Method and related device for optimizing the number and quality of TR components in a subarray of a phased array radar

Abstract

This invention discloses a method, system, and related equipment for optimizing the quantity and quality of TR components in a phased array radar subarray, relating to the field of phased array radar technology. The method includes: obtaining a set of qualified TR components after fault diagnosis; orthogonally delimiting the TR component set using a physical channel extraction operator to obtain two disjoint heterogeneous subspaces; constructing a decision tensor for assembling the phased array radar subarray; constructing component quantity ratio constraints, full-band amplitude and phase consistency extreme value constraints, and radio frequency channel mask constraints based on the heterogeneous subspaces and the decision tensor; constructing a hierarchical objective function containing a first priority target and a second priority target under the decision tensor; under the constraints of component quantity ratio, full-band amplitude and phase consistency extreme value constraints, and radio frequency channel mask constraints, introducing a feature verification operator in a heuristic optimization algorithm to solve the hierarchical objective function; and outputting the optimal assembly scheme for the quantity and quality of TR components in the phased array radar subarray when the heuristic optimization algorithm converges.

Description

Technical Field

[0001] This invention relates to the field of phased array radar technology, and more specifically, to a method and related equipment for optimizing the quantity and quality of TR components in a phased array radar subarray. Background Technology

[0002] In the hardware architecture of phased array radar, the transmit / receive (T / R) module, as the core execution unit of the radio frequency front-end, often accounts for the majority of the overall radar system cost in terms of R&D and manufacturing costs. In recent years, with the rapid evolution of low-cost millimeter-wave phased array technology and its large-scale commercial and military applications, the number of T / R modules integrated into a single radar array has increased exponentially. The quality of a single T / R module and the quality of the subarray composed of multiple T / R modules are key factors affecting radar quality. Therefore, ensuring absolute reliability of module quality in high-density, high-volume flexible production lines has become a pressing issue that needs to be addressed in this field.

[0003] In the production and testing pipeline of components, the surge in production data of T / R components and the lack of intelligent diagnostic capabilities, as well as the multi-target extreme phase matching problem under the phased array antenna subarray architecture, are the main problems and key technologies that need to be overcome.

[0004] The challenge of multi-target extreme phase matching in the subarray architecture of phased array antennas. To reduce the complexity and cost of feed networks for large-scale arrays, modern phased array radars generally adopt a "subarray" splicing architecture. In the subarray structure, the microwave amplitude and phase consistency of T / R components within the same subarray directly determines the beam pointing accuracy and sidelobe level suppression capability of the entire radar system. However, due to the inherent discreteness of semiconductor materials and chip processing technology, T / R components from the same batch still have slight differences in amplitude and phase performance. Therefore, before entering final assembly, process engineers need to perform a critical "phase matching" process. This process has typical characteristics of multi-target extreme phase matching: firstly, it is necessary to maximize the utilization rate of T / R components through a group quantity priority strategy to ensure production line capacity; secondly, it is necessary to optimize and solve for the array combination with the smallest amplitude and phase dispersion to eliminate quantization errors and beam distortion.

[0005] Multi-objective extreme phasing in a phased array antenna subarray architecture is a typical nondeterministic polynomial problem (NP-Hard). In actual flexible production lines, phasing is not only constrained by physical resources such as heterogeneous channel ratios and testing equipment, but also requires strict adherence to the process extreme value red line, which has a "veto" nature. If the phasing index of a subarray fails to meet the red line requirement, the subarray cannot be grouped and must be returned for phasing to be redone. Traditional multi-objective optimization algorithms that rely on manual trial and error or simple weighted summation, such as traditional genetic algorithms, have obvious limitations in dealing with such problems: these methods attempt to combine "quantity (production capacity)" and "quality (divergence)" into a unified fitness function by fixing weights. This not only easily falls into the trap of "sacrificing amplitude for phase" in local optima, but also causes the algorithm to sacrifice the global optimal combination in order to solve the local subarray optimum, resulting in a sharp reduction in the number of global groups and serious waste of high-value components. Therefore, designing a multi-objective extreme phase-coordinated global optimization scheduling algorithm that can maximize the number of phase subarrays while ensuring the acquisition of array combinations with the minimum amplitude-phase dispersion is an urgent problem to be solved. Summary of the Invention

[0006] The purpose of this invention is to provide a method and related equipment for optimizing the quantity and quality of TR components in a phased array radar subarray. In the array assembly stage, a dynamic phase matching algorithm based on Monte Carlo lexicographical optimization under extreme value constraints is designed. This algorithm effectively avoids the index neutralization trap of traditional weighted summation, strictly follows the process priority to complete topology optimization, and solves the phase matching problem of large-scale heterogeneous components.

[0007] The above-mentioned technical objective of the present invention is achieved through the following technical solution:

[0008] A first aspect of the present invention provides a method for optimizing the quantity and quality of TR components in a phased array radar subarray, the method comprising:

[0009] Obtain the set of TR components that have passed fault diagnosis;

[0010] The physical channel extraction operator is used to orthogonally delimit the TR component set, resulting in two disjoint heterogeneous subspaces;

[0011] Construct a decision tensor for assembling a phased array radar subarray; wherein the decision tensor represents the mapping relationship between a single TR component and the phased array radar subarray;

[0012] The constraints on the component ratio of the phased array radar subarray, the extreme value constraint of full-band amplitude and phase consistency, and the radio frequency channel mask constraint are constructed based on the heterogeneous subspace and decision tensor.

[0013] Construct a hierarchical objective function under the decision tensor, which includes a first priority objective and a second priority objective; wherein, the first priority objective is the total number of effective ligands, and the second priority objective is the global average root mean square error of all effective ligands;

[0014] Under the constraints of component quantity ratio, full-band amplitude and phase consistency extreme value constraint, and radio frequency channel mask constraint, a feature verification operator is introduced into the heuristic optimization algorithm to solve the hierarchical objective function. When the heuristic optimization algorithm converges, the optimal assembly scheme of TR components in the phased array radar subarray with the optimal quantity and quality is output.

[0015] In one implementation, the set of qualified TR components after fault diagnosis is obtained, including:

[0016] Acquire multimodal feature data generated by microwave testing of the TR component under test;

[0017] Feature extraction is performed on multimodal feature data to obtain feature maps;

[0018] The feature maps are input in parallel to the spectral domain attention branch and the spatial domain attention branch to extract spectral domain features and spatial domain features;

[0019] The spectral domain features and spatial domain features are spliced ​​and fused to obtain the fused features;

[0020] The fused features are input into a multilayer perceptron classifier, which outputs unqualified TR components. Unqualified TR components are removed from the TR components to be tested, resulting in a set of qualified TR components.

[0021] In one implementation, two consecutive 3×3 convolutional layers are used to extract features from multimodal feature data to obtain feature maps.

[0022] In one implementation, the feature map is input to the spectral domain attention branch to extract spectral domain features, including:

[0023] The feature map is folded in terms of element dimensions using a spectral domain dimensionality reduction and rearrangement operator, and the query matrix, key matrix, and value matrix generated by the linear mapping are used to calculate the first energy correlation matrix representing the frequency points after the element dimensions are folded.

[0024] The first energy correlation matrix is ​​dynamically reconstructed using the normalized attention mask to obtain spectral domain features.

[0025] In one implementation, the feature map is input to a spatial attention branch to extract spectral domain features, including:

[0026] The feature map is folded in the frequency domain using a rearrangement operator, and the second energy correlation matrix representing the interaction between array elements is calculated using the spatial query matrix and key matrix generated by the linear mapping.

[0027] The second energy correlation matrix is ​​dynamically reconstructed using the normalized attention mask to obtain spatial features.

[0028] In one implementation, the process of determining the feature verification operator is as follows:

[0029] Map the phase components of the channel to a direction vector on a two-dimensional complex unit circle;

[0030] Calculate the analytical mean vector of the phase of all channels in the phase matching group of the phased array radar subarray, and extract the absolute magnitude of the analytical mean vector;

[0031] The feature verification operator is calculated based on the absolute modulus.

[0032] In one implementation, the heuristic optimization algorithm is implemented using the Monte Carlo random rearrangement algorithm; wherein, in the process of solving the heuristic optimization algorithm, the first priority objective is greater than the second priority objective.

[0033] A second aspect of the present invention provides an electronic device, including a memory and a processor;

[0034] A memory for storing computer programs, the computer programs including program instructions;

[0035] A processor is configured to execute the program instructions to cause the electronic device to perform the steps of a method for optimizing the quantity and quality of TR components in a phased array radar subarray, as provided in the first aspect of the invention.

[0036] A third aspect of the present invention provides a computer program product comprising program instructions that, when executed by an electronic device, cause the electronic device to perform the steps of a method for optimizing the quantity and quality of TR components in a phased array radar subarray as provided in the first aspect of the present invention.

[0037] A fourth aspect of the present invention provides a computer-readable storage medium comprising a computer program that, when executed by one or more processors, implements a method for optimizing the quantity and quality of TR components in a phased array radar subarray as provided in the first aspect of the present invention.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] This invention addresses the problem of traditional Pareto algorithms easily generating invalid solutions when dealing with veto-type extreme value red lines and discrete component combination explosions in multi-objective phase coordination scheduling of T / R component subarrays. It proposes a Monte Carlo lexicographical optimization (MC-LEX)-based sorting scheduling algorithm. A physical channel masking mechanism to suppress edge effects and a feature verification operator to overcome phase entanglement are designed, eliminating mathematical misjudgments caused by microwave non-Euclidean characteristics from the ground up. A lexicographical evolution rule is established, prioritizing maximizing the number of subarray groups and optimizing subarray quality as the second priority. This invention overcomes the local deadlock barrier of traditional scheduling, achieving production line efficiency while strictly ensuring microwave consistency. Attached Figure Description

[0040] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:

[0041] Figure 1 This is a flowchart illustrating a method for optimizing the quantity and quality of TR components in a phased array radar subarray, as provided in an embodiment of the present invention.

[0042] Figure 2 This is a framework diagram for two-stage fault diagnosis provided in an embodiment of the present invention.

[0043] Figure 3 This is a framework diagram of the initial fault screening provided in an embodiment of the present invention.

[0044] Figure 4 This is a framework diagram of a two-stage spatial spectrum attention network provided in an embodiment of the present invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0046] It should be noted that the terms "comprising" or "may include" used in the various embodiments of this application indicate the presence of the claimed function, operation, or element, and do not limit the addition of one or more functions, operations, or elements. Furthermore, as used in the various embodiments of this application, the terms "comprising," "having," and their cognates are intended only to indicate a specific feature, number, step, operation, element, component, or combination of the foregoing, and should not be construed as primarily excluding the presence of one or more other features, numbers, steps, operations, elements, components, or combinations of the foregoing, or adding one or more combinations of the foregoing.

[0047] It should be understood that terms such as "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0048] Please refer to Figure 1 This invention provides a method for optimizing the quantity and quality of TR components in a phased array radar subarray, the method comprising the following steps:

[0049] S101, Obtain the set of TR components that have passed the fault diagnosis.

[0050] In this embodiment, in the flexible production line of phased array radar TR components, quality control of the TR components is a crucial preliminary step to ensure the final beamforming accuracy of the radar. Fault diagnosis of the TR components, i.e., the identification and detection of faults in the TR components, is key to controlling the quality of the T / R components. Accordingly, this chapter addresses the problems in existing quality control methods, such as the difficulty in identifying hidden "soft faults" due to complex multivariate nonlinear coupling relationships, the difficulty in fault diagnosis due to the tight coupling between the quality of the T / R components and multiple disturbances in each process, and the difficulty in fault detection due to human judgment and inspection during the testing process. It proposes an intelligent fault diagnosis algorithm framework for TR components based on a dual-stage space-spectral attention network (DSSA-Net). This chapter first addresses the heterogeneous test data generated by multiple models of equipment at the production line's bottom layer, designing a preprocessing mechanism that includes bilinear spatial alignment and multimodal coordinate reconstruction to construct standardized feature data. Second, considering the reality that the losses and costs of missed faulty parts in the production line far outweigh those of falsely reported normal parts, a Lagrange asymmetric risk bias is introduced to construct a one-stage rapid initial screening module that integrates physical prior rules and a lightweight network. This module enables the identification of damaged hard faults and removes a large amount of normal background data. Finally, for fine-grained failure modes that are difficult to identify after initial screening, a deep feature decoupling network incorporating spatial-spectral dual-domain attention is designed to accurately extract and dynamically fuse complex electromagnetic coupling variation features across frequency bands and spaces, achieving precise decoupling and classification of various hidden soft faults.

[0051] In actual operation, the T / R components of a phased array radar subarray do not operate in a single steady state. Instead, they achieve flexible pointing and energy control of the radar beam by rapidly switching internal numerically controlled phase shifters and numerically controlled attenuators. Therefore, to diagnose complex, fine-grained faults within the components, the primary task is to transform the discrete, high-dimensional microwave test response under multi-control word conditions into a structured feature space that is easy for deep learning networks to process.

[0052] Specifically, after fault diagnosis of the TR component under test, a qualified set of TR components can be obtained. The specific diagnostic process is as follows:

[0053] S1011, acquire the multimodal feature data generated by the microwave test of the TR component under test.

[0054] Specifically, let the number of bits of the digitally controlled phase shifter in the T / R component be... The number of bits in the digitally controlled attenuator is In fully automated testing, the set of phase-shift control words can be represented as follows: The attenuation control word set can be represented as .

[0055] In any given combination of control words Under excitation, the microwave response of the component captured by the test instrument contains four core characteristics: transmit channel phase error. Attenuation deviation of the transmission channel Receive channel phase error and receiving channel attenuation deviation .

[0056] Due to the involvement of multimodal feature data, there are dimensional mismatches in the data matrices of the transmit and receive channels, missing data at some frequency points, or misaligned physical coordinates. Since deep learning models strictly rely on the structured and dimensionally unified input features, failure to spatially align heterogeneous data will disrupt the "joint physical representation" of the transmit and receive states of components at the same physical frequency point, causing the model to be unable to accurately capture physical coupling faults across channels. Therefore, tensor alignment and reconstruction of the multimodal feature data are necessary. This embodiment introduces a bilinear interpolation algorithm as a spatial alignment operator to eliminate multi-source heterogeneity. Compared to simple zero-padding or nearest-neighbor interpolation, bilinear interpolation, by performing linear mappings on two orthogonal coordinate axes, ensures the continuity of the reconstructed data and preserves the smooth gradient changes of the original microwave features to the greatest extent possible.

[0057] Suppose the coordinates of four adjacent real discrete sampling points on the known feature map and their corresponding microwave eigenvalues ​​are as follows: , , , For any target mapping point within the rectangular area. Its unknown eigenvalues The bilinear interpolation process can be equivalently reconstructed into the following matrix multiplication form to facilitate tensor acceleration computation within deep learning frameworks:

[0058] .

[0059] Based on the above bilinear spatial mapping logic, let the dimension of the original measured emission phase shift matrix be... This allows it to be mapped to a uniform standard feature map size. The spatial mapping relationship is defined as follows: ,in, This is the spatial scaling factor for adaptive calculation.

[0060] Subsequently, by phase shifting the transmission Emission attenuation Receive phase shift With receive attenuation The four physical modalities simultaneously perform the above denoising and alignment operations. Drawing on the feature fusion logic of multi-channel images in computer vision, the four aligned two-dimensional matrices are concatenated and stitched along the channel dimension, ultimately reconstructing the complete physical state of the component into a dense multimodal feature tensor. Number of channels in this system ,Right now This quantization process not only endows the originally discrete test data with spatial connectivity and cross-modal synchronization similar to a multi-channel image, but also enables abrupt changes in local gradients within the matrix. It can directly and implicitly map the physical damage of specific internal components, laying a data foundation for high-dimensional feature extraction in subsequent fault diagnosis models.

[0061] In actual production lines, the challenges of fault diagnosis for T / R components mainly include two aspects: First, normal samples account for the vast majority of a large number of test data, while the number of fault samples is extremely small, resulting in an extremely unbalanced data distribution in the solution space; Second, when the minute physical damage of the core components inside the component is mapped to the peripheral test signals, it often exhibits highly similar frequency and spatial domain confusion characteristics, making it difficult to locate fine-grained faults.

[0062] For fault detection of such complex RF components, relying solely on control thresholds set by manual experience can easily lead to a large number of false negatives and false negatives. Directly using deep learning models for end-to-end diagnosis of all data is not only computationally time-consuming, but also often gets stuck in local optima when dealing with fine-grained faults with extremely similar features, failing to meet the stringent requirements of high real-time performance and high accuracy in actual production lines. Therefore, this embodiment proposes to achieve a balance between efficient initial fault screening and fine-grained localization accuracy by integrating prior knowledge of microwave physics with the powerful feature decoupling capabilities of a dual-domain attention network. Addressing the difficulties in dynamic feature extraction and multi-scale spatial modeling encountered by deep reinforcement learning or traditional convolutional networks when solving multimodal microwave features, a solution is proposed... Figure 2The illustrated two-stage fault diagnosis algorithm framework, for the first stage, in the lean quality control system of TR components, does not simply pursue overall classification accuracy, but must follow the principle of prioritizing extreme recall. In engineering practice, missed detections can lead to faulty components being mixed into the array, causing spatial beamforming distortion or even transmitter burnout; the engineering cost of this fatal hidden danger is exponentially higher than the computational and time costs of sending normal components to the second stage for re-inspection due to false alarms. This extreme asymmetric risk requires the initial screening model to have bias capabilities.

[0063] Define the input measure space as The sample pairs are ,in These represent normal and faulty conditions, respectively. The traditional empirical risk minimization of cross-entropy can be expressed as:

[0064] ,in, Assuming empirical risk for the model on the current dataset, This represents all learnable weights and biases in a deep network. Given the mathematical expectation of input feature X and its corresponding label y, The predicted probability output by the model represents the confidence level of the input X as a faulty component. This indicates a normal sample. This indicates a faulty sample.

[0065] However, in the actual production line of T / R components, normal samples account for over 98% of the probability measure. This extremely imbalanced distribution is... This will cause the gradient descent direction of the model to change during parameter optimization. The model is almost entirely dominated by a large number of normal samples. In order to minimize the overall loss, the model tends to predict all samples as normal, thus producing false negatives.

[0066] Therefore, this embodiment introduces an asymmetric cost matrix to reconstruct the unconstrained cross-entropy empirical risk into a restricted risk minimization problem with strict boundary constraints:

[0067] ; ;in, The optimization objective is to minimize the false alarm loss of the model on the normal sample set while ensuring the constraints are met. The loss function; Let be the indicator function, when the model outputs the predicted probability of the real fault sample as That is, when a false negative occurs, the function value is 1, otherwise it is 0; The maximum tolerable false negative measure.

[0068] After establishing the aforementioned objective of minimizing the constrained risk, the following dilemma remains in actual end-to-end training: the objective function cannot be directly solved using the conventional gradient-based backpropagation algorithm. The fundamental reason is that the constraint condition *st* includes a discrete characteristic function. From a mathematical perspective, the indicator function is essentially a 0-1 step loss, which is non-convex and discontinuous throughout its domain. At the decision threshold, its derivative does not exist; at all other locations, its derivative is always zero.

[0069] If this is directly used as a penalty term in the loss function of a deep neural network, it will cause the network to encounter absolute "gradient stagnation" during backpropagation. Since the returned gradient is zero, the large number of weight parameters in the network will be affected. Without effective updates, the model optimization process will instantly stall. To address this issue and allow the constraint to be integrated into the neural network, this embodiment introduces the generalized Lagrange multiplier method and a convex relaxation strategy: First, using the Lagrange multiplier method, the original boundary constraint is transformed into a soft, high-penalty term in the objective function. For any attempt to breach the false negative threshold... The predictive behavior is transformed into an unconstrained minimax biduplex problem by applying a dynamic multiplier coefficient as a penalty weight.

[0070] To address the issue of non-differentiability of the indicator function, a smooth and continuously differentiable logarithmic loss is used as the convex surrogate loss for the indicator function. Through convex relaxation, the originally non-convex 0-1 step penalty is smoothly mapped to a monotonically increasing cross-entropy cost.

[0071] After the Lagrange transformation and relaxation solution described above, the originally complex restricted integral equation is equivalently derived into a weighted binary cross-entropy loss function that can be directly and efficiently backpropagated during network training, as shown below:

[0072] ,in, The final objective loss function of the first-stage hybrid anomaly screening network is denoted as . This represents the total number of samples in the current training batch. For the first The true label of each sample; For the network to the first Fault prediction probability of each sample output; multiplier Given a high weight.

[0073] From the perspective of backpropagation dynamics, for fault samples Its loss is related to network output The partial derivatives are .

[0074] because During backpropagation, faulty samples release enormous gradient energy, which exerts a strong "repulsive force" in the parameter space, forcing the nonlinear decision boundary to retreat towards the core region of normal samples. Although this bias sacrifices local specificity, it mathematically guarantees spatial inclusiveness for sparse fault points, thus improving the accuracy of fault diagnosis.

[0075] like Figure 3 As shown, regarding the physical prior rule operator, for hard faults such as broken gold wires on bare chips and burnt-out bias circuits in the RF link, a computational overhead of [missing information] is adopted. The rule set is used for pre-identification.

[0076] 1) Second-order flatness degradation operator: The ideal frequency response of a T / R component should have extremely high flatness within a specified bandwidth. Calculate the root mean square error at 81 frequency points: ,in, The set production line amplitude threshold; if there exists any channel c satisfying If so, it is determined that the network matching is severely imbalanced.

[0077] 2) Zero-gradient dead value operator: For the specific "locked" state caused by control chip failure, extract the feature range:

[0078] ,in, For characteristic range, This is a microwave characteristic observation sequence. This is the threshold for determining a dead value. If... This means that the control command is completely unable to modulate the radio frequency carrier, directly triggering a hard block.

[0079] Based on low-latency diagnostics, the network removes deep stacking and adopts a flattened structure, such as... Figure 3 As shown, it includes the following structure:

[0080] 1) Microwave time-delay convolutional layer: using a scale of The one-dimensional dilated convolution specifically slides along the frequency sequence direction. Its forward equation is: in, The input microwave feature tensor is the standardized value. For channel indexing, This indicates the frequency point or control state sequence position currently covered by the sliding window; is the learnable convolutional kernel weight matrix; m is the local spatial sliding window index, indicating the operator's performance in extracting the _th ... When examining the features at a given location, the microwave response state of its neighbors was also considered to capture the local gradient changes of the features. To correct linear units, it provides extremely strong nonlinear fitting capabilities, while its "negative value zeroing" characteristic helps the network filter out microwave noise floor.

[0081] 2) Global Mean Aggregation: Global mean aggregation is used at the end of feature extraction to replace the traditional fully connected layer. This is achieved by averaging the 64 spatial channels of the array elements. in, This refers to the high-dimensional feature vector output from the network's end. This is the result of global aggregation.

[0082] Global mean aggregation reduces the number of network parameters while giving the model robust smoothing capabilities against burst noise from a single array element. Finally, it outputs a highly confident soft probability boundary through Sigmoid activation. .

[0083] At the end of the hybrid prediction engine, the core challenge is to efficiently align and deeply fuse the physical prior domain characterizing microwave attenuation with the data-driven domain capturing the latent features of high-dimensional data, ensuring both physical interpretability and prediction accuracy. Addressing the feature cancellation and information bottleneck problems that easily arise when processing heterogeneous multimodal data using traditional simple weighted summation or single feature concatenation, this embodiment proposes a dual-domain feature fusion mechanism based on Hadamard product and high-dimensional cascaded composite operations: First, to achieve hard constraints and nonlinear modulation of the forward features of the neural network by physical rules, Hadamard product is introduced to construct a physical attention gating. Its algebraic mapping equation is as follows: In the formula, This is the one-dimensional deep data feature vector output by the lightweight convolutional network on the right. The prior bias vector of the same dimension is calculated based on the expert rule base for the left physical branch; Represents the element-wise multiplication operation of a matrix; These are the interaction features modulated by physical gating. In terms of physical semantics, the prior bias vector... It acts as a set of adaptive feature masks. When a certain feature extracted by the deep network violates the extreme operating condition envelope of the T / R component or microwave common sense, the corresponding physical prior weight will approach 0, thereby forcibly suppressing the activation of the pseudo feature through multiplication operation; conversely, if the feature evolution conforms to physical laws, it is allowed or strengthened.

[0084] S1012, extract features from multimodal feature data to obtain feature maps.

[0085] Specifically, in the diagnostic framework of the phased array radar T / R component, after the efficient initial screening of the first stage of fault diagnosis mentioned above, the sample set entering the second stage consists of faults that have been screened by physical rules and lightweight sensing boundaries. These faults cover 8 types of fine-grained faults.

[0086] To effectively overcome the discrimination bottleneck of traditional convolutional networks on such highly coupled datasets, a two-stage spatial-spectral attention network is proposed and constructed, with the network structure as follows: Figure 4 As shown, traditional deep learning backbone networks, such as ResNet or VGG, often employ large-stride convolutions or max pooling in shallow layers to rapidly expand the receptive field and reduce computational cost. However, in microwave feature tensors, there is an extremely tight Markov physical continuity between adjacent frequency points and adjacent array elements. Pooling operations irreversibly disrupt the high-frequency ripple structure of the frequency response curve, and these ripples are the core criteria for distinguishing faults.

[0087] Based on this, a high-fidelity isodimensional convolution stem was designed. This high-fidelity isodimensional convolution stem consists of two continuous layers. Convolutional layers consist of: ;

[0088] ;in, The input is the multimodal microwave feature tensor. and These are the weight matrices for the first and second equal-dimensional convolutional layers, respectively. and This is the bias term for the convolution operations between the first and second layers; For batch normalization operations; It is a non-linear activation function; This is the intermediate transition feature map output from the first convolutional layer; This is a two-dimensional random deactivation operation; This is the high-fidelity feature map that is the final output of the convolutional stem module.

[0089] First Convolutional layer and second Between convolutional layers, a deactivation rate is introduced. The two-dimensional random deactivation mechanism of Dropout2D. Unlike the traditional one-dimensional Dropout, Dropout2D randomly erases information across the entire channel plane of the feature map. Its basic idea is to force the network to learn orthogonal representations in the 64 basic channels, preventing the model from overfitting to specific microwave noise when faced with a difficult dataset amplified from 3200 samples. The final output... The isolated amplitude and phase measures are transformed into a high-dimensional continuous field containing the first-order derivatives of adjacent spacetime.

[0090] S1013, the feature map is input in parallel to the spectral domain attention branch and the spatial domain attention branch to extract spectral domain features and spatial domain features.

[0091] Specifically, the feature map is input into the spectral domain attention branch to extract spectral domain features, including: using the spectral domain dimensionality reduction and rearrangement operator to fold the element dimensions of the feature map, and using the query matrix, key matrix and value matrix generated by the linear mapping to calculate the first energy correlation matrix representing the frequency points after the element dimensions are folded; and using the normalized attention mask to dynamically reconstruct the first energy correlation matrix to obtain the spectral domain features.

[0092] If the global self-attention mechanism of the standard visual Transformer is directly applied to a two-dimensional feature map containing array elements and frequency points, its space and time complexity will exhibit a quadratic explosion, making it difficult to meet the needs of online diagnosis. Furthermore, this method performs an indiscriminate cross-inner product of features from different array elements and features from different frequency points, lacking interpretability from the perspective of microwave physics mechanisms. Therefore, this embodiment proposes a spectral domain dimensionality reduction and rearrangement operator. :

[0093] in, This is the number of sample groups sent to the network in the current batch; Represents the dimension of the TR component array; Represents discrete frequency points in production data; This represents the number of microwave feature channels extracted by the network. It is the four-dimensional basic microwave feature tensor output by the equal-dimensional convolution stem.

[0094] This operation will change the array element dimension. Along with batch dimensions Deep folding was implemented, which, in principle, forcibly cut off the spatial physical crosstalk between adjacent array elements, strictly isolating and limiting the computational field of the attention mechanism to an independent single array element. This forces the network to eliminate common-mode noise between multi-channel components, allowing it to focus on evaluating the autocorrelation and evolution trend of 81 frequency points along the sweep axis within a single component.

[0095] After completing dimensional decoupling, the network obtains independent query matrices in the low-rank mapping space through linear projection. Key matrix Value matrix Subsequently, the first energy correlation matrix between frequency points is calculated. : ;

[0096] Furthermore, the features are dynamically reconstructed using a normalized attention mask, and residual fusion is performed with a learnable physical scaling parameter to obtain spectral domain features: in, , , The query matrix, key matrix, and value matrix generated for the linear mapping; The first energy correlation matrix of the frequency points is used to characterize the feature inner product scores between each pair of the 81 frequency points; This is a normalized activation function used to transform discrete energy scores into attention weight masks that sum to 1; This is the inverse rearrangement operator, used for summation during dimension alignment; The learnable physical scaling factor is initialized to zero, allowing the network to maintain the stability of the original features in the early stages of training, and then adaptively injecting spectral domain attention residuals. The final output feature tensor after spectral domain attention fusion and spatial restoration.

[0097] First energy correlation matrix In this context, the main diagonal elements reflect the absolute energy intensity of each frequency point, while the off-diagonal elements precisely quantify the "frequency point". Response distortion at frequency points The resulting "traction effect" occurs when components experience soft faults such as filter minor losses or phase shifter jumps, disrupting the originally smooth and continuous transition on the frequency response curve. Through the dynamic weight allocation of this attention module, these isolated singularities that cause local band collapse or edge distortion are given high attention weights and amplified exponentially in the high-dimensional feature representation space. This mechanism endows the model with global perception of subtle deformations across frequency bands, achieving precise decoupling of fine-grained physical failure modes at the spectral domain level.

[0098] The feature map is input into the spatial attention branch to extract spectral domain features, including: using a rearrangement operator to fold the feature map in the frequency domain dimension, and using the spatial query matrix and key matrix generated by the linear mapping to calculate the second energy correlation matrix representing the interaction between array elements after frequency domain dimension folding; and using the normalized attention mask to dynamically reconstruct the second energy correlation matrix to obtain spatial domain features.

[0099] In actual troubleshooting of phased array radar, focusing only on the frequency response of a single channel will inevitably miss system-level faults such as loose RF interfaces or abnormal power supply to array busbars. The typical physical characteristics of such system-level faults are: multiple T / R component channels that are physically adjacent or in the same power supply quadrant will exhibit highly consistent microwave performance drops or phase drift at specific test frequencies. To accurately capture these common failure characteristics in the array topology, a spatial attention module completely parallel to the spectral domain mechanism was designed in DSSA-Net.

[0100] This module is mathematically and logically orthogonal to the spectral domain, and it incorporates the frequency dimension. Along with batch dimension folding, i.e., rearrangement operators : ;in, The four-dimensional basic microwave feature tensor output by the equal-dimensional convolution stem; This is the number of sample groups sent to the network in the current batch; Represents the dimension of the TR component array; Represents discrete frequency points in production data; This represents the number of microwave feature channels extracted by the network; This is the three-dimensional spatial feature tensor generated after frequency domain folding.

[0101] From the perspective of tensor algebra, this operator will reduce the microwave frequency domain dimension. Along with batch dimensions It has undergone deep folding. Its physical essence lies in forcibly cutting off cross-band crosstalk between different frequency points.

[0102] After completing frequency domain masking, the network generates a query matrix for the spatial domain in the mapping space. AND key matrix And a second energy correlation matrix representing the interaction between array elements is generated through dot product operation. Finally, residual fusion is used to output spatial features. :

[0103] ;

[0104] ;in, Spatial query matrix, key matrix, and value matrix generated for linear mapping; It is a spatially consistent aggregation matrix used to quantify the feature similarity and coupling force of any two T / R components within a subarray at a specific frequency point; It is an inverse rearrangement operator used to restore the calculated spatial features to the original four-dimensional tensor; These are learnable spatial physical scaling parameters.

[0105] S1014 splices and fuses spectral and spatial features to obtain fused features.

[0106] Specifically, the spectral features of two independent parallel branches are combined. With spatial characteristics If a simple matrix element-wise addition is used, it will forcibly confuse the independent semantics of spatial topology and frequency domain response, thereby destroying the orthogonality of the extracted features within their respective domains. Therefore, firstly, the dual-domain features are directly concatenated along the channel dimension, expanding them into a robust 128-channel tensor; subsequently, a... Point convolution performs non-linear recombination across dimensions:

[0107] ;in, The high-dimensional feature tensor representing the output of the spectral and spatial domain attention modules is... This is a channel-level splicing operation; The weight matrix is ​​the point convolution kernel weight matrix; For batch normalization operations and nonlinear activation functions; This is a spatial random deactivation operation; It is the high-fidelity output tensor after deep fusion of features from two domains of the spatial spectrum.

[0108] in, The purpose of convolution is to act as a linear combiner, adaptively weighting the fundamental basis vectors from the spectral and spatial domains at each pair of "space-frequency" coordinates. Simultaneously, a deactivation rate is introduced into the network. The intermediate layer random deactivation mechanism effectively suppresses the overfitting of deep networks to a few atypical difficult samples by randomly masking the activation state of some neurons during training, thereby enhancing the generalization ability of the model's dual-domain feature fusion.

[0109] S1015 inputs the fused features into the multilayer perceptron classifier, outputs unqualified TR components, removes unqualified TR components from the TR components to be tested, and obtains a set of qualified TR components.

[0110] In this embodiment, at the classification end of the network, traditional CNN architectures often use the Flatten operation to directly flatten high-dimensional feature maps into one-dimensional vectors. This instantly destroys the "space-frequency" topology within the features and generates extremely large fully connected parameters. Therefore, this network abandons the flattening operation that destroys spatial structure and instead employs adaptive global average pooling to rigorously and losslessly compress multi-channel feature maps into a dense global context vector of length 64. Then, it is fed into a multilayer perceptron classifier to generate the final diagnostic decision: ;in, This is a 64-dimensional dense feature vector generated after global average pooling compression. This is the weight matrix in the MLP classifier; This is a one-dimensional batch normalization operation; For inactivation rate A one-dimensional brute-force random deactivation operator; This is the posterior probability distribution vector of the network's final output.

[0111] S102, the physical channel extraction operator is used to orthogonally delimit the TR component set, resulting in two disjoint heterogeneous subspaces.

[0112] In this embodiment, the total set of diagnostically qualified T / R components to be matched is defined as follows: Its total size is According to the physical design specifications of the radar array, the components are divided into full-scale components with 24 effective radio frequency channels and edge components with only 16 effective radio frequency channels.

[0113] Operators are extracted via physical channels. Orthogonally delimit the set. The mapping operator... The physical essence of this is to read the non-zero element dimensions of the underlying test tensor of the microwave component. Based on this mapping, two disjoint heterogeneous subspaces are constructed. and : ;in, and They represent the first in the set. The and the first A physical microwave component. This equation satisfies... and .

[0114] This division of heterogeneous space directly determines that subsequent search algorithms need to perform joint optimization in two independent combined subspaces.

[0115] Specifically, because components are continuously consumed and cannot be reused during assembly, the system's state space dynamically shrinks as assembly progresses. To describe the algorithm in the... The system state after the first successful grouping defines the dynamic remaining state space. : ; where, initial state If in the first In the next step, the algorithm successfully found a phase group that satisfies the physical consistency constraint. ,in Includes subsets and The state transition equation of the system can then be rigorously defined as the set difference operation: ;in, This represents the number of iterations required for the algorithm to successfully complete the assembly. and They represent the first time. At step, the set of all components that have not yet been grouped and the remaining components of the edge-masked components; "\" is the difference operation in the set.

[0116] S103, construct the decision tensor of the assembled phased array radar subarray; wherein, the decision tensor represents the mapping relationship between a single TR component and the phased array radar subarray.

[0117] To evaluate the final state of the algorithm's output, explicit decision variables need to be defined. Assume that throughout the entire iteration cycle, the algorithm attempts to construct... A subarray group was constructed. A two-dimensional Boolean decision tensor was created. Its matrix elements This represents the mapping relationship between a single component and a specific assembly group, namely:

[0118] .

[0119] Based on the principle of exclusivity of indivisible physical devices, the decision matrix The row vectors must satisfy strict bounds, meaning that each physical component can be assigned to at most one hardware group: In the formula, The total number of phase-matching components in the component pool; This is the theoretically maximum number of radar subarrays that can be constructed, preset for the algorithm. This inequality constitutes the mathematical prerequisite for the subsequent heterogeneous assembly topology mask constraint. In summary, the phase matching process of the phased array radar subarray has been defined as based on... The mathematical optimization problem of matrices.

[0120] S104, based on the heterogeneous subspace and decision tensor, construct the component ratio constraints, full-band amplitude and phase consistency extreme value constraints, and radio frequency channel mask constraints of the phased array radar subarray.

[0121] Specifically, each independent physical subarray must be composed of exactly 6 T / R components, and the ratio of its full components to edge components must be strictly constrained to 4:2. To enforce this topological ratio within the vast decision space and significantly prune illegal combinations, a Boolean indicator function is introduced. For decision tensors The column vectors construct a simultaneous linear equality constraint: ;

[0122] In the formula, This represents the total size of the components in the component pool. The maximum number of groups is preset for the algorithm; To characterize the first Does the first component belong to the first...? Group decision variables; This is an indicator function that returns 1 if the logical condition within the parentheses is true, and 0 otherwise. For individual physical components; and These represent the heterogeneous subspaces of the full component and the edge-shielded component, respectively.

[0123] Since edge shielding components are often deployed at the edge of the array during physical assembly, in order to suppress edge diffraction effects and cut off clutter coupling in passive areas, 8 specific channels out of its 24 physical channels are designed as inactive passive isolation terminals at the hardware level.

[0124] If the background thermal noise and static bias generated by these eight channels are included in the root mean square calculation when evaluating microwave conformance, components that are actually of high quality will be wrongly judged and rejected. Therefore, a physical shielding index set for passive channels was constructed. A binary mask vector operating in the 24-dimensional channel space was defined. :

[0125] ;

[0126] ;

[0127] For those assigned to a group Any edge shielding component in Assuming that the original multidimensional feature tensor extracted from its underlying layer is This framework introduces the Hadamard product (denoted as...). Perform a channel-by-channel hard truncation projection on it: In the formula, This indicates the specific serial number of the physical channel for the microwave signal; An invalid channel sequence pre-set for the project; mask vector In the The specific values ​​that a dimension can take; and Let represent the high-dimensional feature tensors before and after masking, respectively. By introducing... The algorithm transforms complex code logic into pure logic. The complexity tensor dot multiplication operator safely eliminates interference from invalid data during the underlying feature aggregation stage, ensuring the physical fidelity of subsequent discreteness evaluation.

[0128] Under the premise of satisfying the macroscopic topology and microscopic channel mask, the lowest-level verification criterion of the phase-coordinated algorithm is to determine the consistency divergence of the microwave measures within the group. Since the T / R module is a broadband device, a discrete representative frequency set covering its operating frequency band is extracted. It covers low-frequency, mid-frequency and high-frequency extreme points.

[0129] The beamforming quality of a phased array radar is determined by the consistency of the worst frequency point within the array; therefore, this is a typical minimization-maximization tolerance control problem. For any decision matrix... The phase group calibrated as 1 At any characteristic frequency point The root mean square error of amplitude and the root mean square error of phase must be simultaneously and strictly controlled within the tolerance threshold of the system link budget. This nonlinear envelope constraint is formalized as follows: ;

[0130] In the formula, This represents the core discrete test frequency point set used for feature verification; and They represent the frequency points respectively. The intra-group amplitude and phase consistency divergence statistic is calculated below, and this statistic has a built-in mask vector before calculation. Perform invalid dimension removal; and These are the hard extreme value tolerance boundaries derived from the radar system's overall linkage, which are strictly set as amplitude boundaries according to actual mass production specifications. Phase boundary Any decision tensor combination that exceeds the extreme value boundary will be determined as an infeasible solution and will be immediately pruned and eliminated during the evaluation period.

[0131] S105, construct a hierarchical objective function under the decision tensor that includes a first priority objective and a second priority objective; wherein, the first priority objective is the total number of effective phase groups, and the second priority objective is the global average root mean square error of all effective phase groups.

[0132] In this embodiment, a Pareto objective game exists between the number of groups and the quality of subarrays in the actual production scheduling of phased array radar component pairing. Typically, to forcibly add a marginally qualified radar subarray, components with high dispersion may be introduced, thereby lowering the average performance of the overall assembly batch. If the traditional linear weighted sum method is used, improper setting of the weight hyperparameters can easily lead to the algorithm sacrificing the number of groups in pursuit of higher quality.

[0133] To eliminate the sensitivity of weight hyperparameters and ensure microwave consistency while maximizing capacity, a hierarchical objective function with absolute priority was constructed.

[0134] First, define the decision tensor. First priority objective function This represents the total number of effective phase groups. A Boolean discriminant function is introduced. When the k-th group strictly satisfies all the above topological ratios and extreme value mask hard constraints, Otherwise, it is 0. The first priority objective can be formalized as: .

[0135] Then define the second priority goal. The global average root mean square error is given for all legal phase groups. To comprehensively consider the dispersion of the amplitude and phase in two dimensions, the overall consistency measure of a single group is used. Defined as the arithmetic combination of the amplitude and phase standardized RMS. The global divergence objective function can be expressed as: In the formula, The maximum possible group size for the search; To test the first Logical discriminant variables for group feasibility; The first decision matrix is ​​determined by One physical assembly group; Representative phase capacity (to be maximized); This represents the microwave performance dispersion of the phase batch.

[0136] S106, under the constraints of component quantity ratio, full-band amplitude and phase consistency extreme value constraint and radio frequency channel mask constraint, introduces a feature verification operator in the heuristic optimization algorithm to solve the hierarchical objective function. When the heuristic optimization algorithm converges, it outputs the assembly scheme with the optimal number and quality of TR components in the phased array radar subarray.

[0137] In this embodiment, based on the aforementioned hierarchical objective function, a lexicographically ordered multi-objective evaluation tuple is constructed. In the dynamic iteration of Monte Carlo heuristic search, in order to determine the newly generated working state... Is it better than the historical global optimal state? The algorithm does not calculate a scalar weighted sum of the two objectives; instead, it performs a strict lexicographical comparison. For simplicity, the new solution is compared to the historical best solution at the [missing value]. The objective function values ​​in terms of priority are respectively ,in .

[0138] Non-dominated solution superiority operator defined in the coordination optimization space as follows: In the formula, the symbol It is read as "strictly superior to in lexicographical order"; and These represent logical OR and logical AND, respectively. This formula logically closes the optimization rule: the algorithm first compares the first priority objective; only when the number of groups in the new solution is strictly greater than the historical solutions, or when the number of groups is exactly equal but the microwave average divergence is smaller, will the algorithm proceed. Only then can Pareto dominance occur, replacing... It becomes the new global optimum. This mechanism ensures that the coordinate algorithm has a stable and rational convergence direction in the combinatorial explosion solution space.

[0139] After establishing the multi-objective hierarchical optimization model, the core driving force of the algorithm shifts to the accurate and efficient evaluation of the internal consistency of the bottom-level assembly group G_k, i.e., the root mean square error (RMS). Since the test data of microwave components is stored as a high-dimensional feature tensor, blindly calculating the discreteness directly in the high-dimensional full space would lead to unacceptable time complexity. In particular, the "amplitude" and "phase" of microwave radio frequency are physically located in completely different topologies. Therefore, a dimensionality reduction aggregation mechanism and an anti-winding loop operator were designed for the amplitude and phase characteristics, respectively.

[0140] The amplitude characteristics of microwave signals are distributed in a standard linear Euclidean space. In the high-frequency iterative search of heuristic algorithms, a dimensionality reduction aggregation operator with channel masks is introduced to quickly evaluate the amplitude consistency of a set of components. Under the premise of satisfying the channel mask constraint, the phase group is extracted. All active physical channels within the network, and at specific discrete frequency points. The standardized Euclidean root mean square error is calculated below: In the formula, The specific microwave frequency point being verified; For the mask vector The filtered set of active channel indices that actually participated in the evaluation; This represents the total number of valid physical channels in the set. For the first Each channel at frequency point Extract the absolute magnitude value below; This is the arithmetic mean of the amplitudes of all activated channels within the current coordination group. Using this operator, the algorithm can perform a linear measurement of amplitude divergence within the group in a relatively short computation cycle.

[0141] In stark contrast to amplitude characteristics, the phase characteristics of microwave signals Defined on a closed topological torus The data exhibits typical characteristics of "non-Euclidean data." In practical engineering, this leads to a severe "phase wrapping effect." For example, when the measured phases of two channels in the array are respectively... and At that time, its phase difference in real physical space is only extremely small. However, if the traditional Euclidean variance calculation logic of formula (4-15) is directly applied, its mathematical absolute deviation will be mistakenly amplified to... The failure of this metric matrix will directly cause the matching algorithm to fall into a minimum trap, resulting in the misjudgment and discarding of a large number of originally highly consistent high-quality component combinations.

[0142] In one embodiment, the phase components of the channel are mapped to direction vectors on a two-dimensional complex unit circle; the analytical mean vector of the phase of all channels in the phase matching group of the phased array radar subarray is calculated, and the absolute magnitude of the analytical mean vector is extracted; the feature verification operator is calculated based on the absolute magnitude.

[0143] Specifically, to fundamentally eliminate the loss of group quantity caused by phase winding, directional statistics theory is introduced to map discrete phase scalars to the complex plane, and a nonlinear cyclic consistency operator is designed. First, Euler's formula is used to map the phase components of the channel. Mapped to a direction vector on the two-dimensional complex unit circle : .

[0144] Subsequently, the phase group is calculated in the complex plane. The analytical mean vector of the phases of all activated channels And extract the absolute magnitude of the mean vector. :

[0145] ; In the formula, j is the imaginary unit; For the first The radian phase value of each active channel. The magnitude of the analytical mean vector. Strictly constrained Within a closed interval, its physical meaning characterizes the phase convergence tendency: when all channels have absolutely consistent phase, When the channel phase is divergent, that is, when it is uniformly and symmetrically distributed in the complex plane, .

[0146] Based on the centripetal concentration modulus A cyclic standard deviation nonlinear operator, completely unaffected by winding artifacts, was derived and used as a benchmark for evaluating the phase divergence of the algorithm at this frequency. ; processed at the mathematical level Faults at the boundary. By embedding this complex plane mapping operator into a lexicographically ordered multi-objective function, Monte Carlo heuristic search enables unbiased gradient sliding and extremum optimization in the real microwave data space.

[0147] The heterogeneous combination and phasing of microwave components is essentially an NP-hard combinatorial optimization problem with complex nonlinear constraints. As the total size of the components to be assembled in the radar array increases, the space of valid combinatorial solutions grows exponentially. Traditional exact solution algorithms, such as branch and bound or dynamic programming, are prone to the "curse of dimensionality" when dealing with such high-dimensional discrete data spaces, and cannot meet the stringent real-time requirements of modern radar industrial production lines.

[0148] To address this, the law of large numbers is introduced, and a heuristic optimization algorithm integrating Monte Carlo random rearrangement and local sliding greedy strategy is designed. This algorithm seamlessly integrates the complex plane cyclic verification operator derived above with the multi-objective lexicographical evaluation function. It should be noted that during the algorithm's solution process, the first priority objective in the hierarchical objective function is greater than the second priority objective.

[0149] This invention also provides an electronic device. The electronic device includes a processor, a memory, a communication interface, and at least one communication bus for connecting the processor, the memory, and the communication interface. The memory includes, but is not limited to, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (PROM), or portable read-only memory (CD-ROM), and is used for related instructions and data.

[0150] The communication interface is used to receive and send data. The processor can be one or more CPUs; if the processor is a single CPU, it can be a single-core CPU or a multi-core CPU. A processor in an electronic device reads one or more programs stored in memory and performs the following operations: acquiring a qualified TR component set after fault diagnosis; orthogonally delimiting the TR component set using a physical channel extraction operator to obtain two disjoint heterogeneous subspaces; constructing a decision tensor for assembling a phased array radar subarray; wherein the decision tensor characterizes the mapping relationship between a single TR component and the phased array radar subarray; constructing component quantity ratio constraints, full-band amplitude-phase consistency extremum constraints, and radio frequency channel mask constraints for the phased array radar subarray based on the heterogeneous subspaces and the decision tensor; constructing a hierarchical objective function under the decision tensor containing a first priority objective and a second priority objective; wherein the first priority objective is the total number of effective phase-matching groups, and the second priority objective is the global average root mean square error of all effective phase-matching groups; under the constraints of component quantity ratio, full-band amplitude-phase consistency extremum constraints, and radio frequency channel mask constraints, introducing a feature verification operator in a heuristic optimization algorithm to solve the hierarchical objective function; and outputting the optimal assembly scheme for the number and quality of TR components in the phased array radar subarray when the heuristic optimization algorithm converges.

[0151] It should be noted that the specific implementation of each operation can be described above. Figure 1 The corresponding description of the method embodiments shown indicates that the electronic device can be used to execute a method for optimizing the quantity and quality of TR components in a phased array radar subarray according to the above method embodiments of this application, which will not be described in detail here.

[0152] This invention also provides a computer-readable storage medium, which is a memory device in a computer device for storing programs and data. It is understood that the computer-readable storage medium here can include both built-in storage media in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the operating system of a terminal. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the method for optimizing the quantity and quality of TR components in a phased array radar subarray in the above embodiments. Those skilled in the art should understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0153] This invention also provides a computer program product containing program instructions. The computer program product may be software or program products containing program instructions, capable of running on a computing device or stored on any usable medium. When the computer program product is run on at least one electronic device, it causes the at least one electronic device to perform a method for optimizing the quantity and quality of TR components in a phased array radar subarray.

[0154] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for optimizing the quantity and quality of TR components in a phased array radar subarray, characterized in that the method... include: Obtain the set of TR components that have passed fault diagnosis; The physical channel extraction operator is used to orthogonally delimit the TR component set, resulting in two disjoint heterogeneous subspaces; Construct a decision tensor for assembling a phased array radar subarray; wherein the decision tensor represents the mapping relationship between a single TR component and the phased array radar subarray; The constraints on the component ratio of the phased array radar subarray, the extreme value constraint of full-band amplitude and phase consistency, and the radio frequency channel mask constraint are constructed based on the heterogeneous subspace and decision tensor. Construct a hierarchical objective function under the decision tensor, which includes a first priority objective and a second priority objective; wherein, the first priority objective is the total number of effective ligands, and the second priority objective is the global average root mean square error of all effective ligands; Under the constraints of component quantity ratio, full-band amplitude and phase consistency extreme value constraint, and radio frequency channel mask constraint, a feature verification operator is introduced into the heuristic optimization algorithm to solve the hierarchical objective function. When the heuristic optimization algorithm converges, the optimal assembly scheme of TR components in the phased array radar subarray with the optimal quantity and quality is output.

2. The method according to claim 1, characterized in that, Obtain the set of TR components that have passed fault diagnosis, including: Acquire multimodal feature data generated by microwave testing of the TR component under test; Feature extraction is performed on multimodal feature data to obtain feature maps; The feature maps are input in parallel to the spectral domain attention branch and the spatial domain attention branch to extract spectral domain features and spatial domain features; The spectral domain features and spatial domain features are spliced ​​and fused to obtain the fused features; The fused features are input into a multilayer perceptron classifier, which outputs unqualified TR components. Unqualified TR components are removed from the TR components to be tested, resulting in a set of qualified TR components.

3. The method according to claim 2, characterized in that, Two consecutive 3×3 convolutional layers are used to extract features from multimodal feature data to obtain feature maps.

4. The method according to claim 2, characterized in that, The feature map is input into the spectral domain attention branch to extract spectral domain features, including: The feature map is folded in terms of element dimensions using a spectral domain dimensionality reduction and rearrangement operator, and the query matrix, key matrix, and value matrix generated by the linear mapping are used to calculate the first energy correlation matrix representing the frequency points after the element dimensions are folded. The first energy correlation matrix is ​​dynamically reconstructed using the normalized attention mask to obtain spectral domain features.

5. The method according to claim 2, characterized in that, The feature map is input into the spatial attention branch to extract spatial features, including: The feature map is folded in the frequency domain using a rearrangement operator, and the second energy correlation matrix representing the interaction between array elements is calculated using the spatial query matrix and key matrix generated by the linear mapping. The second energy correlation matrix is ​​dynamically reconstructed using the normalized attention mask to obtain spatial features.

6. The method according to claim 1, characterized in that, The process for determining the feature verification operator is as follows: Map the phase components of the channel to a direction vector on a two-dimensional complex unit circle; Calculate the analytical mean vector of the phase of all channels in the phase matching group of the phased array radar subarray, and extract the absolute magnitude of the analytical mean vector; The feature verification operator is calculated based on the absolute modulus.

7. The method according to claim 1, characterized in that, The heuristic optimization algorithm is implemented using the Monte Carlo random rearrangement algorithm; wherein, in the process of solving the heuristic optimization algorithm, the first priority objective is greater than the second priority objective.

8. An electronic device, characterized in that, Including memory and processor; A memory for storing computer programs, the computer programs including program instructions; A processor for executing the program instructions to cause the electronic device to perform the steps of a method for optimizing the quantity and quality of TR components in a phased array radar subarray as described in any one of claims 1 to 7.

9. A computer program product containing program instructions, characterized in that, When the program instructions are executed by the electronic device, the electronic device performs the steps of a method for optimizing the quantity and quality of TR components in a phased array radar subarray as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a computer program that, when executed by one or more processors, implements a method for optimizing the quantity and quality of TR components in a phased array radar subarray as described in any one of claims 1 to 7.

Images