Dynamic iteration based radar system integrity assessment parameter adaptive optimization method

Abstract

The application provides a radar system integrity evaluation parameter adaptive optimization method based on dynamic iteration, relates to the technical field of radar system evaluation, and comprises the following steps: collecting radar system operation data in real time, performing dimension reduction processing to extract key evaluation parameters and performance benchmark values; using performance deviation as a state variable and a parameter value range as an action variable to optimize the evaluation parameters; and initializing a sliding time window to calculate confidence and cluster, extract time sequence change characteristics, and dynamically correct the evaluation parameters, so that the adaptive optimization of the evaluation parameters is realized, and the accuracy and timeliness of radar system integrity evaluation are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of radar system evaluation, and in particular to a radar system integrity evaluation parameter adaptive optimization method based on dynamic iteration. BACKGROUND

[0002] The radar system is an important detection and monitoring device in modern military and civilian fields, and the integrity of the radar system directly affects the detection efficiency and task execution success rate. The traditional radar system integrity evaluation mainly relies on fixed threshold judgment, artificial experience analysis and periodic maintenance detection. With the improvement of the complexity and functional diversity of the radar system, the radar operating state presents the characteristics of high dimensionality, nonlinearity and dynamic change.

[0003] However, the existing radar system integrity evaluation technology still has the problems of lacking adaptive adjustment ability for different radar working modes and environmental conditions, leading to inaccurate evaluation results in complex electromagnetic environments and variable task scenarios, being unable to timely reflect the change trend of the actual operating state of the radar, being difficult to effectively extract and utilize the deep performance degradation characteristics and rules contained in the historical data, lacking effective learning and prediction ability for the long-term operating state of the radar system, and being unable to optimize in time according to the performance changes of the radar system and the influence of environmental factors, resulting in lagging evaluation results or misjudgment. SUMMARY

[0004] The radar system integrity evaluation parameter adaptive optimization method based on dynamic iteration provided by the embodiments of the present application can at least solve some of the problems existing in the prior art.

[0005] In a first aspect, the radar system integrity evaluation parameter adaptive optimization method based on dynamic iteration comprises:

[0006] Real-time collection of current operating data and historical operating data of the radar system, conversion of the historical operating data into a multi-dimensional evaluation parameter space and execution of convolution operation and pooling operation to obtain reduced dimension data, extraction of key evaluation parameters from the reduced dimension data and calculation of radar performance benchmark values;

[0007] Comparison of the current operating data with the radar performance benchmark values to generate performance deviation, taking the performance deviation as a state variable and taking the value range corresponding to the key evaluation parameters as an action variable, calculation of a reward value and optimization of a radar training network based on the reward value to obtain integrity evaluation parameters;

[0008] The initialization sliding time window and the confidence corresponding to the integrity evaluation parameter in each sliding time window are calculated, the integrity evaluation parameter is clustered based on the confidence, and a weighted evaluation value is calculated based on the clustering result; the time sequence variation feature of the weighted evaluation value is extracted, and the integrity evaluation parameter is dynamically corrected to obtain an integrity evaluation result.

[0009] In an optional implementation,

[0010] The historical operation data is converted into a multi-dimensional evaluation parameter space, and a convolution operation and a pooling operation are performed to obtain reduced dimension data; key evaluation parameters are extracted from the reduced dimension data, and a radar performance benchmark value is calculated.

[0011] The historical operation data is mapped to a multi-dimensional evaluation parameter space to obtain multi-dimensional evaluation parameters; the multi-dimensional evaluation parameters are used as vertices, and the correlation between the multi-dimensional evaluation parameters is used as edges; a splicing operation and an activation are performed on the multi-dimensional evaluation parameters to determine edge weight values, and an initial dynamic heterogeneous graph is obtained.

[0012] Based on the pre-set learnable parameters and the multi-dimensional evaluation parameters, a graph attention coefficient is calculated by combining a pre-set bidirectional linear rectifier function; the feature values of the vertices are updated based on the graph attention coefficient and the multi-dimensional evaluation parameters to obtain an optimized dynamic heterogeneous graph, and a convolution operation and a pooling operation are performed on the updated feature values of the vertices to obtain first reduced dimension data.

[0013] The time sequence correlation coefficient between the multi-dimensional evaluation parameters is calculated, the edge weight values are updated based on the time sequence correlation coefficient, the optimized dynamic heterogeneous graph is reconstructed using the updated edge weight values to obtain a reconstructed dynamic heterogeneous graph, and important features are extracted from the reconstructed dynamic heterogeneous graph based on a pre-set importance threshold to obtain second reduced dimension data.

[0014] The first reduced dimension data and the second reduced dimension data are linearly weighted and fused to obtain feature fusion data, principal component analysis is performed on the feature fusion data to determine key evaluation parameters, and a radar performance benchmark value is solved.

[0015] In an optional implementation,

[0016] The principal component analysis is performed on the feature fusion data to determine the key evaluation parameters, and the radar performance benchmark value is solved.

[0017] An adaptive weight matrix is obtained by calculating the Euclidean distance between feature vectors in the feature fusion data and an adaptive bandwidth parameter, the feature fusion data is projected based on a pre-set projection matrix to obtain projected data, the projected data is reconstructed to obtain reconstructed data, and a reconstruction error between the reconstructed data and the feature fusion data is calculated, and a target optimization value is obtained based on the reconstruction error and the adaptive weight matrix;

[0018] Principal components and corresponding eigenvalues are obtained by performing principal component analysis on the projected data, a variance contribution rate of the principal components is calculated based on the eigenvalues, and a key evaluation parameter is determined in combination with a pre-set variance contribution rate threshold;

[0019] An inter-class dispersion matrix and an intra-class dispersion matrix corresponding to the projected data are calculated, a Fisher discriminant value is determined based on the inter-class dispersion matrix and the intra-class dispersion matrix, and a discriminability of the key evaluation parameter is calculated based on the Fisher discriminant value and a projection coefficient in the projection matrix;

[0020] The projection matrix is iteratively updated by an alternating optimization strategy based on the target optimization value and the Fisher discriminant value, an importance coefficient of the key evaluation parameter is calculated based on the updated projection matrix and the discriminability, and a radar performance benchmark value is obtained based on the importance coefficient.

[0021] In an optional implementation,

[0022] The current running data is compared with the radar performance benchmark value to generate a performance deviation, the performance deviation is taken as a state variable, a value range of the key evaluation parameter is taken as an action variable, a reward value is calculated, and a radar training network is optimized based on the reward value to obtain an integrity evaluation parameter, including:

[0023] A difference between the current running data and the radar performance benchmark value in multiple performance indicators is calculated to obtain a performance deviation, the performance deviation is constructed as a performance deviation node, the key evaluation parameter is constructed as an evaluation parameter node, a causal relationship between the performance deviation node and the evaluation parameter node is determined, a directed acyclic graph is constructed based on the causal relationship, and the performance deviation is converted into a state variable based on the directed acyclic graph;

[0024] A value range of the key evaluation parameter is determined based on the state variable, and multiple adjustment values are selected as action variables, a causal sensitivity matrix is constructed by a gradient calculation method, and an intervention step is determined, a prediction confidence is determined by intervention operation on the state variable based on the intervention step, a decision space is determined by Monte Carlo sampling, and an evaluation decision is determined in combination with the prediction confidence;

[0025] Based on the evaluation decision, an average processing effect value is calculated, and a reward value is calculated in combination with a performance change amount of the state variable, the reward value is input into a preset radar training network to calculate a cumulative reward expectation, and parameters of the radar training network are optimized according to the cumulative reward expectation to obtain an optimized radar training network, the reliability weight of each key evaluation parameter is determined based on the optimized radar training network and the causal relationship, and the integrity evaluation parameter is calculated.

[0026] In an optional implementation,

[0027] A causal sensitivity matrix is constructed by a gradient calculation method, and an intervention step is determined, a prediction confidence is determined by performing intervention operation on the state variable based on the intervention step, a decision space is determined by Monte Carlo sampling, and an evaluation decision is determined in combination with the prediction confidence, including:

[0028] A causal sensitivity matrix is constructed by analyzing the change relationship between the state variable and the action variable through a gradient calculation method, and an intervention step of each action variable is generated, the state variable is intervened by an intervention operation, an intervention state variable value is obtained, and a deviation analysis is performed on the intervention state variable value and a preset target state value to obtain an intervention effect value;

[0029] A prediction confidence interval is determined based on the intervention state variable value by using an interval estimation method, a prediction uncertainty is generated according to a characteristic parameter of the prediction confidence interval and a historical sample number obtained in advance, and a prediction confidence is determined based on the prediction uncertainty and a time decay characteristic of the intervention effect value;

[0030] A historical experience vector is extracted based on historical decision data obtained in advance, and a decision space is determined, a matching analysis is performed on the intervention state variable value and the historical experience vector by a similarity measurement method, and a similarity value of the prediction result and the historical experience is generated by a distance mapping mechanism;

[0031] A candidate decision set is extracted in the decision space by a Monte Carlo sampling method, the candidate decisions are screened according to the prediction confidence, and an effective candidate decision is determined according to the similarity value, an execution effect of each effective candidate decision is determined according to the intervention effect value, and an effective candidate decision with an optimal execution effect is selected as an evaluation decision and output.

[0032] In an optional implementation,

[0033] A sliding time window is initialized, and a confidence corresponding to the integrity evaluation parameter is calculated in each sliding time window, the integrity evaluation parameter is clustered based on the confidence, and a weighted evaluation value is calculated based on the clustering result, a time sequence change characteristic of the weighted evaluation value is extracted, and the integrity evaluation parameter is dynamically corrected to obtain an integrity evaluation result, including:

[0034] initializing a sliding time window and collecting integrity evaluation parameters in the sliding time window and calculating mutual information and information entropy between different integrity evaluation parameters, calculating parameter correlation weight according to the mutual information and the information entropy, extracting state feature indexes of the integrity evaluation parameters and calculating the state feature indexes to determine confidence of the integrity evaluation parameters;

[0035] calculating offset of the integrity evaluation parameters relative to pre-set clustering centers in a gradient iteration manner and updating the clustering centers according to the correlation weight, and clustering the integrity evaluation parameters according to the updated clustering centers to obtain a plurality of clustering groups;

[0036] calculating variance of the integrity evaluation parameters in the clustering groups and determining in-group weight of the integrity evaluation parameters based on the variance, and calculating weighted evaluation values in combination with clustering group weight corresponding to the clustering groups;

[0037] calculating difference of the weighted evaluation values in adjacent windows to obtain adjacent change features, calculating global features of the weighted evaluation values based on time sequence change of the weighted evaluation values, calculating correction amount based on the adjacent change features and the global features and dynamically correcting the integrity evaluation parameters, calculating corrected deviation and updating correction coefficient based on the deviation and a decay coefficient, and obtaining integrity evaluation results.

[0038] In an optional implementation,

[0039] calculating offset of the integrity evaluation parameters relative to pre-set clustering centers in a gradient iteration manner and updating the clustering centers according to the correlation weight, and clustering the integrity evaluation parameters according to the updated clustering centers to obtain a plurality of clustering groups includes:

[0040] collecting state values of the integrity evaluation parameters and constructing a space-time feature matrix based on the state values, determining time sequence weight coefficients based on the space-time feature matrix, and calculating weighted offset between the integrity evaluation parameters and pre-set clustering centers to obtain time sequence weighted offset;

[0041] extracting evolution trend features of the integrity evaluation parameters and combining the time sequence weighted offset to obtain comprehensive offset features, updating the clustering centers based on the comprehensive offset features, constructing dynamic grouping boundaries based on the updated clustering centers, determining boundary deformable parameters according to distance distribution of the state values to the updated clustering centers, and calculating grouping attribution degrees based on the dynamic grouping boundaries and the boundary deformable parameters;

[0042] The continuity constraint is performed on the group attribution, and an initial clustering group result is obtained by smoothing the continuity constraint result through a preset timing smoothing factor; a stability index of the initial clustering group result is calculated; if the stability index is greater than a preset stability threshold, the initial clustering group result is taken as a clustering group output; otherwise, a correction direction is calculated based on a change trend of the stability index, and the clustering center is updated along the correction direction until the stability index is greater than the stability threshold, and a plurality of clustering groups are obtained.

[0043] In a second aspect of the embodiment of the application, a radar system integrity evaluation parameter adaptive optimization system based on dynamic iteration is provided, comprising:

[0044] A first unit is configured to collect current operation data and historical operation data of a radar system in real time, convert the historical operation data into a multi-dimensional evaluation parameter space, perform convolution operation and pooling operation to obtain reduced dimension data, extract key evaluation parameters from the reduced dimension data, and calculate a radar performance benchmark value;

[0045] A second unit is configured to compare the current operation data with the radar performance benchmark value to generate a performance deviation, take the performance deviation as a state variable, take a value range corresponding to the key evaluation parameters as an action variable, calculate a reward value, and optimize a radar training network based on the reward value to obtain an integrity evaluation parameter;

[0046] A third unit is configured to initialize a sliding time window, calculate a confidence degree corresponding to the integrity evaluation parameter in each sliding time window, cluster the integrity evaluation parameter based on the confidence degree, calculate a weighted evaluation value based on the clustering result, extract a timing change feature of the weighted evaluation value, and dynamically correct the integrity evaluation parameter to obtain an integrity evaluation result.

[0047] In a third aspect of the embodiment of the application, an electronic device is provided, comprising:

[0048] A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke the instructions stored in the memory to perform the method described above.

[0049] In a fourth aspect of the embodiment of the application, a computer readable storage medium is provided, which stores computer program instructions, and the computer program instructions are executed by a processor to implement the method described above.

[0050] In the present application, the intelligentization and adaptive optimization of radar system integrity evaluation are realized by the radar system integrity evaluation parameter adaptive optimization method based on dynamic iteration, the accuracy and reliability of the evaluation result are improved, the system abnormality can be more accurately identified by comparing the performance benchmark value with the current running data, the dynamic adjustment and optimization of the evaluation parameter are realized, the evaluation process is more in line with the actual operation, the sliding time window and confidence calculation mechanism are introduced, the evaluation parameter is clustered and weighted, and the time sequence change characteristics are extracted for dynamic correction, so that the evaluation result is more comprehensive and stable, the real-time performance and adaptability of the radar system integrity evaluation are effectively improved, and more reliable technical support is provided for the maintenance and decision of the radar system. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 A flowchart of the radar system integrity evaluation parameter adaptive optimization method based on dynamic iteration of the embodiment of the present application is shown in

[0052] Figure 2 A parameter clustering dynamic evaluation flowchart of the radar system integrity evaluation parameter adaptive optimization method based on dynamic iteration of the embodiment of the present application is shown in DETAILED DESCRIPTION

[0053] To make the purpose, technical scheme and advantages of the embodiment of the present application clearer, the technical scheme in the embodiment of the present application will be described clearly and completely in combination with the drawings in the embodiment of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0054] The technical scheme of the present application will be described in detail in specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in some embodiments.

[0055] Figure 1 A flowchart of the method of the embodiment of the present application is shown in Figure 1 As shown, the method comprises:

[0056] Real-time acquisition of current running data and historical running data of the radar system, conversion of the historical running data into a multi-dimensional evaluation parameter space and execution of convolution operation and pooling operation to obtain reduced dimension data, extraction of key evaluation parameters from the reduced dimension data and calculation of radar performance benchmark value;

[0057] The current operation data is compared with the radar performance benchmark value to generate a performance deviation, the performance deviation is taken as a state variable, a value range corresponding to the key evaluation parameter is taken as an action variable, a reward value is calculated, and a radar training network is optimized based on the reward value to obtain the integrity evaluation parameter;

[0058] A sliding time window is initialized, and a confidence corresponding to the integrity evaluation parameter is calculated in each sliding time window. The integrity evaluation parameter is clustered based on the confidence, and a weighted evaluation value is calculated based on the clustering result. A time sequence variation feature of the weighted evaluation value is extracted, and the integrity evaluation parameter is dynamically corrected to obtain an integrity evaluation result.

[0059] In an optional implementation,

[0060] The historical operation data is converted into a multi-dimensional evaluation parameter space, and convolution operation and pooling operation are performed to obtain reduced dimension data. Key evaluation parameters are extracted from the reduced dimension data, and a radar performance benchmark value is calculated.

[0061] The historical operation data is mapped to a multi-dimensional evaluation parameter space to obtain multi-dimensional evaluation parameters. The multi-dimensional evaluation parameters are taken as vertices, and the correlation between the multi-dimensional evaluation parameters is taken as edges. The multi-dimensional evaluation parameters are spliced and activated to determine edge weight values to obtain an initial dynamic heterogeneous graph.

[0062] Based on the pre-set learnable parameters and the multi-dimensional evaluation parameters, a graph attention coefficient is calculated by combining a pre-set bidirectional linear rectifier function. The feature values of the vertices are updated based on the graph attention coefficient and the multi-dimensional evaluation parameters to obtain an optimized dynamic heterogeneous graph. Convolution operation and pooling operation are performed on the updated feature values of the vertices to obtain first reduced dimension data.

[0063] A time sequence correlation coefficient between the multi-dimensional evaluation parameters is calculated, and the edge weight values are updated based on the time sequence correlation coefficient. The optimized dynamic heterogeneous graph is reconstructed using the updated edge weight values to obtain a reconstructed dynamic heterogeneous graph. Important features are extracted from the reconstructed dynamic heterogeneous graph based on a pre-set importance threshold to obtain second reduced dimension data.

[0064] The first reduced dimension data and the second reduced dimension data are linearly weighted and fused to obtain feature fusion data. Principal component analysis is performed on the feature fusion data to determine key evaluation parameters, and a radar performance benchmark value is solved.

[0065] The historical operation data generated in the operation process of the radar system is collected, including signal strength, detection distance, azimuth angle error, interference suppression capability and other multi-dimensional original indexes. The collected historical operation data is mapped to a multi-dimensional evaluation parameter space after normalization processing to obtain multi-dimensional evaluation parameters, including but not limited to transmission power stability index, receiving sensitivity index, angle accuracy index, distance accuracy index and target resolution index. Each multi-dimensional evaluation parameter represents the evaluation result of a specific performance dimension of the radar system. The multi-dimensional evaluation parameters are used as vertices, and the correlation between the multi-dimensional evaluation parameters is used as edges. The characteristics of adjacent vertices are connected through splicing operation, and the edge weight value is determined by using a leaky rectified linear unit as an activation function to obtain an initial dynamic heterogeneous graph. Specifically, for two vertices, a first vertex and a second vertex, the feature vectors thereof are a first feature vector and a second feature vector, respectively. The two feature vectors are connected through splicing operation and the weight value of the edge is calculated through the activation function.

[0066] Based on the pre-set learnable parameters and multi-dimensional evaluation parameters, the graph attention coefficient is calculated by combining the pre-set bidirectional linear rectification function. The learnable parameters are a weight matrix, and the dimension is the output feature dimension multiplied by the input feature dimension, wherein the input feature dimension is the original feature dimension, and the output feature dimension is the target feature dimension. For each vertex, linear transformation is performed through the weight matrix, and for the vertices connected thereto, the attention coefficient is calculated. After calculating the attention coefficient, the feature value of the vertex is updated based on the graph attention coefficient and the multi-dimensional evaluation parameter to obtain an optimized dynamic heterogeneous graph. The update formula is to weight sum all the features of the vertices adjacent to a certain vertex, and the weight is the attention coefficient. Convolution operation is performed on the updated vertex feature value, a 3*3 convolution kernel is used to extract local features, and then maximum pooling operation is performed, the pooling window size is 2*2 and the step is 2, which reduces the feature dimension by 75% to obtain first reduced dimension data.

[0067] A time series correlation coefficient between the multi-dimensional evaluation parameters is calculated, a historical change trend of the same parameter is analyzed with a time window of 30 days, a Pearson correlation coefficient between the parameters is calculated, and the time series correlation coefficient between the transmit power stability index and the receive sensitivity index is 0.82, for example, indicating that the two have a strong positive correlation. The edge weight value is updated based on the time series correlation coefficient. When the absolute value of the correlation coefficient is greater than 0.7, the edge weight value is enhanced by 30%. When the absolute value of the correlation coefficient is between 0.4 and 0.7, the edge weight value remains unchanged. When the absolute value of the correlation coefficient is less than 0.4, the edge weight value is weakened by 50%. The optimized dynamic heterogeneous graph is reconstructed using the updated edge weight value to obtain a reconstructed dynamic heterogeneous graph. The second reduced dimension data is obtained by extracting important features from the reconstructed dynamic heterogeneous graph in combination with a preset importance threshold. The importance threshold is set to 0.65. When the edge weight value is greater than the threshold, the edge and the vertex features connected thereto are retained; otherwise, they are pruned and removed. After the pruning operation, the feature dimension is further reduced by about 60% to obtain the second reduced dimension data.

[0068] The first reduced dimension data and the second reduced dimension data are linearly weighted and fused to obtain feature fusion data. The fusion weight coefficients are 0.4 and 0.6, respectively, that is, the calculation formula of the feature fusion data is: 0.4*first reduced dimension data+0.6*second reduced dimension data. Principal component analysis is performed on the feature fusion data, and the first few principal components with a cumulative contribution rate of 95% are selected to determine the key evaluation parameters. In an actual application case, it is found through principal component analysis that the contribution rates of the transmit power stability index, the receive sensitivity index, and the angle accuracy index are 45%, 30%, and 20%, respectively, and the cumulative contribution rate reaches 95%. These three parameters are determined as the key evaluation parameters, and the key evaluation parameters are weighted and averaged according to their contribution proportions in the principal component analysis to obtain the radar performance benchmark value.

[0069] In this embodiment, by dynamically updating the edge weight value and the vertex feature value in the graph structure, the complex nonlinear dependence relationship between the evaluation parameters and its time series change characteristics can be adaptively captured, the sensitivity and representation ability to the running state change are improved, the fidelity of the local structure features and the global time series features is considered through the dual-channel dimension reduction and linear weighted fusion, the key evaluation parameters are more representative and discriminative, the precision and stability of the radar performance evaluation are improved, and the high-reliability identification and dynamic optimization of the radar performance benchmark value under complex running states are realized.

[0070] In an alternative embodiment,

[0071] Performing principal component analysis on the feature fusion data to determine the key evaluation parameters and solving the radar performance benchmark value includes:

[0072] An adaptive weight matrix is obtained by calculating the Euclidean distance between feature vectors in the feature fusion data and an adaptive bandwidth parameter, the feature fusion data is projected based on a preset projection matrix to obtain projected data, the projected data is reconstructed to obtain reconstructed data, and a reconstruction error between the reconstructed data and the feature fusion data is calculated, and a target optimization value is obtained based on the reconstruction error and the adaptive weight matrix;

[0073] Principal components and corresponding eigenvalues are obtained by performing principal component analysis on the projected data, a variance contribution rate of the principal components is calculated based on the eigenvalues, and a key evaluation parameter is determined in combination with a preset variance contribution rate threshold;

[0074] An inter-class dispersion matrix and an intra-class dispersion matrix corresponding to the projected data are calculated, a Fisher discriminant value is determined based on the inter-class dispersion matrix and the intra-class dispersion matrix, and a degree of discrimination of the key evaluation parameter is calculated based on the Fisher discriminant value and a projection coefficient in the projection matrix;

[0075] The projection matrix is iteratively updated by an alternating optimization strategy based on the target optimization value and the Fisher discriminant value, an importance coefficient of the key evaluation parameter is calculated based on the updated projection matrix and the degree of discrimination, and a radar performance benchmark value is obtained based on the importance coefficient.

[0076] An adaptive weight matrix is obtained by calculating the Euclidean distance between feature vectors in the feature fusion data and an adaptive bandwidth parameter, for any two feature vectors in the feature fusion data, the Euclidean distance is calculated, that is, the square sum of the difference between the corresponding elements of the two vectors is square rooted. The adaptive bandwidth parameter is obtained by local density estimation, and the bandwidth parameter can be automatically adjusted for different working states of the radar system. For example, in a normal working state, the bandwidth parameter takes a value of 0.3; in a high load state, the bandwidth parameter takes a value of 0.5. The calculated Euclidean distance is substituted into a Gaussian kernel function to obtain an attenuation coefficient, forming an adaptive weight matrix.

[0077] The feature fusion data is projected based on a preset projection matrix to obtain projected data. An initial value of the projection matrix is generated by a random initialization method, and has a dimension of a number of features of the feature fusion data multiplied by a dimension of a target dimension reduction space. For example, for fusion data including 20 features, if the target dimension reduction is to an 8-dimensional space, the projection matrix has a dimension of 20*8=160. The projection process is implemented by multiplication of the feature fusion data and the projection matrix. The projected data is reconstructed to obtain reconstructed data. The reconstruction process uses multiplication of a transpose of the projection matrix and the projected data. A reconstruction error between the reconstructed data and the feature fusion data is calculated, and a mean square error is used as a measurement, i.e., a sum of squares of differences between elements of the reconstructed data and the original feature fusion data is divided by a total number of elements. For example, in a certain radar system instance, the calculated reconstruction error is 0.083, indicating a high reconstruction quality. A target optimization value is obtained based on the reconstruction error and an adaptive weight matrix. The target optimization value is calculated by weighted summation of the reconstruction error and the adaptive weight matrix, with a weight coefficient of 0.7, i.e., the target optimization value is equal to 0.7*reconstruction error+0.3*norm of the adaptive weight matrix.

[0078] Principal component analysis is performed on the projected data to obtain principal components and corresponding eigenvalues. The principal component analysis is implemented by calculating a covariance matrix of the projected data, and then performing eigenvalue decomposition on the covariance matrix. The eigenvalues are sorted in descending order, and corresponding eigenvectors are the principal components. A variance contribution rate of the principal components is calculated based on the eigenvalues, and the variance contribution rate is equal to a single eigenvalue divided by a sum of all eigenvalues. A key evaluation parameter is determined in combination with a preset variance contribution rate threshold. The variance contribution rate threshold is set to 85%, and when a cumulative variance contribution rate exceeds the threshold, an original evaluation parameter associated with the corresponding principal component is determined as the key evaluation parameter. For example, it is found through principal component analysis that the variance contribution rates of the first three principal components are 62%, 18%, and 7% respectively, and the cumulative contribution rate is 87%, which exceeds the preset threshold of 85%. Therefore, the corresponding antenna pattern accuracy index, transmit power stability index, and receiver noise coefficient index are determined as the key evaluation parameters.

[0079] The inter-class dispersion matrix and the intra-class dispersion matrix corresponding to the projection data are calculated. The inter-class dispersion matrix describes the difference between the evaluation results of different performance level radar systems, and is calculated by the weighted outer product sum of the mean vectors of different categories. The intra-class dispersion matrix describes the dispersion of samples within the same performance level, and is calculated by the weighted outer product sum of the difference between the samples within each category and the class mean. The Fisher discriminant value is determined based on the inter-class dispersion matrix and the intra-class dispersion matrix, and the Fisher discriminant value is equal to the trace of the inter-class dispersion matrix divided by the trace of the intra-class dispersion matrix. For example, in a certain radar system instance, the calculated Fisher discriminant value is 8.65, indicating that the key evaluation parameter has strong discrimination ability for different performance levels. The discrimination degree of the key evaluation parameter is calculated based on the Fisher discriminant value and the projection coefficients in the projection matrix. The discrimination degree is calculated by multiplying the Fisher discriminant value by the absolute value of the corresponding dimension coefficient in the projection matrix. The higher the discrimination degree, the more significant the influence of the parameter on the radar performance.

[0080] The projection matrix is iteratively updated based on the target optimization value and the Fisher discriminant value through an alternating optimization strategy, which includes fixing the Fisher discriminant value, optimizing the target optimization value, fixing the target optimization value, and optimizing the Fisher discriminant value. In the iteration process, the update step of the projection matrix is initially set to 0.01 and gradually decreases as the number of iterations increases. When the variation amplitude of the projection matrix in three consecutive iterations is less than 0.001, it is considered that the optimization converges, and the iteration is stopped. In practical applications, 30-50 iterations are usually required to achieve convergence. The importance coefficient of the key evaluation parameter is calculated based on the updated projection matrix and the discrimination degree. The importance coefficient is equal to the discrimination degree multiplied by the norm of the corresponding dimension of the updated projection matrix. For example, in a certain radar system instance, the importance coefficients of the antenna pattern accuracy index, the transmit power stability index, and the receiver noise coefficient index are 0.42, 0.35, and 0.22, respectively. The radar performance benchmark value is solved based on the importance coefficient, and the calculation method is the weighted sum of the normalized scores of each key evaluation parameter and the corresponding importance coefficient. In practical applications, the antenna pattern accuracy index score of a certain radar system is 89.5, the transmit power stability index score is 93.2, and the receiver noise coefficient index score is 90.8, and the calculated radar performance benchmark value is 90.9.

[0081] In the embodiment, by setting a learnable projection matrix and combining with a reconstruction error minimization constraint, the features maintain high fidelity to express the global structure relationship of the original data in the low-dimensional space, which significantly improves the representation quality and separability of the feature after dimensionality reduction. The Fisher discriminant value is obtained by calculating the inter-class dispersion and intra-class dispersion, which quantifies the feature distribution difference under different operating conditions, improves the parameter distinguishability and model discriminability, and updates the projection matrix through the alternating optimization of the target optimization value and the Fisher discriminant value, which realizes the dynamic balance between reconstruction accuracy and classification distinguishability, and improves the accuracy, stability and adaptive optimization ability of radar performance evaluation.

[0082] In an optional implementation,

[0083] Comparing the current operating data with the radar performance benchmark value generates a performance deviation, taking the performance deviation as a state variable, taking the value range corresponding to the key evaluation parameter as an action variable, calculating a reward value and optimizing the radar training network based on the reward value to obtain the integrity evaluation parameter, including:

[0084] Calculate the difference between the current operating data and the radar performance benchmark value to obtain the performance deviation, construct the performance deviation as a performance deviation node, construct the key evaluation parameter as an evaluation parameter node, determine the causal relationship between the performance deviation node and the evaluation parameter node, construct a directed acyclic graph based on the causal relationship, and convert the performance deviation into a state variable based on the directed acyclic graph;

[0085] Determine the value range of the key evaluation parameter based on the state variable and select a plurality of adjustment values as action variables, construct a causal sensitivity matrix by a gradient calculation method and determine an intervention step, determine a prediction confidence by performing intervention operation on the state variable based on the intervention step, determine a decision space by Monte Carlo sampling and determine an evaluation decision based on the prediction confidence;

[0086] Calculate the average processing effect value based on the evaluation decision and the performance change amount of the state variable to calculate the reward value, input the reward value into a preset radar training network to calculate the cumulative reward expectation, and optimize the parameters of the radar training network according to the cumulative reward expectation to obtain an optimized radar training network, determine the reliability weight of each key evaluation parameter based on the optimized radar training network and the causal relationship, and calculate the integrity evaluation parameter.

[0087] The current operating data of the radar system is obtained and compared with multiple performance indicators in the radar performance benchmark value, and the performance deviation is calculated. For example, the measured value of the antenna pattern accuracy index of the current radar system is 86.7, compared with the benchmark value 89.5, the deviation is -2.8; the measured value of the transmit power stability index is 90.1, compared with the benchmark value 93.2, the deviation is -3.1; the measured value of the receiver noise coefficient index is 88.9, compared with the benchmark value 90.8, the deviation is -1.9. The performance deviation is constructed as a performance deviation node, and the key evaluation parameters are constructed as evaluation parameter nodes. By analyzing the time sequence relationship between parameter changes and performance deviations in historical data, the causal relationship between the performance deviation node and the evaluation parameter node is determined. Based on the causal discovery algorithm, the changes of performance indicators before and after the changes of parameters are analyzed. For example, after the change of a certain evaluation parameter, the performance indicator changes accordingly, and the change has statistical significance, so it is considered that there is a causal relationship between the evaluation parameter and the performance indicator. Based on the determined causal relationship, a directed acyclic graph is constructed, in which the nodes represent evaluation parameters or performance indicators, and the directed edges represent the direction and strength of the causal relationship. Based on the directed acyclic graph, the performance deviation is converted into a state variable, which includes three dimensions of the amplitude, duration and change trend of the performance deviation.

[0088] Based on the state variable, the value range of the key evaluation parameter is determined. For the antenna pattern accuracy index, the value range is ±5%; for the transmit power stability index, the value range is ±3%; for the receiver noise coefficient index, the value range is ±4%. Within the value range of each parameter, five adjustment values are uniformly selected as action variables. For example, for the transmit power stability index, the selected adjustment values are 3% reduction, 1.5% reduction, no change, 1.5% increase, and 3% increase. A causal sensitivity matrix is constructed by a gradient calculation method, and the matrix elements represent the degree of influence of the change of the evaluation parameter on the performance indicator. In practical applications, the sensitivity of the antenna pattern accuracy index to the overall performance of the system is 0.46, the sensitivity of the transmit power stability index is 0.38, and the sensitivity of the receiver noise coefficient index is 0.25. Based on the sensitivity, the intervention step is determined. The higher the sensitivity, the smaller the intervention step, to ensure the stability of the adjustment. For the antenna pattern accuracy index, the intervention step is set to 1.5%; for the transmit power stability index, the intervention step is set to 2%; for the receiver noise coefficient index, the intervention step is set to 2.5%. Based on the intervention step, the state variable is intervened and operated, the changes of system performance under different adjustment schemes are predicted, and the prediction confidence is calculated. The prediction confidence is determined by the prediction accuracy in similar scenarios in the history, and the higher the confidence, the more reliable the prediction result. By the Monte Carlo sampling method, 1000 sample points are generated in the parameter space, the performance indicator values corresponding to each sample point are evaluated, and a decision space is formed. Combined with the prediction confidence, the adjustment scheme with the maximum performance improvement and the confidence exceeding 85% is selected in the decision space as the evaluation decision.

[0089] The average treatment effect value is calculated based on the evaluation decision, that is, the causal effect of a specific parameter adjustment on system performance is estimated. The average treatment effect value is calculated by comparing the expected changes in performance indicators before and after the parameter adjustment. The reward value is calculated by combining the performance change amount of the state variable, which is calculated by multiplying the performance improvement amplitude by the confidence and then by the timeliness coefficient. The timeliness coefficient represents the durability of the adjustment effect, and the value range is 0 to 1. The stronger the durability, the greater the coefficient. For example, in a certain radar system instance, the performance improvement of the antenna pattern accuracy index after adjustment is 2.3, the confidence is 88%, and the timeliness coefficient is 0.92. The calculated reward value is 1.86. The reward value is input into the preset radar training network, which is composed of three layers of fully connected neural networks. The number of nodes in the input layer is the dimension of the state variable, the number of nodes in the hidden layer is 32, and the number of nodes in the output layer is the number of action variables. The cumulative reward expectation is calculated through the radar training network, that is, the expected effect of long-term performance improvement. The parameters of the radar training network are optimized according to the cumulative reward expectation. The random gradient descent method is used, and the learning rate is initially set to 0.005 and gradually reduced as the training rounds increase. When the cumulative reward expectation changes of five consecutive training rounds are less than 0.5%, it is considered that the network optimization converges, and the optimized radar training network is obtained.

[0090] The reliability weight of each key evaluation parameter is determined based on the optimized radar training network and the causal relationship. The reliability weight is composed of two parts: the sensitivity of the parameter to the performance and the stability of the parameter adjustment. The higher the sensitivity, the better the stability of the adjustment effect, and the greater the reliability weight. The specific calculation method is the weighted average of the sensitivity and the stability coefficient, and the weight ratio is 7:3. For example, in a certain radar system instance, the sensitivity of the antenna pattern accuracy index is 0.46, the stability coefficient is 0.89, and the calculated reliability weight is 0.58; the reliability weight of the transmit power stability index is calculated to be 0.26; and the reliability weight of the receiver noise coefficient index is calculated to be 0.16. The integrity evaluation parameter is calculated according to the reliability weight, and the calculation method is the weighted deviation sum of the actual value and the reference value of each key evaluation parameter. The closer the integrity evaluation parameter to zero, the closer the system performance to the ideal state.

[0091] In this embodiment, the causal relationship between the performance deviation node and the key evaluation parameter node is used to construct a directed acyclic graph, the performance deviation relationship is converted into a quantifiable state variable, the causal explainable modeling of the radar operating state is realized, the result opacity problem caused by the dependence on empirical weight or black box model in the traditional method is avoided, the influence intensity of different evaluation parameters on the performance deviation is dynamically identified through the intervention calculation and causal sensitivity matrix construction based on the state variable, and the optimal intervention step is adaptively determined, the average processing effect and the performance change amount are converted into reward signals inputting the radar training network by introducing the reinforcement learning mechanism, the parameter configuration is optimized in the continuous interaction, the transformation from passive evaluation to active optimization is realized, the reliability weight of each key evaluation parameter is calculated based on the optimized training network and the causal relationship, the real contribution of different parameters to the integrity of the radar system is comprehensively reflected, and the explainability, dynamic adaptability and intelligent decision level of the radar performance integrity evaluation are significantly improved.

[0092] In an alternative embodiment,

[0093] The causal sensitivity matrix is constructed and the intervention step is determined by the gradient calculation method, the state variable is intervened and operated based on the intervention step to determine the prediction confidence, the decision space is determined by the Monte Carlo sampling, and the evaluation decision is determined in combination with the prediction confidence, including:

[0094] The causal sensitivity matrix is constructed and the intervention step of each action variable is generated by analyzing the change relationship between the state variable and the action variable through the gradient calculation method, the state variable is intervened and operated in combination with the counterfactual intervention method, the intervention state variable value is obtained, and the intervention effect value is obtained by analyzing the deviation between the intervention state variable value and the preset target state value;

[0095] The prediction confidence interval is determined based on the intervention state variable value by using the interval estimation method, the prediction uncertainty is generated according to the characteristic parameters of the prediction confidence interval and the historical sample number obtained in advance, and the prediction confidence is determined based on the prediction uncertainty and the time decay characteristics of the intervention effect value;

[0096] The historical experience vector is extracted based on the historical decision data obtained in advance, and the decision space is determined, the intervention state variable value and the historical experience vector are matched and analyzed by the similarity measurement method, and the similarity value of the prediction result and the historical experience is generated in combination with the distance mapping mechanism;

[0097] The candidate decision set is extracted in the decision space by the Monte Carlo sampling method, the candidate decisions are screened according to the prediction confidence, and the effective candidate decisions are determined according to the similarity value, the execution effect of each effective candidate decision is determined according to the intervention effect value, and the effective candidate decision with the optimal execution effect is selected as the evaluation decision output.

[0098] The causal sensitivity matrix is constructed by analyzing the change relationship between the state variables and the action variables through the gradient calculation method. A small disturbance is made to each action variable, and the change amplitude of the state variable is recorded. The change rate is calculated as the sensitivity value. For example, when the antenna pattern accuracy index is increased by 1%, the detection distance index is increased by 0.85%, and the corresponding sensitivity value is 0.85. For the key evaluation parameters of the radar system, the constructed causal sensitivity matrix contains multi-dimensional evaluation results. For example, in a certain radar system instance, the sensitivity of the antenna pattern accuracy index to the detection distance is 0.85, and the sensitivity to the angle accuracy is 0.73; the sensitivity of the transmit power stability index to the detection distance is 0.62, and the sensitivity to the clutter suppression capability is 0.41; the sensitivity of the receiver noise coefficient index to the signal-to-noise ratio is 0.88, and the sensitivity to the false alarm probability is 0.56. The intervention step of each action variable is generated based on the causal sensitivity matrix. The higher the sensitivity, the smaller the intervention step, to ensure the stability and accuracy of the adjustment. For the antenna pattern accuracy index, the intervention step is set to 0.8%; for the transmit power stability index, the intervention step is set to 1.2%; for the receiver noise coefficient index, the intervention step is set to 1%.

[0099] The intervention state variable value is obtained by combining the counterfactual intervention method for state variable intervention operation. The counterfactual intervention method modifies the value of a specific node while keeping other causal mechanisms unchanged, and predicts the response. In actual operation, the values of each evaluation parameter are adjusted in turn to predict the performance change. For example, when the antenna pattern accuracy index is increased by 2%, the detection distance is predicted to be increased by 1.7%, and the angle accuracy is predicted to be increased by 1.46%. The intervention effect value is obtained by deviation analysis of the intervention state variable value and the preset target state value. The target state value is set based on the radar system design specifications and task requirements, for example, the detection distance target value is 120 kilometers, and the angle accuracy target value is 0.2°. The intervention effect value is calculated as the weighted difference value of the intervention state variable value and the target state value. The weight is determined according to the importance of the index. For example, in a certain radar system instance, after intervention on the antenna pattern accuracy index, the intervention effect value is 0.32, indicating that the adjustment has a significant improvement on the performance.

[0100] The interval estimation method is used to determine the prediction confidence interval based on the intervention state variable value. Interval estimation calculates the mean and standard deviation of the intervention state variable, and determines the upper and lower limits combined with the confidence level. In practical applications, a confidence level of 95% is used, and the standard deviation is calculated based on historical data. For the adjusted detection distance of the antenna pattern accuracy index, the predicted value is 118.2 km, the standard deviation is 2.5 km, and the prediction confidence interval is 113.3 to 123.1 km. The prediction uncertainty is generated according to the characteristic parameters of the prediction confidence interval and the pre-acquired historical sample number. The characteristic parameters include interval width, skewness and kurtosis, and the historical sample number affects the reliability of the confidence interval. The prediction uncertainty is calculated as a function of the interval width and the sample number, and the more samples, the lower the uncertainty. In the optimization process of the radar system, based on 1200 historical samples, the prediction uncertainty of the antenna pattern accuracy adjustment is calculated to be 0.14. The prediction confidence is determined based on the prediction uncertainty and the time decay characteristics of the intervention effect value. The time decay characteristics represent the trend of the intervention effect weakening over time, and an exponential decay model is used to describe it. The prediction confidence is calculated as 1 minus the prediction uncertainty multiplied by the time decay coefficient. In the actual application, the time decay coefficient of the antenna pattern accuracy adjustment is 0.93, and the calculated prediction confidence is 0.79.

[0101] Based on the pre-acquired historical decision data, the historical experience vector is extracted and the decision space is determined. The historical experience vector contains the intervention measures and their effects in similar scenarios in the past, and the decision space is the set of all possible adjustment schemes. In the optimization practice of the radar system, 365 maintenance adjustment data in the past two years are collected to form a historical experience library. The intervention state variable value is matched and analyzed with the historical experience vector through the similarity measurement method. The cosine similarity algorithm is used to calculate the cosine value of the vector angle, and the closer the value is to 1, the more similar it is. The similarity value of the prediction result and the historical experience is generated combined with the distance mapping mechanism. The distance mapping converts the distance in the Euclidean space into the value in the similarity space, and the Gaussian kernel function is used to achieve it. In the adjustment case of a certain radar system, the similarity value of the intervention state variable value and the most similar case in the historical experience library is 0.88, indicating that it has high reference value.

[0102] The candidate decision set is extracted in the decision space by a Monte Carlo sampling method. The Monte Carlo sampling randomly generates 1000 parameter combination schemes, each of which contains adjustment values of multiple key evaluation parameters. The candidate decisions are screened according to the prediction confidence, and the screening standard is that the prediction confidence is greater than 0.75. In the radar system optimization example, 327 of the 1000 candidate schemes meet the screening conditions. The effective candidate decisions are determined according to the similarity value, and the scheme with a similarity value greater than 0.8 to the historical successful experience is selected as the effective candidate decision. After screening, 86 effective candidate decisions are obtained. The execution effect of each effective candidate decision is determined according to the intervention effect value, and the execution effect is calculated as the intervention effect value multiplied by the prediction confidence and then multiplied by the similarity value. In actual application, the scheme with the optimal execution effect in the 86 effective candidate decisions is that the antenna pattern accuracy index is improved by 1.5%, the transmitter power stability index is improved by 1%, and the receiver noise coefficient index is reduced by 0.8%. The execution effect value of the scheme is 0.283, and the prediction system overall performance is improved by 2.7%. The effective candidate decision with the optimal execution effect is selected as the evaluation decision output.

[0103] In this embodiment, the gradient calculation method is used to construct the causal sensitivity matrix, which can accurately depict the response relationship between the state variable and the action variable, realize the quantitative analysis and interpretable modeling of the key influencing factors, and combine the counterfactual intervention method to intervene in the operation of the state variable. In the absence of actual experiments, the potential performance under different intervention conditions is inferred, which significantly improves the reasoning ability and causal recognition accuracy. By introducing interval estimation and time decay mechanism, the uncertainty propagation and effect decay characteristics are considered in the prediction process, so that the prediction confidence is more consistent with the dynamic change characteristics in the actual running environment, and the robustness and credibility of the evaluation result are improved. Through similarity matching analysis of the intervention state variable value and the historical experience vector, combined with the distance mapping mechanism, auxiliary decision based on experience knowledge is realized, and the generalization ability in the case of sparse or abnormal historical data is improved.

[0104] In an alternative embodiment,

[0105] The sliding time window is initialized, and the confidence corresponding to the integrity evaluation parameter in each sliding time window is calculated. The integrity evaluation parameter is clustered based on the confidence, and a weighted evaluation value is calculated based on the clustering result. The time sequence variation characteristics of the weighted evaluation value are extracted, and the integrity evaluation parameter is dynamically corrected to obtain the integrity evaluation result, including:

[0106] initializing a sliding time window, collecting integrity evaluation parameters in the sliding time window, calculating mutual information and information entropy between different integrity evaluation parameters, calculating correlation weight between parameters according to the mutual information and the information entropy, extracting state feature indicators of the integrity evaluation parameters and calculating the state feature indicators to determine the confidence of the integrity evaluation parameters;

[0107] calculating the offset of the integrity evaluation parameters relative to the pre-set clustering center in a gradient iteration manner and updating the clustering center according to the correlation weight, clustering the integrity evaluation parameters according to the updated clustering center to obtain a plurality of clustering groups;

[0108] calculating the variance of the integrity evaluation parameters in the clustering group and determining the intra-group weight of the integrity evaluation parameters based on the variance, calculating a weighted evaluation value in combination with the clustering group weight corresponding to the clustering group;

[0109] calculating the difference of the weighted evaluation value in adjacent windows to obtain an adjacent change feature, calculating a global feature of the weighted evaluation value based on the time sequence change of the weighted evaluation value, calculating a correction amount based on the adjacent change feature and the global feature and dynamically correcting the integrity evaluation parameters, calculating the corrected deviation and updating the correction coefficient based on the deviation and a decay coefficient to obtain an integrity evaluation result.

[0110] initializing a sliding time window, the window size is set to 60 sampling periods and the sliding step is 10 sampling periods. Collecting integrity evaluation parameters in the sliding time window, including antenna pattern accuracy indicators, transmit power stability indicators, receiver noise coefficient indicators, signal processing gain indicators, target detection probability indicators, etc. Calculating mutual information and information entropy between different integrity evaluation parameters, the mutual information reflects the correlation between two parameters, and the information entropy represents the uncertainty of the parameter itself. Mutual information is calculated by the logarithmic ratio of the joint distribution and the marginal distribution of the parameters in the statistical window, and information entropy is calculated by the negative logarithmic expectation value of the parameter probability distribution. In practical applications, the mutual information value between the antenna pattern accuracy indicator and the transmit power stability indicator is 0.63, the information entropy of the antenna pattern accuracy indicator is 2.45, and the information entropy of the transmit power stability indicator is 2.28. Calculate the correlation weight between parameters according to the mutual information and the information entropy, and the calculation method is to divide the mutual information value by the geometric mean of the information entropy of the corresponding parameters. The correlation weight calculation result of the aforementioned two indicators is 0.25. Extract the state feature indicators of the integrity evaluation parameters, including mean, standard deviation, skewness, kurtosis, trend slope, etc. Calculate the confidence of the integrity evaluation parameters by calculating the state feature indicators, and the calculation method is a weighted combination of the feature indicators, and the weight is determined according to the performance of the historical data. The calculated value of the confidence of the antenna pattern accuracy indicator is 0.89, indicating that the reliability of the measured value of this indicator is high.

[0111] The integrity assessment parameters' offsets relative to pre-set cluster centers are calculated using a gradient iterative approach. Initial cluster centers are initialized using the k-means++ algorithm, with 3 clusters representing "good," "normal," and "abnormal" states. The distance from each integrity assessment parameter to its cluster center is calculated using Euclidean distance. Cluster centers are updated based on association weights, with the update method being the original cluster center plus the weighted average of all parameter offsets multiplied by the learning rate. The learning rate is initially set to 0.05 and gradually decreases with each iteration. For example, in a radar system instance, the initial antenna pattern accuracy value corresponding to the "good" state cluster centers is 95, which becomes 93.8 after twenty iterations, better reflecting the actual operating conditions. The integrity assessment parameters are then clustered based on the updated cluster centers, resulting in multiple cluster groups. The cluster grouping follows the minimum distance principle, assigning each parameter to the group containing the nearest cluster center. In practical applications, among 60 sampling points within a certain time window, 37 were assigned to the "good" group, 20 to the "normal" group, and 3 to the "abnormal" group.

[0112] The variance of the integrity assessment parameters within each cluster group is calculated; a smaller variance indicates higher consistency of parameters within the group. Taking the "Good" cluster group as an example, the variance within the group for antenna pattern accuracy is 2.64, and the variance for transmit power stability is 1.87. The weights of the integrity assessment parameters within each group are determined based on the variance, calculated by normalizing the inverse of the variance. A smaller variance results in a larger weight within the group. The calculated weight for antenna pattern accuracy is 0.32, and for transmit power stability is 0.45. A weighted assessment value is then calculated by combining the weights of the corresponding cluster groups. The cluster group weights are determined based on the number of samples in each group; a higher percentage of samples results in a larger weight. In the aforementioned example, the weight for the "Good" group is 0.62, for the "Normal" group it is 0.33, and for the "Abnormal" group it is 0.05. The weighted assessment value is obtained by multiplying each parameter's assessment value within the group by its corresponding weight and then summing the results. The calculated weighted assessment value is 88.7.

[0113] The difference between the weighted evaluation values of adjacent windows is obtained as the adjacent change feature. Taking two consecutive windows as an example, the weighted evaluation value of the previous window is 88.7, and the current window is 87.5, the difference is -1.2, indicating that the system performance has slightly decreased. The global feature of the weighted evaluation value is calculated based on the time sequence change of the weighted evaluation value. The global feature is obtained by trend analysis on the weighted evaluation values of the past five windows, including the average slope, the fluctuation amplitude and the periodicity feature. For example, in actual application, the weighted evaluation values of a radar system in the past five windows are 89.3, 88.9, 89.1, 88.7 and 87.5 respectively, and the calculated average slope is -0.45, indicating that the system performance shows a slow downward trend. The correction amount is calculated based on the adjacent change feature and the global feature. The correction amount is calculated by weighted combination of the adjacent change feature and the global feature, and the weight ratio is 4:6. When the global feature shows a stable downward trend, the correction amount will increase to improve the sensitivity; when the fluctuation is large but there is no obvious trend, the correction amount will decrease to increase the stability. The integrity evaluation parameter is dynamically corrected, and the correction method is to add the correction amount multiplied by the correction coefficient to the original parameter value. The initial value of the correction coefficient is 0.5, and then it is dynamically adjusted according to the correction effect.

[0114] The corrected deviation is calculated, which is the difference between the corrected evaluation value and the actual observation value. The correction coefficient is updated based on the deviation and the attenuation coefficient, and the update method is to subtract the deviation multiplied by the attenuation coefficient from the original correction coefficient. The attenuation coefficient is set to 0.1 to ensure smooth adjustment of the correction coefficient. If the corrected evaluation value is larger than the actual observation value, the correction coefficient decreases; otherwise, it increases. Through multiple iterations, the correction coefficient gradually converges to the optimal value. For example, in a certain radar system instance, after 10 iterations, the correction coefficient is adjusted from 0.5 to 0.42, and the average error between the corrected evaluation value and the actual observation value is reduced by 67%, obtaining the integrity evaluation result, including the weighted evaluation value, the performance trend judgment and the abnormal warning.

[0115] In this embodiment, the sliding time window mechanism is used to dynamically collect and continuously analyze the integrity evaluation parameters, which can capture the time sequence change feature of the radar running state in real time, significantly improve the timeliness and continuity of the integrity evaluation, introduce the mutual information and information entropy to calculate the correlation weight between parameters, realize the accurate quantification of the nonlinear dependence relationship between multi-dimensional evaluation parameters, effectively enhance the reliability and interpretability of parameter correlation modeling, use gradient iteration to update the cluster center, which can adaptively adjust the cluster structure according to the dynamic distribution characteristics of the parameters, improve the accuracy and stability of clustering, avoid the problem that the traditional static clustering method is sensitive to abnormal or mutant data, and realize the balance of importance of different levels of features through intra-group variance calculation and weighted evaluation value generation, so that the evaluation result can comprehensively reflect the local stability and global trend change.

[0116] Figure 2A parameter clustering dynamic evaluation flowchart of an embodiment of the present application based on a dynamic iteration radar system integrity evaluation parameter adaptive optimization method.

[0117] In an alternative embodiment,

[0118] The gradient iteration method is used to calculate the offset of the integrity evaluation parameter relative to the preset clustering center, and the clustering center is updated according to the correlation weight. The integrity evaluation parameter is clustered and grouped according to the updated clustering center to obtain a plurality of clustering groups, including:

[0119] The state value of the integrity evaluation parameter is collected, and a space-time feature matrix is constructed based on the state value. The time sequence weight coefficient is determined based on the space-time feature matrix, and the offset between the integrity evaluation parameter and the preset clustering center is weighted and calculated to obtain a time sequence weighted offset.

[0120] The evolution trend feature of the integrity evaluation parameter is extracted, and the time sequence weighted offset is combined to obtain a comprehensive offset feature. The clustering center is updated based on the comprehensive offset feature. The dynamic grouping boundary is constructed based on the updated clustering center. The boundary deformable parameter is determined according to the distance distribution of the state value to the updated clustering center. The grouping attribution degree is calculated based on the dynamic grouping boundary and the boundary deformable parameter.

[0121] The grouping attribution degree is continuously constrained, and the continuous constraint result is smoothed by a preset time sequence smoothing factor to obtain an initial clustering grouping result. The stability index of the initial clustering grouping result is calculated. If the stability index is greater than a preset stability threshold, the initial clustering grouping result is output as a clustering group. Otherwise, the correction direction is calculated based on the change trend of the stability index, and the clustering center is updated along the correction direction until the stability index is greater than the stability threshold, and a plurality of clustering groups are obtained.

[0122] The state values of the integrity evaluation parameters are collected, including the antenna pattern accuracy index, the transmitting power stability index, the receiver noise coefficient index, and other radar key performance indicators. The collection period is set to be once per hour, and 72 time series data points of each parameter are obtained by continuous collection for 72 hours. A space-time feature matrix is constructed based on the state values, with the rows representing different time points, the columns representing different evaluation parameters, and the matrix elements being the parameter state values at the corresponding time points. In practical applications, the space-time feature matrix has a dimension of 72*8, representing 8 integrity evaluation parameters at 72 time points. Time series weight coefficients are determined based on the space-time feature matrix, and a sliding window weighted average method is used for calculation. For each time point, the data closer to the current time has a higher weight, and an exponential decay function is used to determine the weight value. In practical applications, the weight coefficient of the data within the last 4 hours is 0.6, the weight coefficient of the data within 4-24 hours is 0.3, and the weight coefficient of the data before 24 hours is 0.1. The weighted offset is calculated by weighting the offset between the integrity evaluation parameters and the preset clustering center. The preset clustering center is initialized as three categories: excellent (90 points or more), good (75-90 points), and abnormal (75 points or less). The offset is calculated as the difference between the parameter state value and each clustering center, and the time series weighted offset is obtained by weighting through the time series weight coefficient.

[0123] The evolution trend feature of the integrity evaluation parameters is extracted, and the trend of the parameter state value is analyzed by piecewise linear regression. The 72-hour data is divided into three 24-hour segments, and the slope of each segment is calculated to form a trend vector. For example, in a certain radar system, the trend slopes of the antenna pattern accuracy index in the three time periods are -0.08, -0.12, and -0.21, respectively, indicating that the performance decline trend is gradually intensifying. The comprehensive offset feature is obtained by combining the time series weighted offset, and the calculation method is the weighted sum of the time series weighted offset and the trend feature, with a weight ratio of 7:3. The clustering center is updated based on the comprehensive offset feature, and the update formula is the original clustering center plus the comprehensive offset feature multiplied by the learning rate. The learning rate is set to 0.15 to ensure smooth updating of the clustering center. For example, the antenna pattern accuracy index clustering center of the excellent category is adjusted from the initial value of 95 points to 92.3 points, which is more consistent with the actual system state. The dynamic grouping boundary is constructed based on the updated clustering center, and a soft boundary definition method is used, with the boundary width being adaptively adjusted according to the data distribution density. The boundary deformable parameter is determined according to the distance distribution of the state value to the updated clustering center, and the standard deviation of each distance value is calculated. The larger the standard deviation, the wider the boundary. In practical applications, the standard deviation of the distance distribution of the antenna pattern accuracy index is 3.45, and the corresponding boundary deformable parameter is 1.8. The grouping attribution degree is calculated based on the dynamic grouping boundary and the boundary deformable parameter. The attribution degree represents the probability that the parameter belongs to a certain group, and the distance is mapped to an attribution degree value between 0 and 1 through a Gaussian kernel function.

[0124] The continuity constraint of the group attribution is ensured to ensure that the group attribution of adjacent time points does not change too drastically, and the continuity constraint is realized by introducing a time smoothing penalty term, and if the attribution of two adjacent time points changes by more than 0.25, a penalty adjustment is performed. The initial clustering grouping result is obtained by smoothing the continuity constraint result through a pre-set time smoothing factor. The time smoothing factor is set to 0.65, which means that the current grouping result is 65% affected by the previous moment and 35% determined by the current measurement. The stability index of the initial clustering grouping result is calculated, and the stability index is defined by the ratio of the intra-group variance to the inter-group distance. The smaller the intra-group variance, the greater the inter-group distance, and the higher the stability index. For example, in a certain radar system instance, the calculated stability index is 0.78. If the stability index is greater than the pre-set stability threshold, the initial clustering grouping result is taken as the clustering group output. The stability threshold is set to 0.75, so the grouping result is considered stable and can be directly output. If the stability index is less than the stability threshold, the correction direction is calculated based on the change trend of the stability index. The correction direction is determined by analyzing the gradient of the stability index to each element of the clustering center, and adjusting the clustering center in the gradient rising direction can improve the stability. The clustering center is updated along the correction direction, and the update step is 0.08 times the difference between the current stability index and the threshold. Iterative updating is performed until the stability index is greater than the stability threshold, and a plurality of clustering groups are obtained.

[0125] In the embodiment, the offset between the parameters and the clustering center is weighted and calculated by introducing a time weight coefficient, which effectively enhances the sensitivity of the clustering process to the historical change trend and the time-dependent feature, so that the clustering result can reflect the true law of the evolution of the parameters with time. By extracting the evolution trend feature and updating the clustering center combined with the comprehensive offset feature, the adaptive adjustment of the clustering center in the time sequence dimension is realized, which significantly improves the adaptability of the clustering model to non-stationary data. Based on the introduction of the dynamic grouping boundary and the deformable parameter, the clustering boundary can automatically adjust the shape according to the state value distribution, which can effectively solve the problem of insufficient clustering accuracy in the transition area of the traditional clustering algorithm. Through the joint processing of the continuity constraint and the time smoothing factor, the time sequence consistency and robustness of the clustering result are enhanced, and the clustering jump caused by short-term fluctuations is avoided.

[0126] In a second aspect, the embodiment of the present application provides a system, comprising:

[0127] The first unit is configured to collect current operation data and historical operation data of the radar system in real time, convert the historical operation data into a multi-dimensional evaluation parameter space, perform convolution operation and pooling operation to obtain reduced dimension data, extract key evaluation parameters from the reduced dimension data, and calculate a radar performance benchmark value.

[0128] The second unit is configured to compare the current operation data with the radar performance benchmark value to generate a performance deviation, take the performance deviation as a state variable, take a value range corresponding to the key evaluation parameter as an action variable, calculate a reward value, and optimize a radar training network based on the reward value to obtain the integrity evaluation parameter.

[0129] The third unit is configured to initialize a sliding time window, calculate a confidence degree corresponding to the integrity evaluation parameter in each sliding time window, cluster the integrity evaluation parameter based on the confidence degree, calculate a weighted evaluation value based on a clustering result, extract a time sequence variation feature of the weighted evaluation value, and dynamically correct the integrity evaluation parameter to obtain an integrity evaluation result.

[0130] In a third aspect, an electronic device is provided, including:

[0131] A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0132] In a fourth aspect, a computer-readable storage medium is provided, which stores computer program instructions, and the computer program instructions are executed by a processor to implement the method described above.

[0133] The present application can be a method, device, system and / or computer program product. The computer program product can include a computer readable storage medium having computer readable program instructions stored therein, which are used to perform various aspects of the present application.

[0134] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for adaptive optimization of radar system integrity assessment parameters based on dynamic iteration, characterized in that, The method comprises the following steps: Real-time acquisition of current operation data and historical operation data of a radar system, conversion of the historical operation data into a multi-dimensional evaluation parameter space, and execution of convolution and pooling operations to obtain reduced dimension data, extraction of key evaluation parameters from the reduced dimension data, and calculation of a radar performance benchmark value, comprising: Mapping the historical operation data to a multi-dimensional evaluation parameter space to obtain multi-dimensional evaluation parameters, taking the multi-dimensional evaluation parameters as vertices, taking the correlation between the multi-dimensional evaluation parameters as edges, performing concatenation and activation operations on the multi-dimensional evaluation parameters to determine edge weight values, and obtaining an initial dynamic heterogeneous graph; Based on pre-set learnable parameters and the multi-dimensional evaluation parameters, a graph attention coefficient is calculated using a pre-set bidirectional linear rectifier function, and the feature values of the vertices are updated based on the graph attention coefficient and the multi-dimensional evaluation parameters to obtain an optimized dynamic heterogeneous graph, and convolution and pooling operations are performed on the updated feature values of the vertices to obtain first reduced dimension data; A time series correlation coefficient between the multi-dimensional evaluation parameters is calculated, the edge weight values are updated based on the time series correlation coefficient, the optimized dynamic heterogeneous graph is reconstructed using the updated edge weight values to obtain a reconstructed dynamic heterogeneous graph, and important features are extracted from the reconstructed dynamic heterogeneous graph based on a pre-set importance threshold to obtain second reduced dimension data; The first reduced dimension data and the second reduced dimension data are linearly weighted and fused to obtain feature fusion data, principal component analysis is performed on the feature fusion data to determine key evaluation parameters, and a radar performance benchmark value is solved; The current operation data and the radar performance benchmark value are compared to generate a performance deviation, the performance deviation is taken as a state variable, the value range corresponding to the key evaluation parameters is taken as an action variable, a reward value is calculated, and a radar training network is optimized based on the reward value to obtain an integrity evaluation parameter; A sliding time window is initialized, and the confidence of the integrity evaluation parameter corresponding to each sliding time window is calculated, the integrity evaluation parameter is clustered based on the confidence, and a weighted evaluation value is calculated based on the clustering result, the time series variation characteristics of the weighted evaluation value are extracted, and the integrity evaluation parameter is dynamically corrected to obtain an integrity evaluation result.

2. The method of claim 1, wherein, The principal component analysis on the feature fusion data to determine the key evaluation parameters and solve the radar performance benchmark value comprises: The Euclidean distance between the feature vectors in the feature fusion data and an adaptive bandwidth parameter are calculated to obtain an adaptive weight matrix, the feature fusion data is projected based on a pre-set projection matrix to obtain projected data, the projected data is reconstructed to obtain reconstructed data, the reconstruction error between the reconstructed data and the feature fusion data is calculated, and a target optimization value is solved based on the reconstruction error and the adaptive weight matrix; The principal component analysis on the projected data obtains principal components and corresponding feature values, the variance contribution rate of the principal components is calculated based on the feature values, and the key evaluation parameters are determined in combination with a pre-set variance contribution rate threshold. The inter-class dispersion matrix and the intra-class dispersion matrix corresponding to the projection data are calculated, a Fisher discriminant value is determined based on the inter-class dispersion matrix and the intra-class dispersion matrix, and a discrimination degree of the key evaluation parameter is calculated based on the Fisher discriminant value and a projection coefficient in the projection matrix; The projection matrix is iteratively updated by an alternating optimization strategy based on the target optimization value and the Fisher discriminant value, an importance coefficient of the key evaluation parameter is calculated based on the updated projection matrix and the discrimination degree, and a radar performance benchmark value is solved based on the importance coefficient.

3. The method of claim 1, wherein, The current running data is compared with the radar performance benchmark value to generate a performance deviation, the performance deviation is taken as a state variable, a value range of the key evaluation parameter is taken as an action variable, a reward value is calculated, and a radar training network is optimized based on the reward value to obtain an integrity evaluation parameter, including: A difference between the current running data and a plurality of performance indicators in the radar performance benchmark value is calculated to obtain a performance deviation, the performance deviation is constructed as a performance deviation node, the key evaluation parameter is constructed as an evaluation parameter node, a causal relationship between the performance deviation node and the evaluation parameter node is determined, a directed acyclic graph is constructed based on the causal relationship, and the performance deviation is converted into a state variable based on the directed acyclic graph; A value range of the key evaluation parameter is determined based on the state variable, a plurality of adjustment values are selected as action variables, a causal sensitivity matrix is constructed by a gradient calculation method, and an intervention step is determined, a prediction confidence is determined by intervention operation on the state variable based on the intervention step, a decision space is determined by Monte Carlo sampling, and an evaluation decision is determined in combination with the prediction confidence; An average processing effect value is calculated based on the evaluation decision, a reward value is calculated in combination with a performance change amount of the state variable, a cumulative reward expectation is calculated by inputting the reward value into a preset radar training network, and parameters of the radar training network are optimized according to the cumulative reward expectation to obtain an optimized radar training network, a reliability weight of each key evaluation parameter is determined based on the optimized radar training network and the causal relationship, and an integrity evaluation parameter is calculated.

4. The method of claim 3, wherein, A causal sensitivity matrix is constructed by a gradient calculation method, and an intervention step is determined, a prediction confidence is determined by intervention operation on the state variable based on the intervention step, a decision space is determined by Monte Carlo sampling, and an evaluation decision is determined in combination with the prediction confidence, including: A causal sensitivity matrix is constructed by a gradient calculation method based on the change relationship between the state variable and the action variable, and an intervention step of each action variable is generated, the state variable is intervened by an intervention operation in combination with a counterfactual intervention method, an intervention state variable value is obtained, and a deviation analysis is performed on the intervention state variable value and a preset target state value to obtain an intervention effect value; A prediction confidence interval is determined based on the intervention state variable value by an interval estimation method, a prediction uncertainty is generated according to a feature parameter of the prediction confidence interval and a historical sample number obtained in advance, and a prediction confidence is determined based on the prediction uncertainty and a time decay characteristic of the intervention effect value; The historical experience vector is extracted based on the historical decision data acquired in advance, and a decision space is determined. The intervention state variable value and the historical experience vector are matched and analyzed by a similarity measurement method, and a similarity value of a prediction result and historical experience is generated by combining a distance mapping mechanism. A candidate decision set is extracted by a Monte Carlo sampling method in the decision space. The candidate decisions are screened according to the prediction confidence, and effective candidate decisions are determined according to the similarity value. The execution effect of each effective candidate decision is determined according to the intervention effect value, and an effective candidate decision with the optimal execution effect is selected as an evaluation decision output.

5. The method of claim 1, wherein, A sliding time window is initialized, and a confidence corresponding to a perfectness evaluation parameter is calculated in each sliding time window. The perfectness evaluation parameters are clustered based on the confidence, and a weighted evaluation value is calculated based on the clustering result. The time sequence variation characteristics of the weighted evaluation value are extracted, and the perfectness evaluation parameters are dynamically corrected to obtain a perfectness evaluation result, including: A sliding time window is initialized, and perfectness evaluation parameters are collected in the sliding time window. Mutual information and information entropy between different perfectness evaluation parameters are calculated. The parameter interrelation weight is calculated according to the mutual information and the information entropy. The state characteristic indicators of the perfectness evaluation parameters are extracted, and the confidence of the perfectness evaluation parameters is calculated according to the state characteristic indicators. The offset of the perfectness evaluation parameters relative to a pre-set clustering center is calculated in a gradient iteration manner, and the clustering center is updated according to the correlation weight. The perfectness evaluation parameters are clustered and grouped into a plurality of cluster groups according to the updated clustering center. The variance of the perfectness evaluation parameters in the cluster group is calculated, and the intra-group weight of the perfectness evaluation parameters is determined based on the variance. The weighted evaluation value is calculated in combination with the cluster group weight corresponding to the cluster group. The difference value of the weighted evaluation value in adjacent windows is calculated to obtain an adjacent variation characteristic. The global feature of the weighted evaluation value is calculated based on the time sequence variation of the weighted evaluation value. The correction amount is calculated based on the adjacent variation characteristic and the global feature, and the perfectness evaluation parameters are dynamically corrected. The corrected deviation is calculated, and the correction coefficient is updated based on the deviation and a decay coefficient to obtain a perfectness evaluation result.

6. The method of claim 5, wherein, The offset of the perfectness evaluation parameters relative to a pre-set clustering center is calculated in a gradient iteration manner, and the clustering center is updated according to the correlation weight. The perfectness evaluation parameters are clustered and grouped into a plurality of cluster groups according to the updated clustering center, including: The state value of the perfectness evaluation parameter is collected, and a space-time feature matrix is constructed based on the state value. The time sequence weight coefficient is determined based on the space-time feature matrix, and the offset between the perfectness evaluation parameter and the pre-set clustering center is weighted calculated to obtain a time sequence weighted offset. extracting an evolution trend feature of the integrity evaluation parameter and combining the time-series weighted offset to obtain a comprehensive offset feature, updating the clustering center based on the comprehensive offset feature, constructing a dynamic grouping boundary based on the updated clustering center, determining a boundary deformable parameter according to a distance distribution of the state value to the updated clustering center, and calculating a grouping attribution degree based on the dynamic grouping boundary and the boundary deformable parameter; continuously constraining the grouping attribution degree and smoothing the continuously constrained result by a pre-set time-series smoothing factor to obtain an initial clustering grouping result, calculating a stability index of the initial clustering grouping result, if the stability index is greater than a pre-set stability threshold, taking the initial clustering grouping result as a clustering group output, otherwise, calculating a correction direction based on a change trend of the stability index and updating the clustering center along the correction direction until the stability index is greater than the stability threshold, and obtaining a plurality of clustering groups.

7. A system for dynamic iteration based radar system health assessment parameter adaptive optimization for implementing the method of any of the preceding claims 1-6, characterized by, Comprising: a first unit configured to collect current operation data and historical operation data of a radar system in real time, convert the historical operation data into a multi-dimensional evaluation parameter space, and perform convolution operation and pooling operation to obtain reduced dimension data, extract key evaluation parameters from the reduced dimension data, and calculate a radar performance benchmark value; a second unit configured to compare the current operation data with the radar performance benchmark value to generate a performance deviation, take the performance deviation as a state variable, take a value range corresponding to the key evaluation parameters as an action variable, calculate a reward value, and optimize a radar training network based on the reward value to obtain an integrity evaluation parameter; a third unit configured to initialize a sliding time window, calculate a confidence degree corresponding to the integrity evaluation parameter in each sliding time window, cluster the integrity evaluation parameter based on the confidence degree, calculate a weighted evaluation value based on the clustering result, extract a time-series change feature of the weighted evaluation value, and dynamically correct the integrity evaluation parameter to obtain an integrity evaluation result.

8. An electronic device, comprising: Comprising: 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 6.

9. 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 6. The computer program instructions, when executed by the processor, implement the method of any one of claims 1 to 6.

Images