A Dynamic Adjustment Method for Radar Training Parameters Based on Adaptive Algorithms

Abstract

This invention provides a method for dynamically adjusting radar training parameters based on an adaptive algorithm, relating to the field of radar technology. The method includes: acquiring waveform parameters to construct feature vectors; calculating a feature correlation matrix and a parameter adjustment direction matrix; mapping to the parameter feature space to generate a priority adjustment sequence; constructing a set of adjustment actions, executing the adjustments and collecting response characteristics; updating the waveform parameters and triggering a new round of training. This invention enables precise dynamic adjustment of radar training parameters, improving training efficiency and system adaptability, and enhancing radar performance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of radar, and in particular to a radar training parameter dynamic adjustment method based on an adaptive algorithm. BACKGROUND

[0002] As the core sensor system of different task platforms, the performance of radar directly affects the task completion ability and effect. In a complex and changeable task environment, radar needs to optimize its working parameters through continuous training to adapt to different task scenarios and target features. The traditional radar training process usually relies on a preset fixed parameter configuration scheme or manual adjustment of related parameters by experienced operators.

[0003] With the increasing complexity of the task environment and the higher requirement for the rapid adaptation ability of the radar system, the optimization and adjustment of the radar training parameters face many technical challenges. The existing radar training parameter adjustment technology still has some defects and deficiencies. The traditional parameter adjustment method lacks adaptability and cannot dynamically adjust the parameter configuration according to the state features of different training stages, resulting in low parameter optimization efficiency in the training process and the inability to effectively cope with complex and changeable target scenarios and electromagnetic environments. The parameter optimization method in the prior art often processes each waveform parameter independently, ignoring the mutual coupling relationship and time sequence correlation between parameters, which makes it difficult to find a globally optimal parameter combination and easily causes parameter conflicts or mutual constraints in the adjustment process. The current radar training system lacks an intelligent parameter priority evaluation mechanism, making it difficult to determine the key parameters with the greatest impact for optimization within a limited training time, resulting in unreasonable allocation of training resources and affecting the overall performance improvement effect of the radar system. SUMMARY

[0004] The embodiment of the present application provides a radar training parameter dynamic adjustment method based on an adaptive algorithm, which can solve the problems in the prior art.

[0005] In a first aspect, the embodiment of the present application provides a radar training parameter dynamic adjustment method based on an adaptive algorithm, comprising:

[0006] Obtaining the waveform parameters of the radar in the training process, and constructing a training state feature vector based on the waveform parameters;

[0007] Calculating the feature correlation matrix of the training state feature vector to obtain a connection strength initial matrix, generating a modulation instruction according to the time sequence correlation analysis of the feature nodes, and performing weighted processing to obtain a feature modulation representation matrix, and calculating a parameter adjustment direction matrix according to the node state change;

[0008] The parameter adjustment direction matrix is ​​mapped to a multi-dimensional parameter feature space, and multiple subspace units are obtained through adaptive grid partitioning. Based on the parameter distribution characteristics, a priority adjustment sequence of waveform parameters is generated.

[0009] Based on the priority adjustment sequence, a set of adjustment actions for waveform parameters is constructed, the adjustment range of waveform parameters is divided into multiple adjustment levels, the adjustment actions are executed and the response characteristic curve is collected, the adjustment strategy selection matrix is ​​updated based on the response characteristic curve, and the next round of adjustment scheme is generated.

[0010] The waveform parameters are updated according to the adjustment scheme, the updated waveform parameters are written into the radar control unit and a new round of training tasks is triggered, and the data records of this round of parameter adjustment process are saved to the training database.

[0011] In one optional embodiment, the initial connection strength matrix is ​​obtained by calculating the feature correlation matrix of the training state feature vectors; modulation instructions are generated based on the temporal correlation analysis of feature nodes and weighted to obtain the feature modulation representation matrix; and the parameter adjustment direction matrix is ​​calculated based on the node state changes, including:

[0012] Calculate the correlation coefficients between features of each dimension in the training state feature vector, construct a feature correlation matrix, and calculate the initial connection strength matrix based on the numerical values ​​of each element in the feature correlation matrix.

[0013] Temporal correlation analysis is performed on each feature node in the initial connection strength matrix to extract the state information of adjacent feature nodes at different times, calculate the joint activation probability between nodes, compare the joint activation probability with a preset adaptive control threshold, and generate a strengthening modulation command or a suppressing modulation command.

[0014] The modulation direction of information transmission is determined according to the enhancement modulation command and the suppression modulation command. The adaptive modulation intensity factor is calculated based on the real-time activity of each feature node. The adaptive modulation intensity factor is input into the recursive nonlinear mapping unit to iteratively calculate the modulation coefficient.

[0015] The training state feature vector is weighted according to the modulation coefficient and input into a neural network with memory function. Through the selective modulation of the neural network, the feature modulation representation matrix is ​​calculated.

[0016] The state change relationship between the preceding and succeeding feature nodes in the feature modulation representation matrix is ​​calculated. The state transition law is extracted to obtain the weight update direction. The initial matrix of connection strength is adjusted, and the parameter adjustment direction matrix is ​​calculated.

[0017] In one optional embodiment, the information transmission modulation direction is determined according to the enhancement modulation command and the suppression modulation command. An adaptive modulation intensity factor is calculated based on the real-time activity of each feature node. The adaptive modulation intensity factor is then input into a recursive nonlinear mapping unit to iteratively calculate the modulation coefficients, including:

[0018] A feature modulation decision table is constructed based on enhancement modulation instructions and suppression modulation instructions. The information transmission modulation direction is obtained by calculating the weight distribution of each element in the feature modulation decision table.

[0019] The real-time activity of each feature node is input into the feature extraction layer of a multi-layer self-calibration network, and a node feature fingerprint sequence is obtained through convolution operation. The feature fingerprint sequence contains the time-varying characteristics of node activity.

[0020] Based on the node feature fingerprint sequence, a time-series correlation analysis is performed at the calibration calculation layer. The historical modulation result data is divided into time windows and the local correlation is calculated. The calibration deviation value is obtained by weighted accumulation.

[0021] The calibration compensation factor is obtained by performing a compensation operation on the calibration deviation value using a feedback adjustment layer. An initial modulation factor is calculated based on the information transmission modulation direction and the topological connection relationship between nodes. The initial modulation factor and the calibration compensation factor are then fused to obtain an adaptive modulation intensity factor.

[0022] Using the adaptive modulation intensity factor as the initial input, iterative calculations are performed within the recursive nonlinear mapping unit. When the output residual of two adjacent iterations is less than a preset convergence threshold, the final modulation coefficient is output.

[0023] In one optional embodiment, the parameter adjustment direction matrix is ​​mapped to a multi-dimensional parameter feature space, and multiple subspace units are obtained through adaptive grid partitioning. The priority adjustment sequence of waveform parameters generated based on parameter distribution characteristics includes:

[0024] The parameter adjustment direction matrix is ​​mapped to a multidimensional parameter feature space. Multiple subspace units are obtained through adaptive mesh partitioning. The probability density function of the parameter distribution within the subspace unit is calculated using kernel density estimation. The mesh boundary is adjusted according to the corresponding gradient information to obtain the initial space partitioning structure.

[0025] Based on the initial spatial partitioning structure, the kernel function bandwidth is adaptively adjusted according to the dispersion of the parameter distribution, and the probability density function is recalculated. The time-varying characteristic value of the parameter density is obtained by weighting the operation with the time decay factor. The local maximum points and jump points of the probability density function are extracted to form a set of spatial repartitioning feature points.

[0026] Based on the spatial re-division feature point set, an optimization objective function is constructed using the density difference between adjacent subspace units. The boundary position adjustment is iteratively calculated using the gradient descent method with inertia and momentum terms. When the density difference between adjacent subspace units is less than a preset difference threshold, the optimized spatial layering structure is output.

[0027] To optimize the spatial hierarchical structure, the importance evaluation index value is calculated by combining the probability density function, the time-varying characteristic value of parameter density and density difference, and the importance weight of subspace unit is determined. Based on the importance weight, the subspace units are sorted to generate a priority adjustment sequence of waveform parameters, and the priority adjustment sequence of waveform parameters is dynamically optimized and updated according to the parameter adjustment feedback results.

[0028] In an optional embodiment, the step of adaptively adjusting the kernel function bandwidth and recalculating the probability density function based on the initial spatial partitioning structure and the degree of dispersion of the parameter distribution includes:

[0029] Based on the parameter distribution within the subspace unit, a local density distribution matrix is ​​constructed. The local density distribution matrix is ​​then subjected to eigenvalue decomposition to obtain the eigenvalue sequence, and the degree of dispersion of the parameter distribution is calculated.

[0030] Based on the degree of dispersion, the upper and lower limits are determined by combining the maximum and minimum density values ​​of the local density distribution matrix, thus obtaining the initial adjustment range.

[0031] Within the initial adjustment interval, multiple kernel function bandwidth value points are divided. The degree of dispersion is used as a weighting coefficient. The weighted combination of the distribution entropy constraint value, local density preservation value, and smoothness constraint value corresponding to each value point is calculated to obtain the bandwidth optimization score.

[0032] The kernel function bandwidth value with the highest bandwidth optimization score is selected as the optimal bandwidth value. The kernel function bandwidth is then adaptively adjusted based on the optimal bandwidth value, and the probability density function within the subspace cell is recalculated.

[0033] In one optional embodiment, a set of adjustment actions for waveform parameters is constructed based on the priority adjustment sequence, the adjustment range of waveform parameters is divided into multiple adjustment levels, adjustment actions are executed and response characteristic curves are collected, the adjustment strategy selection matrix is ​​updated based on the response characteristic curves, and the next round of adjustment scheme is generated, including:

[0034] Based on the priority adjustment sequence, a set of adjustment actions for waveform parameters is constructed, the adjustment range of waveform parameters is divided into multiple adjustment levels, the rate of change of the current level is calculated in real time, and the adjustment level is dynamically adjusted according to the rate of change.

[0035] The waveform parameter adjustment action is executed, the response characteristic curve of the waveform parameter is collected in real time, the overshoot, settling time and steady-state error of the response characteristic curve are extracted, and the adjustment evaluation value is calculated by combining the historical adjustment trajectory of the waveform parameter.

[0036] The adjustment strategy selection matrix is ​​updated based on the adjustment evaluation value and the response characteristic curve. The transient response mode in the waveform parameter adjustment process is analyzed by the time series feature extraction method. The correspondence between the transient response mode and the adjustment evaluation value is stored in the adjustment strategy selection matrix.

[0037] Based on the response mode characteristics stored in the adjustment strategy selection matrix, and combined with the coupling constraint relationship of the waveform parameters, an adjustment objective function is constructed. The optimal adjustment gear combination is solved using a dynamic programming algorithm to generate the next round of adjustment scheme.

[0038] In one optional embodiment, a set of adjustment actions for waveform parameters is constructed based on a priority adjustment sequence, the adjustment range of waveform parameters is divided into multiple adjustment levels, the rate of change of the current level is calculated in real time, and the adjustment level is dynamically adjusted according to the rate of change, including:

[0039] Based on the priority adjustment sequence, a set of adjustment actions for waveform parameters is constructed, and the adjustment range of waveform parameters is divided into multiple adjustment levels;

[0040] Obtain the time-domain waveform of the waveform parameters, calculate the corresponding noise level, fluctuation trend and historical adjustment accuracy, and combine them to construct a quality assessment vector;

[0041] The quality assessment vector is compared with the preset benchmark feature vector. The gear splitting threshold correction coefficient and the gear merging threshold correction coefficient are calculated based on the deviation value of the comparison. They are then multiplied by the preset benchmark threshold to obtain the gear splitting threshold and the gear merging threshold.

[0042] The current gear is determined based on adjacent adjustment points in the gear adjustment range. The rate of change of the current gear is calculated in real time. When the rate of change is greater than or equal to the gear splitting threshold, the midpoint of the current gear is selected to generate two sub-gears. When the rate of change is less than the gear merging threshold, adjacent gears with parameter differences less than a preset difference value are merged to dynamically optimize the gear adjustment range.

[0043] A second aspect of this invention provides a radar training parameter dynamic adjustment system based on an adaptive algorithm, comprising:

[0044] The first unit is used to acquire the waveform parameters of the radar during the training process and construct a training state feature vector based on the waveform parameters.

[0045] The second unit is used to calculate the feature correlation matrix of the training state feature vector to obtain the initial connection strength matrix, generate modulation instructions based on the temporal correlation analysis of feature nodes, and perform weighted processing to obtain the feature modulation representation matrix, and calculate the parameter adjustment direction matrix based on the node state changes.

[0046] The third unit is used to map the parameter adjustment direction matrix to a multi-dimensional parameter feature space, obtain multiple subspace units through adaptive grid partitioning, and generate a priority adjustment sequence of waveform parameters based on parameter distribution characteristics;

[0047] The fourth unit is used to construct a set of adjustment actions for waveform parameters based on the priority adjustment sequence, divide the adjustment range of waveform parameters into multiple adjustment levels, execute adjustment actions and collect response characteristic curves, update the adjustment strategy selection matrix based on the response characteristic curves, and generate the next round of adjustment scheme.

[0048] The fifth unit is used to update the waveform parameters according to the adjustment scheme, write the updated waveform parameters into the radar control unit and trigger a new round of training tasks, and at the same time save the data records of this round of parameter adjustment process to the training database.

[0049] A third aspect of the present invention provides an electronic device, comprising:

[0050] processor;

[0051] Memory used to store processor-executable instructions;

[0052] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0053] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0054] In this embodiment of the invention, the radar training parameter dynamic adjustment method based on adaptive algorithms achieves precise control of waveform parameter adjustment by constructing training state feature vectors and analyzing feature correlation, thereby improving the efficiency and accuracy of parameter optimization. By employing adaptive grid partitioning technology and a priority adjustment sequence construction mechanism, the method solves the problems of strong blindness and redundant adjustment steps in traditional parameter adjustment, enabling rapid adaptive adjustment of radar waveform parameters according to training task requirements and shortening the training cycle. Through response characteristic curve analysis and dynamic updating of adjustment strategies, this method establishes a complete closed-loop feedback mechanism for parameter adjustment, enhancing the radar's adaptability in complex electromagnetic environments. Simultaneously, the continuous accumulation of training data provides data support for long-term optimization of radar performance. Attached Figure Description

[0055] Figure 1 This is a flowchart illustrating the dynamic adjustment method for radar training parameters based on an adaptive algorithm according to an embodiment of the present invention.

[0056] Figure 2 A flowchart for generating adaptive hierarchical and priority adjustment sequences for parameter space. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0059] Figure 1 This is a flowchart illustrating the dynamic adjustment method for radar training parameters based on an adaptive algorithm according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0060] Obtain the waveform parameters of the radar during the training process, and construct a training state feature vector based on the waveform parameters;

[0061] The initial connection strength matrix is ​​obtained by calculating the feature correlation degree matrix of the training state feature vectors. Modulation instructions are generated based on the temporal correlation analysis of feature nodes and weighted to obtain the feature modulation representation matrix. The parameter adjustment direction matrix is ​​calculated based on the node state changes.

[0062] The parameter adjustment direction matrix is ​​mapped to a multi-dimensional parameter feature space, and multiple subspace units are obtained through adaptive grid partitioning. Based on the parameter distribution characteristics, a priority adjustment sequence of waveform parameters is generated.

[0063] Based on the priority adjustment sequence, a set of adjustment actions for waveform parameters is constructed, the adjustment range of waveform parameters is divided into multiple adjustment levels, the adjustment actions are executed and the response characteristic curve is collected, the adjustment strategy selection matrix is ​​updated based on the response characteristic curve, and the next round of adjustment scheme is generated.

[0064] The waveform parameters are updated according to the adjustment scheme, the updated waveform parameters are written into the radar control unit and a new round of training tasks is triggered, and the data records of this round of parameter adjustment process are saved to the training database.

[0065] In one specific implementation, waveform parameters during the training process are acquired from the radar control unit, including technical parameters such as pulse repetition frequency, pulse width, antenna scan rate, and receiver gain. These acquired parameters are used to construct a training state feature vector, which comprehensively reflects the radar's current operating state and parameter configuration, providing a data foundation for subsequent analysis.

[0066] An initial connection strength matrix is ​​generated through correlation analysis between eigenvectors, representing the degree of mutual influence among parameters. Modulation instructions are generated based on temporal correlation analysis of feature nodes, guiding how parameters are adjusted sequentially over time. The modulation instructions are weighted to obtain a feature modulation representation matrix, with weights reflecting the importance of different parameter adjustments. A parameter adjustment direction matrix is ​​calculated based on node state changes, clarifying the adjustment direction and magnitude of each parameter.

[0067] The parameter adjustment direction matrix is ​​mapped to a multi-dimensional parameter feature space, which is composed of various waveform parameters, visually displaying the relationships between parameters. Multiple sub-space units are obtained through adaptive mesh partitioning, making the complex parameter space structured and manageable. A priority adjustment sequence for waveform parameters is generated based on parameter distribution characteristics, ensuring that important parameters are adjusted first, thus improving adjustment efficiency.

[0068] Constructing a set of waveform parameter adjustment actions based on a priority adjustment sequence is the specific step in executing the adjustment. Dividing the waveform parameter adjustment range into multiple adjustment levels makes the adjustment process more structured and controllable. Adjustment actions are executed, and response characteristic curves are acquired, recording the impact of parameter changes on radar performance. The adjustment strategy selection matrix is ​​updated based on the response characteristic curves; this matrix stores historical adjustment experience to guide future adjustment decisions. By analyzing the response curves and historical data, the next round of adjustment schemes is generated, achieving closed-loop optimization of parameter adjustment.

[0069] The calculated new parameter values ​​are written to the radar control unit to ensure the parameter update takes effect. After the update is complete, a new round of training is triggered to verify the effect of the parameter adjustment. Simultaneously, the data from this round of parameter adjustment is saved to the training database to provide data support for future parameter optimization. Through continuous iterative parameter adjustment and data accumulation, the radar's training effect is continuously optimized, achieving a steady improvement in performance.

[0070] In one optional implementation, the initial connection strength matrix is ​​obtained by calculating the feature correlation matrix of the training state feature vectors; modulation instructions are generated based on the temporal correlation analysis of feature nodes and weighted to obtain the feature modulation representation matrix; and the parameter adjustment direction matrix is ​​calculated based on the node state changes, including:

[0071] Calculate the correlation coefficients between features of each dimension in the training state feature vector, construct a feature correlation matrix, and calculate the initial connection strength matrix based on the numerical values ​​of each element in the feature correlation matrix.

[0072] Temporal correlation analysis is performed on each feature node in the initial connection strength matrix to extract the state information of adjacent feature nodes at different times, calculate the joint activation probability between nodes, compare the joint activation probability with a preset adaptive control threshold, and generate a strengthening modulation command or a suppressing modulation command.

[0073] The modulation direction of information transmission is determined according to the enhancement modulation command and the suppression modulation command. The adaptive modulation intensity factor is calculated based on the real-time activity of each feature node. The adaptive modulation intensity factor is input into the recursive nonlinear mapping unit to iteratively calculate the modulation coefficient.

[0074] The training state feature vector is weighted according to the modulation coefficient and input into a neural network with memory function. Through the selective modulation of the neural network, the feature modulation representation matrix is ​​calculated.

[0075] The state change relationship between the preceding and succeeding feature nodes in the feature modulation representation matrix is ​​calculated. The state transition law is extracted to obtain the weight update direction. The initial matrix of connection strength is adjusted, and the parameter adjustment direction matrix is ​​calculated.

[0076] In one specific implementation, the correlation coefficients between the features of each dimension in the training state feature vector are calculated based on the correlation analysis. Specifically, for any two features in the training state feature vector, the correlation degree is obtained using the Pearson correlation coefficient calculation method. During calculation, sample data for the two features are obtained, their respective averages are calculated, and then the ratio of the product of the covariance and standard deviation of the two features is calculated. After calculating the correlation coefficients for all features pairwise, a feature correlation matrix is ​​constructed. For example, for a training state feature vector containing 10 dimensions such as radar target echo intensity, clutter suppression level, and target detection probability, a 10×10 feature correlation matrix can be obtained, where the matrix element values ​​range from [-1, 1]. Values ​​closer to 1 indicate a higher degree of positive correlation, and values ​​closer to -1 indicate a higher degree of negative correlation.

[0077] Based on the feature correlation matrix, an initial connection strength matrix is ​​further calculated. Each element in the initial connection strength matrix represents the initial connection strength between the corresponding feature nodes, obtained by performing a nonlinear transformation on the absolute values ​​of the corresponding elements in the feature correlation matrix. The nonlinear transformation uses the sigmoid function to map the absolute values ​​of the feature correlation, ensuring that the element values ​​in the initial connection strength matrix are all distributed within the interval (0, 1). For example, when the correlation coefficient of two radar training parameter features is 0.75, the initial connection strength value after the nonlinear transformation is approximately 0.68; when the correlation coefficient is -0.82, the initial connection strength value is approximately 0.72.

[0078] When performing temporal correlation analysis on each feature node in the initial connection strength matrix, the state change data of each feature node over multiple consecutive training periods are extracted. For each pair of adjacent feature nodes, the ratio of the number of times they are simultaneously activated to the number of times they are individually activated within the time window is calculated to determine the joint activation probability between the nodes. Taking radar target detection features and tracking accuracy features as an example, within 100 training periods, the number of times the target detection feature value exceeds the preset threshold of 0.6 is 65 times, the number of times the tracking accuracy feature value exceeds the preset threshold of 0.7 is 58 times, and the number of times both exceed their respective thresholds is 45 times. Therefore, the joint activation probability is 45 / (65+58-45)≈0.58.

[0079] The calculated joint activation probability is compared with a preset adaptive adjustment threshold. This threshold is dynamically adjusted based on the current training stage, historical training results, and task difficulty, and is typically initially set to 0.5. When the joint activation probability is higher than the adaptive adjustment threshold, a strengthening modulation command is generated; when it is lower, a suppressing modulation command is generated. Continuing with the previous example, the joint activation probability of 0.58 is greater than the preset threshold of 0.5, therefore a strengthening modulation command is generated, indicating a positive promoting relationship between target detection features and tracking accuracy features, requiring enhanced information transfer between them.

[0080] The modulation direction of information transmission is determined based on the enhancement modulation command and the suppression modulation command. For feature node pairs that generate enhancement modulation commands, information is transmitted from the node with higher activity to the node with lower activity; for feature node pairs that generate suppression modulation commands, the information interaction between the two nodes is reduced. The real-time activity of a feature node is calculated by the frequency with which the node exceeds the activation threshold in the most recent N training epochs. For example, if the target detection feature exceeds the activation threshold 15 times in the most recent 20 epochs, its activity is 0.75; if the tracking accuracy feature exceeds the activation threshold 12 times in the same period, its activity is 0.6, and the information transmission direction is from the target detection feature to the tracking accuracy feature.

[0081] The adaptive modulation intensity factor is calculated based on the real-time activity of each feature node. The modulation intensity factor is directly proportional to the activity of the source node and inversely proportional to the activity of the target node, while also considering the joint activation probability of the two nodes. The calculated adaptive modulation intensity factor is input to a recursive nonlinear mapping unit, which iterates multiple times to obtain the final modulation coefficient. The recursive nonlinear mapping unit uses a hyperbolic tangent function for multiple iterations, with the result of each iteration serving as the input for the next, until convergence or a preset number of iterations is reached. For example, with a source node activity of 0.75, a target node activity of 0.6, and a joint activation probability of 0.58, the calculated initial modulation intensity factor is approximately 0.73. After 5 rounds of recursive iteration, the final modulation coefficient converges to 0.65.

[0082] The training state feature vectors are weighted according to modulation coefficients. For each feature dimension, its weight in the overall representation is adjusted based on its modulation coefficients with other feature dimensions. The adjusted feature vectors are then input into a neural network with memory capabilities. This network employs a long short-term memory (LSTM) structure, enabling it to retain historical training information and perform selective modulation. The neural network outputs a feature modulation representation matrix, which reflects the modulated states of each feature dimension and their interrelationships.

[0083] The computational feature modulation representation matrix calculates the state change relationship between preceding and subsequent feature nodes. For each pair of causally related feature nodes, the influence pattern of the preceding node's state change on the subsequent node's state is analyzed, and the state transition law is extracted. The state transition law is obtained by statistically analyzing the response change trend of the subsequent node before and after the preceding node's state change. For example, when the target detection feature value increases from 0.65 to 0.85, the tracking accuracy feature value increases by an average of 0.15; while when the target detection feature value decreases from 0.85 to 0.65, the tracking accuracy feature value decreases by an average of 0.12. Based on this state transition law, the weight update direction is determined, the initial connection strength matrix is ​​adjusted, and the parameter adjustment direction matrix is ​​finally calculated.

[0084] Each element in the parameter adjustment direction matrix represents the direction and magnitude of adjustment for the corresponding training parameter. Positive values ​​indicate an increase in the corresponding parameter, negative values ​​indicate a decrease in the corresponding parameter, and the absolute value indicates the adjustment magnitude. For example, a value of 0.25 for radar transmit power in the parameter adjustment direction matrix indicates that the transmit power needs to be moderately increased; a value of -0.15 for clutter suppression threshold indicates that the clutter suppression threshold needs to be slightly decreased. Based on the parameter adjustment direction matrix, the radar training system automatically adjusts various parameters to achieve dynamic optimization of the training effect.

[0085] In one optional implementation, the information transmission modulation direction is determined based on the enhancement modulation command and the suppression modulation command. An adaptive modulation intensity factor is calculated based on the real-time activity of each feature node. The adaptive modulation intensity factor is then input into a recursive nonlinear mapping unit for iterative calculation to obtain modulation coefficients, including:

[0086] A feature modulation decision table is constructed based on enhancement modulation instructions and suppression modulation instructions. The information transmission modulation direction is obtained by calculating the weight distribution of each element in the feature modulation decision table.

[0087] The real-time activity of each feature node is input into the feature extraction layer of a multi-layer self-calibration network, and a node feature fingerprint sequence is obtained through convolution operation. The feature fingerprint sequence contains the time-varying characteristics of node activity.

[0088] Based on the node feature fingerprint sequence, a time-series correlation analysis is performed at the calibration calculation layer. The historical modulation result data is divided into time windows and the local correlation is calculated. The calibration deviation value is obtained by weighted accumulation.

[0089] The calibration compensation factor is obtained by performing a compensation operation on the calibration deviation value using a feedback adjustment layer. An initial modulation factor is calculated based on the information transmission modulation direction and the topological connection relationship between nodes. The initial modulation factor and the calibration compensation factor are then fused to obtain an adaptive modulation intensity factor.

[0090] Using the adaptive modulation intensity factor as the initial input, iterative calculations are performed within the recursive nonlinear mapping unit. When the output residual of two adjacent iterations is less than a preset convergence threshold, the final modulation coefficient is output.

[0091] In one specific implementation, when constructing the feature modulation decision table based on the enhanced modulation command and the suppressed modulation command, all feature nodes need to be divided according to their functional categories, such as target detection, signal processing, and jamming countermeasures. Radar training parameters typically include multiple dimensions such as waveform selection, pulse repetition frequency, beam scanning strategy, and signal processing threshold. The feature modulation decision table is in matrix form, with rows and columns corresponding to source and target nodes, respectively. A matrix element value of 1 indicates enhanced modulation, -1 indicates suppressed modulation, and 0 indicates no modulation relationship. For example, during a radar training process, an enhanced modulation command is generated between the target detection threshold and the false alarm rate, and the corresponding element in the decision table has a value of 1; while a suppressed modulation command is generated between the target detection threshold and the beam switching rate, and the corresponding element has a value of -1. The complete feature modulation decision table reflects the modulation relationship network within the feature vector of the entire training state.

[0092] When calculating the weight distribution of each element in the feature modulation decision table, two dimensions need to be considered: node importance and modulation intensity. Node importance is determined by evaluating the number of connections between the node and other nodes and its historical modulation effects, while modulation intensity is calculated based on the difference between the joint activation probability and a preset threshold. The weights of these two factors are then weighted to obtain the weight value of each element in the decision table. For example, if the target detection threshold node has an importance score of 0.85, its joint activation probability with the false alarm control rate node is 0.72, and the preset threshold is 0.5, the calculated weight value is 0.85 × (0.72 - 0.5) × 10 ≈ 1.87. After the weight distribution is calculated, the modulation direction of information transmission can be determined; elements with higher weight values ​​have higher priority in the modulation direction from the source node to the target node.

[0093] When inputting the real-time activity of each feature node into the feature extraction layer of the multi-layer self-calibration network, the activity data of each node is first sampled in time series. The sampling frequency is dynamically adjusted according to the complexity of the training task, typically once every 10 seconds. For radar target tracking function training, there may be multiple feature nodes such as target acquisition rate, tracking accuracy, and multi-target resolution, and the activity data of each node constitutes a time series. After these time series are input into the feature extraction layer, convolution operations are performed through multi-channel convolution kernels. The convolution kernel size is set to 3×3, the number of channels is 16, the stride is 1, and the padding strategy is homomorphic padding. The time-varying characteristics of node activity are extracted through convolution operations to obtain the node feature fingerprint sequence. For the target tracking accuracy node, its feature fingerprint sequence may show a feature pattern of rapid increase in activity when highly maneuverable targets appear, while the activity remains stable in simple target scenarios.

[0094] Based on the node feature fingerprint sequences, temporal correlation analysis is performed at the calibration calculation layer, dividing the historical modulation results data into time windows. The time window size is typically set to 10 to 30 minutes, dynamically adjusted according to the training task duration. Within each time window, the cross-correlation function between different node feature fingerprint sequences is calculated to obtain the local correlation. The local correlation reflects the degree of mutual influence between different feature nodes within a specific time window. For example, in a complex electromagnetic environment training scenario, the local correlation between the interference suppression feature node and the target detection feature node within a 30-minute window is 0.63, indicating that the interference suppression capability has a significant impact on target detection performance. The local correlations of all time windows are weighted and accumulated, with higher weights for recent windows and lower weights for longer windows, and the accumulated result yields the calibration deviation value. In a certain training iteration, the accumulated calibration deviation value between target tracking and fire control was 0.42, indicating a certain deviation in the relationship between the two in the historical modulation results, requiring calibration.

[0095] When performing compensation calculations on calibration deviation values ​​using the feedback adjustment layer, an adaptive feedback control algorithm is employed. This algorithm dynamically adjusts the compensation level based on the magnitude and sign of the calibration deviation value. High-gain positive compensation is used when the calibration deviation value is positive and large; high-gain negative compensation is used when the calibration deviation value is negative and large; and low-gain fine-tuning compensation is used when the calibration deviation value is close to zero. For example, when the calibration deviation value is 0.42, the calculated calibration compensation factor is 0.35; when the calibration deviation value is -0.28, the calibration compensation factor is -0.21. The initial modulation factor is calculated based on the information transmission modulation direction and the topological connections between nodes. The topological connections are determined through the functional relevance of feature nodes and the data flow direction. For example, there is a direct connection between the target detection node and the target tracking node, while the target detection node and the fire control node require an indirect connection through the target tracking node. For directly connected node pairs, the initial modulation factor is higher, typically between 0.6 and 0.9; for indirectly connected node pairs, the initial modulation factor is lower, typically between 0.3 and 0.6. The initial modulation factor and the calibration compensation factor are weighted and fused to obtain the adaptive modulation intensity factor. The fusion weights are dynamically adjusted based on historical training results, with a typical value of 0.7 for the initial modulation factor and 0.3 for the calibration compensation factor. In a certain radar target tracking training, the initial modulation factor from the target detection node to the target tracking node was 0.78, and the calibration compensation factor was 0.35. The resulting adaptive modulation intensity factor was 0.78 × 0.7 + 0.35 × 0.3 ≈ 0.65.

[0096] Using an adaptive modulation intensity factor as the initial input, iterative calculations are performed within a recursive nonlinear mapping unit (RNN). The RNN employs a multilayer perceptron structure with two hidden layers, each containing 32 and 16 nodes respectively, and uses the ReLU activation function. During iteration, the output of each round serves as the input for the next round, while a random perturbation is introduced to prevent getting trapped in local optima. The perturbation strength decreases with the number of iterations, initially set to 0.05 with a decay rate of 0.85. The final modulation coefficient is output when the output residual of two adjacent iterations is less than a preset convergence threshold. The preset convergence threshold is typically set between 0.01 and 0.001 and is dynamically adjusted according to training accuracy requirements. With an adaptive modulation intensity factor of 0.65 as the initial input, after iterative calculation by the recursive nonlinear mapping unit, the output of the first round is 0.72, the second round is 0.68, the third round is 0.67, the fourth round is 0.666, and the fifth round is 0.664. At this time, the output residual between two adjacent rounds is 0.002, which is less than the preset convergence threshold of 0.005. Therefore, the final modulation coefficient is determined to be 0.664.

[0097] In practical applications of dynamic adjustment of radar training parameters, the above methods can effectively improve training results. They are particularly suitable for rapid optimization and adaptive training of radar performance under variable mission environments, providing technical support for radar systems to maintain optimal working conditions under complex conditions.

[0098] In one optional implementation, the parameter adjustment direction matrix is ​​mapped to a multi-dimensional parameter feature space, and multiple subspace units are obtained through adaptive mesh partitioning. The priority adjustment sequence of waveform parameters generated based on parameter distribution characteristics includes:

[0099] The parameter adjustment direction matrix is ​​mapped to a multidimensional parameter feature space. Multiple subspace units are obtained through adaptive mesh partitioning. The probability density function of the parameter distribution within the subspace unit is calculated using kernel density estimation. The mesh boundary is adjusted according to the corresponding gradient information to obtain the initial space partitioning structure.

[0100] Based on the initial spatial partitioning structure, the kernel function bandwidth is adaptively adjusted according to the dispersion of the parameter distribution, and the probability density function is recalculated. The time-varying characteristic value of the parameter density is obtained by weighting the operation with the time decay factor. The local maximum points and jump points of the probability density function are extracted to form a set of spatial repartitioning feature points.

[0101] Based on the spatial re-division feature point set, an optimization objective function is constructed using the density difference between adjacent subspace units. The boundary position adjustment is iteratively calculated using the gradient descent method with inertia and momentum terms. When the density difference between adjacent subspace units is less than a preset difference threshold, the optimized spatial layering structure is output.

[0102] To optimize the spatial hierarchical structure, the importance evaluation index value is calculated by combining the probability density function, the time-varying characteristic value of parameter density and density difference, and the importance weight of subspace unit is determined. Based on the importance weight, the subspace units are sorted to generate a priority adjustment sequence of waveform parameters, and the priority adjustment sequence of waveform parameters is dynamically optimized and updated according to the parameter adjustment feedback results.

[0103] In one specific implementation, when mapping the parameter adjustment direction matrix to a multi-dimensional parameter feature space, the adjustment direction and amplitude of each parameter in the matrix need to be converted into vector representations in the feature space. For radar training parameter adjustment, the feature space typically includes dimensions such as pulse repetition frequency, pulse width, beam scanning rate, and clutter suppression threshold. During the mapping process, the adjustment direction of each parameter corresponds to the direction of the corresponding dimension in the feature space, and the adjustment amplitude corresponds to the length of the vector in that dimension. For example, in a certain training iteration, the pulse repetition frequency adjustment direction value is 0.32, and the pulse width adjustment direction value is -0.25. The corresponding vector components in the feature space are 0.32 and -0.25, respectively, indicating a moderate increase in the pulse repetition frequency and a slight decrease in the pulse width.

[0104] When obtaining multiple subspace cells through adaptive mesh generation, the mesh density is dynamically determined based on the parameter distribution characteristics. A dense mesh is used in regions with drastic parameter changes, while a sparse mesh is used in regions with gradual parameter changes. The initial mesh generation uses a uniform distribution, which is then refined based on parameter gradient information. Taking pulse repetition frequency and beam scanning rate as examples, when both are in a medium-level range, the region is divided into a 5×5 mesh; when the pulse repetition frequency is close to the upper limit and the beam scanning rate is low, the region is further subdivided into an 8×3 mesh to more precisely capture parameter changes.

[0105] Kernel density estimation was used to calculate the probability density function of the parameter distribution within each subspace cell. A Gaussian kernel was chosen as the kernel function, and the initial bandwidth was set to 0.15. The probability density of each parameter within different subspace cells was calculated based on historical training data. For the radar target detection parameter set, within a certain subspace cell, the calculated probability density was 0.28 for the detection threshold in the range of 0.45 to 0.55, and 0.42 for the range of 0.55 to 0.65, indicating that the parameter performs better in the range of 0.55 to 0.65. The grid boundary was adjusted based on the gradient information of the parameters within each subspace cell. The grid boundary was moved in the gradient direction where the gradient was larger, so that the grid better adapted to the parameter distribution characteristics. In one training iteration, it was found that the gradient value at the intersection of the pulse width and the suppression threshold reached 0.35, which was higher than the preset threshold of 0.2. Therefore, the grid boundary was moved 0.08 units in the gradient direction to obtain an initial spatial partitioning structure that better matched the parameter distribution characteristics.

[0106] Based on the initial spatial partitioning structure, the kernel function bandwidth is adaptively adjusted according to the dispersion of the parameter distribution. Regions with concentrated parameter distributions use a smaller bandwidth to improve estimation accuracy, while regions with dispersed parameter distributions use a larger bandwidth to enhance generalization ability. The dispersion of the parameter distribution is assessed using the coefficient of variation (COP). A COP greater than 0.5 increases the bandwidth by 25%, while a COP less than 0.2 decreases it by 20%. For example, in radar target tracking training, if the COP of the tracking threshold parameter is found to be 0.65 within a certain subspace cell, and the original bandwidth is 0.15, the adjusted bandwidth increases to 0.1875. The probability density function is recalculated using the adjusted bandwidth and weighted with a time decay factor. The time decay factor decreases as the data time interval increases: data from the most recent hour has a weight of 1, data from 1 to 3 hours has a weight of 0.8, data from 3 to 6 hours has a weight of 0.5, and data older than 6 hours has a weight of 0.2. The weighted calculation yields the time-varying characteristic value of the parameter density, reflecting the trend and stability of parameter changes over time.

[0107] When extracting local maxima and transition points of the probability density function, a peak detection algorithm is used to identify local maxima, and a transition point detection algorithm is used to identify transition points. The minimum peak height for peak detection is set to 0.2, and the minimum peak spacing to 0.1; the significance level for transition point detection is set to 0.05. In the radar waveform selection parameter space, the detected local maxima include a density peak of 0.38 corresponding to a pulse repetition frequency of 10kHz and a density peak of 0.42 corresponding to a beam dwell time of 15ms; the detected transition points include a density change rate of 0.25 when the clutter suppression threshold jumps from 0.35 to 0.4. These local maxima and transition points constitute a spatial repartition feature point set, providing key reference points for subsequent optimization.

[0108] Based on the spatial repartition feature point set, an optimization objective function is constructed using the density difference between adjacent subspace units. The optimization objective is to minimize the parameter density difference between adjacent subspace units while maintaining the clustering of parameter distribution within each subspace unit. The objective function comprehensively considers three aspects: density difference, clustering, and smoothness, with weights of 0.5, 0.3, and 0.2, respectively. Gradient descent with inertia and momentum terms is used to iteratively calculate the boundary position adjustment. The inertia coefficient is set to 0.7, the momentum coefficient to 0.3, and the initial learning rate to 0.05, decreasing with the number of iterations. In each iteration, the gradient of the objective function with respect to the boundary position is calculated, and the direction and magnitude of the boundary adjustment are determined by combining the inertia and momentum terms. When the density difference between adjacent subspace units is less than a preset difference threshold of 0.05 or the maximum number of iterations (100) is reached, the optimized spatial hierarchical structure is output.

[0109] To optimize the spatial hierarchical structure, importance evaluation index values ​​are calculated by combining the probability density function, time-varying parameter density characteristics, and density differences. The importance evaluation index comprehensively considers three dimensions: parameter influence range, sensitivity to change, and time-varying stability. The influence range has a weight of 0.4, evaluated by the integral value of the probability density function of the parameter within the subspace cell; the sensitivity to change has a weight of 0.4, evaluated by the degree of impact of parameter adjustment on training effect; and the time-varying stability has a weight of 0.2, evaluated by the degree of fluctuation of the time-varying parameter density characteristics. In radar target recognition training, the calculated importance index values ​​are 0.75 for waveform selection parameters, 0.68 for pulse compression ratio parameters, and 0.62 for Doppler processing parameters.

[0110] When determining the importance weights of subspace units, a soft maximum function is used to normalize the importance evaluation index values, ensuring that the sum of the weights of all subspace units is 1. Subspace units are ranked based on their importance weights, with higher-weighted units adjusted first. The generated priority adjustment sequence for waveform parameters guides the order of parameter adjustments and resource allocation during training. For example, in a certain training iteration, the priority adjustment sequence might be: waveform selection parameter (weight 0.28), pulse compression ratio parameter (weight 0.25), Doppler processing parameter (weight 0.22), beam management parameter (weight 0.15), and clutter processing parameter (weight 0.10). According to this sequence, the waveform selection parameter is adjusted first, followed by the pulse compression ratio parameter, and so on. During parameter adjustment, the adjustment effect is continuously monitored, and feedback data is collected. Based on the parameter adjustment feedback results, the priority adjustment sequence for waveform parameters is dynamically optimized and updated; the weight of adjustment strategies with good feedback results is increased, while the weight of adjustment strategies with poor feedback results is decreased. After multiple rounds of training, the priority adjustment sequence may be adjusted to: Doppler processing parameters (weight 0.30), waveform selection parameters (weight 0.27), pulse compression ratio parameters (weight 0.23), beam management parameters (weight 0.12), and clutter processing parameters (weight 0.08). Through this dynamic optimization mechanism, the radar training parameter adjustment strategy continuously improves itself, its adaptability is constantly enhanced, and the training effect is continuously improved.

[0111] like Figure 2 As shown, a flowchart illustrating the parameter space adaptive hierarchical and priority adjustment sequence generation process is presented.

[0112] In an optional implementation, the step of adaptively adjusting the kernel function bandwidth and recalculating the probability density function based on the initial spatial partitioning structure and the degree of dispersion of the parameter distribution includes:

[0113] Based on the parameter distribution within the subspace unit, a local density distribution matrix is ​​constructed. The local density distribution matrix is ​​then subjected to eigenvalue decomposition to obtain the eigenvalue sequence, and the degree of dispersion of the parameter distribution is calculated.

[0114] Based on the degree of dispersion, the upper and lower limits are determined by combining the maximum and minimum density values ​​of the local density distribution matrix, thus obtaining the initial adjustment range.

[0115] Within the initial adjustment interval, multiple kernel function bandwidth value points are divided. The degree of dispersion is used as a weighting coefficient. The weighted combination of the distribution entropy constraint value, local density preservation value, and smoothness constraint value corresponding to each value point is calculated to obtain the bandwidth optimization score.

[0116] The kernel function bandwidth value with the highest bandwidth optimization score is selected as the optimal bandwidth value. The kernel function bandwidth is then adaptively adjusted based on the optimal bandwidth value, and the probability density function within the subspace cell is recalculated.

[0117] In one specific implementation, a local density distribution matrix is ​​constructed based on the parameter distribution within a subspace cell. This local density distribution matrix reflects the parameter distribution within the subspace cell and can be constructed by calculating the local density values ​​of each sample point within the subspace cell. Specifically, assuming the subspace cell contains N sample points, the density relationship between each sample point and other sample points is calculated using an initial kernel function, forming an N×N local density distribution matrix M. Each element M[i, j] in the matrix represents the density relationship value between the i-th sample point and the j-th sample point, with a value ranging from 0 to 1; a larger value indicates a stronger correlation between the two points. For example, in a subspace cell containing 5 sample points, the possible local density distribution matrix is: [[1.0, 0.85, 0.32, 0.15, 0.08], [0.85, 1.0, 0.45, 0.22, 0.11], [0.32, 0.45, 1.0, 0.76, 0.39], [0.15, 0.22, 0.76, 1.0, 0.91], [0.08, 0.11, 0.39, 0.91, 1.0]].

[0118] After obtaining the local density distribution matrix, eigenvalue decomposition is performed to obtain the eigenvalue sequence. Eigenvalue decomposition can be achieved through iterative calculation, outputting a set of eigenvalues ​​arranged in descending order. For example, eigenvalue decomposition of the above matrix might yield the eigenvalue sequence: [3.24, 1.58, 0.12, 0.04, 0.02]. The dispersion of the parameter distribution is calculated based on this set of eigenvalues. The dispersion D can be measured by the ratio of the variance to the mean of the eigenvalue sequence; a larger D indicates a more dispersed parameter distribution, requiring corresponding adjustment of the kernel function bandwidth. Using the above eigenvalue sequence as an example, the calculated dispersion D might be 0.78.

[0119] Based on the calculated dispersion level D, the upper and lower limits of bandwidth adjustment are determined by combining the maximum density value MaxDen and the minimum density value MinDen of the local density distribution matrix, resulting in the initial adjustment interval [LowerBound, UpperBound]. The upper limit UpperBound can be obtained by multiplying MaxDen by (1+D), and the lower limit LowerBound can be obtained by multiplying MinDen by (1-D / 2). Using the data in the previous example, assuming MaxDen is 1.0, MinDen is 0.02, and the dispersion level D is 0.78, then UpperBound = 1.0 × (1+0.78) = 1.78, LowerBound = 0.02 × (1-0.78 / 2) = 0.0122, and the initial adjustment interval is [0.0122, 1.78].

[0120] Within the defined initial adjustment interval [0.0122, 1.78], 10 kernel function bandwidth values ​​are evenly divided, such as [0.0122, 0.2095, 0.4068, 0.6041, 0.8014, 0.9987, 1.1960, 1.3933, 1.5906, 1.78]. For each value, the corresponding distribution entropy constraint value (Entropy), local density preservation value (Density), and smoothness constraint value (Smooth) are calculated. These three indicators are then weighted and combined to obtain the bandwidth optimization score (Score). The calculation formula is: Score = D×Entropy + (1-D / 2)×Density + (1-D)×Smooth. Wherein, the distribution entropy constraint value (Entropy) measures the information content of the probability density distribution under the given bandwidth value, the local density preservation value (Density) measures the degree to which the bandwidth adjustment preserves the original local density characteristics, and the smoothness constraint value (Smooth) measures the smoothness of the probability density function.

[0121] For each bandwidth value, calculate its optimization score. For example, for the value 0.4068, the calculated Entropy = 0.85, Density = 0.92, and Smooth = 0.78. Substituting these values ​​into the formula, we get Score = 0.78 × 0.85 + (1 - 0.78 / 2) × 0.92 + (1 - 0.78) × 0.78 = 1.3911. For the value 0.8014, the calculated Entropy = 0.92, Density = 0.88, and Smooth = 0.85. Substituting these values ​​into the formula, we get Score = 0.78 × 0.92 + (1 - 0.78 / 2) × 0.88 + (1 - 0.78) × 0.85 = 1.4739. By calculating the scores for all value points, the score sequence [0.9821, 1.1052, 1.2783, 1.3911, 1.4739, 1.4256, 1.3871, 1.2534, 1.1267, 0.9823] is obtained.

[0122] From the scoring sequence, the bandwidth value with the highest score is selected as the optimal bandwidth value. In this example, the score of 1.4739 corresponding to the value 0.8014 is the highest, so 0.8014 is selected as the optimal bandwidth value. The kernel function bandwidth is adaptively adjusted based on this optimal bandwidth value, and the probability density function within the subspace cell is recalculated. Specifically, the kernel function bandwidth is set to 0.8014, and the density relationship between each sample point and other sample points within the subspace cell is recalculated to construct a new probability density function. Through this adaptive adjustment method, the most suitable kernel function bandwidth is automatically selected based on the dispersion of the parameter distribution, improving the accuracy and expressive power of the probability density function.

[0123] The adjusted probability density function can more accurately reflect the distribution characteristics of parameters within the subspace cells, effectively improving the accuracy of subsequent analysis and applications. For example, in a certain parameter space, the unadjusted probability density function may not be prominent enough in high-density regions and may be too smooth in low-density regions, while the adjusted probability density function can more clearly distinguish between different density regions, providing a more reliable basis for subsequent parameter optimization.

[0124] In one optional implementation, a set of adjustment actions for waveform parameters is constructed based on the priority adjustment sequence, the adjustment range of waveform parameters is divided into multiple adjustment levels, adjustment actions are executed and response characteristic curves are collected, the adjustment strategy selection matrix is ​​updated based on the response characteristic curves, and the next round of adjustment scheme is generated, including:

[0125] Based on the priority adjustment sequence, a set of adjustment actions for waveform parameters is constructed, the adjustment range of waveform parameters is divided into multiple adjustment levels, the rate of change of the current level is calculated in real time, and the adjustment level is dynamically adjusted according to the rate of change.

[0126] The waveform parameter adjustment action is executed, the response characteristic curve of the waveform parameter is collected in real time, the overshoot, settling time and steady-state error of the response characteristic curve are extracted, and the adjustment evaluation value is calculated by combining the historical adjustment trajectory of the waveform parameter.

[0127] The adjustment strategy selection matrix is ​​updated based on the adjustment evaluation value and the response characteristic curve. The transient response mode in the waveform parameter adjustment process is analyzed by the time series feature extraction method. The correspondence between the transient response mode and the adjustment evaluation value is stored in the adjustment strategy selection matrix.

[0128] Based on the response mode characteristics stored in the adjustment strategy selection matrix, and combined with the coupling constraint relationship of the waveform parameters, an adjustment objective function is constructed. The optimal adjustment gear combination is solved using a dynamic programming algorithm to generate the next round of adjustment scheme.

[0129] In one specific implementation, a priority adjustment sequence is first constructed. Based on parameter influence analysis, waveform parameters are sorted from highest to lowest importance to form a priority adjustment sequence. For example, for pulse Doppler radar, parameters such as pulse repetition frequency, pulse width, antenna scan rate, and receiver gain can be included in the adjustment sequence, prioritizing the adjustment of parameters that affect the accuracy of target detection range and velocity.

[0130] The waveform parameter adjustment range is divided into multiple adjustment levels. Taking the pulse repetition frequency as an example, if its adjustment range is 1000Hz to 5000Hz, it can be divided into 10 levels, each 400Hz apart. The level change rate is calculated in real time, which is the ratio of the parameter value change per unit time to the level width. When the change rate exceeds the preset threshold of 0.8, it indicates that the parameter changes too quickly, and the level width needs to be increased; when the change rate is below the threshold of 0.2, it indicates that the parameter changes too slowly, and the level width needs to be decreased to improve accuracy. The dynamically adjusted level width is the original level width multiplied by the change rate adjustment coefficient, which is determined by the nonlinear mapping function of the change rate.

[0131] When performing waveform parameter adjustment, adjustment commands are sent to the radar via the control interface, and the response characteristic curves of the waveform parameters are acquired in real time. Taking antenna scan rate adjustment as an example, the change in scan rate over time is recorded after each adjustment, with a sampling frequency of 100Hz and continuous sampling for 5 seconds. Key indicators are extracted from the response characteristic curves: overshoot, settling time, and steady-state error. Overshoot is calculated as the percentage difference between the maximum overshoot value and the target value relative to the target value. For example, if the target scan rate is 30° / s and the maximum response value is 33° / s, then the overshoot is 10%. Settling time is the time required for the parameter value to enter and remain within ±5% of the target value. For example, it takes 0.8 seconds from the start of adjustment to stabilization within the range of 28.5° / s to 31.5° / s. Steady-state error is the percentage deviation between the average value of the stabilized parameter and the target value. For example, if the average value after stabilization is 29.5° / s, then the steady-state error is 1.67%.

[0132] The adjustment evaluation value is calculated by combining the historical adjustment trajectory of the waveform parameters. This evaluation value comprehensively considers the time efficiency, accuracy, and stability of the current adjustment. The evaluation value consists of three weighted components: adjustment time score, overshoot score, and steady-state error score. The adjustment time score is calculated using an exponential decay function, with a higher score for shorter time. The overshoot and steady-state error scores are both calculated using reciprocal functions, with a higher score for smaller errors. The weighting ratio of the three components is 4:3:3, and the final score ranges from 0 to 100. For example, if the adjustment time is 0.8 seconds, the overshoot is 10%, and the steady-state error is 1.67%, the calculated adjustment time score is 82, the overshoot score is 75, and the steady-state error score is 88, resulting in a weighted adjustment evaluation value of 82.

[0133] The adjustment strategy selection matrix is ​​updated based on the adjustment evaluation values ​​and response characteristic curves. A time-series feature extraction method is used to analyze the transient response patterns during waveform parameter adjustment, including time-domain features such as rise time, peak time, and settling time. For each response pattern, the corresponding adjustment evaluation value is recorded and stored in the adjustment strategy selection matrix. This matrix is ​​a two-dimensional table; rows represent different response pattern types, columns represent the adjustment level change, and cell values ​​are historical adjustment evaluation values. For example, for the "rapid rise - slight oscillation - rapid stabilization" response pattern, the historical adjustment evaluation value is 85 points when the level change is +2, and 78 points when the level change is +3. Through continuous accumulation of adjustment experience, the evaluation values ​​in the matrix are constantly updated, reflecting the statistical regularity of the adjustment effect.

[0134] Based on the response mode characteristics stored in the adjustment strategy selection matrix, an adjustment objective function is constructed by combining the coupling constraints of waveform parameters. These coupling constraints include physical limitations and performance requirements, such as the product of pulse repetition frequency and pulse width not exceeding the maximum duty cycle limit of 0.2, and an inverse relationship between antenna scan rate and receiver gain. The adjustment objective function comprises three parts: the expected adjustment evaluation value, a parameter coupling penalty term, and a range change penalty term. The weights of each part are dynamically adjusted according to the current task requirements.

[0135] Dynamic programming is used to solve for the optimal combination of adjustment levels. The adjustment process is viewed as a multi-stage decision problem, with each stage corresponding to a parameter to be adjusted and each decision corresponding to a level selection. The state space is defined by the current combination of levels for each parameter, and the transition equation is defined by the change in the evaluation value caused by the level change. The optimal level path is solved by backtracking to generate the next round of adjustment scheme. For example, if the current pulse repetition frequency is at level 5, the pulse width is at level 3, and the antenna scan rate is at level 4, the dynamic programming algorithm calculates the optimal adjustment scheme as follows: adjust the pulse repetition frequency to level 6, maintain the pulse width at level 3, and adjust the antenna scan rate to level 5.

[0136] In actual adjustment, by iteratively executing the above steps, the adjustment accuracy and efficiency of waveform parameters are significantly improved. Taking a certain type of radar as an example, after adopting the method of this embodiment, the parameter adjustment time is shortened, the overshoot is reduced, the steady-state error is lowered, the tracking accuracy is improved, and the overall radar performance evaluation score is further improved. The method of this embodiment is applicable to various radar systems that require real-time and precise adjustment, and has strong versatility and practical value.

[0137] In one optional implementation, a set of adjustment actions for waveform parameters is constructed based on a priority adjustment sequence, the adjustment range of waveform parameters is divided into multiple adjustment levels, the rate of change of the current level is calculated in real time, and the adjustment level is dynamically adjusted according to the rate of change, including:

[0138] Based on the priority adjustment sequence, a set of adjustment actions for waveform parameters is constructed, and the adjustment range of waveform parameters is divided into multiple adjustment levels;

[0139] Obtain the time-domain waveform of the waveform parameters, calculate the corresponding noise level, fluctuation trend and historical adjustment accuracy, and combine them to construct a quality assessment vector;

[0140] The quality assessment vector is compared with the preset benchmark feature vector. The gear splitting threshold correction coefficient and the gear merging threshold correction coefficient are calculated based on the deviation value of the comparison. They are then multiplied by the preset benchmark threshold to obtain the gear splitting threshold and the gear merging threshold.

[0141] The current gear is determined based on adjacent adjustment points in the gear adjustment range. The rate of change of the current gear is calculated in real time. When the rate of change is greater than or equal to the gear splitting threshold, the midpoint of the current gear is selected to generate two sub-gears. When the rate of change is less than the gear merging threshold, adjacent gears with parameter differences less than a preset difference value are merged to dynamically optimize the gear adjustment range.

[0142] In one specific implementation, before the adjustment process begins, a priority adjustment sequence for waveform parameters is defined according to the actual application scenario. For example, for radar signal processing, waveform parameters such as pulse width, pulse amplitude, and repetition frequency can be sorted by importance to form a priority adjustment sequence. Based on this sequence, a set of adjustment actions for the waveform parameters is constructed, and each adjustment action includes a parameter identifier, adjustment direction, and adjustment step size.

[0143] For each waveform parameter, its adjustment range is divided into multiple adjustment levels. Taking the pulse width parameter as an example, if its value range is 10μs to 100μs, it can be initially divided into four levels: [10μs, 25μs], [25μs, 50μs], [50μs, 75μs], and [75μs, 100μs]. The upper and lower limits and the midpoint value of each level are stored as adjustment reference points.

[0144] During operation, real-time time-domain waveform data of waveform parameters is acquired. Taking a pulse signal collected at a sampling frequency of 1MHz as an example, 1000 waveform sampling points are collected in each adjustment cycle to form a time-domain waveform sequence. Noise levels are calculated on the acquired waveform data, specifically by taking the standard deviation of the waveform's stable segment; for example, the noise level in a certain acquisition is 0.05V. Simultaneously, the fluctuation trend is calculated using the rate of change of the values ​​from the most recent 5 samples; for example, the current fluctuation trend is +0.02. Historical adjustment accuracy is calculated using the average error of the most recent 10 adjustments; for example, the current historical adjustment accuracy is 98.5%.

[0145] The calculated noise level, fluctuation trend, and historical adjustment accuracy are combined to construct a quality assessment vector [0.05, 0.02, 0.985]. A preset baseline feature vector is [0.03, 0.01, 0.95], representing the ideal reference value. The quality assessment vector is compared with the baseline feature vector, and the weighted sum of the deviations in each dimension is calculated as the comprehensive deviation value. Assume the currently calculated comprehensive deviation value is 0.15.

[0146] Based on the comprehensive deviation value, the gear splitting threshold correction coefficient and the gear merging threshold correction coefficient are calculated. When the deviation value is 0.15, according to the preset mapping relationship, the splitting threshold correction coefficient is 1.2, and the merging threshold correction coefficient is 0.85. The preset gear splitting baseline threshold is 0.25, and the merging baseline threshold is 0.1. After correction, the actual gear splitting threshold is 0.3 (0.25 × 1.2), and the gear merging threshold is 0.085 (0.1 × 0.85).

[0147] During real-time adjustment, the current adjustment level is determined. Assuming the current pulse width is 35 μs, the current level is [25 μs, 50 μs]. The parameter change rate within this level is calculated by taking the ratio of the adjustment amplitude to the number of adjustments in the last 5 adjustments within this level. Assuming the average adjustment amplitude of the last 5 adjustments is 5 μs and the number of adjustments is 10, the change rate is 0.5.

[0148] Since the calculated rate of change of 0.5 is greater than the gear splitting threshold of 0.3, the gear splitting operation will be triggered. The midpoint of the current gear [25μs, 50μs], 37.5μs, will be selected as the splitting point to generate two sub-gears [25μs, 37.5μs] and [37.5μs, 50μs], thereby improving the adjustment accuracy of this range.

[0149] In another scenario, assuming the current gear is [75μs, 100μs], the calculated rate of change is 0.08, which is less than the gear merging threshold of 0.085. Meanwhile, the parameter characteristics of the adjacent gear [50μs, 75μs] differ from the current gear by 0.03, which is less than the preset difference threshold of 0.05. In this case, the system will merge these two gears into [50μs, 100μs], reducing unnecessary adjustment precision and improving adjustment efficiency.

[0150] The rate of change of each adjustment level is continuously monitored, and the level division is dynamically adjusted according to the above rules. Taking a certain operation as an example, after 10 adjustment cycles, the initial 4 levels may evolve into new level distributions of [10μs, 25μs], [25μs, 37.5μs], [37.5μs, 50μs], and [50μs, 100μs]. The high rate of change area is given a more refined level division, while the low rate of change area is given a coarser level division.

[0151] In practical applications, the priority adjustment sequence is continuously optimized based on feedback from the adjustment results, such as increasing the priority of parameters with significant adjustment effects. The entire adjustment process forms a closed-loop feedback mechanism, continuously optimizing the waveform parameter adjustment strategy and improving the system's adaptability and adjustment efficiency to different application scenarios.

[0152] By using the above-mentioned dynamic adjustment method, the accuracy distribution is automatically adjusted according to actual needs during waveform parameter adjustment. More precise adjustment is provided in critical areas, and more efficient adjustment strategies are adopted in non-critical areas, thereby achieving a balance between the accuracy and efficiency of waveform parameter adjustment.

[0153] The radar training parameter dynamic adjustment system based on adaptive algorithm according to this invention includes:

[0154] The first unit is used to acquire the waveform parameters of the radar during the training process and construct a training state feature vector based on the waveform parameters.

[0155] The second unit is used to calculate the feature correlation matrix of the training state feature vector to obtain the initial connection strength matrix, generate modulation instructions based on the temporal correlation analysis of feature nodes, and perform weighted processing to obtain the feature modulation representation matrix, and calculate the parameter adjustment direction matrix based on the node state changes.

[0156] The third unit is used to map the parameter adjustment direction matrix to a multi-dimensional parameter feature space, obtain multiple subspace units through adaptive grid partitioning, and generate a priority adjustment sequence of waveform parameters based on parameter distribution characteristics;

[0157] The fourth unit is used to construct a set of adjustment actions for waveform parameters based on the priority adjustment sequence, divide the adjustment range of waveform parameters into multiple adjustment levels, execute adjustment actions and collect response characteristic curves, update the adjustment strategy selection matrix based on the response characteristic curves, and generate the next round of adjustment scheme.

[0158] The fifth unit is used to update the waveform parameters according to the adjustment scheme, write the updated waveform parameters into the radar control unit and trigger a new round of training tasks, and at the same time save the data records of this round of parameter adjustment process to the training database.

[0159] A third aspect of the present invention provides an electronic device, comprising:

[0160] processor;

[0161] Memory used to store processor-executable instructions;

[0162] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0163] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0164] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0165] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for dynamic adjustment of radar training parameters based on adaptive algorithm, characterized in that, The method comprises the following steps: acquiring waveform parameters of the radar in a training process, and constructing a training state feature vector based on the waveform parameters; calculating a feature correlation matrix of the training state feature vector to obtain an initial connection strength matrix, generating modulation instructions based on a time sequence correlation analysis of feature nodes, and performing weighted processing to obtain a feature modulation representation matrix, and calculating a parameter adjustment direction matrix based on a state change of the nodes; mapping the parameter adjustment direction matrix to a multi-dimensional parameter feature space, obtaining a plurality of subspace units through adaptive grid division, and generating a priority adjustment sequence of the waveform parameters based on parameter distribution characteristics; constructing an adjustment action set of the waveform parameters based on the priority adjustment sequence, dividing an adjustment range of the waveform parameters into a plurality of adjustment gears, executing the adjustment action, collecting a response characteristic curve, updating an adjustment strategy selection matrix based on the response characteristic curve, and generating a next round of adjustment scheme; updating the waveform parameters according to the adjustment scheme, writing the updated waveform parameters into a radar control unit and triggering a new round of training task, and saving data records of the parameter adjustment process of the current round to a training database.

2. The method of claim 1, wherein, The method comprises the following steps: calculating a feature correlation matrix of the training state feature vector to obtain an initial connection strength matrix, generating modulation instructions based on a time sequence correlation analysis of feature nodes, and performing weighted processing to obtain a feature modulation representation matrix, and calculating a parameter adjustment direction matrix based on a state change of the nodes; calculating a correlation coefficient between each dimension feature in the training state feature vector, constructing a feature correlation matrix, and calculating an initial connection strength matrix based on the numerical value of each element in the feature correlation matrix; performing a time sequence correlation analysis on each feature node in the initial connection strength matrix, extracting state information of adjacent feature nodes at different times, calculating a joint activation probability between the nodes, comparing the joint activation probability with a preset adaptive control threshold, and generating a reinforcement modulation instruction or an inhibition modulation instruction; determining an information transmission modulation direction according to the reinforcement modulation instruction and the inhibition modulation instruction, calculating an adaptive modulation intensity factor based on the real-time activity of each feature node, inputting the adaptive modulation intensity factor into a recursive nonlinear mapping unit to iteratively calculate a modulation coefficient; weighting the training state feature vector according to the modulation coefficient, inputting the training state feature vector into a neural network with memory function, and calculating a feature modulation representation matrix through the selective modulation of the neural network; 3. The method of claim 2, wherein, calculating a state change relationship between a previous feature node and a subsequent feature node in the feature modulation representation matrix, extracting a state transition rule to obtain a weight update direction, adjusting the initial connection strength matrix, and calculating a parameter adjustment direction matrix. The method comprises the following steps: constructing a feature modulation decision table based on the reinforcement modulation instruction and the inhibition modulation instruction, and obtaining an information transmission modulation direction by calculating the weight distribution of each element in the feature modulation decision table; The real-time activity of each feature node is input into a feature extraction layer of a multi-layer self-calibration network, and a node feature fingerprint sequence is obtained through convolution operation, wherein the feature fingerprint sequence contains time-varying characteristics of the node activity; According to the node feature fingerprint sequence, time correlation analysis is performed in a calibration calculation layer, historical modulation result data is divided according to a time window and local correlation is calculated, and a calibration deviation value is obtained through weighted accumulation; A calibration compensation factor is obtained by performing compensation operation on the calibration deviation value through a feedback adjustment layer, an initial modulation factor is calculated according to the information transmission modulation direction and the topological connection relationship between nodes, and the initial modulation factor and the calibration compensation factor are fused to obtain an adaptive modulation intensity factor; The adaptive modulation intensity factor is taken as an initial input, and iterative calculation is performed in a recursive nonlinear mapping unit, and when the output residual error of adjacent two iterations is less than a preset convergence threshold, a final modulation coefficient is output.

4. The method of claim 1, wherein, The parameter adjustment direction matrix is mapped to a multi-dimensional parameter feature space, a plurality of subspace units are obtained through adaptive grid division, and a priority adjustment sequence of the waveform parameter is generated based on the parameter distribution characteristics, including: The parameter adjustment direction matrix is mapped to a multi-dimensional parameter feature space, a plurality of subspace units are obtained through adaptive grid division, the probability density function of the parameter distribution in the subspace unit is calculated by using kernel density estimation, the grid boundary is adjusted according to the corresponding gradient information, and an initial space division structure is obtained; Based on the initial space division structure, the bandwidth of the kernel function is adaptively adjusted according to the discrete degree of the parameter distribution, and the probability density function is recalculated, and the parameter density time-varying characteristic value is obtained by weighted operation with a time decay factor, the local maximum value point and the jump point of the probability density function are extracted, and a space redivision feature point set is formed; Based on the space redivision feature point set, the density difference between adjacent subspace units is used to construct an optimization objective function, and the boundary position adjustment amount is iteratively calculated by using the gradient descent method with an inertia term and a momentum term, and when the density difference between adjacent subspace units is less than a preset difference threshold, an optimized space hierarchical structure is output. For the optimized space hierarchical structure, the importance evaluation index value is calculated by combining the probability density function, the parameter density time-varying characteristic value and the density difference, the importance weight of the subspace unit is determined, the priority adjustment sequence of the waveform parameter is generated based on the importance weight of the subspace unit, and the priority adjustment sequence of the waveform parameter is dynamically optimized and updated according to the parameter adjustment feedback result.

5. The method of claim 4, wherein, The step of adaptively adjusting the bandwidth of the kernel function and recalculating the probability density function based on the initial space division structure according to the discrete degree of the parameter distribution includes: Based on the parameter distribution in the subspace unit, a local density distribution matrix is constructed, the local density distribution matrix is feature-decomposed to obtain a feature value sequence, and the discrete degree of the parameter distribution is calculated; Based on the discrete degree, the upper limit and the lower limit are determined by combining the maximum density value and the minimum density value of the local density distribution matrix, and an initial adjustment interval is obtained. The value points of the plurality of kernel function bandwidths are divided in the initial adjustment interval, the dispersion degree is taken as a weight coefficient, and a weighted combination of a distribution entropy constraint value, a local density preservation value and a smoothing constraint value corresponding to each value point is respectively calculated to obtain a bandwidth optimization score; The kernel function bandwidth value with the highest bandwidth optimization score is selected as an optimal bandwidth value, and the kernel function bandwidth is adaptively adjusted according to the optimal bandwidth value, and the probability density function in the subspace unit is recalculated.

6. The method of claim 1, wherein, Based on the priority adjustment sequence, an adjustment action set of the waveform parameter is constructed, the adjustment range of the waveform parameter is divided into a plurality of adjustment gears, the adjustment action is executed and the response characteristic curve is collected, the adjustment strategy selection matrix is updated based on the response characteristic curve, and the next round of adjustment scheme is generated, including: Based on the priority adjustment sequence, an adjustment action set of the waveform parameter is constructed, the adjustment range of the waveform parameter is divided into a plurality of adjustment gears, the change rate of the current gear is calculated in real time, and the adjustment gears are dynamically adjusted according to the change rate; The adjustment action of the waveform parameter is executed, the response characteristic curve of the waveform parameter is collected in real time, the overshoot, adjustment time and steady-state error of the response characteristic curve are extracted, and the adjustment evaluation value is calculated in combination with the historical adjustment trajectory of the waveform parameter; Based on the adjustment evaluation value and the response characteristic curve, the adjustment strategy selection matrix is updated, the time sequence feature extraction method is used to analyze the transient response mode in the waveform parameter adjustment process, and the corresponding relationship between the transient response mode and the adjustment evaluation value is stored in the adjustment strategy selection matrix; Based on the response mode features stored in the adjustment strategy selection matrix, the adjustment target function is constructed in combination with the coupling constraint relationship of the waveform parameter, the optimal adjustment gear combination is solved by using a dynamic programming algorithm, and the next round of adjustment scheme is generated.

7. The method of claim 6, wherein, Based on the priority adjustment sequence, an adjustment action set of the waveform parameter is constructed, the adjustment range of the waveform parameter is divided into a plurality of adjustment gears, the change rate of the current gear is calculated in real time, and the adjustment gears are dynamically adjusted according to the change rate, including: Based on the priority adjustment sequence, an adjustment action set of the waveform parameter is constructed, the adjustment range of the waveform parameter is divided into a plurality of adjustment gears; The time-domain waveform of the waveform parameter is obtained, the corresponding noise level, fluctuation trend and historical adjustment accuracy are calculated, and a quality evaluation vector is constructed by combination; The quality evaluation vector is compared with a preset reference feature vector, the gear splitting threshold correction coefficient and the gear merging threshold correction coefficient are calculated according to the deviation value of the comparison, and the gear splitting threshold and the gear merging threshold are obtained by multiplying the preset reference threshold respectively; Based on the adjacent adjustment nodes in the adjustment gears, the current gear is determined, the change rate of the current gear is calculated in real time, two sub-gears are generated by selecting the midpoint of the current gear when the change rate is greater than or equal to the gear splitting threshold, and the adjacent gears with a parameter difference less than a preset difference value are merged when the change rate is less than the gear merging threshold, so that the adjustment gears are dynamically optimized.

8. A radar training parameter dynamic adjustment system based on an adaptive algorithm for implementing the method of any of the preceding claims 1-7, characterized in that, It includes: A first unit is configured to obtain a waveform parameter of a radar in a training process, and construct a training state feature vector based on the waveform parameter; The second unit is configured to calculate a feature correlation matrix of the training state feature vector to obtain an initial connection strength matrix, generate a modulation instruction according to a time sequence correlation analysis of the feature nodes, and perform weighted processing to obtain a feature modulation representation matrix, and calculate a parameter adjustment direction matrix according to a node state change; The third unit is configured to map the parameter adjustment direction matrix to a multi-dimensional parameter feature space, obtain a plurality of subspace units through adaptive grid division, and generate a priority adjustment sequence of the waveform parameters based on parameter distribution features; The fourth unit is configured to construct an adjustment action set of the waveform parameters based on the priority adjustment sequence, divide an adjustment range of the waveform parameters into a plurality of adjustment gears, execute the adjustment action and collect a response characteristic curve, update an adjustment strategy selection matrix based on the response characteristic curve, and generate a next round of adjustment scheme; The fifth unit is configured to update the waveform parameters according to the adjustment scheme, write the updated waveform parameters into a radar control unit and trigger a new round of training task, and save data records of the parameter adjustment process in the current round to a training database.

9. An electronic device, comprising: comprise: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to invoke the instructions stored in the memory to execute the method of any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon computer program instructions, wherein, The computer program instructions, when executed by the processor, implement the method of any one of claims 1 to 7.

Images