Laser array adaptive control method and system based on artificial intelligence

Abstract

The invention provides a laser array adaptive control method and system based on artificial intelligence, and relates to the technical field of laser control, and the method comprises the steps: obtaining real-time and target light field distribution, constructing a coupling interference topological graph through nonlinear feature extraction, and achieving the decoupling of a contribution component and an interference component; a control strategy containing global coupling constraint is generated in combination with a graph neural network and a dual decomposition mechanism, the control strategy is decomposed into a local control target and coupling consistency constraint, and iterative optimization is performed to generate a modulation control instruction, so that the array light field control precision is improved, and the adaptability of a system to environmental disturbance is enhanced.

Description

Technical Field

[0001] This invention relates to the field of laser control technology, and in particular to an adaptive control method and system for laser arrays based on artificial intelligence. Background Technology

[0002] Laser arrays, as high-power light sources, have wide applications in materials processing, laser communication, lidar, and directed-energy weapons. By coherently combining multiple laser units, a significant increase in output power can be achieved. However, due to the complex thermal, mechanical, and electromagnetic coupling relationships between the laser units, mutual interference during actual operation can cause the output light field distribution to deviate from the ideal state, affecting the overall system performance.

[0003] Existing methods struggle to effectively identify and characterize the complex coupling relationships between laser elements, especially in large-scale array systems where the thermal, electromagnetic, and mechanical couplings between elements exhibit highly nonlinear characteristics, making accurate description by traditional models difficult. Traditional control strategies lack consideration for the global constraints of the laser array system; methods that control each element independently cannot effectively coordinate coupling interference between elements, resulting in poor system stability, slow response speed, and difficulty in handling complex dynamic environments. Existing algorithms have limitations in handling the difference between high-dimensional output and target light fields, often focusing only on intensity or phase distribution, lacking multi-scale, multi-domain joint optimization capabilities, and thus failing to achieve accurate light field synthesis and efficient control. Summary of the Invention

[0004] This invention provides an adaptive control method and system for laser arrays based on artificial intelligence, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides an artificial intelligence-based adaptive control method for laser arrays, comprising: Acquire the real-time output light field distribution and target light field distribution of the laser array; Nonlinear feature extraction is performed on the real-time output light field distribution to obtain the array state feature vector. Based on the array state feature vector, the mutual coupling relationship between each laser unit in the laser array is identified through a causal inference mechanism, and a coupling interference topology diagram is constructed. Based on the coupling interference topology diagram, the independent contribution component and the coupling interference component of each laser unit are separated and decoupled to obtain the decoupled state vector. Based on the decoupled state vector, a decoupled optical field is constructed and its difference is measured in a multi-scale spatial domain and frequency domain joint measure with the target optical field distribution to obtain the optical field deviation. Based on the optical field deviation, information is propagated and aggregated on the nodes and edges of the coupled interference topology graph through a graph neural network. The independent contribution components of each laser unit and the coupled interference components of adjacent units are fused and encoded to generate a control strategy representation containing global coupling constraints. Based on the control strategy characterization, the global control objective of the laser array is decomposed into the local control objective and coupling consistency constraint of each laser unit through the dual decomposition mechanism. Under the condition of the coupling consistency constraint, the local control objective is iteratively optimized to generate modulation control commands for each laser unit in the laser array.

[0006] Nonlinear feature extraction is performed on the real-time output light field distribution to obtain the array state feature vector. Based on the array state feature vector, the mutual coupling relationship between each laser unit in the laser array is identified through a causal inference mechanism, and a coupling interference topology diagram is constructed, including: The real-time output light field distribution is extracted using a convolutional neural network to obtain spatial feature maps of multiple different receptive fields. These maps are then adaptively fused to obtain fused spatial features. The real-time output light field distribution is transformed to the frequency domain by Fourier transform, and the frequency domain amplitude distribution features and frequency domain phase distribution features are extracted and concatenated with the fused spatial features to obtain the array state feature vector. Extract the state sub-vectors corresponding to each laser unit from the array state feature vector and perform time-series sampling to obtain the state observation sequence of each laser unit at multiple time points; based on the state observation sequence, quantify the statistical dependency relationship of state changes by calculating the mutual information between each laser unit; Each laser unit is treated as a node, and the statistical dependencies between laser units are treated as directed edges to obtain a directed acyclic graph. The direct and indirect causal relationships between laser units are identified in the directed acyclic graph through conditional independence tests. The directed edges corresponding to the direct causal relationships are retained, and the redundant directed edges are deleted based on the indirect causal relationships to obtain the coupling interference topology graph.

[0007] Based on the aforementioned coupling interference topology, the independent contribution components and coupling interference components of each laser unit are separated and decoupled to obtain the decoupled state vector, which includes: Construct the causal dependency matrix of the laser array, and fill the matrix element values ​​at the corresponding positions in the causal dependency matrix based on the directed edge weights of the coupled interference topology graph. Set the matrix element values ​​between unconnected laser unit pairs to zero. The coupling interference topology graph is sorted topologically, and each laser unit is decoupled sequentially according to the topology sort. For the target laser unit being processed, the non-zero elements of the corresponding row of the target laser unit are extracted from the causal dependency matrix. The non-zero elements are multiplied and accumulated element by element with the independent contribution components of the corresponding predecessor nodes to obtain the coupling interference components received by the target laser unit. The observation state value of the target laser unit is extracted from the array state feature vector, and the coupling interference component is subtracted from the observation state value to obtain the independent contribution component of the target laser unit. Traverse all nodes in the coupled interference topology graph and sequentially separate and decouple the independent contribution component and the coupled interference component of each laser unit; then concatenate the independent contribution components of each laser unit according to the index order of the nodes in the coupled interference topology graph to obtain the decoupling state vector.

[0008] Based on the decoupled state vector, a decoupled optical field is constructed and subjected to a multi-scale spatial-frequency domain joint difference measurement with the target optical field distribution to obtain the optical field deviation, including: Based on the decoupled state vector, the complex amplitude distribution formed by each laser unit on the far-field observation plane is calculated and coherently superimposed and modulus squared to obtain the spatial light intensity distribution of the decoupled light field. Two-dimensional spatial Fourier transforms are performed on the target light field distribution and the spatial light intensity distribution respectively to obtain the target frequency domain spectrum of the target light field distribution and the decoupled frequency domain spectrum of the spatial light intensity distribution; the frequency domain difference between the target frequency domain spectrum and the decoupled frequency domain spectrum is calculated to obtain the frequency domain deviation component; and the spatial domain point-by-point difference is calculated on the target light field distribution and the spatial light intensity distribution to obtain the spatial domain deviation component. The frequency domain energy of the frequency domain deviation component is accumulated in order of increasing spatial frequency. The boundary spatial frequency corresponding to the frequency domain energy accumulation ratio reaching the preset energy boundary ratio is determined. If the spatial frequency is lower than the boundary spatial frequency, the frequency domain deviation component is set as a low-frequency deviation component; otherwise, it is set as a high-frequency deviation component. The spatial domain deviation component is segmented at multiple scales. The radial distance from each spatial location point in the far-field observation plane to the energy centroid of the light field is calculated. If the radial distance is less than a preset segmentation threshold, the spatial domain deviation component is divided into a core deviation component; otherwise, it is divided into an outer deviation component. The optical field deviation is obtained by fusing the low-frequency deviation component, the high-frequency deviation component, the core deviation component, and the peripheral deviation component.

[0009] Based on the optical field deviation, information is propagated and aggregated at the nodes and edges of the coupled interference topology graph using a graph neural network. The independent contribution components of each laser unit are fused and encoded with the coupled interference components of adjacent units to generate a control strategy representation containing global coupling constraints, including: The light field deviation is input into a graph neural network, and the light field deviation is spatially decomposed to obtain the local deviation components corresponding to each node in the coupled interference topology graph. At each node in the coupled interference topology graph, the independent contribution component and the local deviation component are concatenated to form a node feature vector, and the coupled interference components between two interconnected nodes form an edge feature vector. In the information propagation and aggregation process of the graph neural network, for the target node in the coupled interference topology graph, the node feature vectors of all neighboring nodes and the edge feature vectors of the connecting edges of the target node are extracted, and element-wise multiplication is performed to obtain the neighbor message vectors transmitted by each neighboring node to the target node; the current node feature vector of the target node and the neighbor message vector are concatenated and nonlinearly mapped to obtain the updated node feature vector of the target node. After performing a preset number of rounds of information propagation and aggregation iteration, the updated node feature vectors of all nodes in the coupled interference topology graph are obtained and global pooling is performed to obtain the global graph representation vector of the laser array; the global graph representation vector is nonlinearly decoded to obtain the control strategy representation containing global coupling constraints.

[0010] Based on the aforementioned control strategy characterization, the global control objective of the laser array is decomposed into the local control objectives and coupling consistency constraints of each laser element through a dual decomposition mechanism, including: The global control objective of the laser array is extracted and decomposed from the control strategy representation. The independent control performance index of a single laser unit is taken as the local optimization objective, and the coupling consistency condition that must be satisfied between adjacent laser units is taken as the coupling constraint term. By introducing dual variables and performing Lagrange relaxation on the coupling constraint terms, the global control objective is transformed into an augmented form to obtain the augmented objective. For each laser unit, a local control objective and a coupling consistency constraint are separated from the augmented objective. The local control objective consists of the local optimization objective and a local coupling penalty term weighted by the dual variable. The coupling consistency constraint consists of the state consistency condition between adjacent laser units and the update rule of the dual variable. The dual variable is iteratively updated, and the deviation between the current state and the state consistency condition between adjacent laser units is calculated as the coupling constraint violation amount. The update direction of the dual variable is determined by the coupling constraint violation amount between adjacent laser units. The iterative update of the dual variable drives the local control objective of each laser unit to gradually satisfy the coupling consistency constraint.

[0011] Under the constraint of coupling consistency, iterative optimization of the local control objective generates modulation control commands for each laser unit in the laser array, including: The negative gradient direction of the local control target is projected onto the tangent space of the feasible region defined by the coupling consistency constraint to obtain the gradient projection direction. The control parameters of each laser unit are updated along the gradient projection direction to obtain the updated control parameters. Based on the updated control parameters, the coupling constraint violation amount between each laser unit and its adjacent laser units is recalculated. The coupling constraint violation amount is combined with the dual variable to obtain the correction magnitude. The correction direction is determined according to the state consistency condition. A correction vector is constructed based on the correction magnitude and the correction direction. The correction vector is superimposed on the updated control parameters to obtain the corrected control parameters. Repeat the iteration until the amount of the coupling constraint violation and the amount of the decrease in the local control objective simultaneously satisfy the convergence condition; The phase control parameters, power control parameters, and beam quality control parameters are separated from the modified control parameters. The phase control parameters are mapped to phase modulation commands, the power control parameters are mapped to power modulation commands, and the beam quality control parameters are mapped to wavefront control commands, thereby obtaining the modulation control commands for each laser unit in the laser array.

[0012] A second aspect of the present invention provides an artificial intelligence-based adaptive control system for a laser array, comprising: The first unit is used to acquire the real-time output light field distribution and target light field distribution of the laser array; The second unit is used to extract nonlinear features from the real-time output light field distribution to obtain an array state feature vector. Based on the array state feature vector, the mutual coupling relationship between each laser unit in the laser array is identified through a causal inference mechanism to construct a coupling interference topology diagram. Based on the coupling interference topology diagram, the independent contribution components and coupling interference components of each laser unit are separated and decoupled to obtain a decoupled state vector. The third unit is used to construct a decoupled optical field based on the decoupled state vector and perform a multi-scale spatial domain and frequency domain joint difference measurement with the target optical field distribution to obtain the optical field deviation. The fourth unit is used to propagate and aggregate information on the nodes and edges of the coupled interference topology graph through a graph neural network according to the optical field deviation, and to fuse and encode the independent contribution components of each laser unit with the coupled interference components of the adjacent units to generate a control strategy representation containing global coupling constraints. The fifth unit is used to decompose the global control objective of the laser array into the local control objective and coupling consistency constraint of each laser unit based on the control strategy characterization through a dual decomposition mechanism, and to iteratively optimize the local control objective under the coupling consistency constraint to generate modulation control commands for each laser unit in the laser array.

[0013] A third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0014] Fourth aspect of the present invention, A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0015] The beneficial effects of this application are as follows: By employing nonlinear feature extraction and causal inference mechanisms, the mutual coupling relationships between units in a laser array can be accurately identified, and a coupling interference topology map can be constructed. This enables accurate characterization of complex coupling interference, solving the technical problem that traditional methods struggle to effectively describe complex coupling relationships between laser units. Based on the coupling interference topology map, the independent contribution components and coupling interference components of each laser unit are separated and decoupled, overcoming the limitation of traditional control methods that treat laser units as independent entities and ignore coupling effects. An innovative multi-scale spatial-frequency domain joint difference measurement method is adopted to compare the decoupled optical field distribution with the target optical field distribution, improving the accuracy and comprehensiveness of optical field deviation assessment. Through graph neural networks, information propagation and aggregation on the coupling interference topology map are performed, achieving the fusion encoding of independent contribution components and coupling interference components. This generates a control strategy representation containing global coupling constraints, improving the coordination and overall performance of the control strategy. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the adaptive control method for laser arrays based on artificial intelligence, as described in an embodiment of the present invention. Figure 2 A schematic diagram of the control process for iterative optimization of a laser array. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0019] Figure 1 This is a flowchart illustrating the adaptive control method for laser arrays based on artificial intelligence, as described in an embodiment of the present invention. Figure 1 As shown, the method includes: Acquire the real-time output light field distribution and target light field distribution of the laser array; Nonlinear feature extraction is performed on the real-time output light field distribution to obtain the array state feature vector. Based on the array state feature vector, the mutual coupling relationship between each laser unit in the laser array is identified through a causal inference mechanism, and a coupling interference topology diagram is constructed. Based on the coupling interference topology diagram, the independent contribution component and the coupling interference component of each laser unit are separated and decoupled to obtain the decoupled state vector. Based on the decoupled state vector, a decoupled optical field is constructed and its difference is measured in a multi-scale spatial domain and frequency domain joint measure with the target optical field distribution to obtain the optical field deviation. Based on the optical field deviation, information is propagated and aggregated on the nodes and edges of the coupled interference topology graph through a graph neural network. The independent contribution components of each laser unit and the coupled interference components of adjacent units are fused and encoded to generate a control strategy representation containing global coupling constraints. Based on the control strategy characterization, the global control objective of the laser array is decomposed into the local control objective and coupling consistency constraint of each laser unit through the dual decomposition mechanism. Under the condition of the coupling consistency constraint, the local control objective is iteratively optimized to generate modulation control commands for each laser unit in the laser array.

[0020] In one optional implementation, nonlinear feature extraction is performed on the real-time output light field distribution to obtain an array state feature vector. Based on the array state feature vector, the mutual coupling relationship between each laser unit in the laser array is identified through a causal inference mechanism, and a coupling interference topology diagram is constructed, including: The real-time output light field distribution is extracted using a convolutional neural network to obtain spatial feature maps of multiple different receptive fields. These maps are then adaptively fused to obtain fused spatial features. The real-time output light field distribution is transformed to the frequency domain by Fourier transform, and the frequency domain amplitude distribution features and frequency domain phase distribution features are extracted and concatenated with the fused spatial features to obtain the array state feature vector. Extract the state sub-vectors corresponding to each laser unit from the array state feature vector and perform time-series sampling to obtain the state observation sequence of each laser unit at multiple time points; based on the state observation sequence, quantify the statistical dependency relationship of state changes by calculating the mutual information between each laser unit; Each laser unit is treated as a node, and the statistical dependencies between laser units are treated as directed edges to obtain a directed acyclic graph. The direct and indirect causal relationships between laser units are identified in the directed acyclic graph through conditional independence tests. The directed edges corresponding to the direct causal relationships are retained, and the redundant directed edges are deleted based on the indirect causal relationships to obtain the coupling interference topology graph.

[0021] The real-time output light field distribution of the laser array is acquired by a high-speed camera or light field detector to form a two-dimensional or three-dimensional data matrix, recording the light intensity and phase information of each spatial point. The acquisition frequency is generally set to 100-1000 frames per second to ensure that the details of the dynamic changes of the laser array are captured.

[0022] Nonlinear feature extraction is performed on the acquired real-time output light field distribution to obtain the array state feature vector. This process is divided into two branches: spatial feature extraction and frequency domain feature extraction.

[0023] In the spatial feature extraction branch, a multi-layer convolutional neural network is used to extract multi-scale features from the light field distribution. A network structure with five convolutional layers is constructed, with each layer using different kernel sizes (3×3, 5×5, and 7×7) to extract features from different receptive fields. The first layer uses 64 3×3 kernels to extract local texture features; the second layer uses 128 3×3 and 5×5 kernels to extract medium-scale features; the third to fifth layers progressively increase the number and size of kernels to capture a wider range of spatial correlations. A ReLU activation function is applied after each convolutional operation to enhance non-linear expressiveness. To adaptively fuse multi-scale features, an attention mechanism is designed to weight the feature maps of each layer. Channel attention weights are calculated for each feature map, with higher-importance channels assigned larger weights. The fusion method uses weighted summation, where each layer's feature map is multiplied by its corresponding weight and then summed to obtain a fused spatial feature map with dimensions H×W×C, where H and W are the spatial dimensions and C is the number of feature channels.

[0024] In the frequency domain feature extraction branch, a two-dimensional fast Fourier transform is performed on the real-time output light field distribution to convert the spatial domain information to the frequency domain. Frequency domain amplitude distribution features and frequency domain phase distribution features are extracted from the transformation results. The frequency domain amplitude features reflect the energy distribution of the light field's spatial frequency, while the phase features contain the coherence information of the light field. During the extraction process, a logarithmic transform is used to process the amplitude to enhance contrast, and the phase information is normalized to the [-π, π] interval. The fused spatial features are concatenated with the frequency domain amplitude and phase features for encoding. A fully connected layer maps the three features to a unified feature space, and nonlinear transformations enhance the expressive power, ultimately generating a fixed-dimensional array state feature vector (e.g., 1024-dimensional).

[0025] Extract the state sub-vectors corresponding to each laser unit from the array state feature vector. Assuming the laser array contains N units, the feature vector is uniformly divided into N sub-vectors, each with a dimension of 1024 / N. Perform time-series sampling on each sub-vector, collecting state observation data at T consecutive time points to form N time-series matrices, each with a dimension of T×(1024 / N).

[0026] For any two laser units i and j, a nonparametric method based on kernel density estimation is used to calculate the mutual information MI(i,j) between their corresponding state vector sequences to capture nonlinear dependencies. A higher mutual information value indicates a stronger coupling relationship between the two units.

[0027] An initial directed acyclic graph is constructed, with each laser unit as a node and the mutual information value as the edge weight. A greedy search algorithm based on the information criterion is used to determine the direction of the edges. If MI(i,j) is greater than a preset threshold (e.g., 0.3), a directed edge is established between units i and j, with the direction determined by the chronological order.

[0028] Direct and indirect causal relationships are identified through conditional independence tests. For any three units i, j, and k, if edges i→j and j→k exist, it is necessary to test whether i→k is a redundant edge. The conditional mutual information CMI(i, k|j) is calculated. If CMI(i, k|j) is significantly less than MI(i, k) (the reduction exceeds 70%), then i→k is considered an indirect causal relationship, and the edge should be deleted.

[0029] Through the above processing, a simplified coupling interference topology diagram is finally obtained, which clearly shows the direct coupling relationship between laser units. This topology diagram can be used to guide the coherent control and performance optimization of laser arrays, such as phase-locked enhancement and suppression of unstable modes.

[0030] In one optional implementation, the independent contribution components and coupling interference components of each laser unit are separated and decoupled based on the coupling interference topology diagram to obtain a decoupled state vector including: Construct the causal dependency matrix of the laser array, and fill the matrix element values ​​at the corresponding positions in the causal dependency matrix based on the directed edge weights of the coupled interference topology graph. Set the matrix element values ​​between unconnected laser unit pairs to zero. The coupling interference topology graph is sorted topologically, and each laser unit is decoupled sequentially according to the topology sort. For the target laser unit being processed, the non-zero elements of the corresponding row of the target laser unit are extracted from the causal dependency matrix. The non-zero elements are multiplied and accumulated element by element with the independent contribution components of the corresponding predecessor nodes to obtain the coupling interference components received by the target laser unit. The observation state value of the target laser unit is extracted from the array state feature vector, and the coupling interference component is subtracted from the observation state value to obtain the independent contribution component of the target laser unit. Traverse all nodes in the coupled interference topology graph and sequentially separate and decouple the independent contribution component and the coupled interference component of each laser unit; then concatenate the independent contribution components of each laser unit according to the index order of the nodes in the coupled interference topology graph to obtain the decoupling state vector.

[0031] Construct a causal dependency matrix for the laser array. This matrix represents the influence relationships between the laser units. Based on the directed edge weights of the coupling interference topology graph, fill the corresponding matrix element values ​​in the causal dependency matrix. For unconnected laser unit pairs, set the corresponding matrix element value to zero. For example, for an array containing five laser units, if the coupling interference weight of unit 1 to unit 3 is 0.15, then fill in 0.15 in the 3rd row and 1st column of the causal dependency matrix; if there is no direct coupling between unit 2 and unit 4, then fill in 0 in the corresponding position of the causal dependency matrix.

[0032] The topological sorting of the coupled interference topology graph is achieved through depth-first search or breadth-first search, determining the order of decoupling processing. The result ensures that each node is processed after all its predecessor nodes. For example, for a topology graph, the sorting result is [1, 2, 3, 4, 5], which means that decoupling processing is performed in the order of unit 1, unit 2, unit 3, unit 4, and unit 5.

[0033] Each laser unit is processed sequentially according to its topological order. For the target laser unit being processed, the non-zero elements in the corresponding row of the causal dependency matrix are extracted. These elements represent the influence weights of the predecessor nodes on the current unit. For example, when processing unit 3, if the third row of the causal dependency matrix contains non-zero elements 0.15 and 0.22, located in the first and second columns respectively, it indicates that unit 1 and unit 2 have coupling interference with unit 3.

[0034] The coupling interference component received by the target laser unit is calculated by multiplying and accumulating these non-zero elements with the independent contribution components of the corresponding predecessor nodes element by element. Continuing the example above, if the independent contribution components of unit 1 and unit 2 are 2.5 and 3.1 respectively, then the coupling interference component received by unit 3 is 0.15×2.5 + 0.22×3.1 = 1.057.

[0035] The observed state value of the target laser element is extracted from the array state feature vector. This value includes the independent contribution component and the coupling interference component. The independent contribution component of the target laser element is obtained by subtracting the calculated coupling interference component from the observed state value. For example, if the observed state value of element 3 is 5.2, then its independent contribution component is 5.2 - 1.057 = 4.143.

[0036] Following the steps described above, all nodes in the coupled interference topology are traversed, and the independent contribution components and coupled interference components of each laser unit are separated and decoupled in turn. In practical applications, the decoupling process needs to consider multiple physical quantities, such as light intensity, frequency, and phase. Therefore, the state of each laser unit is a multidimensional vector.

[0037] After decoupling all laser units, the independent contribution components of each unit are concatenated according to the index order of the nodes in the coupling interference topology diagram to obtain the decoupling state vector. For example, for five laser units, if the independent contribution components of each unit are [3.2, 2.7, 4.143, 3.8, 2.9], then the decoupling state vector is [3.2, 2.7, 4.143, 3.8, 2.9].

[0038] The decoupling state vector reflects the inherent state of each laser unit in the absence of external coupling interference. It can be used for subsequent anomaly detection and performance evaluation. By comparing the decoupling state vectors at different time points, the performance change trends of the laser units can be identified. Analyzing the difference between the decoupling state vector and the expected value can detect potential faults or anomalies. The decoupling process can also be extended to dynamic scenarios. By repeatedly performing the decoupling operation at different time points, the state changes of each laser unit and the dynamic evolution of coupling relationships can be tracked. In complex laser array systems, this decoupling analysis is of great significance for understanding system behavior, optimizing control strategies, and improving system reliability.

[0039] In one optional implementation, a decoupled optical field is constructed based on the decoupled state vector, and a multi-scale spatial-frequency domain joint difference metric is performed with the target optical field distribution to obtain the optical field deviation, including: Based on the decoupled state vector, the complex amplitude distribution formed by each laser unit on the far-field observation plane is calculated and coherently superimposed and modulus squared to obtain the spatial light intensity distribution of the decoupled light field. Two-dimensional spatial Fourier transforms are performed on the target light field distribution and the spatial light intensity distribution respectively to obtain the target frequency domain spectrum of the target light field distribution and the decoupled frequency domain spectrum of the spatial light intensity distribution; the frequency domain difference between the target frequency domain spectrum and the decoupled frequency domain spectrum is calculated to obtain the frequency domain deviation component; and the spatial domain point-by-point difference is calculated on the target light field distribution and the spatial light intensity distribution to obtain the spatial domain deviation component. The frequency domain energy of the frequency domain deviation component is accumulated in order of increasing spatial frequency. The boundary spatial frequency corresponding to the frequency domain energy accumulation ratio reaching the preset energy boundary ratio is determined. If the spatial frequency is lower than the boundary spatial frequency, the frequency domain deviation component is set as a low-frequency deviation component; otherwise, it is set as a high-frequency deviation component. The spatial domain deviation component is segmented at multiple scales. The radial distance from each spatial location point in the far-field observation plane to the energy centroid of the light field is calculated. If the radial distance is less than a preset segmentation threshold, the spatial domain deviation component is divided into a core deviation component; otherwise, it is divided into an outer deviation component. The optical field deviation is obtained by fusing the low-frequency deviation component, the high-frequency deviation component, the core deviation component, and the peripheral deviation component.

[0040] Based on the decoupled state vector, the complex amplitude distribution formed by each laser unit on the far-field observation plane is calculated. For each laser unit, its complex amplitude can be expressed as a combination of the amplitude and phase information of that unit. Suppose the laser array contains N units, and the complex amplitude generated by each unit on the far-field observation plane can be calculated by the optical field transmission model, which is usually approximated by the Fresnel-Kirchhoff diffraction formula. For the i-th unit, its complex amplitude at the position (x, y) on the far-field observation plane can be expressed as Ai(x, y).

[0041] The complex amplitudes generated by all laser units are coherently superimposed, i.e., the complex amplitudes of each unit are added together to obtain the total complex amplitude field A(x, y). The modulus square operation is then performed on this complex amplitude field to obtain the spatial intensity distribution I(x, y) of the decoupled optical field. This step can be expressed as summing the complex amplitudes of N laser units and then taking the modulus square, i.e., I(x, y) = |∑Ai(x, y)| 2 .

[0042] Two-dimensional spatial Fourier transforms are performed on the target light field distribution T(x, y) and the spatial light intensity distribution I(x, y) decoupled into a light field, respectively, to transform the light intensity distribution in the spatial domain to the frequency domain, thus obtaining the target frequency domain spectrum T(fx, fy) of the target light field distribution and the decoupled frequency domain spectrum I(fx, fy) of the spatial light intensity distribution, where fx and fy represent the spatial frequencies in the x and y directions, respectively.

[0043] The frequency domain difference between the target frequency domain spectrum and the decoupled frequency domain spectrum is usually calculated by taking the square of the amplitude difference as the frequency domain deviation component, i.e., DF(fx, fy) = |T(fx, fy) - I(fx, fy)| 2 Furthermore, the spatial domain point-by-point difference between the target light field distribution T(x,y) and the spatial light intensity distribution I(x,y) is calculated to obtain the spatial domain deviation component DS(x,y)=|T(x,y)-I(x,y)|.

[0044] The frequency domain energy of the frequency domain deviation components is accumulated in ascending order of spatial frequency. The total energy EF_total of the frequency domain deviation components DF(fx, fy) is calculated. The frequency domain deviation components are integrated with the origin in ascending order of radial distance to obtain the accumulated energy EF_cum(f) at different spatial frequencies. When EF_cum(f) / EF_total reaches a preset energy threshold ratio (e.g., 0.7), the corresponding spatial frequency f is determined as the threshold spatial frequency fc.

[0045] Based on the boundary spatial frequency fc, the frequency domain deviation component DF(fx, fy) is divided into two parts: when the spatial frequency is lower than the boundary spatial frequency, DF(fx, fy) is set as the low-frequency deviation component DF_low(fx, fy); when the spatial frequency is higher than or equal to the boundary spatial frequency, DF(fx, fy) is set as the high-frequency deviation component DF_high(fx, fy).

[0046] The spatial domain bias component DS(x, y) is segmented into multiple scales, and the coordinates (xc, yc) of the energy centroid of the light field in the far-field observation plane are calculated. The energy centroid can be obtained by calculating the first moment of the spatial light intensity distribution I(x, y). The radial distance r from each spatial point (x, y) to the energy centroid of the light field is calculated as r = ((x - xc)). 2 +(y-yc) 2 ) 1 / 2 If the radial distance r is less than the preset segmentation threshold rt (e.g., half the half-width of the light spot), then the spatial domain deviation component DS(x, y) is divided into the core deviation component DS_core(x, y); otherwise, it is divided into the peripheral deviation component DS_surr(x, y).

[0047] The optical field deviation D is obtained by fusing the low-frequency deviation component DF_low, the high-frequency deviation component DF_high, the core deviation component DS_core, and the peripheral deviation component DS_surr using a weighted summation method.

[0048] In practical applications, the boundary spatial frequency fc and the segmentation threshold rt can be adjusted according to the characteristics of the laser array and the requirements of the target optical field. For example, when there are many laser units, a higher spatial frequency optical field can be synthesized. In this case, the value of fc can be increased to control the high-frequency components more precisely. When the energy of the target optical field distribution is mainly concentrated in the central region, the value of rt can be appropriately reduced to more accurately assess the deviation of the core region.

[0049] In one optional implementation, based on the optical field deviation, information is propagated and aggregated at the nodes and edges of the coupled interference topology graph using a graph neural network. The independent contribution components of each laser unit are fused and encoded with the coupled interference components of adjacent units to generate a control strategy representation containing global coupling constraints, including: The light field deviation is input into a graph neural network, and the light field deviation is spatially decomposed to obtain the local deviation components corresponding to each node in the coupled interference topology graph. At each node in the coupled interference topology graph, the independent contribution component and the local deviation component are concatenated to form a node feature vector, and the coupled interference components between two interconnected nodes form an edge feature vector. In the information propagation and aggregation process of the graph neural network, for the target node in the coupled interference topology graph, the node feature vectors of all neighboring nodes and the edge feature vectors of the connecting edges of the target node are extracted, and element-wise multiplication is performed to obtain the neighbor message vectors transmitted by each neighboring node to the target node; the current node feature vector of the target node and the neighbor message vector are concatenated and nonlinearly mapped to obtain the updated node feature vector of the target node. After performing a preset number of rounds of information propagation and aggregation iteration, the updated node feature vectors of all nodes in the coupled interference topology graph are obtained and global pooling is performed to obtain the global graph representation vector of the laser array; the global graph representation vector is nonlinearly decoded to obtain the control strategy representation containing global coupling constraints.

[0050] The optical field deviation is input into a graph neural network for spatial decomposition. This process can be achieved through multi-layer convolution operations. The error matrix between the measured optical field and the ideal optical field of each unit in the laser array is used as input, and local features are extracted through convolutional filters. For an M×N laser array, the optical field deviation can be represented as a D-dimensional tensor. After processing by K convolutional kernels, the local deviation component corresponding to each laser unit is obtained.

[0051] After spatial decomposition, the node and edge features of the coupling interference topology are constructed. For each node i in the topology, its feature vector is formed by concatenating independent contribution components and local deviation components. The independent contribution components reflect the internal parameters of a single laser, such as phase, amplitude, and polarization state; the local deviation components describe the error between the actual optical field and the target optical field of the unit. The concatenation of these two components forms a node feature vector Fi of length d. For the edge between nodes i and j, its edge feature vector Eij is composed of the coupling interference components between the two nodes, containing information on thermal, mechanical, and electromagnetic coupling effects.

[0052] For a target node v in the topological graph, determine the set N(v) of all its neighboring nodes. In each iteration, extract the node feature vector Fu and the edge feature vector Euv of the connecting edge from each neighboring node u∈N(v), and perform element-wise multiplication to obtain the neighbor message vector Muv = Fu ⊗ Euv.

[0053] The current feature vector Fv of the target node v is aggregated and concatenated with all neighbor message vectors Muv. A non-linear mapping is then performed using a multilayer perceptron (MLP) to update the node's feature vector. The aggregated vector Av is obtained by summing or averaging all neighbor message vectors. Fv and Av are concatenated to form a long vector [Fv || Av]. A non-linear transformation is then performed using the MLP to obtain the updated feature vector Fv' of node v. The MLP contains two fully connected layers with a ReLU activation function in between to ensure that the feature vector fully extracts coupling relationship information.

[0054] After the graph neural network iterates for a preset number of rounds (e.g., 3 rounds), it collects the updated feature vectors of all nodes and performs a global pooling operation. The global pooling can be achieved using average pooling or max pooling, merging the feature vectors of all nodes into a single global graph representation vector G. This vector contains the global coupling constraint information of the entire laser array.

[0055] The global graph representation vector G is nonlinearly decoded by a decoding network consisting of multiple fully connected neural networks. The high-dimensional graph representation vector is mapped to a control parameter space. The output control strategy representation includes the adjustment parameters of each laser unit, such as phase compensation value, drive current adjustment amount, temperature control parameters, etc. These parameters have already taken into account the global coupling constraints.

[0056] In practical applications, if the output light field of a laser unit deviates, this method can comprehensively consider the adjustment needs of the unit itself and the coupling effects between neighboring units to generate optimal control commands. For example, for a 16-unit fiber laser array, when the phase of the four central units drifts, traditional methods only adjust the abnormal units, which can easily lead to cascading deviations. However, the control strategy generated by this method not only compensates for the abnormal units but also pre-adjusts the surrounding units, effectively suppressing the propagation of coupling interference and reducing the recovery time of the synthesized spot quality from 150 milliseconds to 35 milliseconds.

[0057] By utilizing the message passing mechanism of graph neural networks on the coupled interference topology, this method effectively integrates the independent contributions and coupled interference information of each laser unit, generating a control strategy with global optimality and realizing high-precision coherent synthesis control of the laser array.

[0058] In one optional implementation, based on the control strategy characterization, the global control objective of the laser array is decomposed into local control objectives and coupling consistency constraints for each laser unit through a dual decomposition mechanism, including: The global control objective of the laser array is extracted and decomposed from the control strategy representation. The independent control performance index of a single laser unit is taken as the local optimization objective, and the coupling consistency condition that must be satisfied between adjacent laser units is taken as the coupling constraint term. By introducing dual variables and performing Lagrange relaxation on the coupling constraint terms, the global control objective is transformed into an augmented form to obtain the augmented objective. For each laser unit, a local control objective and a coupling consistency constraint are separated from the augmented objective. The local control objective consists of the local optimization objective and a local coupling penalty term weighted by the dual variable. The coupling consistency constraint consists of the state consistency condition between adjacent laser units and the update rule of the dual variable. The dual variable is iteratively updated, and the deviation between the current state and the state consistency condition between adjacent laser units is calculated as the coupling constraint violation amount. The update direction of the dual variable is determined by the coupling constraint violation amount between adjacent laser units. The iterative update of the dual variable drives the local control objective of each laser unit to gradually satisfy the coupling consistency constraint.

[0059] The global control objective of the laser array is extracted from the control strategy representation. For an array containing N laser elements, the global control objective can typically be represented as the sum of the control performance indicators of each element and the coupling consistency requirements. The independent control performance indicators of individual laser elements (such as output power stability, phase control accuracy, etc.) are used as local optimization objectives. At the same time, the coupling consistency conditions that adjacent laser elements must satisfy, such as phase locking requirements and power equalization conditions, are identified as coupling constraints.

[0060] For example, in a phase-locked laser array, the local optimization objective focuses on the phase stability of individual lasers, while coupling constraints require that the phase difference between adjacent lasers be maintained within a preset range. This decomposition transforms the complex global control problem into a more manageable local problem and constraint.

[0061] We introduce dual variables to perform Lagrangian relaxation on the coupling constraint terms. For each pair of adjacent laser units i and j, we introduce a corresponding dual variable λij, which represents the "price" or "penalty" for violating the coupling constraint. Specifically, the original constraint optimization problem requires each laser unit to optimize its control objective while strictly satisfying the coupling constraint. This hard constraint makes the problem difficult to decompose and solve. By introducing dual variables, we transform the coupling constraint from a hard constraint to a soft constraint, that is, the degree of constraint satisfaction is added to the objective function in the form of a penalty term.

[0062] The coupling constraints between all adjacent laser unit pairs are identified, and a dual variable is introduced for each coupling constraint. A dual penalty term is constructed, which is equal to the product of the dual variable and the amount of coupling constraint violation. All dual penalty terms are accumulated onto the original local optimization objective to form an augmented objective function. Specifically, for laser unit i, its augmented objective is equal to its original local optimization objective, plus the sum of the dual penalty terms corresponding to all laser units j adjacent to unit i. Each dual penalty term is calculated as the dual variable λij multiplied by the current state difference between unit i and unit j.

[0063] When the coupling constraint between a pair of laser units is violated, the corresponding state difference increases, leading to a larger dual penalty term. This worsens the augmented objective, driving the optimization process towards satisfying the constraints. Through this transformation, the coupling constraints that originally required collaborative satisfaction by each unit are relaxed into penalty terms in the local objective function of each unit. This allows each laser unit to optimize independently based on the augmented objective, while gradually satisfying the coupling consistency constraint through the adjustment mechanism of the dual variable. The key to this method is to centralize the distributed constraints in the local objectives of each unit, maintaining the distributed nature of the computation while gradually satisfying the global constraints during the optimization process.

[0064] Based on this, for each laser unit, local control objectives and coupling consistency constraints are separated from the augmented objective. For laser unit i, its local control objective consists of the original local optimization objective and a local coupling penalty term weighted by the dual variable. The local coupling penalty term reflects the difference in coupling state between this unit and its neighboring units, and the weighting coefficients are determined by the dual variable. Meanwhile, the coupling consistency constraints include two parts: state consistency conditions between adjacent laser units (such as phase difference, power ratio, and other technical indicators) and update rules for the dual variable.

[0065] State consistency conditions can be set according to specific application requirements. For example, for phase locking requirements, the phase difference between adjacent lasers needs to be kept near a specific value; for power equalization requirements, the output power ratio of adjacent units needs to be close to a preset value.

[0066] In each iteration, the deviation between the current state and the state consistency condition between adjacent laser units is calculated as a coupling constraint violation. For example, for phase locking requirements, the deviation between the current phase difference and the target phase difference can be calculated; for power equalization requirements, the deviation between the current power ratio and the target power ratio can be calculated.

[0067] The update direction of the dual variable is determined based on the degree of constraint violation. When the degree of constraint violation increases, the corresponding dual variable increases, which increases the weight of the penalty term in the local control objective and prompts each unit to pay more attention to satisfying the coupling constraints. When the constraints are gradually satisfied, the dual variable decreases, which allows each unit to pay more attention to its own local optimization objective.

[0068] The specific dual variable update rule can be expressed as: the new dual variable value equals the current dual variable value plus the step size factor multiplied by the coupling constraint violation. The step size factor controls the convergence speed and can be adjusted according to the system's dynamic characteristics. Through iterative updates of the dual variables, the local control objectives of each laser unit are driven to gradually satisfy the coupling consistency constraints.

[0069] In practical applications, a distributed controller can be designed to operate independently on the processors of each laser unit, requiring only the exchange of state information and dual variables with adjacent units. This design significantly reduces communication overhead and improves system robustness. When a laser unit fails, only the units directly connected to it are affected, preventing the entire system from collapsing.

[0070] For specific laser array systems, the parameters of the dual decomposition mechanism can be adjusted according to actual needs. For example, in large fiber laser arrays, more attention needs to be paid to phase-locking accuracy; in semiconductor laser arrays, more attention needs to be paid to the power imbalance caused by temperature fluctuations.

[0071] Through this dual decomposition mechanism, the control of the laser array achieves the global optimization goal while maintaining the distributed characteristics of the control architecture, balancing the relationship between system performance and implementation complexity, and providing an effective solution for the collaborative control of large-scale laser arrays.

[0072] In one optional implementation, iterative optimization of the local control objective under the coupling consistency constraint to generate modulation control commands for each laser unit in the laser array includes: The negative gradient direction of the local control target is projected onto the tangent space of the feasible region defined by the coupling consistency constraint to obtain the gradient projection direction. The control parameters of each laser unit are updated along the gradient projection direction to obtain the updated control parameters. Based on the updated control parameters, the coupling constraint violation amount between each laser unit and its adjacent laser units is recalculated. The coupling constraint violation amount is combined with the dual variable to obtain the correction magnitude. The correction direction is determined according to the state consistency condition. A correction vector is constructed based on the correction magnitude and the correction direction. The correction vector is superimposed on the updated control parameters to obtain the corrected control parameters. Repeat the iteration until the amount of the coupling constraint violation and the amount of the decrease in the local control objective simultaneously satisfy the convergence condition; The phase control parameters, power control parameters, and beam quality control parameters are separated from the modified control parameters. The phase control parameters are mapped to phase modulation commands, the power control parameters are mapped to power modulation commands, and the beam quality control parameters are mapped to wavefront control commands, thereby obtaining the modulation control commands for each laser unit in the laser array.

[0073] like Figure 2 As shown, the method includes: The negative gradient direction of the local control objective is projected onto the tangent space of the feasible region defined by the coupling consistency constraint to obtain the gradient projection direction. For each laser unit, the gradient of its local control objective function with respect to the control parameters is calculated. Due to the coupling relationship between laser units, these parameters cannot be arbitrarily adjusted and must satisfy the coupling consistency constraint. Therefore, projecting the negative gradient direction onto the tangent space of the feasible region defined by the coupling consistency constraint yields a gradient projection direction that both improves the control objective and satisfies the constraint conditions.

[0074] Taking phase control as an example, suppose there are N laser units, the phase parameter of the i-th unit is φi, and the local control objective function is Li(φi). After calculating the negative gradient direction -∇Li(φi), it needs to be projected onto a constrained space that satisfies the condition that the phase difference between adjacent units does not exceed the allowable value Δφmax. This can be achieved through the projection matrix P, and the projected direction is P·(-∇Li(φi)).

[0075] The control parameters of each laser unit are updated along the gradient projection direction to obtain the updated control parameters. A step size α is used, and the update formula is: New parameter = Old parameter + α × Gradient projection direction. The step size α can be determined by methods such as line search to ensure that the objective function decreases sufficiently. For example, for the power control parameter Pi, it is updated to Pi' = Pi + α·(P·(-∇Li(Pi))).

[0076] Based on the updated control parameters, the coupling constraint violation amounts between each laser element and its adjacent laser elements are recalculated. For example, for the phase coupling constraint |φi-φj|≤Δφmax, where i and j are adjacent elements, the violation amount can be expressed as max(0, |φi'-φj'|-Δφmax). The violation amounts for power and beam quality parameters are calculated similarly.

[0077] The correction magnitude is obtained by combining the coupling constraint violation amount with the dual variable. The dual variable reflects the importance of the constraint, is initially set to zero, and is dynamically adjusted during the iteration process. Let λij represent the dual variable related to the phase constraint |φi-φj|≤Δφmax, and the correction magnitude can be expressed as λij×violation amount. Similarly, power and beam quality parameters also have corresponding correction magnitudes.

[0078] The state consistency condition requires that the parameter differences between adjacent units be within an allowable range. If φi' is greater than φj' and the difference exceeds the threshold, then φi needs to be decreased while φj needs to be increased; conversely, the opposite is also true, thus determining the direction of correction.

[0079] A correction vector is constructed based on the correction magnitude and correction direction. This correction vector is then superimposed onto the updated control parameters to obtain the corrected control parameters. For example, if the phase constraint between the i-th unit and its adjacent j-th unit is violated, the component of the correction vector for the i-th unit is -λij × violation amount (decreasing φi), and the component for the j-th unit is +λij × violation amount (increasing φj).

[0080] Repeat the above iterative process until both the amount of coupling constraint violation and the decrease in the local control objective simultaneously satisfy the convergence condition. The convergence condition can be set as follows: all constraint violations are less than a preset threshold ε1, and the relative change of the objective function is less than a threshold ε2 in several consecutive iterations.

[0081] From the revised control parameters, phase control parameters, power control parameters, and beam quality control parameters are separated. Phase control parameters are mapped to phase modulation commands: the optimized phase parameter values ​​are converted into digital commands acceptable to the phase modulator, such as converting radian values ​​into digital control values ​​for the modulator. Power control parameters are mapped to power modulation commands: the optimized power parameter values ​​are converted into control commands for the pump source or attenuator, such as current values ​​or attenuation ratios. Beam quality control parameters are mapped to wavefront control commands: the optimized beam quality parameters are converted into control commands for the adaptive optics system, such as the driving voltage of the deformable mirror.

[0082] Through the above steps, iterative optimization of the local control objective under coupling consistency constraints was completed, and modulation control commands for each laser unit in the laser array were generated. These commands can be directly applied to the real-time control of the laser array to optimize the overall performance.

[0083] In practical applications, this method can be used for coherent combining of high-power laser systems. By precisely controlling the phase, power, and beam quality parameters of each laser unit, the beam output by the laser array can achieve the best combining effect. It is suitable for laser array control in fields such as lidar, laser processing, and laser communication.

[0084] This invention relates to an artificial intelligence-based adaptive control system for laser arrays, the system comprising: The first unit is used to acquire the real-time output light field distribution and target light field distribution of the laser array; The second unit is used to extract nonlinear features from the real-time output light field distribution to obtain an array state feature vector. Based on the array state feature vector, the mutual coupling relationship between each laser unit in the laser array is identified through a causal inference mechanism to construct a coupling interference topology diagram. Based on the coupling interference topology diagram, the independent contribution components and coupling interference components of each laser unit are separated and decoupled to obtain a decoupled state vector. The third unit is used to construct a decoupled optical field based on the decoupled state vector and perform a multi-scale spatial domain and frequency domain joint difference measurement with the target optical field distribution to obtain the optical field deviation. The fourth unit is used to propagate and aggregate information on the nodes and edges of the coupled interference topology graph through a graph neural network according to the optical field deviation, and to fuse and encode the independent contribution components of each laser unit with the coupled interference components of the adjacent units to generate a control strategy representation containing global coupling constraints. The fifth unit is used to decompose the global control objective of the laser array into the local control objective and coupling consistency constraint of each laser unit based on the control strategy characterization through a dual decomposition mechanism, and to iteratively optimize the local control objective under the coupling consistency constraint to generate modulation control commands for each laser unit in the laser array.

[0085] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0086] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0087] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0088] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An adaptive control method for laser arrays based on artificial intelligence, characterized in that, include: Acquire the real-time output light field distribution and target light field distribution of the laser array; Nonlinear feature extraction is performed on the real-time output light field distribution to obtain the array state feature vector. Based on the array state feature vector, the mutual coupling relationship between each laser unit in the laser array is identified through a causal inference mechanism, and a coupling interference topology diagram is constructed. Based on the coupling interference topology diagram, the independent contribution component and the coupling interference component of each laser unit are separated and decoupled to obtain the decoupled state vector. Based on the decoupled state vector, a decoupled optical field is constructed and its difference is measured in a multi-scale spatial domain and frequency domain joint measure with the target optical field distribution to obtain the optical field deviation. Based on the optical field deviation, information is propagated and aggregated on the nodes and edges of the coupled interference topology graph through a graph neural network. The independent contribution components of each laser unit and the coupled interference components of adjacent units are fused and encoded to generate a control strategy representation containing global coupling constraints. Based on the control strategy characterization, the global control objective of the laser array is decomposed into the local control objective and coupling consistency constraint of each laser unit through the dual decomposition mechanism. Under the condition of the coupling consistency constraint, the local control objective is iteratively optimized to generate modulation control commands for each laser unit in the laser array.

2. The method according to claim 1, characterized in that, Nonlinear feature extraction is performed on the real-time output light field distribution to obtain the array state feature vector. Based on the array state feature vector, the mutual coupling relationship between each laser unit in the laser array is identified through a causal inference mechanism, and a coupling interference topology diagram is constructed, including: The real-time output light field distribution is extracted using a convolutional neural network to obtain spatial feature maps of multiple different receptive fields. These maps are then adaptively fused to obtain fused spatial features. The real-time output light field distribution is transformed to the frequency domain by Fourier transform, and the frequency domain amplitude distribution features and frequency domain phase distribution features are extracted and concatenated with the fused spatial features to obtain the array state feature vector. Extract the state sub-vectors corresponding to each laser unit from the array state feature vector and perform time-series sampling to obtain the state observation sequence of each laser unit at multiple time points; based on the state observation sequence, quantify the statistical dependency relationship of state changes by calculating the mutual information between each laser unit; Each laser unit is treated as a node, and the statistical dependencies between laser units are treated as directed edges to obtain a directed acyclic graph. The direct and indirect causal relationships between laser units are identified in the directed acyclic graph through conditional independence tests. The directed edges corresponding to the direct causal relationships are retained, and the redundant directed edges are deleted based on the indirect causal relationships to obtain the coupling interference topology graph.

3. The method according to claim 1, characterized in that, Based on the aforementioned coupling interference topology, the independent contribution components and coupling interference components of each laser unit are separated and decoupled to obtain the decoupled state vector, which includes: Construct the causal dependency matrix of the laser array, and fill the matrix element values ​​at the corresponding positions in the causal dependency matrix based on the directed edge weights of the coupled interference topology graph. Set the matrix element values ​​between unconnected laser unit pairs to zero. The coupling interference topology graph is sorted topologically, and each laser unit is decoupled sequentially according to the topology sort. For the target laser unit being processed, the non-zero elements of the corresponding row of the target laser unit are extracted from the causal dependency matrix. The non-zero elements are multiplied and accumulated element by element with the independent contribution components of the corresponding predecessor nodes to obtain the coupling interference components received by the target laser unit. The observation state value of the target laser unit is extracted from the array state feature vector, and the coupling interference component is subtracted from the observation state value to obtain the independent contribution component of the target laser unit. Traverse all nodes in the coupled interference topology graph and sequentially separate and decouple the independent contribution component and the coupled interference component of each laser unit; then concatenate the independent contribution components of each laser unit according to the index order of the nodes in the coupled interference topology graph to obtain the decoupling state vector.

4. The method according to claim 1, characterized in that, Based on the decoupled state vector, a decoupled optical field is constructed and subjected to a multi-scale spatial-frequency domain joint difference measurement with the target optical field distribution to obtain the optical field deviation, including: Based on the decoupled state vector, the complex amplitude distribution formed by each laser unit on the far-field observation plane is calculated and coherently superimposed and modulus squared to obtain the spatial light intensity distribution of the decoupled light field. Two-dimensional spatial Fourier transforms are performed on the target light field distribution and the spatial light intensity distribution respectively to obtain the target frequency domain spectrum of the target light field distribution and the decoupled frequency domain spectrum of the spatial light intensity distribution; the frequency domain difference between the target frequency domain spectrum and the decoupled frequency domain spectrum is calculated to obtain the frequency domain deviation component; and the spatial domain point-by-point difference is calculated on the target light field distribution and the spatial light intensity distribution to obtain the spatial domain deviation component. The frequency domain energy of the frequency domain deviation component is accumulated in order of increasing spatial frequency. The boundary spatial frequency corresponding to the frequency domain energy accumulation ratio reaching the preset energy boundary ratio is determined. If the spatial frequency is lower than the boundary spatial frequency, the frequency domain deviation component is set as a low-frequency deviation component; otherwise, it is set as a high-frequency deviation component. The spatial domain deviation component is segmented at multiple scales. The radial distance from each spatial location point in the far-field observation plane to the energy centroid of the light field is calculated. If the radial distance is less than a preset segmentation threshold, the spatial domain deviation component is divided into a core deviation component; otherwise, it is divided into an outer deviation component. The optical field deviation is obtained by fusing the low-frequency deviation component, the high-frequency deviation component, the core deviation component, and the peripheral deviation component.

5. The method according to claim 1, characterized in that, Based on the optical field deviation, information is propagated and aggregated at the nodes and edges of the coupled interference topology graph using a graph neural network. The independent contribution components of each laser unit are fused and encoded with the coupled interference components of adjacent units to generate a control strategy representation containing global coupling constraints, including: The light field deviation is input into a graph neural network, and the light field deviation is spatially decomposed to obtain the local deviation components corresponding to each node in the coupled interference topology graph. At each node in the coupled interference topology graph, the independent contribution component and the local deviation component are concatenated to form a node feature vector, and the coupled interference components between two interconnected nodes form an edge feature vector. In the information propagation and aggregation process of the graph neural network, for the target node in the coupled interference topology graph, the node feature vectors of all neighboring nodes and the edge feature vectors of the connecting edges of the target node are extracted, and element-wise multiplication is performed to obtain the neighbor message vectors transmitted by each neighboring node to the target node; the current node feature vector of the target node and the neighbor message vector are concatenated and nonlinearly mapped to obtain the updated node feature vector of the target node. After performing a preset number of rounds of information propagation and aggregation iteration, the updated node feature vectors of all nodes in the coupled interference topology graph are obtained and global pooling is performed to obtain the global graph representation vector of the laser array; the global graph representation vector is nonlinearly decoded to obtain the control strategy representation containing global coupling constraints.

6. The method according to claim 1, characterized in that, Based on the aforementioned control strategy characterization, the global control objective of the laser array is decomposed into the local control objectives and coupling consistency constraints of each laser element through a dual decomposition mechanism, including: The global control objective of the laser array is extracted and decomposed from the control strategy representation. The independent control performance index of a single laser unit is taken as the local optimization objective, and the coupling consistency condition that must be satisfied between adjacent laser units is taken as the coupling constraint term. By introducing dual variables and performing Lagrange relaxation on the coupling constraint terms, the global control objective is transformed into an augmented form to obtain the augmented objective. For each laser unit, a local control objective and a coupling consistency constraint are separated from the augmented objective. The local control objective consists of the local optimization objective and a local coupling penalty term weighted by the dual variable. The coupling consistency constraint consists of the state consistency condition between adjacent laser units and the update rule of the dual variable. The dual variable is iteratively updated, and the deviation between the current state and the state consistency condition between adjacent laser units is calculated as the coupling constraint violation amount. The update direction of the dual variable is determined by the coupling constraint violation amount between adjacent laser units. The iterative update of the dual variable drives the local control objective of each laser unit to gradually satisfy the coupling consistency constraint.

7. The method according to claim 6, characterized in that, Under the constraint of coupling consistency, iterative optimization of the local control objective generates modulation control commands for each laser unit in the laser array, including: The negative gradient direction of the local control target is projected onto the tangent space of the feasible region defined by the coupling consistency constraint to obtain the gradient projection direction. The control parameters of each laser unit are updated along the gradient projection direction to obtain the updated control parameters. Based on the updated control parameters, the coupling constraint violation amount between each laser unit and its adjacent laser units is recalculated. The coupling constraint violation amount is combined with the dual variable to obtain the correction magnitude. The correction direction is determined according to the state consistency condition. A correction vector is constructed based on the correction magnitude and the correction direction. The correction vector is superimposed on the updated control parameters to obtain the corrected control parameters. Repeat the iteration until the amount of the coupling constraint violation and the amount of the decrease in the local control objective simultaneously satisfy the convergence condition; The phase control parameters, power control parameters, and beam quality control parameters are separated from the modified control parameters. The phase control parameters are mapped to phase modulation commands, the power control parameters are mapped to power modulation commands, and the beam quality control parameters are mapped to wavefront control commands, thereby obtaining the modulation control commands for each laser unit in the laser array.

8. An artificial intelligence-based adaptive control system for laser arrays, used to implement the method as described in any one of claims 1-7, characterized in that, include: The first unit is used to acquire the real-time output light field distribution and target light field distribution of the laser array; The second unit is used to extract nonlinear features from the real-time output light field distribution to obtain an array state feature vector. Based on the array state feature vector, the mutual coupling relationship between each laser unit in the laser array is identified through a causal inference mechanism, and a coupling interference topology diagram is constructed. Based on the aforementioned coupling interference topology, the independent contribution components and coupling interference components of each laser unit are separated and decoupled to obtain a decoupled state vector. The third unit is used to construct a decoupled optical field based on the decoupled state vector and perform a multi-scale spatial domain and frequency domain joint difference measurement with the target optical field distribution to obtain the optical field deviation. The fourth unit is used to propagate and aggregate information on the nodes and edges of the coupled interference topology graph through a graph neural network according to the optical field deviation, and to fuse and encode the independent contribution components of each laser unit with the coupled interference components of the adjacent units to generate a control strategy representation containing global coupling constraints. The fifth unit is used to decompose the global control objective of the laser array into the local control objective and coupling consistency constraint of each laser unit based on the control strategy characterization through a dual decomposition mechanism, and to iteratively optimize the local control objective under the coupling consistency constraint to generate modulation control commands for each laser unit in the laser array.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images