Motion error compensation and signal coherent accumulation method based on deep network

Abstract

The invention discloses a motion error compensation and signal coherent accumulation method based on a deep network, which is applied to the field of radar signal processing, and aims at solving the problems of motion error compensation and echo signal accumulation of echo signals, and solving the problem that complex motion error and echo signal coherent accumulation compensated through the deep network is not considered in the prior art. Motion error compensation and signal accumulation of echoes containing airborne platform motion errors are difficult to deal with; according to the method, motion phase components in echoes are eliminated through a PWGRT algorithm, when search parameters are matched with a target, motion error phase components containing noise are output, then a 1D-Inception-ResNet-V2 structure capable of adapting to phase sequence input is designed, phase change characteristics between pulses are mined, and phase information of motion errors is further extracted. And finally, phase information of the motion error is extracted by using the deep network, and compensation and signal accumulation of the motion error are realized.

Description

Technical Field

[0001] This invention belongs to the field of radar signal processing, and specifically relates to a motion error compensation and signal coherent accumulation technique. Background Technology

[0002] With the continuous emergence of targets possessing high speed, high maneuverability, long range, and strong stealth characteristics, such as stealth fighters, the detection capabilities of modern radar are facing unprecedented challenges. The core challenge in addressing these challenges lies in how to achieve efficient and reliable detection of high-speed maneuvering targets under low SNR conditions. Airborne bistatic platforms have advantages over ground-based platforms in terms of high mobility and strong anti-jamming capabilities. However, during actual flight, airborne bistatic radar platforms often encounter unpredictable interferences such as airflow disturbances, mechanical control deviations, and platform vibrations. These factors directly cause the airborne platform's trajectory to deviate from the ideal state, thereby disrupting the signal consistency required for coherent accumulation and ultimately causing a severe decline in the performance of traditional coherent accumulation and detection methods.

[0003] To achieve coherent accumulation of high-speed maneuvering targets for airborne bistatic radar in the presence of motion errors, it is necessary not only to correct and compensate for the range migration (RM) and Doppler frequency migration (DFM) caused by the time-varying relative motion between the target and the radar, but also to consider the amplitude-phase shift caused by the motion errors of the airborne platform and implement corresponding compensation measures. In existing research, Liu Hui introduced the exterior orientation element error model from optical photogrammetry to analyze the impact of motion errors generated during the actual motion of the airborne platform on the phase center shift of the radar antenna. Zhang T et al. constructed a motion error equation based on Taylor approximation, but did not provide an analytical expression for the motion error. Li L et al. constructed a three-dimensional time-varying trajectory deviation equation for the attitude angle under a linear array MIMO SAR platform through coordinate system transformation, but lacked consideration and analysis of the coupling effects of motion errors and the high-speed maneuvering characteristics of the platform and the target. Yang Fan et al. studied an improved generalized Radon Fourier transform algorithm (Improved GRFT, IGRFT), which achieved high-precision processing of motion error compensation and signal accumulation detection for airborne bistatic radar. However, in order to reduce the computational cost, this accumulation method reduced the dimensionality of the complex motion error model, resulting in some distortion in the model and a loss of accumulation performance.

[0004] In the field of radar target detection, some research has been conducted on estimating target motion parameters using deep neural networks. C. Wang et al. used a fully convolutional neural network (FCN) to predict the target trajectory and then used delinear frequency modulation (DMFM) to accumulate the target energy on the predicted trajectory. However, this method struggles to handle targets with high-order motion. J. Zheng et al. introduced a data model-driven subspace generalized Radon–Fourier Transform (SGRFT) algorithm, which can use a deep network to predict the subspace of target velocity and acceleration, and accumulate energy within the predicted subspace. However, both of these methods struggle to handle echo signals containing accumulated platform motion errors.

[0005] In summary, regarding the issues of motion error compensation and echo signal accumulation in echo signals, the above methods do not consider compensating for complex motion errors and coherent accumulation of echo signals through deep networks, making it difficult to address motion error compensation and signal accumulation in echoes with motion errors from airborne platforms. Summary of the Invention

[0006] To address the aforementioned issues, this invention proposes a motion error compensation and signal coherent accumulation method based on deep networks, which achieves effective coherent accumulation for high-speed moving targets with motion errors.

[0007] The technical solution adopted in this invention is: a method for motion error compensation and signal coherent accumulation based on deep networks, comprising:

[0008] S1. Construct a motion error model for the airborne platform;

[0009] S2. Construct an echo model based on the airborne platform motion error model from step S1;

[0010] S3. Input the echo signal represented by the echo model constructed in step S2 into the PWGRT algorithm, and output a phase sequence containing only motion error;

[0011] S4. Construct the 1D-Inception-ResNet-V2 structure;

[0012] S5. Construct training set data based on the phase sequence containing only motion error output by the PWGRT algorithm;

[0013] S6. Design the loss function; and train the 1D-Inception-ResNet-V2 built in step S4 based on the training dataset constructed in step S5. The output of 1D-Inception-ResNet-V2 is the motion error phase information.

[0014] S7. Input the echo signal to be processed, represented by the echo model constructed in step S2, into the PWGRT algorithm, output a phase sequence containing only motion error, and then input the phase sequence into the trained 1D-Inception-ResNet-V2 to obtain the motion error phase information corresponding to the echo signal to be processed.

[0015] The output of the PWGRT algorithm corresponding to the echo signal to be processed is compensated based on the motion error phase information of the echo signal to be processed.

[0016] Finally, the process is accumulated based on the compensated echo signal to be processed.

[0017] The beneficial effects of this invention are as follows: This invention eliminates the motion phase component in the echo using the PWGRT algorithm. When the search parameters match the target, it outputs the noisy motion error phase component. Then, a 1D-Inception-ResNet-V2 structure capable of adapting to phase sequence input is designed to mine the phase change characteristics between pulses and further extract the phase information of the motion error. Finally, the phase information of the motion error extracted by the deep network is used to achieve motion error compensation and signal accumulation. This invention can significantly improve the target echo signal-to-noise ratio, thereby enhancing the radar's detection capability. Attached Figure Description

[0018] Figure 1 This is a flowchart for the invention process.

[0019] Figure 2 The pulse compression results and the target trajectory extracted by the PWGRT algorithm;

[0020] (a) is the pulse compression result, and (b) is the target trajectory extracted by the PWGRT algorithm.

[0021] Figure 3 It uses a 1D-Inception-ResNet-V2 architecture.

[0022] Figure 4 The motion error phase sequence;

[0023] Wherein, (a) is the motion error phase sequence extracted by the network, (b) is the actual motion error phase sequence, and (c) is the network extraction error.

[0024] Figure 5 This is the result of accumulating the algorithm proposed in this invention.

[0025] Figure 6 Accumulate results for traditional methods;

[0026] Among them, (a) is the accumulation result of the GRFT algorithm, and (b) is the accumulation result of the IGRFT algorithm.

[0027] Figure 7 This is for detecting performance curves. Detailed Implementation

[0028] To facilitate understanding of the technical content of this invention by those skilled in the art, the following description, in conjunction with the accompanying drawings, further illustrates the invention.

[0029] This invention is verified using Matlab simulation experiments. The correctness and effectiveness of this invention are verified in the scientific computing software Matlab R2022b. The technical solution of this invention is further described below with reference to the accompanying drawings.

[0030] like Figure 1 As shown, the present invention proposes a motion error compensation and signal coherent accumulation method based on deep networks, which includes the following steps:

[0031] S1. Modeling of Airborne Platform Motion Error. To describe the motion error, this invention first needs to model the motion error of the airborne platform and analyze its variation law, laying the foundation for subsequent establishment of the echo signal model and analysis of echo characteristics. The specific implementation method is as follows:

[0032] In a Cartesian three-dimensional coordinate system, the airborne platform Airborne platform The initial positions of the target and the target are respectively , and The initial velocities are respectively , and The initial accelerations are respectively , and ,in, This indicates the transpose operation. They represent airborne platforms respectively. Coordinates on the x, y, z axes They represent airborne platforms respectively. Coordinates on the x, y, z axes These represent the target's coordinates on the x, y, and z axes, respectively. For airborne platform initial velocity, for Projection on the y-axis For airborne platform initial velocity, for Projection on the y-axis Let the initial velocity of the target be... They are respectively Projection onto the x, y, z axes For airborne platform initial acceleration, for Projection on the y-axis For airborne platform initial acceleration, for Projection on the y-axis The initial acceleration of the target. These are the projections of the target acceleration onto the x, y, and z axes, respectively.

[0033] The motion error of an airborne platform can be modeled as rotational error and translational error. This invention uses... , Indicates airborne platform The yaw and pitch angles in the rotation error, , Indicates airborne platform The yaw and pitch angles in the rotation error can be used to determine the airborne platform. Yaw angle Pitch angle They are respectively represented as

[0034] (1)

[0035] (2)

[0036] in, Indicates slow time. , The number of pulses. and They represent and The range, and They represent , frequency, and They represent and The first appearance.

[0037] Airborne platform Translational error along axis, shaft and The components of the axis are represented as Translational error of airborne platform l along axis, shaft and The components of the axis are represented as Taking airborne platform k as an example, the translational error along... axis, shaft and The components of the axis can be expressed as

[0038] (3)

[0039] in, , and They represent the first individual airborne platforms along axis, axis, The magnitude of the translational error component of the shaft. , and They represent the first individual airborne platforms along axis, axis, The frequency of the translational error components of the shaft. , and They represent the first individual airborne platforms along axis, axis, The initial phase of the translational error of the shaft.

[0040] Airborne platform With airborne platform The variation law of motion error between them over time is as follows

[0041] (4)

[0042] in,

[0043] (5)

[0044] (6)

[0045] in, The relative positions of the centers of the transmitting and receiving antennas with respect to the center of mass of the airborne platform. and They represent airborne platforms respectively. Equivalent initial radial distance to the target, airborne platform The equivalent initial radial distance to the target. and They represent airborne platforms respectively. Equivalent initial radial velocity to the target, airborne platform The equivalent initial radial velocity to the target. and Airborne platforms Equivalent initial radial acceleration to the target, airborne platform The equivalent initial radial acceleration to the target.

[0046] The nodes mentioned in this step are specifically explained as follows: an airborne bistatic radar includes two nodes that share both transmitting and receiving capabilities. Shared transmission and receiving means that it can both transmit and receive signals.

[0047] S2. Echo Characteristic Analysis. To achieve efficient accumulation of echo signals under motion errors, this invention requires establishing an echo model and analyzing the phase and envelope characteristics of the echo, laying the foundation for subsequent data-model dual-driven accumulation of echo signals containing motion errors. The specific implementation method is as follows:

[0048] Considering the radar's transmitting node is The receiving node is The radar transmits a linear frequency modulated signal, i.e.

[0049] (7)

[0050] in

[0051] (8)

[0052] To save time, The pulse width. For frequency modulation slope, This is the initial carrier frequency for the radar.

[0053] Without considering noise interference, the echo signal with motion error is:

[0054] (9)

[0055] in, The amplitude of the echo signal. It is the speed of light. The distance history model, which takes into account the motion error of the airborne platform, is expressed as follows:

[0056] (10)

[0057] in, , Radar transmission node Reaching the target and then reaching the receiving node The initial distance, Radar transmission node Reaching the target and then reaching the receiving node The equivalent two-way speed, Radar transmission node Reaching the target and then reaching the receiving node The equivalent two-way speed.

[0058] Next, by performing down-conversion and pulse compression processing on equation (9), the pulse compression echo signal can be expressed as follows:

[0059] (11)

[0060] in, The amplitude of the pulse pressure signal. For bandwidth, For wavelength. When At that time, the echo signal showed distance movement.

[0061] As can be seen from formula (11), the inter-pulse phase change of the radar echo signal includes the motion phase component and the motion error phase component, which are respectively composed of the relative motion terms of the target and the airborne platform. and the motion error term of the airborne platform Brought. Among them. It can be modeled as a simple second-order linear motion model. Common inter-pulse accumulation algorithms such as the Generalized Radon-Fourier Transform (GRFT) and KT-MFP can compensate for this phase change well, while the motion error term... The model is quite complex and requires the use of deep network methods for compensation.

[0062] The radar parameters used in this example are set as follows: initial carrier frequency signal bandwidth Pulse repetition period Sampling rate Pulse width Pulse repetition time pulse number The training set pulse-compressed echo signal-to-noise ratio is -8dB. The target parameters are set as follows: the initial position of the airborne launch platform is (794.8, 596.1, 5960.7) m, the velocity is (0, 25, 0) m / s, and the acceleration is (0, 3.5, 0) m / s². 2 The initial position of the airborne receiving platform is (998.8, 799, 5992.7) m, its velocity is (0, 30, 0) m / s, and its acceleration is (0, 3, 0) m / s². 2The target's initial position is (359.8, 6996.4, 599.7) m, its velocity is (-13, 77.95, -20) m / s, and its acceleration is (0, 8, 6) m / s². 2 .

[0063] S3. The Phase Weighted Generalized Radon Transform (PWGRT) algorithm is used to extract the echo target trajectory and compensate for the phase caused by motion. To extract the echo target trajectory and compensate for the phase introduced by motion, this invention utilizes the PWGRT algorithm to perform phase compensation on the echo signal, ensuring that the echo phase only contains the motion error phase and noise phase, thus laying the foundation for subsequent echo signal accumulation. The specific implementation method is as follows:

[0064] The PWGRT algorithm takes the echo signal shown in formula (11) as input, compensates for the echo phase change caused by the relative motion between the radar and the target between pulses, and outputs a phase sequence containing only motion error. This method only extracts the target trajectory and constructs the phase compensation sequence. It is assumed that the search parameter space of the PWGRT algorithm for the target's distance, velocity, and acceleration relative to the radar are respectively... , and The formula is as follows:

[0065] (12)

[0066] in, , and These are the distance search parameters, velocity search parameters, and acceleration search parameters, respectively. The phase weighting function is expressed as follows:

[0067] (13)

[0068] The phase sequence in formula (12) is expressed as follows:

[0069] (14)

[0070] When search parameters , and Matching with actual motion parameters, i.e. , , When, formula (14) can be expressed as

[0071] (15)

[0072] As can be seen, the phase change between pulses caused by the relative motion term can be basically eliminated, realizing the decoupling of the motion term, while the phase information of the motion error can be preserved, laying the foundation for the subsequent network to extract the phase information of the motion error and compensate for it.

[0073] This invention sets the PWGRT algorithm's distance search interval to [138.01, 139.21] km, divided into 101 search parameters; the velocity search interval to [-0.21, 199.79] m / s, divided into 101 search parameters; and the acceleration search interval to [0, 10] m / s. 2 It is divided into 21 search parameters.

[0074] The search distance was 138.61m, the search speed was 99.79m / s, and the search acceleration was 5m / s². 2 At that time, the search parameters match the target motion parameters. Figure 2 The result of pulse compression of the echo signal and the target trajectory extracted by the PWGRT algorithm.

[0075] S4. Construct a 1D-Inception-ResNet-V2 structure. To adapt to phase sequence input, this invention constructs a 1D-Inception-ResNet-V2 structure to mine inter-pulse phase change features and further extract phase information of motion errors. The specific implementation method is as follows:

[0076] After the PWGRT algorithm achieves compensation and decoupling of relative motion terms, this invention designs a 1D-Inception-ResNet-V2 structure, which can adapt to the input of phase sequences, mine the phase change features between pulses, and further extract the phase information of motion errors. The network structure is as follows: Figure 3 As shown.

[0077] 1D-Inception-ResNet-V2 takes the phase sequence processed by the noisy PWGRT algorithm as input and extracts the motion error phase sequence as output, which facilitates subsequent motion error phase compensation and accumulation.

[0078] The core idea of ​​1D-Inception-ResNet-V2 is to combine the Inception module with residual connections, leveraging the multi-scale feature extraction capabilities of the Inception module and the gradient propagation advantages of residual connections to improve the network's training efficiency and performance. 1D-Inception-ResNet-V2 can be divided into the following parts: the 1D-Stem module, the 1D-Inception-ResNet module, the Reduction module, and the regression layer. The following sections will introduce each of these modules in detail.

[0079] The 1D-Stem module is the starting part of the network. This invention replaces the two-dimensional convolutional kernels in the traditional Stem module with one-dimensional convolutional kernels, thus constructing a 1D-Stem module to accommodate sequential input. The input to the 1D-Stem module in this invention is a noisy sequence randomly generated in S3 based on formula (14). The module can quickly extract preliminary features of inter-pulse phase information under different time window lengths, laying the foundation for the network to learn short-term and long-term phase patterns.

[0080] This invention replaces the two-dimensional convolutional kernels in the traditional Inception-ResNet module with one-dimensional convolutional kernels to construct the 1D-Inception-ResNet module. This module is the core part of the network and, similar to the traditional structure, is divided into three types: 1D-Inception-ResNet-A, 1D-Inception-ResNet-B, and 1D-Inception-ResNet-C. The difference between these three types lies in their input and output sizes.

[0081] The input to the 1D-Inception-ResNet module is the result of downsampling and preliminary feature extraction performed by the 1D-Stem module. This module further extracts the preliminary results through different operations, enabling it to learn the complex phase patterns extensively and giving the network better learning capabilities.

[0082] The reduction modules are of two types: Reduction-A and Reduction-B. Reduction-A and Reduction-B are inserted between stacked 1D-Inception-ResNet modules. Reduction-A follows 1D-Inception-ResNet-A and performs more aggressive downsampling, rapidly compressing the results extracted by 1D-Inception-ResNet-A to reduce the size of the extracted results and accelerate network convergence. Reduction-B follows 1D-Inception-ResNet-B and performs finer downsampling to preserve phase information as much as possible. The difference between Reduction-A and Reduction-B lies in their input and output sizes.

[0083] The regression layer is the final output layer of the network, which outputs the phase information of the motion error extracted from the phase sequence for subsequent compensation.

[0084] S5. Generate training set data. To facilitate the network's learning of the motion error phase change pattern during iterative training, this invention requires setting up a network training set for network training. The specific implementation method is as follows:

[0085] After designing the structure of the deep network, it is also necessary to build the corresponding training set data so that the network can learn the phase pattern of motion error from it.

[0086] This invention randomly generates several sets of motion parameters and motion error parameters involved in formula (4), and produces motion error terms. Then, based on formula (15), the output features of the training set containing noise are generated. .

[0087] (16)

[0088] in, , It is a complex Gaussian distribution with a mean of 0 and a variance of 1.

[0089] The generated data is divided into training and validation sets in an 8:2 ratio.

[0090] S6. Design the loss function. To evaluate the effectiveness of extracting motion error phase information during network training, this invention requires the design of a loss function. The specific implementation method is as follows:

[0091] After obtaining the training set, it is also necessary to establish a loss function to evaluate the learning effect of the network. In order to adapt to the needs of subsequent inter-pulse accumulation in 1D-Inception-ResNet-V2, this invention adds an accumulation efficiency weight on the basis of the Huber loss function, so that the network can reduce the error between input features and output features in the direction of improving accumulation efficiency during the iteration process.

[0092] The Huber loss function can be expressed as follows:

[0093] (17)

[0094] in, Labels for the network's output features. These are the hyperparameters of the Huber function.

[0095] To improve the optimization capability of accumulation efficiency during network iteration, this invention incorporates an accumulation efficiency weight. First, when the relative motion parameters find the correct value, the inter-pulse phase change caused by motion is compensated. At this point, the echo signal accumulation process can be represented as...

[0096] (18)

[0097] This invention uses the ratio of the actual accumulated peak value to the ideal peak value as the accumulation efficiency weight.

[0098] (19)

[0099] The loss function based on accumulation efficiency assessment can be expressed as:

[0100] (20)

[0101] The Adam optimizer is set to the 1D-Inception-ResNet-V2 optimizer, with an initial learning rate of 0.001, a batch size of 1000, and a maximum number of iterations of 1000.

[0102] After completing the loss function design and training option settings, 1D-Inception-ResNet-V2 is trained using training and validation set data. The loss function value of 1D-Inception-ResNet-V2 is calculated by forward propagation, and then the parameters of the network model are updated by backpropagation algorithm. This process is repeated until the predetermined maximum number of iterations is reached. At the same time, the Adam optimizer is used to reduce the loss function value during the training process, so that the predicted phase change and the actual phase change gradually approach each other, realizing the function of motion error phase sequence extraction.

[0103] After network processing, the phase change caused by motion error was successfully extracted, and the network-predicted phase result basically matches the actual motion error phase, such as... Figure 4 As shown in the figure. Specifically, at the 20th pulse, the difference between the motion error phase information extracted by the network and the actual motion error phase information is 0.436°. This shows that the network can effectively extract the phase information of the motion error, laying the foundation for subsequent cancellation of the phase changes caused by the motion error.

[0104] S7. Motion Error Compensation and Echo Accumulation. Finally, the phase sequence is input into the trained network to compensate for motion errors, and then accumulated based on the network output. The specific implementation method is as follows:

[0105] After the deep network parameters are iterated, the network can extract the corresponding motion error sequence from the phase sequence output by PWGRT for each parameter. Furthermore, it can achieve motion error compensation and signal accumulation. The signal accumulation process can be represented as follows:

[0106] (twenty one)

[0107] get Then, constant false alarm rate (CFAR) detection processing is performed to finally obtain the target detection results including the motion error of the airborne platform.

[0108] For each search parameter The network extracts the motion error phase information corresponding to each search parameter, and further compensates and eliminates motion errors on the PWGRT results to achieve the superposition of inter-pulse signals. Figure 5 The echo accumulation results in the range-velocity domain are shown, indicating that the compensated signal has the range search parameters as follows: At a distance of 138.61km, the speed search parameters are: Since the phase changes between pulses and the phase changes of motion error have been compensated, the target echo energy can be accumulated well.

[0109] The accumulation results of the GRFT and IGRFT algorithms in the range-velocity domain are as follows: Figure 6 As shown, due to the phase change caused by motion error not being effectively compensated, energy cannot be accumulated between pulses.

[0110] Figure 7 As can be seen from the performance curves, the algorithm proposed in this invention, compared to the GRFT algorithm, benefits from the compensation for motion errors, allowing for better accumulation of the echo signal and achieving a 3.88dB improvement at a detection probability of 80%. Since the IGRFT algorithm suffers from some model distortion during the dimensionality reduction process for motion errors, the method proposed in this invention achieves a 2.42dB improvement at a detection probability of 80% compared to the IGRFT method.

[0111] Compared to the high-speed maneuvering target motion parameter estimation and detection method based on Full Convolutional Network (FCN), although this method is also based on dual data model driving, it cannot compensate for the phase change caused by motion error. The method proposed in this invention has a 5.09dB improvement at a detection probability of 80%.

[0112] Those skilled in the art will recognize that the embodiments described herein are for the purpose of helping to understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the scope of the claims of the invention.

Claims

1. A method for motion error compensation and signal coherent accumulation based on deep networks, characterized in that, include: S1. Construct a motion error model for the airborne platform; S2. Construct an echo model based on the airborne platform motion error model from step S1; S3. Input the echo signal represented by the echo model constructed in step S2 into the PWGRT algorithm, and output a phase sequence containing only motion error; S4. Construct the 1D-Inception-ResNet-V2 structure; S5. Construct training set data based on the phase sequence containing only motion error output by the PWGRT algorithm; S6. Design the loss function; and train the 1D-Inception-ResNet-V2 built in step S4 based on the training dataset constructed in step S5. The output of 1D-Inception-ResNet-V2 is the motion error phase information. S7. Input the echo signal to be processed, represented by the echo model constructed in step S2, into the PWGRT algorithm, output a phase sequence containing only motion error, and then input the phase sequence into the trained 1D-Inception-ResNet-V2 to obtain the motion error phase information corresponding to the echo signal to be processed. The output of the PWGRT algorithm corresponding to the echo signal to be processed is compensated based on the motion error phase information of the echo signal to be processed. Finally, the process is accumulated based on the compensated echo signal to be processed.

2. The method for motion error compensation and signal coherent accumulation based on deep networks according to claim 1, characterized in that, The motion error model of the airborne platform mentioned in step S1 is expressed as follows: ； in, For radar nodes With radar nodes Motion error between , ; Indicates slow time. , The number of pulses; Representing radar nodes Coordinates on the x, y, z axes; Representing radar nodes Coordinates on the x, y, z axes; These represent the target's coordinates on the x, y, and z axes, respectively. These represent the target's velocities along the x, y, and z axes, respectively. These represent the target's accelerations along the x, y, and z axes, respectively. They represent airborne platforms respectively. Translational error along axis, shaft and The components of the axis; These represent the translational error along the path of the airborne platform l. axis, shaft and The components of the axis; The relative positions of the center of the transmitting antenna and the center of the receiving antenna with respect to the center of mass of the airborne platform; and They represent airborne platforms respectively. Equivalent initial radial distance to the target, airborne platform The equivalent initial radial distance to the target; and They represent airborne platforms respectively. Equivalent initial radial velocity to the target, airborne platform The equivalent initial radial velocity to the target; and Airborne platforms Equivalent initial radial acceleration to the target, airborne platform The equivalent initial radial acceleration to the target.

3. The method for motion error compensation and signal coherent accumulation based on deep networks according to claim 2, characterized in that, Step S2 includes the following sub-steps: The radar transmits a linear frequency modulated signal, i.e. ； in, , To save time, The pulse width. For frequency modulation slope, This is the initial carrier frequency for the radar. Without considering noise interference, the echo signal with motion error is: ； in, The amplitude of the echo signal. At the speed of light, The distance history model that takes into account the motion error of the airborne platform is expressed as follows: ； in, , Radar transmission node Reach the target and then reach the receiving node The initial distance, Radar transmission node Reach the target and then reach the receiving node The equivalent two-way speed, Radar transmission node Reach the target and then reach the receiving node The equivalent two-way speed. Next, for... Perform down-conversion and pulse compression processing to represent the pulse compression echo signal as... ； in, The amplitude of the pulse pressure signal. For bandwidth, λ is the wavelength.

4. The method for motion error compensation and signal coherent accumulation based on deep networks according to claim 3, characterized in that, The PWGRT algorithm in step S3 is as follows: Assume the PWGRT algorithm has the following search parameter spaces for the target's range, velocity, and acceleration relative to the radar: , and The formula is as follows: ； in, , and These are the distance search parameters, velocity search parameters, and acceleration search parameters, respectively. , and , This is the phase weighting function.

5. The method for motion error compensation and signal coherent accumulation based on deep networks according to claim 4, characterized in that, The 1D-Inception-ResNet-V2 in step S4 includes, in sequence: a 1D-Stem module, a first 1D-Inception-ResNet module, a first Reduction module, a second 1D-Inception-ResNet module, a second Reduction module, a third 1D-Inception-ResNet module, a pooling layer, and a regression layer.

Images