Multi-convolutional neural network-based UWB radar target motion prediction method and system

Abstract

The invention relates to a UWB radar target motion prediction method and system based on a multi-convolutional neural network, and the method comprises the steps: obtaining echo signal features collected by a UWB radar in real time and corresponding target motion parameters, and obtaining an initial data set through time series data division and normalization processing; extracting and splicing the initial data set based on a local mean decomposition algorithm to obtain standard input feature data; constructing a motion parameter prediction model based on the echo signal features, optimizing the motion parameter prediction model by adopting a genetic algorithm, and performing training through standard input feature data; and eliminating multi-path and noise interference from echo signal characteristics corresponding to a to-be-detected target through local mean decomposition, and inputting the echo signal characteristics into the trained motion parameter prediction model to obtain predicted motion parameters of the to-be-detected target. The UWB radar motion parameter prediction precision and generalization ability can be improved through LMD interference removal and CNN combined with attention and gating and genetic algorithm optimization.

Description

Technical Field

[0001] This application relates to the field of UWB radar technology, and in particular to a UWB radar target motion prediction method and system based on multiple convolutional neural networks. Background Technology

[0002] Ultra-wideband (UWB) radar technology, with its core characteristics of wide spectrum and short pulses, boasts advantages such as high temporal resolution, strong anti-interference capability, and high positioning accuracy, and has been widely applied in various fields including indoor positioning, industrial equipment monitoring, security tracking, and missile defense. Its core requirement is to accurately predict key motion parameters of a target, such as distance, velocity, and azimuth, by analyzing the characteristics of the target echo signal, providing data support for subsequent positioning, tracking, or threat assessment. However, in complex real-world scenarios, UWB radar echo signals are susceptible to interference from multipath reflections, random noise, and non-line-of-sight (NLOS) propagation. Furthermore, the dynamic changes in the target's motion state lead to nonlinear and non-stationary characteristics in the signal features, posing a significant challenge to the accuracy and stability of motion parameter prediction.

[0003] In existing technologies, UWB radar target motion parameter prediction methods are mainly divided into two categories: one is based on traditional signal processing algorithms, including Kalman filtering, least squares method, TOA / TDOA positioning algorithm, etc., which estimate target parameters by establishing a signal propagation model or filtering and smoothing; the other is based on conventional neural network models, such as BP neural network (BPNN) and long short-term memory network (LSTM), which use deep learning to fit the mapping relationship between signal features and motion parameters. However, these existing technologies have obvious shortcomings. Traditional signal processing algorithms are poorly adaptable to non-stationary signals and have difficulty effectively separating multipath interference from the real signal components, resulting in large parameter estimation biases. They also rely on accurate dynamic models, which can easily lead to prediction lag when the target's motion state is complex. Furthermore, conventional neural networks cannot dynamically allocate weights to different echo signal features, making it difficult to highlight the key effective features for the prediction results. In addition, some models have imperfect gating mechanisms, making it difficult to accurately capture the temporal dependencies of the target's motion, affecting the consistency and accuracy of parameter prediction. At the same time, the model hyperparameters often rely on manual tuning or simple optimization strategies, which can easily get trapped in local optima, resulting in weak generalization ability and difficulty in adapting to changes in signal characteristics under different monitoring scenarios, leading to unstable prediction accuracy in practical applications.

[0004] Therefore, it is necessary to improve one or more of the problems existing in the above-mentioned related technical solutions.

[0005] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this disclosure is to provide a method and system for predicting the motion of UWB radar targets based on multiple convolutional neural networks, thereby overcoming, to at least some extent, one or more problems caused by the limitations and defects of related technologies.

[0007] In a first aspect, this application provides a UWB radar target motion prediction method based on multiple convolutional neural networks, including: The echo signal characteristics and corresponding target motion parameters acquired in real time by UWB radar are obtained, and an initial dataset is obtained through time-series data partitioning and normalization. The echo signal characteristics include echo intensity, pulse arrival time, signal-to-noise ratio, multipath component energy, and antenna array phase difference. The target motion parameters are range, velocity, and azimuth. The initial dataset is extracted and concatenated based on the Local Mean Decomposition (LMD) algorithm to eliminate multipath and noise interference and obtain standard input feature data. A motion parameter prediction model is constructed based on the echo signal characteristics. The MAE loss function is used as the prediction performance metric. The motion parameter prediction model is optimized using a genetic algorithm and trained using the standard input feature data. The echo signal features corresponding to the target under test are decomposed by local mean to eliminate multipath and noise interference, and then input into the trained motion parameter prediction model to obtain the predicted motion parameters of the target under test.

[0008] In one possible implementation, the step of acquiring the echo signal characteristics and corresponding target motion parameters collected in real time by UWB radar, and obtaining an initial dataset through time-series data partitioning and normalization processing, includes: A UWB radar sensor array is deployed in the monitoring area to collect echo signal characteristic data related to the target's motion parameters, and the target's distance parameters, velocity, and azimuth are recorded by a positioning device. The echo intensity is obtained through the signal strength detection module of the UWB radar sensor, the pulse arrival time is obtained through the time measurement module of the UWB radar sensor, the signal-to-noise ratio is obtained through the noise analysis module of the UWB radar sensor, the multipath component energy is obtained through the multipath resolution module of the UWB radar sensor, and the antenna array phase difference is obtained through the phase difference detection module of the UWB radar sensor. For the echo signal characteristic data per second, the sampling period is set to 0.1 seconds, and 10 uniform samples are performed to obtain the echo signal characteristics at each sampling time point; The echo signal characteristics per second are transformed into the interval [0,1] using the maximum and minimum normalization methods to obtain the initial dataset.

[0009] In one possible implementation, the echo signal characteristics at each sampling time point are as follows: Among them, the Let be the echo signal feature data at the i-th sampling time point, where i∈1,10; The normalization formula is: Among them, the The first normalized data Echo signal characteristic data at each sampling time point.

[0010] In one possible implementation, the step of extracting and concatenating the initial dataset based on the local mean decomposition algorithm to eliminate multipath interference and noise interference, and obtaining standard input feature data, includes: The initial dataset is used as the data to be decomposed to extract local extreme points, and the average value of adjacent local extreme points is calculated. The average values ​​of the local extreme points are connected by a straight line to obtain the local mean function. Based on the local extreme points, the envelope function values ​​of adjacent extreme points are calculated, and the adjacent envelope function values ​​are connected by the moving average method to obtain the envelope estimation function. The local mean function is separated from the initial dataset, the frequency modulation function is calculated based on the envelope estimation function, iterating until the envelope estimation function of the frequency modulation function is always 1, and all the envelope estimation functions are multiplied to obtain the envelope signal. The envelope signal is then multiplied by the frequency modulation function to obtain the first component of the initial dataset. The first component is used as the decomposed data to obtain the second component, the second component is used as the decomposed data to obtain the third component, the third component is used as the decomposed data to obtain the fourth component, and the four components are concatenated to obtain the standard input feature data.

[0011] In one possible implementation, the formula for calculating the average value of the local extrema is: in, It is the average value of the local extreme points. For the first in the initial dataset Local extreme points, For the first in the initial dataset Local extreme points; The formula for calculating the average value of the envelope function is: in, The average value of the envelope function; The formula for calculating the envelope signal is: in, For envelope signal, Let n be the envelope estimation function; The standard input feature data is: in, For standard input feature data, The echo intensity after component splicing. The arrival time of the pulse after component splicing. The signal-to-noise ratio of the signal after component splicing. This refers to the energy of the multipath components after component splicing. This represents the phase difference of the antenna array after component splicing.

[0012] In one possible implementation, the steps of constructing a motion parameter prediction model based on the echo signal features, using the MAE loss function as a metric for prediction performance, optimizing the motion parameter prediction model using a genetic algorithm, and training it with the standard input feature data include: A motion parameter prediction model is constructed based on an attention mechanism, a convolutional neural network, and a gating unit; wherein, the motion parameter prediction model includes: an attention mechanism layer, a convolutional layer, a gating unit layer, a Dropout layer, a regularization layer, and a Sigmoid activation function layer; Multiple initial populations are randomly generated, and after training the multiple initial populations, the MAE value of each population is solved; wherein, each initial population represents a combination scheme of the number of attention layers, convolutional layers and gating units in the model; The population with the lowest MAE value among the multiple initial populations is selected as the optimal population, and the optimal population is updated. When the maximum number of iterations is reached, the final optimal population with the lowest MAE value is obtained, thereby obtaining the motion parameter prediction model with the optimal combination of attention layer number, convolution layer number, and gate unit number. The optimized motion parameter prediction model is trained using the standard input feature data.

[0013] In one possible implementation, the attention mechanism layer is used to calculate the optimal set of weights for each echo signal feature in the prediction of target motion parameters, thereby obtaining weighted input feature data. The convolutional layer is used to extract detailed features from the weighted input feature data; The gated unit layer is used to obtain the long temporal dependency relationship of the captured features based on the detailed features, and to obtain temporal feature data; The Dropout layer is used to deactivate some neurons in the model and randomly suppress some feature values ​​in the temporal features to prevent overfitting. The regularization layer is used to normalize the feature values ​​of time-series feature data that suppresses some feature values. The Sigmoid activation function layer is used to process the normalized temporal feature data through the Sigmoid activation function to obtain a set of predicted motion parameters.

[0014] In one possible implementation, the MAE value is calculated using the MAE loss function, which is: in, The value of the loss function. For the actual target motion parameters, The target motion parameters predicted by the motion parameter prediction model. The number of predicted values. .

[0015] In one possible implementation, the optimal population is updated as follows: in, , For the updated initial population, For the first An initial population, To update the speed weights, , To select an interval random numbers, For the optimal population, In the first Number of attention layers in each population The number of convolutional layers. This refers to the number of layers in the gate control unit.

[0016] Secondly, this application provides a UWB radar target motion prediction system based on a multi-convolutional neural network, the system being used to execute the above-described method, the system comprising: The data acquisition module is used to acquire the echo signal characteristics and corresponding target motion parameters acquired in real time by the UWB radar, and to obtain an initial dataset through time-series data partitioning and normalization. The echo signal characteristics include echo intensity, pulse arrival time, signal-to-noise ratio, multipath component energy and antenna array phase difference, and the target motion parameters are range, velocity and azimuth angle. The decomposition and splicing module is used to extract and splice the initial dataset based on the local mean decomposition algorithm, eliminate multipath and noise interference, and obtain standard input feature data; The model building module is used to build a motion parameter prediction model based on the echo signal features, use the MAE loss function as the prediction performance metric, optimize the motion parameter prediction model using a genetic algorithm, and train it using the standard input feature data. The motion prediction module is used to input the echo signal features corresponding to the target under test into the trained motion parameter prediction model after eliminating multipath and noise interference through local mean decomposition, so as to obtain the predicted motion parameters of the target under test.

[0017] The technical solution provided in this application may include the following beneficial effects: This application presents a UWB radar target motion prediction method and system based on multi-convolutional neural networks. It can accurately extract and stitch together multi-dimensional echo signal features using the LMD algorithm, efficiently eliminating multipath interference and random noise, and significantly improving the purity and reliability of input features. Furthermore, by constructing a multi-convolutional neural network that integrates attention mechanisms and gating units, and combining the outlier robustness of the MAE loss function with the global optimization capability of the genetic algorithm, it accurately captures the nonlinear correlation and time-series dependence between signal features and target motion parameters, greatly improving the prediction accuracy and stability of distance, velocity, and azimuth. Simultaneously, relying on standardized data preprocessing procedures and an optimized model architecture, it ensures that the model has strong generalization capabilities under different monitoring scenarios, adapting to the real-time detection requirements of UWB radar and providing efficient and reliable technical support for accurate target tracking.

[0018] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0019] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure. It is obvious that the drawings described below are merely some embodiments of this disclosure, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0020] Figure 1A flowchart illustrating a UWB radar target motion prediction method based on a multi-convolutional neural network in an exemplary embodiment of this disclosure is shown. Figure 2 A detailed flowchart of step S100 of the UWB radar target motion prediction method based on a multi-convolutional neural network in an exemplary embodiment of this disclosure is shown. Figure 3 A detailed flowchart of step S200 of the UWB radar target motion prediction method based on a multi-convolutional neural network in an exemplary embodiment of this disclosure is shown. Figure 4 A detailed flowchart of step S300 of the UWB radar target motion prediction method based on a multi-convolutional neural network in an exemplary embodiment of this disclosure is shown. Figure 5 This diagram illustrates the structure of a UWB radar target motion prediction system based on a multi-convolutional neural network in an exemplary embodiment of this disclosure. Detailed Implementation

[0021] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that this disclosure will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0022] This example implementation first provides a UWB radar target motion prediction method based on multiple convolutional neural networks. This method can be applied to a terminal device, such as a mobile terminal like a mobile phone, desktop computer, personal digital assistant, laptop, tablet, or smartwatch. (Reference) Figure 1 As shown, the method may include the following steps: Step S100: Obtain the echo signal characteristics and corresponding target motion parameters acquired in real time by UWB radar, and obtain the initial dataset through time series data partitioning and normalization processing; the echo signal characteristics include echo intensity, pulse arrival time, signal-to-noise ratio, multipath component energy and antenna array phase difference, and the target motion parameters are range, velocity and azimuth angle.

[0023] Step S200: Extract and concatenate the initial dataset based on the local mean decomposition algorithm to eliminate multipath and noise interference and obtain standard input feature data.

[0024] Step S300: Construct a motion parameter prediction model based on the echo signal features, use the MAE loss function as the prediction performance metric, optimize the motion parameter prediction model using a genetic algorithm, and train it using the standard input feature data.

[0025] Step S400: After eliminating multipath and noise interference by local mean decomposition, the echo signal features corresponding to the target to be tested are input into the trained motion parameter prediction model to obtain the predicted motion parameters of the target to be tested.

[0026] The aforementioned method provides a high-quality data foundation for motion parameter prediction through comprehensive acquisition and standardized preprocessing of multi-dimensional echo signal features. Combined with the adaptive signal decomposition capability of the local mean decomposition algorithm, it accurately separates multipath interference and random noise, significantly improving the purity and reliability of input features. Simultaneously, by constructing a multi-convolutional neural network integrating attention mechanisms and gating units, it dynamically allocates the weights and priorities of different signal features, efficiently capturing the nonlinear correlation between features and motion parameters, as well as the long-term temporal dependencies of target motion. Coupled with the strong robustness of the MAE loss function to outlier data and the global optimization capability of the genetic algorithm, it effectively avoids the model getting trapped in local optima, optimizing the network structure and parameter configuration. Ultimately, it achieves high-precision and high-stability prediction of target distance, velocity, and azimuth. Furthermore, the standardized preprocessing process and optimized model architecture endow the method with strong generalization ability, enabling it to adapt to complex electromagnetic environments and diverse monitoring scenarios, meeting the core requirements of real-time detection and precise tracking by UWB radar, and providing reliable technical support for subsequent positioning and decision-making.

[0027] Below, we will refer to Figures 2 to 4 The steps of the method described above in this example embodiment will be explained in more detail.

[0028] In step S100, the echo signal characteristics and corresponding target motion parameters acquired in real time by the UWB radar are obtained, and an initial dataset is obtained through time-series data partitioning and normalization.

[0029] The echo signal characteristics include echo intensity, pulse arrival time, signal-to-noise ratio, multipath component energy, and antenna array phase difference; the target motion parameters are distance, velocity, and azimuth.

[0030] It should be noted that this step fully considers the characteristics of the radar echo signal, including echo intensity, in the prediction of UWB radar target motion parameters. Pulse arrival time Signal-to-noise ratio Multipath component energy Antenna array phase difference Distance to the corresponding target motion parameters ,speed Azimuth The impact of this; therefore, in this step, based on the clear radar hardware parameters, a UWB radar sensor is deployed to record echo signal characteristic data in real time and record the corresponding target real parameter data.

[0031] In one embodiment, such as Figure 2 As shown, step S100 may include the following sub-steps.

[0032] In step S110, a UWB radar sensor array is deployed in the monitoring area to collect echo signal characteristic data related to the target's motion parameters, and the target's distance parameters, velocity, and azimuth are recorded by a positioning device.

[0033] It should be noted that the deployment of UWB radar sensor arrays must be based on a pre-defined monitoring area, the possible trajectory of the target's movement, and the radar hardware parameters, ensuring that the sensor array can fully cover the target's movement area without any blind spots, thereby accurately collecting echo signal characteristics from different positions and under different motion states. The positioning equipment used must have high-precision measurement capabilities matching the UWB radar's acquisition frequency. The target's actual distance, velocity, and azimuth data recorded by the equipment will serve as supervisory labels for model training, used for subsequent calculation of prediction errors and optimization of model parameters, ensuring that the model learns the true mapping relationship between echo signal characteristics and target motion parameters.

[0034] The echo intensity is obtained through the signal strength detection module of the UWB radar sensor, the pulse arrival time is obtained through the time measurement module of the UWB radar sensor, the signal-to-noise ratio is obtained through the noise analysis module of the UWB radar sensor, the multipath component energy is obtained through the multipath resolution module of the UWB radar sensor, and the antenna array phase difference is obtained through the phase difference detection module of the UWB radar sensor.

[0035] In step S120, for the echo signal characteristic data per second, the sampling period is set to 0.1 seconds, and 10 uniform samples are performed to obtain the echo signal characteristics at each sampling time point.

[0036] It should be noted that the sampling period is set to 0.1 seconds and the number of samples per session is set to 10. This is a comprehensive consideration of the dynamics of the target motion and the efficiency of data processing: the 0.1-second sampling interval can avoid losing the motion details of the target in a short period of time due to sparse sampling, while avoiding data redundancy due to excessive sampling; the 10 uniform samples can completely preserve the temporal continuity of the target motion within one second, providing a coherent input basis for subsequent model learning based on temporal features.

[0037] The echo signal characteristics at each sampling time point are as follows: Among them, the Let be the echo signal feature data at the i-th sampling time point, where i∈1,10.

[0038] Understandable, For arbitrary echo signal characteristic data, such as ,in, For the first The echo intensity at each sampling time point.

[0039] It should be noted that, apart from echo intensity, pulse arrival time, signal-to-noise ratio, multipath component energy, and antenna array phase difference in echo signal characteristics should all be constructed using the same 10-times uniform sampling method as echo intensity to ensure that the time-series dimensions of different echo signal characteristics are completely consistent, which facilitates the subsequent integration of multiple features and model input.

[0040] In step S130, the maximum and minimum normalization methods are used to transform the echo signal characteristics per second to the interval [0,1] to obtain the initial dataset.

[0041] It should be noted that this normalization operation needs to cover two types of data simultaneously: one is the time-series data of all echo signal features, and the other is the actual motion parameters of the target recorded by the positioning device. Performing normalization uniformly can not only eliminate the difference in the dimensions of different physical quantities, but also ensure that the numerical range of the input features and the label data matches during model training, and avoid a certain type of parameter from dominating the model's learning process due to its excessively large numerical magnitude.

[0042] The normalization formula is as follows: Among them, the The first normalized data Echo signal characteristic data at each sampling time point.

[0043] It should be noted that the echo signal characteristics at each sampling time point after normalized data processing can be expressed as: ,in, The first normalized data The target distance at each sampling time point; ,in, The first normalized data The target velocity at each sampling time point; ,in, The first normalized data The target azimuth angle at each sampling time point; ,in, The first normalized data Pulse arrival time at each sampling time point; ,in, The first normalized data Signal-to-noise ratio at each sampling time point; ,in, The first normalized data Multipath component energy at each sampling time point; ,in, The first normalized data The phase difference of the antenna array at each sampling time point; ,in, The first normalized data The echo intensity at each sampling time point.

[0044] This yields 300 sets of data for subsequent model training: ,in, The echo intensity in the training data; ,in, The pulse arrival time in the training data; ,in, The signal-to-noise ratio in the training data; ,in, Energy of the multipath components in the training data; ,in, The antenna array phase difference in the training data; ,in, The target distance in the training data; ,in, The target velocity in the training data; ,in, The target azimuth angle is in the training data.

[0045] In step S200, the initial dataset is extracted and concatenated based on the local mean decomposition algorithm to eliminate multipath and noise interference, thereby obtaining standard input feature data.

[0046] In one embodiment, such as Figure 3 As shown, step S200 may include the following sub-steps: In step S210, the initial dataset is used as the data to be decomposed to extract local extreme points, and the average value of adjacent local extreme points is calculated. The average values ​​of the local extreme points are connected by a straight line to obtain the local mean function.

[0047] It should be noted that extracting local extrema is the core prerequisite for the local mean decomposition algorithm to separate signal components. Calculating the average value of adjacent extrema and connecting them to form a local mean function is to fit the trend baseline of the original signal. This baseline can reflect the basic fluctuation trend of the signal, avoid the characteristics deviating from the true law due to noise interference when processing the original signal in the later stage, and provide a stable benchmark for separating the effective frequency modulation function.

[0048] The formula for calculating the average value of the local extreme points is as follows: in, It is the average value of the local extreme points. For the first in the initial dataset Local extreme points, For the first in the initial dataset Local extreme points.

[0049] It should be noted that the adjacent local extreme points selected in the formula should be selected sequentially according to the time sequence of the initial dataset, and should include both maximum and minimum points. This calculation method with full coverage of extreme points allows the local mean function to fully fit the fluctuation characteristics of the original data, ensuring the accuracy of subsequent signal decomposition.

[0050] In step S220, based on the local extreme points, the envelope function values ​​of adjacent extreme points are calculated, and the adjacent envelope function values ​​are connected by the moving average method to obtain the envelope estimation function.

[0051] It should be noted that the window size of the moving average method needs to match the sampling frequency of the initial dataset: if the window is too large, the envelope estimation will be too smooth and lose the details of the true amplitude fluctuation of the signal; if the window is too small, it will not be able to effectively suppress noise and will cause meaningless high-frequency jitter in the envelope estimation function.

[0052] The formula for calculating the average value of the envelope function is as follows: in, This represents the average value of the envelope function.

[0053] It should be noted that the envelope function value is calculated in absolute form because the envelope needs to characterize the fluctuation range of the signal amplitude. The envelope function value obtained by this formula can accurately quantify the degree of amplitude change between adjacent extreme points, providing a reliable amplitude reference for subsequent separation of the frequency modulation function.

[0054] In step S230, the local mean function is separated from the initial dataset, the frequency modulation function is calculated based on the envelope estimation function, iterating until the envelope estimation function of the frequency modulation function is always 1, and all the envelope estimation functions are multiplied to obtain the envelope signal, and the envelope signal is multiplied by the frequency modulation function to obtain the first component of the initial dataset.

[0055] It should be noted that the core termination condition for the LMD algorithm to extract pure frequency-modulated signals is that the envelope estimation function of the frequency modulation function is always 1. At this point, the frequency modulation function has eliminated amplitude fluctuation interference and only retains the instantaneous frequency characteristics of the signal, possessing clear physical meaning. Each envelope estimation function generated during the iteration process is a gradual correction to the local amplitude variation law of the signal. The final envelope signal obtained by multiplying these functions can accurately characterize the instantaneous amplitude variation trend of the first component and satisfies the characteristics of non-negativity and slow variation. The first component obtained by multiplying the envelope signal with the pure frequency modulation function is essentially the first product function decomposed by the LMD algorithm. Its core function is to separate the high-frequency and high-amplitude effective signal components in the initial data, initially filter out multipath interference and random noise, and lay the foundation for subsequent extraction of more effective components and generation of high-purity standard input features.

[0056] In step S240, the first component is used as the decomposed data to obtain the second component, the second component is used as the decomposed data to obtain the third component, the third component is used as the decomposed data to obtain the fourth component, and the four components are concatenated to obtain the standard input feature data.

[0057] It should be noted that the data obtained from the initial dataset... The components are separated, and the newly obtained data is used as new original data. Steps S210-S230 are repeated to obtain the next component. The components, since the characteristic information of the data is mainly concentrated in the first few... In this step, the number of decompositions is selected as 4, ultimately yielding 4 components from the initial data. The four components are then concatenated via a channel concatenation operation to obtain data with noise removed and containing new feature information. .

[0058] The standard input feature data is: in, For standard input feature data, The echo intensity after component splicing. The arrival time of the pulse after component splicing. The signal-to-noise ratio of the signal after component splicing. This refers to the energy of the multipath components after component splicing. This represents the phase difference of the antenna array after component splicing.

[0059] It should be noted that each set of timing data in the obtained echo intensity, pulse arrival time, signal signal-to-noise ratio, multipath component energy, and antenna array phase difference is used as the input of S210, and the signal decomposition and splicing steps of S210-S240 are repeated to obtain data with noise removed and containing new feature information, namely the standard input feature data.

[0060] In step S300, a motion parameter prediction model is constructed based on the echo signal characteristics. The MAE loss function is used as the prediction performance metric. The motion parameter prediction model is optimized using a genetic algorithm and trained using the standard input feature data.

[0061] It should be noted that the MAE loss function was chosen because the target motion parameters belong to a continuous regression task. MAE is more robust to outlier data and can more accurately reflect the mean absolute error between the predicted and the true values, which meets the accuracy evaluation requirements of motion parameter prediction. The genetic algorithm is used to optimize the model, which is to globally optimize the hyperparameters of modules such as the attention mechanism layer and convolutional layer, avoiding getting stuck in local optima by manual parameter tuning. During training, the standard input feature data needs to be divided into training set, validation set and test set in a 7:2:1 ratio. The training set is used to update the model parameters, the validation set monitors the overfitting state in real time, and the test set is used to evaluate the generalization performance of the model.

[0062] In one embodiment, such as Figure 4 As shown, step S300 may include the following sub-steps: In step S310, a motion parameter prediction model is constructed based on an attention mechanism, a convolutional neural network, and a gating unit; wherein the motion parameter prediction model includes: an attention mechanism layer, a convolutional layer, a gating unit layer, a Dropout layer, a regularization layer, and a Sigmoid activation function layer.

[0063] It should be noted that the various layers of the model need to form a collaborative logic with interconnected functions: the attention mechanism layer is responsible for assigning differentiated weights to different echo signal features, highlighting features that contribute highly to the prediction of motion parameters; the convolutional layer is used to extract local detail information from the features; the gating unit layer focuses on capturing the long temporal dependencies of features; the dropout layer reduces the risk of overfitting by randomly deactivating some neurons; and the regularization layer further constrains the parameter complexity.

[0064] The attention mechanism layer is used to calculate the optimal weight set for each echo signal feature in the prediction of target motion parameters, and obtain weighted input feature data. The convolutional layer is used to extract detailed features from the weighted input feature data; The gated unit layer is used to obtain the long temporal dependency relationship of the captured features based on the detailed features, and to obtain temporal feature data; The Dropout layer is used to deactivate some neurons in the model and randomly suppress some feature values ​​in the temporal features to prevent overfitting. The regularization layer is used to normalize the feature values ​​of time-series feature data that suppresses some feature values. The Sigmoid activation function layer is used to process the normalized temporal feature data through the Sigmoid activation function to obtain a set of predicted motion parameters.

[0065] It should be noted that, since five different echo signal features can affect the predicted values ​​of target motion parameters, an attention mechanism is used to calculate the optimal weight set for each echo signal feature in the prediction of target motion parameters. , Indicates echo intensity Pulse arrival time Signal-to-noise ratio Multipath component energy Antenna array phase difference The optimal weights corresponding to each echo signal feature are determined, and then the standard input feature data is used. With the optimal weight set Multiply to obtain weighted input feature data ,Right now ; Obtain weighted input feature data Then, this data is fed into the multi-layer convolutional layers of the target motion parameter prediction model designed in this invention to further extract weighted input feature data. Detail features ,Right now ,in, Represents the processing function of the convolutional layer; it extracts the detailed features output by the convolutional layer. After the gating unit layer, we get To prevent overfitting, The data is fed into the Dropout layer, and finally, a regularization layer is added to further adjust the output of the Dropout layer, resulting in a predicted set of target motion parameters. ,Right now , ,in, This represents the gated unit layer processing function. These represent the Dropout layer processing function and the regularization layer processing function, respectively.

[0066] Furthermore, during gating unit training, the gate weights are first initialized and reset. Update gate weights Output gate weights Then, each data point in each group of data, after being processed by a multi-layer convolutional neural network, is sequentially fed into the gating unit for training, continuously updating and resetting the gate weights. Update gate weights Output gate weights After training, the final reset gate weights are obtained. Update gate weights Output gate weights .

[0067] In step S320, multiple initial populations are randomly generated, and after training the multiple initial populations, the MAE value of each population is solved; wherein each initial population represents a combination scheme of the number of attention layers, convolutional layers and gating units in the model.

[0068] It should be noted that the random generation Initial population Each population represents a combination scheme of the number of attention layers Z, the number of convolutional layers U, and the number of gated units M in the model, i.e. ,in, Indicates the first Number of attention layers in each population and the number of convolutional layers and number of gating unit layers The specific number, The number of attention layers determined for each population and the number of convolutional layers and number of gating unit layers The training is performed using combinations of schemes, and the solution for each combination is obtained. value.

[0069] The MAE value is calculated using the MAE loss function, which is: in, The value of the loss function. For the actual target motion parameters, The target motion parameters predicted by the motion parameter prediction model. The number of predicted values. .

[0070] It should be noted that, in order to minimize the prediction error of the target motion parameters and improve the robustness and generalization ability of the prediction model, this invention proposes a target motion parameter prediction model optimization method based on genetic algorithm, in order to find the number of attention layers in the attention mechanism module corresponding to the best prediction effect of the model. and the number of convolutional layers and number of gating unit layers .

[0071] In step S330, the population with the lowest MAE value among the multiple initial populations is selected as the optimal population, and the optimal population is updated. When the maximum number of iterations is reached, the final optimal population with the lowest MAE value is obtained, thereby obtaining the motion parameter prediction model with the optimal combination of attention layer number, convolution layer number, and gating unit number.

[0072] It should be noted that, based on the number of attention layers, convolutional layers, and gating units in the determined optimal combination scheme, these are used as the number of attention layers, convolutional layers, and gating units in the target motion parameter prediction model, in order to obtain the network structure of the target motion parameter prediction model with the best prediction effect.

[0073] The method for updating the optimal population is as follows: in, , For the updated initial population, For the first An initial population, To update the speed weights, , To select an interval random numbers, For the optimal population, In the first Number of attention layers in each population The number of convolutional layers. This refers to the number of layers in the gate control unit.

[0074] It should be noted that the update speed weight z is set to 0.6 here, which is a result of comprehensively considering the efficiency and stability of population evolution: a z value that is too large will cause the population update amplitude to be too high, leading to deviation from the region where the optimal solution is located; a z value that is too small will slow down the population evolution rhythm and prolong the optimization time. A random number in the interval [0,1] is introduced. This is to incorporate randomness into population updates, preventing the population from converging excessively in a single direction, maintaining population diversity, and thus increasing the probability of finding the globally optimal hyperparameter combination. Furthermore, since the number of attention layers, convolutional layers, and gated units are all positive integers, this applies to the updated population... Each component needs to be rounded down to ensure that it meets the actual number of layers required by the model network structure.

[0075] In step S340, the optimized motion parameter prediction model is trained using the standard input feature data.

[0076] It should be noted that, based on the optimal network results, the model is trained using the standard feature dataset processed by LMD to obtain the final trained target motion parameter prediction model.

[0077] In step S400, the echo signal features corresponding to the target under test are decomposed by local mean to eliminate multipath and noise interference, and then input into the trained motion parameter prediction model to obtain the predicted motion parameters of the target under test.

[0078] It should be noted that the preprocessing procedure for the echo signal features of the target to be tested must be completely consistent with the preprocessing procedure for the initial dataset during the training phase. This is to avoid feature distribution shifts caused by differences in preprocessing methods, which could affect the prediction accuracy of the model. If the echo signal to be tested has missing data or outliers, the same interpolation or outlier removal rules as those used in the training set must be applied for preprocessing to ensure that the feature data quality input to the model matches the training data, thereby ensuring the reliability of the prediction results.

[0079] Furthermore, in this example embodiment, a practical deployment of the prediction model is also provided. This model is deployed into a UWB radar system, and the target motion parameters are predicted by using the real-time echo signal characteristics collected by the sensor. Based on the prediction data, the target can be accurately tracked or located.

[0080] A UWB radar sensor array is deployed in the monitoring area, with a sampling period set to 0.1 seconds, to obtain target echo signal characteristic data, including echo intensity. Pulse arrival time Signal-to-noise ratio Multipath component energy Antenna array phase difference .

[0081] The sampled data is fed into a normalization module to eliminate differences in the units of measurement between data points. The normalized data is then fed into a LMD-based data processing module to further extract features and eliminate noise. Finally, the LMD-processed data is fed into a target motion parameter prediction model trained on the dataset, resulting in a predicted set of target motion parameters. Based on the obtained target motion parameters, a prediction set is made. To achieve precise tracking or location of targets, providing decision support for intelligent systems.

[0082] Furthermore, this example embodiment also provides a UWB radar target motion prediction system based on a multi-convolutional neural network. (Reference) Figure 5 As shown, the system may include: The data acquisition module is used to acquire the echo signal characteristics and corresponding target motion parameters acquired in real time by the UWB radar, and to obtain an initial dataset through time-series data partitioning and normalization. The echo signal characteristics include echo intensity, pulse arrival time, signal-to-noise ratio, multipath component energy and antenna array phase difference, and the target motion parameters are range, velocity and azimuth angle. The decomposition and splicing module is used to extract and splice the initial dataset based on the local mean decomposition algorithm, eliminate multipath and noise interference, and obtain standard input feature data; The model building module is used to build a motion parameter prediction model based on the echo signal features, use the MAE loss function as the prediction performance metric, optimize the motion parameter prediction model using a genetic algorithm, and train it using the standard input feature data. The motion prediction module is used to input the echo signal features corresponding to the target under test into the trained motion parameter prediction model after eliminating multipath and noise interference through local mean decomposition, so as to obtain the predicted motion parameters of the target under test.

[0083] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the appended claims.

Claims

1. A UWB radar target motion prediction method based on multi-convolutional neural networks, characterized in that, include: The echo signal characteristics and corresponding target motion parameters acquired in real time by UWB radar are obtained, and the initial dataset is obtained through time series data partitioning and normalization. The echo signal characteristics include echo intensity, pulse arrival time, signal-to-noise ratio, multipath component energy, and antenna array phase difference; the target motion parameters are distance, velocity, and azimuth. The initial dataset is extracted and concatenated based on the local mean decomposition algorithm to eliminate multipath and noise interference and obtain standard input feature data; A motion parameter prediction model is constructed based on the echo signal characteristics. The MAE loss function is used as the prediction performance metric. The motion parameter prediction model is optimized using a genetic algorithm and trained using the standard input feature data. The echo signal features corresponding to the target under test are decomposed by local mean to eliminate multipath and noise interference, and then input into the trained motion parameter prediction model to obtain the predicted motion parameters of the target under test.

2. The UWB radar target motion prediction method based on multi-convolutional neural networks according to claim 1, characterized in that, The steps of acquiring the echo signal characteristics and corresponding target motion parameters collected in real time by UWB radar, and obtaining the initial dataset through time-series data partitioning and normalization, include: A UWB radar sensor array is deployed in the monitoring area to collect echo signal characteristic data related to the target's motion parameters, and the target's distance parameters, velocity, and azimuth are recorded by a positioning device. The echo intensity is obtained through the signal strength detection module of the UWB radar sensor, the pulse arrival time is obtained through the time measurement module of the UWB radar sensor, the signal-to-noise ratio is obtained through the noise analysis module of the UWB radar sensor, the multipath component energy is obtained through the multipath resolution module of the UWB radar sensor, and the antenna array phase difference is obtained through the phase difference detection module of the UWB radar sensor. For the echo signal characteristic data per second, the sampling period is set to 0.1 seconds, and 10 uniform samples are performed to obtain the echo signal characteristics at each sampling time point; The echo signal characteristics per second are transformed into the interval [0,1] using the maximum and minimum normalization methods to obtain the initial dataset.

3. The UWB radar target motion prediction method based on multi-convolutional neural networks according to claim 2, characterized in that, The characteristics of the echo signal at each sampling time point are as follows: Among them, the Let be the echo signal feature data at the i-th sampling time point, where i∈1,10; The normalization formula is: Among them, the The first normalized data Echo signal characteristic data at each sampling time point.

4. The UWB radar target motion prediction method based on multi-convolutional neural networks according to claim 1, characterized in that, The steps of extracting and concatenating the initial dataset based on the local mean decomposition algorithm to eliminate multipath interference and noise interference, and obtaining standard input feature data, include: The initial dataset is used as the data to be decomposed to extract local extreme points, and the average value of adjacent local extreme points is calculated. The average values ​​of the local extreme points are connected by a straight line to obtain the local mean function. Based on the local extreme points, the envelope function values ​​of adjacent extreme points are calculated, and the adjacent envelope function values ​​are connected by the moving average method to obtain the envelope estimation function. The local mean function is separated from the initial dataset, the frequency modulation function is calculated based on the envelope estimation function, iterating until the envelope estimation function of the frequency modulation function is always 1, and all the envelope estimation functions are multiplied to obtain the envelope signal. The envelope signal is then multiplied by the frequency modulation function to obtain the first component of the initial dataset. The first component is used as the decomposed data to obtain the second component, the second component is used as the decomposed data to obtain the third component, the third component is used as the decomposed data to obtain the fourth component, and the four components are concatenated to obtain the standard input feature data.

5. The UWB radar target motion prediction method based on multi-convolutional neural networks according to claim 4, characterized in that, The formula for calculating the average value of the local extreme points is: in, It is the average value of the local extreme points. For the first in the initial dataset Local extreme points, For the first in the initial dataset Local extreme points; The formula for calculating the average value of the envelope function is: in, The average value of the envelope function; The formula for calculating the envelope signal is: in, For envelope signal, Let n be the envelope estimation function; The standard input feature data is: in, For standard input feature data, The echo intensity after component splicing. The arrival time of the pulse after component splicing. The signal-to-noise ratio of the signal after component splicing. This refers to the energy of the multipath components after component splicing. This represents the phase difference of the antenna array after component splicing.

6. The UWB radar target motion prediction method based on multi-convolutional neural networks according to claim 1, characterized in that, The steps of constructing a motion parameter prediction model based on the echo signal features, using the MAE loss function as a metric for prediction performance, optimizing the motion parameter prediction model using a genetic algorithm, and training it with the standard input feature data include: A motion parameter prediction model is constructed based on an attention mechanism, a convolutional neural network, and a gating unit; wherein, the motion parameter prediction model includes: an attention mechanism layer, a convolutional layer, a gating unit layer, a Dropout layer, a regularization layer, and a Sigmoid activation function layer; Multiple initial populations are randomly generated, and after training the multiple initial populations, the MAE value of each population is solved; wherein, each initial population represents a combination scheme of the number of attention layers, convolutional layers and gating units in the model; The population with the lowest MAE value among the multiple initial populations is selected as the optimal population, and the optimal population is updated. When the maximum number of iterations is reached, the final optimal population with the lowest MAE value is obtained, thereby obtaining the motion parameter prediction model with the optimal combination of attention layer number, convolution layer number, and gate unit number. The optimized motion parameter prediction model is trained using the standard input feature data.

7. The UWB radar target motion prediction method based on multiple convolutional neural networks according to claim 6, characterized in that, The attention mechanism layer is used to calculate the optimal weight set of each echo signal feature in the prediction of target motion parameters, and obtain weighted input feature data. The convolutional layer is used to extract detailed features from the weighted input feature data; The gated unit layer is used to obtain the long temporal dependency relationship of the captured features based on the detailed features, and to obtain temporal feature data; The Dropout layer is used to deactivate some neurons in the model and randomly suppress some feature values ​​in the temporal features to prevent overfitting. The regularization layer is used to normalize the feature values ​​of time-series feature data that suppresses some feature values. The Sigmoid activation function layer is used to process the normalized temporal feature data through the Sigmoid activation function to obtain a set of predicted motion parameters.

8. The UWB radar target motion prediction method based on multiple convolutional neural networks according to claim 6, characterized in that, The MAE value is calculated using the MAE loss function, which is: in, The value of the loss function. For the actual target motion parameters, The target motion parameters predicted by the motion parameter prediction model. The number of predicted values. .

9. The UWB radar target motion prediction method based on multiple convolutional neural networks according to claim 6, characterized in that, The optimal population update method is as follows: in, , For the updated initial population, For the first An initial population, To update the speed weights, , To select an interval random numbers, For the optimal population, In the first Number of attention layers in each population The number of convolutional layers. This refers to the number of layers in the gate control unit.

10. A UWB radar target motion prediction system based on multi-convolutional neural networks, characterized in that, The system is used to perform the method as described in any one of claims 1 to 9, the system comprising: The data acquisition module is used to acquire the echo signal characteristics and corresponding target motion parameters acquired in real time by the UWB radar, and to obtain an initial dataset through time-series data partitioning and normalization. The echo signal characteristics include echo intensity, pulse arrival time, signal-to-noise ratio, multipath component energy and antenna array phase difference, and the target motion parameters are range, velocity and azimuth angle. The decomposition and splicing module is used to extract and splice the initial dataset based on the local mean decomposition algorithm, eliminate multipath and noise interference, and obtain standard input feature data; The model building module is used to build a motion parameter prediction model based on the echo signal features, use the MAE loss function as the prediction performance metric, optimize the motion parameter prediction model using a genetic algorithm, and train it using the standard input feature data. The motion prediction module is used to input the echo signal features corresponding to the target under test into the trained motion parameter prediction model after eliminating multipath and noise interference through local mean decomposition, so as to obtain the predicted motion parameters of the target under test.

Images