variables cannot be directly measured, so it is difficult and inconvenient to analyze radar ranging accuracy directly from the influencing factors. Existing radar ranging accuracy evaluation research tends to start from the data itself and analyze the distribution characteristics of variables through statistical methods or machine learning methods.

[0004] In the evaluation process, time series response modeling is a widely used and indispensable part. Through time series response modeling, it is possible to predict the key indicators and performance of equipment, optimize the test plan, and assist in forming more scientific and powerful decisions. For example, when the test cost is high and the test cycle is long, time series response modeling can help determine the number of tests, key nodes, etc., and reduce blindness and waste of resources in the test.

[0005] A sequence of values ​​of the same statistical indicator arranged in the order of their occurrence is called a time series. There are many existing time series prediction methods, which can be divided into traditional time series analysis methods and modern time series analysis methods. The former decomposes the time series into several basic factors for analysis; the latter is based on various random process theories and uses various complex statistical and computational models to analyze time series data. While considering data trends and seasonality, it can capture the complex structure and dynamic relationship in the data. Traditional time series analysis methods believe that time series are generally the result of the combined effects of long-term trends, seasonal changes, cyclic changes, and irregular changes. Common analysis methods include exponential smoothing, seasonal decomposition, moving average, autoregressive difference moving average, etc. Modern time series analysis methods can handle more complex time series relationships and are suitable for high-dimensional and big data environments; among them, deep learning methods have the advantages of strong nonlinear representation capabilities and no reliance on high-quality features. Common methods include convolutional neural networks (CNN), recurrent neural networks (RNN), long short-term memory (LSTM), transformer, etc.

[0006] At present, some people have designed a BP neural network framework model to predict the actual detection distance of the radar, and accurately fit the radar detection distance based on the input of factors such as weather, interference, target height, size, etc. that affect the detection distance. Others have proposed a correction angle algorithm based on LSTM to address the impact of parameters such as pipeline travel distance, acceleration, torque, and speed on the attitude deviation of pipeline bottom radar. They compared the prediction effect with that of the ARIMA model and found that the neural network has higher accuracy. However, most of the existing time series analysis methods only consider dynamic data, that is, time series factors, and ignore static factors that have no order, such as temperature, wind force and other external environmental factors, which often cause inaccurate fitting problems. Summary of the invention

[0007] The technical problem to be solved by the present invention is: When static factors and dynamic factors exist at the same time, how to comprehensively mix the factors to construct a response model for data with long cycles, large data, high dimensions, and complex processes, and thus achieve high-precision response prediction of input variables. In response to the above technical problems, the present invention provides a radar measurement distance fitting method and device based on deep learning.

[0008] In order to achieve the above object, the present invention is implemented by the following technical solutions:

[0009] On the one hand, the present invention proposes a radar measurement distance fitting method based on deep learning, comprising:

[0010] Inputting a radar ranging data sample, the radar ranging data sample includes a true value of the radar measurement distance data and corresponding dynamic factor data affecting the radar ranging and static factor data affecting the radar ranging;

[0011] Extract static data features of static factor data affecting radar ranging from radar ranging data samples;

[0012] Extracting dynamic data features of dynamic factor data affecting radar ranging from radar ranging data samples;

[0013] The static data features and dynamic data features corresponding to the radar ranging data samples are spliced ​​to obtain a high-dimensional target feature vector;

[0014] The radar distance data is predicted based on the high-dimensional target feature vector, and the loss function between the predicted value of the radar distance data and the true value of the radar distance data is constructed to train the model and obtain the optimal training model.

[0015] On the other hand, a radar measurement distance fitting device based on deep learning is provided, comprising:

[0016] An input module is used to input radar ranging data samples, where the radar ranging data samples include the true value of the radar measurement distance data and the corresponding dynamic factor data affecting the radar ranging and the static factor data affecting the radar ranging;

[0017] A static data feature extraction module is used to extract static data features of static factor data that affects radar ranging in radar ranging data samples;

[0018] A dynamic data feature extraction model is used to extract dynamic data features of dynamic factor data that affect radar ranging in radar ranging data samples;

[0019] A splicing module is used to splice the static data features and dynamic data features corresponding to the radar ranging data samples to obtain a high-dimensional target feature vector;

[0020] The prediction module is used to predict the radar measurement distance data based on the high-dimensional target feature vector, and at the same time construct a loss function between the predicted value of the radar measurement distance data and the true value of the radar measurement distance data to realize the training of the model and obtain the optimal training model.

[0021] 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 when the processor executes the computer program, the steps of the above-mentioned radar measurement distance fitting method based on deep learning are implemented.

[0022] On the other hand, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-mentioned radar measurement distance fitting method based on deep learning.

[0023] On the other hand, the present invention provides a computer program product, which is stored on a computer-readable storage medium and includes computer instructions, which, when executed by a processor, enable a computer device to implement the steps of the above-mentioned deep learning-based radar measurement distance fitting method.

[0024] Compared with the prior art, the technical effects of the present invention are:

[0025] The present invention focuses on the distance detection index, and records the distance of the target (such as the target) measured by the radar relative to the radar as the radar measurement distance as the response variable. Two key indicators that have important influencing factors on the distance measurement from the mechanism are selected as the dynamic data input model, including the Radar Cross Section (RCS) and the Signal to Noise Ratio (SNR); in addition, four external influencing factors are selected as the static data input model, including filtering, temperature, wind force level, navigation route and other factors.

[0026] The problem of multivariate time series prediction of comprehensive static factors and dynamic factors is a complex problem. For this, the present invention proposes a radar distance measurement fitting method based on deep learning, including inputting radar ranging data samples, extracting static data features of static factor data and dynamic data features of dynamic factor data in radar ranging data samples; splicing static data features and dynamic data features corresponding to radar ranging data samples to obtain high-dimensional target feature vectors; predicting radar ranging data based on high-dimensional target feature vectors, and constructing a loss function between the predicted value of radar ranging data and the true value of radar ranging data to achieve model training and obtain the optimal training model. The present invention takes into account the trend and periodicity of factors and the dependency between factors, realizes feature extraction of mixed factors, and makes the prediction result more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. 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 the structures shown in these drawings without paying creative work.

[0028] Figure 1 A flowchart of a radar distance measurement fitting method based on deep learning provided by an embodiment;

[0029] Figure 2 It is a dynamic time series feature extraction diagram based on Transformer in one embodiment of the present invention. DETAILED DESCRIPTION

[0030] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. 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 creative work are within the scope of protection of the present invention.

[0031] Reference Figure 1 , an embodiment provides a radar measurement distance fitting method based on deep learning, including:

[0032] Inputting a radar ranging data sample, the radar ranging data sample includes a true value of the radar measurement distance data and corresponding dynamic factor data affecting the radar ranging and static factor data affecting the radar ranging;

[0033] Extract static data features of static factor data affecting radar ranging from radar ranging data samples;

[0034] Extracting dynamic data features of dynamic factor data affecting radar ranging from radar ranging data samples;

[0035] The static data features and dynamic data features corresponding to the radar ranging data samples are spliced ​​to obtain a high-dimensional target feature vector;

[0036] The radar distance data is predicted based on the high-dimensional target feature vector, and the loss function between the predicted value of the radar distance data and the true value of the radar distance data is constructed to train the model and obtain the optimal training model.

[0037] The true value of the radar ranging data in the radar ranging data sample in the present invention is the difference between the true distance of the target from the radar and the measured distance of the target from the radar measured by the radar.

[0038] The present invention focuses on the distance detection index, and records the distance of the target (such as the target) measured by the radar relative to the radar as the radar measurement distance as the response variable. Two key indicators that have important influencing factors on the distance measurement from the mechanism are selected as dynamic factors, including Radar Cross Section (RCS) and Signal to Noise Ratio (SNR); in addition, four external influencing factors are selected as static factors, including filtering, temperature, wind force level, navigation route and other factors. The multivariate time series prediction problem that integrates static factors and dynamic factors is a complex problem.

[0039] Given a lookback window, use the past Predict the future in steps Step data, ,in , is the mixed influencing factor data of the lookback window, The predicted future The radar measures the distance data within the step time. It represents the prediction mapping function, which characterizes the relationship between the six influencing factors and the radar measurement distance. The key to multivariate prediction is to construct an algorithm to characterize this mapping relationship. In modern time series analysis methods, the model is often a black box, that is, the relationship cannot be mathematically represented, but it has high accuracy for complex problems.

[0040] In the present invention, a lookback window is given, and is the static factor data sequence in the lookback window, It is the number of static factors. The data of static factors is independent of time and can be both qualitative and quantitative data. is the dynamic factor data sequence in the lookback window, where is the number of dynamic factors, is the length of the lookback window, The dynamic factor data series is , .

[0041] There are four static factors in the static factor data: N =4, which are filter data, temperature data, wind force level data and navigation route data; They are the first, second, third, and fourth static factor data sequences in the lookback window, namely, the filter data sequence, temperature data sequence, wind force level data sequence, and navigation route data sequence in the lookback window;

[0042] There are two dynamic factors in the dynamic factor data, namely M =2, are radar cross section and signal-to-noise ratio respectively; It represents the first dynamic factor data sequence, i.e. the radar cross-sectional area data sequence in the lookback window. Represents the second dynamic factor data sequence, namely the signal-to-noise ratio data sequence in the look-back window.

[0043] In one embodiment, the static data feature extraction module based on CNN extracts the static data features of the static factor data that affects the radar ranging in the radar ranging data sample, first extracts the local features using the convolution layer, and then reduces the size of the feature map through the pooling layer. The input of the static data feature extraction module based on CNN is the lookback window 4 static factor data , output It is the static data feature extracted by 4 factors, and the calculation formula is as follows:

[0044] ;

[0045] ;

[0046] in, represents the convolution kernel, and They represent the size of the convolution kernel, , , Indicates size The convolution kernel, is the bias term of the convolutional layer, Represents the input matrix Middle Static Factor No. data, input matrix Is Static Factors in Time The spatial area captured by the sliding window within the lookback window time, is the local feature in the input matrix aligned with the convolution kernel, is the activation function, That is, the extracted static data features.

[0047] In the response model, "channel" refers to the way or path for independently processing each input dimension. When extracting features from multi-dimensional dynamic data, each channel processes the data dimension it is responsible for separately, that is, channel independence. Channel independence can enhance the stability and accuracy of the model to a limited extent. However, when there are complex dependencies between high-dimensional dynamic data, channel independence will ignore the interaction and correlation between different variables, resulting in the model being insufficient in capturing the linkage characteristics between variables, thereby causing deviations when processing interdependent time series data, and thus affecting the model's predictive performance. For the trend component of dynamic data, there is often no obvious relationship between different factor channels, while there is a significant correlation between seasonal components. To this end, when constructing the response model, the present invention fully considers the characteristics of the two dynamic factors, and needs to design channel-dependent and channel-independent feature extraction modules for the trend component and the seasonal component, respectively.

[0048] Figure 2 It is a dynamic time series feature extraction diagram based on Transformer in one embodiment of the present invention, which includes: firstly, the trend component and the period component of the input dynamic data enter the trend component intra-attention module and the period component inter-attention module respectively;

[0049] The attention module within the trend component provides the time step position information of the dynamic factor trend component through position embedding to generate the query , by numerical embedding, the input trend component data is numerically represented to generate a key Sum ,Will The input cross-attention mechanism calculates the dependencies between different factors, extracts the dependencies within the trend component, and concatenates the features of different factors into a unified output;

[0050] The periodic component attention module numerically embeds the periodic value of each time step of the input dynamic data periodic component and maps it to the low-dimensional feature space to obtain the query ,key Sum , input it into the self-attention mechanism to capture the long-term step dependency in the periodic component of a single factor, and then flatten the periodic component feature data into data with the same dimension as the trend component feature;

[0051] Finally, the extracted trend features and period features are fused, the feature dimensions are aligned and summed up, and the features of the dynamic data are obtained.

[0052] In another embodiment, extracting dynamic data features of dynamic factor data affecting radar ranging in radar ranging data samples includes the following steps:

[0053] Enter the dynamic factor data sequence in the look-back window;

[0054] Perform channel-independent instance normalization on different dynamic factor data sequences and calculate the mean and standard deviation of each time series channel;

[0055] After normalizing the dynamic factor data in the corresponding dynamic factor data sequence based on the mean and standard deviation of each time series channel, the trend component and cycle component of the dynamic factor data are extracted;

[0056] Based on the trend component of the dynamic factor data, the cross-attention mechanism is used to obtain the output features of the intra-slice attention representing local details;

[0057] Based on the periodic components of the dynamic factor data, the self-attention mechanism is used to calculate the correlation between segments, and the output features of the inter-segment attention representing the global temporal dependency information are obtained;

[0058] The output features of the intra-slice attention and the inter-slice attention are fused to obtain the final dynamic data features.

[0059] For the dynamic factor data sequence in the lookback window , corresponding to the two dynamic factors of RCS and SNR and Perform channel-independent instance normalization and calculate the mean and standard deviation of each time series channel.

[0060] For each channel , No. Dynamic factor data series The mean and variance The calculation formula is as follows:

[0061] ;

[0062] ;

[0063] Normalize the input data using these statistics:

[0064] ;

[0065] in, and is a learnable parameter that allows the model to introduce a certain degree of adjustability during normalization. is a very small value used to prevent the denominator from being 0.

[0066] After normalizing the data, first extract the dynamic time series data The trend component in and periodic components .

[0067] Use the moving average method to normalize the dynamic time series data Smoothing is performed to capture the trend component of dynamic time series data. The formula is:

[0068] ;

[0069] Then we can get the periodic component, the formula is:

[0070] ;

[0071] in, It is a moving average operation that captures trends by smoothing the data. is a padding operation to keep the length of the sequence unchanged. and They are the extracted trend component and cycle component respectively.

[0072] Based on the decoupling of data in each dimension, a dual attention mechanism is proposed to model the temporal dependency of the separated trend component and periodic component. Aiming at the variable independence of the trend component and the variable dependency of the periodic component, a channel-independent cross-attention mechanism and a channel-dependent self-attention mechanism are constructed respectively.

[0073] Intra-trend component attention is used to capture individual channels Dependencies within a time step, Dynamic factor data series The trend component is , embedded into a lower dimensional feature space along the feature length, we get , is the feature dimension after embedding, and embedding is a trainable linear transformation. Then, embedding is used to generate query from trend components , key and value Next, the cross-attention mechanism is used to obtain the output features of the intra-chip attention representing the local details, where the intra-chip attention is calculated through the self-attention mechanism, and the formula is as follows:

[0074] ;

[0075] The output features of the intra-slice attention of each fragment are concatenated to obtain the final output features of the intra-slice attention representing local details:

[0076] ;

[0077] Inter-periodic component attention is used to capture the global relationship between periodic components in different dimensions, also along the feature length Each factor is embedded and rearranged so that the time step and the embedding feature dimension are combined to obtain ,in Then, the self-attention mechanism is used to calculate the correlation between the fragments and obtain the trend attention:

[0078] ;

[0079] in, are the queries, keys, and values ​​obtained through linear transformation.

[0080] Finally, the output features of the intra-chip attention and the output features of the inter-chip attention are fused to obtain the final dynamic data features. And the output of inter-slice attention By performing a linear transformation on the time step of each fragment and then adding the features within the local fragment to the attention results between fragments, the final dynamic data features are obtained as follows:

[0081] ;

[0082] The final dynamic data features contain the local and global temporal dependency information of the dynamic data.

[0083] After obtaining the final dynamic data features, RevIN is used to perform a denormalization operation to remap the dynamic features back to the scale of the original data, making the model results more interpretable:

[0084] ;

[0085] After extracting the static data features and the denormalized dynamic data features, the extracted static data features and dynamic data features are concatenated to form a complete feature vector, i.e., a high-dimensional target feature vector , the formula is as follows:

[0086] ;

[0087] The radar measurement distance data is predicted based on the high-dimensional target feature vector, that is, the concatenated high-dimensional target feature vector is projected through the fully connected layer to map the high-dimensional target feature vector to the target dimension. The formula is expressed as:

[0088] ;

[0089] in, Represents the predicted value of radar measurement distance data, represents the high-dimensional target feature vector, represents a nonlinear fully connected layer, Look back at the historical prediction values ​​of the window and use them as corrections to the prediction results to enhance the model's temporal memory ability.

[0090] At the same time, a loss function between the predicted value of the radar measurement distance data and the true value of the radar measurement distance data is constructed to train the model and obtain the optimal training model.

[0091] The effect quantification indexes in the present invention are mean square error (MSE) and mean absolute error (MAE), where MSE is used as the loss function iteration model, and the index calculation formula is as follows:

[0092] ;

[0093] ;

[0094] in, is the predicted future time length, Indicates the number of radar ranging data samples The true value of the radar distance data, It means that the first The predicted value of the radar measurement distance data.

[0095] The optimizer uses the Adam optimizer to optimize the model through back-propagation and record the loss of the optimal training model.

[0096] To prove the effectiveness of the present invention, a specific simulation example is provided below:

[0097] For the test data of a new type of radar, each model of aircraft was inspected for 4 sorties. The radar evenly sampled the target and recorded the actual distance of the target from the radar, the distance from the radar to itself, RCS and SNR during the inspection process. The data smoothing filtering, the temperature and wind level of the day, and the flight route of the target were also recorded. The distance of the target (such as the target) measured by the radar relative to the radar was recorded as the radar measurement distance as the response variable. The factor data was integrated into the model, the network parameters were adjusted, and MSE and MAE were used as indicators to consider the accuracy of the model fitting.

[0098] First, preprocess the data. In order to ensure the stability and accuracy of radar ranging, and considering the limited ranging range of the radar, only the data of the flight test section is retained. Therefore, the starting point and end point are set for the flight test, and only the ranging data within this distance is quantitatively analyzed. Based on experience, the starting point and end point are set to 30km and 270km. The distance threshold is set to 90 meters, and all ranging data must be within the threshold range. Data exceeding the threshold is deleted, and outliers are removed according to the Laida criterion. The amount of data before and after data preprocessing is shown in Table 1:

[0099] Table 1 Data scale before and after data preprocessing

[0100]

[0101] Then, the data was input into the model and the parameters were adjusted. The network parameter settings of the dynamic network are shown in Table 2, as follows:

[0102] Table 2 Network parameter settings

[0103]

[0104] Finally, the best training results of the four sorties of this model are obtained, as shown in Table 3:

[0105] Table 3 Training results

[0106]

[0107] In one embodiment, a radar measurement distance fitting device based on deep learning is provided, comprising:

[0108] An input module is used to input radar ranging data samples, where the radar ranging data samples include the true value of the radar measurement distance data and the corresponding dynamic factor data affecting the radar ranging and the static factor data affecting the radar ranging;

[0109] A static data feature extraction module is used to extract static data features of static factor data that affects radar ranging in radar ranging data samples;

[0110] A dynamic data feature extraction model is used to extract dynamic data features of dynamic factor data that affect radar ranging in radar ranging data samples;

[0111] A splicing module is used to splice the static data features and dynamic data features corresponding to the radar ranging data samples to obtain a high-dimensional target feature vector;

[0112] The prediction module is used to predict the radar measurement distance data based on the high-dimensional target feature vector, and at the same time construct a loss function between the predicted value of the radar measurement distance data and the true value of the radar measurement distance data to realize the training of the model and obtain the optimal training model.

[0113] The implementation method of each module in the above-mentioned device and the construction of the model can adopt the method described in any of the above-mentioned embodiments, which will not be repeated here.

[0114] On the other hand, the present invention provides a computer device, including a memory and a processor, the memory storing a computer program, and the processor implementing the steps of the radar measurement distance fitting method based on deep learning provided in any of the above embodiments when executing the computer program. The computer device may be a server. The computer device includes a processor, a memory, a network interface and a database connected via a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an 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 the computer program in the non-volatile storage medium. The database of the computer device is used to store sample data. The network interface of the computer device is used to communicate with an external terminal through a network connection.

[0115] On the other hand, the present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the radar measurement distance fitting method based on deep learning provided in any of the above embodiments are implemented.

[0116] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may 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 many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0117] Matters not covered by the present invention are known technologies.

[0118] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, 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.

[0119] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention. It should be pointed out that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

[0120] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A radar measurement distance fitting method based on deep learning, characterized in that: include: Input radar ranging data samples, which include the true value of radar distance measurement data and the corresponding dynamic factor data affecting radar ranging and static factor data affecting radar ranging, including: given lookback window ,set up is the static factor data sequence in the lookback window, It is the number of static factors. The data of static factors is independent of time and can be both qualitative and quantitative data. is the dynamic factor data sequence in the lookback window, where is the number of dynamic factors, is the length of the lookback window, The dynamic factor data series is , d =1,..., L , =1,..., M ; wherein the static factors in the static factor data have four static factors, namely N =4, which are filter data, temperature data, wind force level data and navigation route data; They are the first, second, third and fourth static factor data sequences in the lookback window, namely, the filter data sequence, temperature data sequence, wind force level data sequence and navigation route data sequence in the lookback window; there are two dynamic factors in the dynamic factor data, namely M =2, are radar cross section and signal-to-noise ratio respectively; It represents the first dynamic factor data sequence, i.e. the radar cross-sectional area data sequence in the lookback window. It represents the second dynamic factor data sequence, i.e., the signal-to-noise ratio data sequence in the look-back window; Extract the static data features of the static factor data that affects radar ranging in the radar ranging data sample. The calculation formula is as follows: in, represents the convolution kernel, and They represent the size of the convolution kernel, p =0,..., -1, =0,..., -1, Indicates size The convolution kernel, is the bias term of the convolutional layer, Represents the input matrix Middle Static Factor No. data, input matrix Is Static Factors in Time The spatial area captured by the sliding window within the lookback window time, is the local feature in the input matrix aligned with the convolution kernel, is the activation function, That is, the extracted static data features; Extract dynamic data features of dynamic factor data affecting radar ranging from radar ranging data samples, including: Enter the dynamic factor data sequence in the look-back window; Perform channel-independent instance normalization on different dynamic factor data sequences and calculate the mean and standard deviation of each time series channel; After normalizing the dynamic factor data in the corresponding dynamic factor data sequence based on the mean and standard deviation of each time series channel, the trend component and cycle component of the dynamic factor data are extracted; Based on the trend component of the dynamic factor data, the cross-attention mechanism is used to obtain the output features of the intra-slice attention representing local details; Based on the periodic components of the dynamic factor data, the self-attention mechanism is used to calculate the correlation between segments, and the output features of the inter-segment attention representing the global temporal dependency information are obtained; The output features of the intra-slice attention and the inter-slice attention are combined to obtain the final dynamic data features; The static data features and dynamic data features corresponding to the radar ranging data samples are spliced ​​to obtain a high-dimensional target feature vector; The radar measurement distance data is predicted based on the high-dimensional target feature vector, and the loss function between the predicted value of the radar measurement distance data and the true value of the radar measurement distance data is constructed to realize the training of the model and obtain the optimal training model.

2. The radar distance measurement fitting method based on deep learning according to claim 1 is characterized in that: Prediction of radar ranging data based on high-dimensional target feature vectors, including: in, represents the high-dimensional target feature vector, represents a nonlinear fully connected layer, Look back at the historical prediction values ​​of the window and use them as corrections to the prediction results to enhance the model's temporal memory ability.

3. The radar measurement distance fitting method based on deep learning according to claim 1 is characterized in that: The mean squared error is chosen as the loss function.

4. The radar measurement distance fitting method based on deep learning according to claim 1 is characterized in that: It also includes using the Adam optimizer to optimize the model through back-propagation and recording the loss of the optimal trained model.

5. A radar distance measurement fitting device based on deep learning, characterized in that: include: The input module is used to input radar ranging data samples. The radar ranging data samples include the true value of the radar measurement distance data and the corresponding dynamic factor data affecting the radar ranging and the static factor data affecting the radar ranging, including: a given lookback window ,set up is the static factor data sequence in the lookback window, It is the number of static factors. The data of static factors is independent of time and can be both qualitative and quantitative data. is the dynamic factor data sequence in the lookback window, where is the number of dynamic factors, is the length of the lookback window, The dynamic factor data series is , d =1,..., L , =1,..., M ; wherein the static factors in the static factor data have four static factors, namely N =4, which are filter data, temperature data, wind force level data and navigation route data; They are the first, second, third and fourth static factor data sequences in the lookback window, namely, the filter data sequence, temperature data sequence, wind force level data sequence and navigation route data sequence in the lookback window; there are two dynamic factors in the dynamic factor data, namely M =2, are radar cross section and signal-to-noise ratio respectively; It represents the first dynamic factor data sequence, i.e. the radar cross-sectional area data sequence in the lookback window. It represents the second dynamic factor data sequence, i.e., the signal-to-noise ratio data sequence in the look-back window; The static data feature extraction module is used to extract the static data features of the static factor data that affects the radar ranging in the radar ranging data sample. The calculation formula is as follows: in, represents the convolution kernel, and They represent the size of the convolution kernel, p =0,..., -1, =0,..., -1, Indicates size The convolution kernel, is the bias term of the convolutional layer, Represents the input matrix Middle Static Factor No. data, input matrix Is Static Factors in Time The spatial area captured by the sliding window within the lookback window time, is the local feature in the input matrix aligned with the convolution kernel, is the activation function, That is, the extracted static data features; The dynamic data feature extraction model is used to extract the dynamic data features of the dynamic factor data that affects the radar ranging in the radar ranging data sample, including: Enter the dynamic factor data sequence in the look-back window; Perform channel-independent instance normalization on different dynamic factor data sequences and calculate the mean and standard deviation of each time series channel; After normalizing the dynamic factor data in the corresponding dynamic factor data sequence based on the mean and standard deviation of each time series channel, the trend component and cycle component of the dynamic factor data are extracted; Based on the trend component of the dynamic factor data, the cross-attention mechanism is used to obtain the output features of the intra-slice attention representing local details; Based on the periodic components of the dynamic factor data, the self-attention mechanism is used to calculate the correlation between segments, and the output features of the inter-segment attention representing the global temporal dependency information are obtained; The output features of intra-slice attention and inter-slice attention are combined to obtain the final dynamic data features; A splicing module is used to splice the static data features and dynamic data features corresponding to the radar ranging data samples to obtain a high-dimensional target feature vector; The prediction module is used to predict the radar measurement distance data based on the high-dimensional target feature vector, and at the same time construct a loss function between the predicted value of the radar measurement distance data and the true value of the radar measurement distance data to realize the training of the model and obtain the optimal training model.

6. The radar distance measurement fitting device based on deep learning according to claim 5 is characterized in that: In the prediction module, the radar ranging data is predicted based on the high-dimensional target feature vector, including: in, represents the high-dimensional target feature vector, represents a nonlinear fully connected layer, Look back at the historical prediction values ​​of the window and use them as corrections to the prediction results to enhance the model's temporal memory ability.

7. The radar distance measurement fitting device based on deep learning according to claim 5 is characterized in that: In the prediction module, mean square error is selected as the loss function.

Images