A large reflector antenna pointing deviation prediction method

Abstract

The application discloses a large-scale reflector antenna pointing deviation prediction method, adopts a training-pushing decoupling architecture: only a single-branch lightweight CNN student network is reserved in the reasoning stage, zero additional calculation overhead is realized; a heterogeneous double-teacher distillation mechanism is introduced in the training stage, local space structure supervision is provided by using DDRNet, global timing correlation supervision is provided by using SegFormer, and progressive three-stage distillation strategies are used for sequentially performing local semantic learning, global correlation rule learning and collaborative fine-tuning. A multi-prototype dynamic center alignment module is designed for depicting multi-modal distribution, a structured geometric alignment module is used for modeling high-order geometric topology, and a dynamic weight balancing module is used for adaptively adjusting the contribution of the double teachers. The application can realize high-precision pointing deviation prediction while maintaining millisecond-level real-time reasoning speed, and is suitable for short-time pointing deviation prediction in the wind-induced disturbance scene of a large-scale reflector antenna.

Description

Technical Field

[0001] This invention relates to the field of pointing control and deep learning technology for large reflector antennas, specifically to a method for predicting wind-induced pointing deviation of large reflector antennas. Background Technology

[0002] Predicting short-time pointing deviations for large reflector antennas (such as radio telescopes, deep space exploration antennas, and satellite communication antennas) under wind disturbance scenarios is a key technology for achieving high-precision and stable tracking. With the evolution of antenna apertures to the hundreds of meters in diameter, gusts of wind have become a major bottleneck restricting the improvement of pointing accuracy. The core challenge of this type of prediction task lies in achieving high-precision short-time pointing deviation prediction based on multi-source heterogeneous sensor data (including antenna attitude variables, angle error variables, motion state variables, servo state variables, wind field variables, environmental variables, etc.), while simultaneously meeting the millisecond-level real-time inference response requirements of the antenna control system.

[0003] In recent years, data-driven deep learning methods have made significant progress in this type of prediction task. Among them, lightweight convolutional neural networks (CNNs) have become the preferred architecture for practical deployment due to their high computational efficiency and fast inference speed. However, due to the constraints of antenna front-end computing resources, lightweight models often struggle to simultaneously handle local transient feature extraction and global temporal correlation modeling, resulting in prediction accuracy that is difficult to achieve ideal levels.

[0004] Knowledge distillation (KD) technology provides an effective way to overcome the performance bottleneck of lightweight networks. Its core idea is to utilize a high-performance teacher network with a large number of parameters and strong representational capabilities to generate soft labels or feature representations during the training phase, guiding the optimization process of the student network. However, traditional distillation methods have the following shortcomings when facing tasks involving predicting wind-induced pointing deviations of large reflector antennas:

[0005] (1) Traditional distillation methods typically employ a single teacher network. However, the task of predicting wind-induced pointing deviation for large reflector antennas has a dual requirement: on the one hand, it needs to capture the local transient spatial response patterns under gust impact (such as the differences in wind pressure distribution in different areas of the antenna panel); on the other hand, it needs to model the global long-range temporal dependence brought about by the evolution of meteorological conditions (such as the continuous change trend of wind speed and direction within a few seconds). A single teacher network cannot simultaneously meet these two requirements, resulting in the prediction model failing to achieve a good balance between local details and global patterns.

[0006] (2) When using heterogeneous architectures (such as CNNs which are good at extracting spatial features and Transformers which are good at modeling long-range dependencies), the feature distributions of different architectures are significantly different, and direct feature alignment can easily lead to gradient conflicts. For antenna pointing deviation prediction, gradient conflicts can cause the model to have difficulty converging to a stable spatial feature extractor and temporal correlation modeler at the same time during training, which in turn increases the variance of the prediction results and causes extreme errors, seriously affecting the reliability of antenna pointing control.

[0007] (3) Existing distillation methods mostly use the class mean as the alignment target, which implicitly assumes that "features within the domain follow a unimodal distribution". However, in actual wind disturbance scenarios, even at the same wind speed level, due to the coupling effect of factors such as wind direction changes, antenna elevation angle differences, and local turbulence intensity fluctuations, the response mode of antenna pointing deviation often exhibits multimodal distribution characteristics. Existing distillation methods using single-center alignment will force all response modes to be compressed into a single representation, resulting in the loss of differentiated information between different response modes under wind disturbance excitation, weakening the model's adaptability to complex wind field conditions.

[0008] (4) Existing methods lack explicit modeling of the geometric topological relationships between elements in the feature space, making it difficult to characterize high-order structural information. In the prediction of wind-induced pointing deviation of large reflector antennas, there are physical correlations between different sensor signals, different time steps, and different spatial location features (such as the propagation path of wind pressure on the antenna aperture surface, the spatiotemporal evolution sequence of gust impact, etc.). Existing methods cannot effectively model and transfer this high-order structural knowledge, which limits the ability of lightweight models to express complex wind field-structure coupling relationships.

[0009] In summary, existing knowledge distillation methods, when faced with the specific task of predicting wind-induced pointing deviation of large reflector antennas, cannot simultaneously provide dual supervision of local transients and global temporal data, struggle to achieve stable feature alignment across heterogeneous architectures, and lack the ability to model multimodal response patterns and physical topology in wind disturbance data. Therefore, there is an urgent need for a lightweight distillation learning framework for predicting wind-induced pointing deviation of large reflector antennas, capable of effectively integrating and aligning global and local information from heterogeneous dual teachers during the training phase, while maintaining the millisecond-level real-time prediction capability of a single-branch CNN structure during the inference phase. Summary of the Invention

[0010] To address the technical challenges of existing methods for predicting wind-induced pointing deviation of large reflector antennas, such as the difficulty for lightweight convolutional neural networks to simultaneously handle local transient feature extraction and global temporal correlation modeling due to computational resource constraints, and the shortcomings of existing knowledge distillation methods for this specific task, including insufficient single-teacher supervision, gradient conflicts in heterogeneous architectures, inability of single-center alignment to characterize multimodal distributions, and lack of geometric topology modeling, this invention proposes a new method for predicting pointing deviation of large reflector antennas. This method adopts a decoupled training-inference architecture: during the inference phase, a single-branch lightweight CNN structure is maintained to achieve zero additional computational overhead; during the training phase, a heterogeneous dual-teacher distillation mechanism is introduced, and a progressive three-stage distillation strategy is adopted. Through modules such as multi-prototype dynamic center alignment, structured geometric alignment, and dynamic weight balancing, the full transfer and fusion of heterogeneous teacher knowledge is achieved.

[0011] The technical solution of this invention is as follows:

[0012] A method for predicting pointing deviation of a large reflector antenna includes the following steps:

[0013] Step S1: Construct a heterogeneous dual-teacher network, student network, and training dataset;

[0014] The heterogeneous dual-teacher network includes:

[0015] Local Semantic Teacher Network: Adopting the DDRNet architecture, it is used to extract the fine spatial structure and geometric topological features of antenna pointing deviation under wind disturbance, and provide supervision signals for local transient spatial response patterns;

[0016] Global Semantic Teacher Network: Adopting the SegFormer architecture, it utilizes the multi-head self-attention mechanism of its Transformer to model the long-range dependencies between multi-sensor time-series data, providing a supervisory signal for global time-series correlation patterns;

[0017] The student network is a single-branch lightweight CNN network (LAGNet). The student network embeds a global-local balance module (GLBBlock) at multiple layers. During the inference stage, only the student network is retained for forward prediction, and it outputs short-term pointing bias prediction values. The student network deploys an adaptive feature refinement module (AFRBlock) in the deep layers to compensate for the degradation of deep information.

[0018] Constructing a training dataset: Obtain stable tracking operation records of a large reflector antenna. These records include the following seven types of variables: antenna attitude variables (azimuth, elevation, command angle), angle error variables (actual azimuth angle difference, actual elevation angle difference), motion state variables (azimuth angular velocity, elevation angular velocity), servo state variables (motor current, motor speed), wind field variables (wind speed, wind direction), environmental variables (temperature, humidity, air pressure), and pointing deviation annotation values. , The multi-source heterogeneous sensor data is serialized and sampled using a fixed time window length T. Each window contains the first six variable values ​​of T consecutive time points as input samples, and the pointing deviation label at the last time point of the same window is used as a supervision label. The window is slidably sampled on the time axis with a sliding step size S to obtain multiple window samples. All window samples are divided into training set, validation set and test set according to a preset ratio. The training set is used for training the student network at each stage, the validation set is used for early stopping judgment and hyperparameter selection between stages, and the test set is used for final performance evaluation.

[0019] Step S2: First stage training: Local semantic learning

[0020] Enable the Local Semantic Teacher Network (DDRNet) and freeze the Global Semantic Teacher Network (SegFormer); set the backbone and decoder head of the student network to a trainable state.

[0021] The window samples in the training dataset are sequentially input into the student network and the local semantic teacher network. The specified layer features of the student network and the local semantic teacher network are extracted. The local distillation loss is calculated by the multi-prototype dynamic center alignment module (DCA) and the structured geometric alignment module (SGA).

[0022] The DCA module is used to characterize the multimodal response distribution under the same physical conditions in wind disturbance data, avoiding information loss caused by single-center alignment;

[0023] The SGA module is used to model the geometric topological relationships between elements in the feature space, guiding students to learn higher-order structural information such as wind pressure propagation paths and gust evolution sequences online.

[0024] The regression loss is combined to form the total loss for the first stage, and all parameters of the student network are updated to establish a robust spatial prior and geometric detail foundation.

[0025] Step S3: Second Stage Training: Global Association Pattern Learning

[0026] Using window samples from the training dataset, after completing local semantic learning, the global semantic teacher network (SegFormer) is introduced. The first part of the student network backbone near the input end is frozen to protect the established geometric feature manifold from being destroyed. Only the last part of the layer near the output end and the decoding head are unfrozen to adapt to the global semantic signal.

[0027] The backbone feature alignment loss (BFA) and shared decoder head alignment loss (SDHA) are calculated to form the global distillation loss, which makes up for the inherent shortcomings of CNN in modeling large-scale temporal data.

[0028] The regression loss is combined to form the total loss for the second stage. The unfrozen parameters in the student network are then updated, and long-range temporal dependencies are injected.

[0029] Step S4: Third Phase Training: Co-fusion Fine-tuning

[0030] Using window samples from the training dataset, the entire student network is unfrozen for full parameter fine-tuning, while the local semantic teacher network and the global semantic teacher network are enabled.

[0031] Introducing a Dynamic Weight Balancing Module (DWBM), which contains two learnable parameters. and Local teacher weights are obtained by mapping using the Softmax function. and global teacher weight :

[0032]

[0033] The total loss is the sum of the regression loss, the weighted local distillation loss, and the global distillation loss; all parameters of the student network are updated simultaneously through backpropagation. and When the gradient direction of a teacher deviates significantly from the optimization direction of the main task, DWBM automatically reduces the weight of that teacher, achieving synergistic complementarity between local and global knowledge.

[0034] The final student network is obtained after training convergence.

[0035] Step S5: Real-time Inference:

[0036] The real-time acquired multi-source heterogeneous sensor data is input into the trained student network, which is a single-branch CNN structure, and outputs short-time pointing deviation prediction results within millisecond delay.

[0037] In a further preferred embodiment, step S2, calculating the local distillation loss using a multi-prototype dynamic center alignment module (DCA), includes the following sub-steps:

[0038] In the specified layer, the feature maps of the student network and the local semantic teacher network are clustered element-wise according to category c, based on physical condition categories; the physical condition categories are discrete labels obtained by discretizing continuous physical quantities into intervals.

[0039] Perform K-Means clustering on the feature set of each category c to generate K representative prototype centers, each prototype representing a response sub-pattern under that category;

[0040] Using cosine similarity as a metric, an optimal matching mapping is established between student prototypes and teacher prototypes. :

[0041]

[0042] in It is a permutation mapping from the teacher prototype to the student prototype, that is, a permutation of length K; The set of all possible K permutations of the mapping. For the first Category in layer The A teacher prototype center For the first Category in layer The The formula represents finding an optimal permutation mapping among all possible prototype pairings, centered on a single student prototype. This maximizes the sum of cosine similarities between K pairs of matched prototypes, thereby establishing the optimal prototype correspondence between teachers and students.

[0043] Calculate the match between the student prototype and the teacher prototype Distance, accumulated to form a multi-prototype dynamic center alignment loss:

[0044]

[0045] in For multi-prototype dynamic center alignment loss value, For deploying the network layer set of DCA modules, This represents the total number of physical condition categories. The number of cluster prototypes under each category. For the first Category in layer The A student prototype center.

[0046] This module is only enabled during the training phase and does not introduce any additional computational overhead during the inference phase.

[0047] In a further preferred embodiment, step S2, calculating the local distillation loss using a Structured Geometry Alignment (SGA) module, includes the following sub-steps:

[0048] Extract the intermediate layer feature maps of the student network and the local semantic teacher network, expand each feature map into N element features, and then perform... Normalization;

[0049] The semantic association strength matrix R between any two elements in the N element features is defined using a non-negative Gaussian kernel function, where This represents the semantic association strength between the i-th element and the j-th element;

[0050]

[0051] in The i-th and j-th are respectively Normalized elemental characteristics The Gaussian kernel bandwidth parameter is defined; the non-negative Gaussian kernel function ensures that the relation matrix R is non-negative and symmetric, and that its degree matrix D is a non-degenerate matrix.

[0052] Construct a degree matrix D based on the relation matrix R, whose diagonal elements are: ;

[0053] Calculate the normalized graph Laplacian matrix ,in It is the identity matrix. The inverse square root matrix of the degree matrix D is used to normalize the degree of different nodes.

[0054] The alignment loss of the student-teacher relationship matrix and the alignment loss of the Laplacian matrix are calculated separately, and then weighted to obtain the structured geometric alignment loss:

[0055]

[0056] in These are the element relationship matrices for the student network and the teacher network, respectively. These are the normalized graph Laplacian matrices for the student network and the teacher network, respectively. Denotes the Frobenius norm. is a weighting coefficient used to balance the contributions of the relation matrix alignment term and the Laplacian matrix alignment term to the total loss.

[0057] This module operates only in the mid-level semantic feature space and introduces additional computational overhead only during the training phase.

[0058] In a further preferred embodiment, the adaptive feature refinement module (AFRBlock) is deployed deep within the student network to refine the deep feature maps of the student network; the adaptive feature refinement module performs the following operations on the deep feature maps:

[0059] Global context modeling: Channel-level description vectors are extracted through global average pooling, and channel attention weights are generated through two layers of 1×1 convolution and Sigmoid mapping;

[0060] Restore operation: The channel attention weights are broadcast in the spatial dimension through bilinear interpolation to reconstruct a low-resolution spatial mesh, generating a weak spatial prior structure, making the network more sensitive to wind pressure distribution boundaries and fine-grained textures;

[0061] Resize operation: restores the spatial grid to the resolution of the input feature map and performs element-wise modulation;

[0062] Local feature refinement: Output after 3×3 convolution, synchronous batch normalization, and ReLU activation.

[0063] This module uses only 1×1 and 3×3 convolution operations, eliminating the need to compute a self-attention matrix and resulting in extremely low computational overhead.

[0064] In a further preferred embodiment, the global-local balancing module (GLBBlock) adopts a Transformer-like metastructure, containing three parallel branches that process the input feature map X and output the feature map Y.

[0065] Local Perception Branch: Processes the input feature map X using convolutional attention and spatial attention to output local features. It is used to preserve the fine spatial topology of features and capture the local transient response patterns under gust impact;

[0066] Global modeling branch: Processes the input feature map X using channel compression and depthwise separable convolution to output global features. It is used to realize long-range dependency modeling, simulate the global receptive field of SegFormer with extremely low computing power cost, and capture the global trend of weather condition evolution.

[0067] Adaptive gated branch: The input feature map X is processed by global average pooling and 1×1 convolution, and the output spatial weights are obtained. This is used for weighted arbitration of global features;

[0068] The output feature map of the module ,in This indicates element-wise multiplication.

[0069] This module uses Group Normalization instead of Batch Normalization to align with the feature statistics distribution of the Transformer.

[0070] In a further preferred embodiment, the Dynamic Weight Balancing Module (DWBM) operates as follows:

[0071] initialization Initial weights During training, two learnable parameters and Automatic updates via gradient descent:

[0072] ,

[0073] When the loss of some teachers is significantly less than the loss of all teachers. Automatically increases, and vice versa; this mechanism enables DWBM to adaptively adjust its contribution weight according to the degree of deviation between the real-time gradient direction of each teacher and the optimization direction of the main task, effectively alleviating gradient conflicts among heterogeneous teachers and achieving synergistic complementarity between local transient knowledge and global temporal knowledge.

[0074] In a further preferred embodiment, the physical condition category includes at least one of the following: dividing the wind speed into P wind speed level intervals according to a preset threshold, dividing the wind direction into Q wind direction sector intervals according to angle, and dividing the antenna elevation angle into U elevation angle intervals according to angle, where P, Q, and U are preset positive integers.

[0075] In a further preferred embodiment, the regression loss adopts the Smooth L1 loss function, defined as:

[0076]

[0077] in For the regression loss value, The vector of prediction bias values ​​output by the student network. This is the vector of true pointing deviation values. The number of samples in a training batch. For the first The prediction bias value for each sample. For the first The true pointing deviation label value for each sample.

[0078] Beneficial effects:

[0079] Compared with the prior art, the present invention has the following beneficial effects:

[0080] (1) This invention addresses the dual requirements of "local transient response" and "global temporal dependence" in the wind-induced pointing deviation prediction task of large reflector antennas. It designs a heterogeneous dual-teacher architecture of DDRNet+SegFormer and constructs a lightweight student network LAGNet for antenna scenarios. The student network embeds GLBBlock at multiple levels to achieve a dynamic balance between local and global features. The local teacher (DDRNet) provides fine spatial structure and geometric topology supervision, while the global teacher (SegFormer) injects rich long-range temporal correlation information, enabling the student network to simultaneously capture local details and model global trends while maintaining a single-branch CNN structure during the inference stage.

[0081] (2) The progressive three-stage distillation strategy proposed in this invention establishes robust spatial priors in the first stage, adapts global semantics in the second stage under the condition of freezing the backbone near the input end, and adaptively adjusts the dual-teacher weights through DWBM in the third stage. Through the path of "local semantic learning → global association rule learning → collaborative fusion fine-tuning", it effectively solves the gradient conflict problem caused by significant differences in feature distribution between heterogeneous architectures of CNN and Transformer. Experimental results verify the significant superiority of the progressive strategy over direct joint training.

[0082] (3) The DCA multi-prototype dynamic center alignment module designed in this invention generates multiple prototype centers by performing K-Means clustering on each physical condition category, and establishes the correspondence between teacher and student prototypes with optimal matching mapping. It fully depicts the intra-class multimodal distribution generated by the coupling effect of factors such as wind direction change, pitch angle difference, and turbulence intensity fluctuation in wind disturbance data, and avoids the information loss caused by the forced compression of all response modes in traditional single-center alignment.

[0083] (4) The SGA structured geometric alignment module designed in this invention uses a non-negative Gaussian kernel function to define the semantic association matrix between elements, avoiding the mathematical defect of degree matrix degradation caused by traditional Softmax row normalization. Through the bi-branch alignment of the normalized graph Laplacian matrix, it explicitly models the second-order geometric topological relationship between elements in the feature space, including boundary directionality consistency, regional connectivity and global smoothness structure, effectively characterizing the physical association topological structure between different sensor signals, between different time steps and between different spatial location features, such as the propagation path of wind pressure on the antenna aperture surface, the spatiotemporal evolution sequence relationship of gust impact, etc.

[0084] (5) The DWBM dynamic weight balancing module designed in this invention uses two learnable parameters. and Automatic gradient updates adaptively adjust the contribution weights of the two teachers to the total loss. When the gradient direction of a certain teacher deviates significantly from the optimization direction of the main task, the weight of that teacher is automatically reduced, which effectively alleviates gradient conflicts when heterogeneous teachers jointly optimize and realizes the synergistic complementarity of local and global knowledge.

[0085] (6) All training-specific modules (DCA, SGA, DWBM) of this invention are enabled only during the training phase; during the inference phase, only the student network is retained, and all distillation enhancement modules are completely removed. AFRBlock and GLBBlock, as inherent structures of the student network, remain in the inference phase without introducing additional computational overhead. The student network maintains a pure single-branch CNN structure during deployment, which meets the millisecond-level real-time inference requirements of the antenna front-end control system.

[0086] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0087] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0088] Figure 1 The diagram shows the overall network structure of an embodiment of the present invention, illustrating the architectural differences between the training and inference phases. The training phase includes a Local Semantic Teacher Network (DDRNet), a Global Semantic Teacher Network (SegFormer), and a Student Network (LAGNet). The Student Network receives distillation supervision from the two teacher networks through a multi-prototype dynamic center alignment module and a structured geometric alignment module. The inference phase retains only the trained Student Network.

[0089] Figure 2 The flowchart of the progressive three-stage heterogeneous distillation strategy in this embodiment of the invention illustrates the training process of the three-stage progressive distillation, including the first stage of local semantic learning, the second stage of global association rule learning, and the third stage of collaborative fusion fine-tuning.

[0090] Figure 3 The diagram shows the design of the adaptive feature refinement module (AFRBlock) according to an embodiment of the present invention, illustrating the four processing stages of the module: global context modeling, Restore operation, Resize operation, and local feature refinement.

[0091] Figure 4The diagram shows the global-local balance module (GLBBlock) network of this invention, where (a) is the overall structure and (b) is the internal structure of LGLBlock, showing the Transformer-like metastructure and three parallel branches (local perception branch, global modeling branch, and adaptive gating branch).

[0092] Figure 5 This is a schematic diagram of the Multi-Prototype Dynamic Center Alignment (DCA) module in an embodiment of the present invention, illustrating the processing flow of input and multi-layer feature extraction, layer-by-layer element aggregation, K-Means clustering, optimal matching mapping, and alignment loss calculation. Detailed Implementation

[0093] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in further detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of this invention.

[0094] (I) Overall Network Architecture:

[0095] like Figure 1 As shown, the pointing deviation prediction method for large reflector antennas provided in this embodiment of the invention employs a heterogeneous dual-teacher distillation architecture during the training phase, including a local semantic teacher network (DDRNet), a global semantic teacher network (SegFormer), and a student network (LAGNet). During training, the student network receives distillation supervision from both teacher networks and performs feature alignment through modules such as the Global-Local Balancing Module (GLBBlock), the Multi-Prototype Dynamic Center Alignment Module (DCA), and the Structured Geometric Alignment Module (SGA). During the inference phase, only the trained student network (single-branch CNN) is retained, and real-time multi-source heterogeneous sensor data (including antenna attitude, angle error, motion state, servo state, wind field state, environmental state, etc.) is input, outputting short-term pointing deviation prediction results.

[0096] The backbone of the student network LAGNet adopts a lightweight CNN structure, deploys an adaptive feature refinement module (AFRBlock) in deep layers to compensate for the degradation of deep information, and embeds GLBBlock in multiple layers to achieve a dynamic balance between local and global features.

[0097] like Figure 1 As shown, the Local Semantic Teacher (DDRNet) provides fine-grained spatial structure and geometric supervision, while the Global Semantic Teacher (SegFormer) provides consistency between long-range temporal data and global patterns. Let the student network and teacher network be defined on the [missing information - likely a date or time interval]. The feature maps output by the layers are as follows:

[0098]

[0099] in Indicates the number of channels. Indicates spatial resolution.

[0100] In this architecture, two types of pre-trained models with distinct architectures were selected as teachers. The local semantic teacher uses DDRNet, leveraging its powerful spatial modeling capabilities to extract fine spatial structures and geometric topological features from the antenna pointing deviation feature map under wind disturbance. For example, it captures the local variation patterns of antenna azimuth and elevation errors within a short timeframe when gusts impact the antenna panel, as well as the local coupling relationship between sudden wind speed changes and angle errors. The global correlation teacher uses SegFormer, which, with its Transformer architecture's long-range dependency modeling capabilities, injects rich global temporal information and correlation consistency, such as the continuous evolution trend of meteorological conditions on timescales of several seconds to tens of seconds, and the cumulative impact of gust sequences on antenna pointing. Logically, DDRNet focuses on guiding students to capture the intra-domain feature distribution and geometric relationships under various physical conditions in wind disturbance data through DCA and SGA, while SegFormer compensates for the inherent shortcomings of CNNs in large-scale temporal data modeling through backbone feature alignment and shared decoder head alignment.

[0101] To bridge the feature gap between heterogeneous architectures and alleviate gradient conflicts, this invention further designs a Transformer-like convolutional semantic block GLBBlock and a TriKD-Seg progressive three-stage distillation strategy. Through the path of "local learning-global injection-cooperative fine-tuning", it ensures that the student network can still obtain information representation capabilities close to multi-branch or Transformer models while maintaining a single-branch CNN structure during the inference stage. Thus, under the limited computing resource constraints of the large reflector antenna front-end controller, it achieves high-precision and low-latency wind-induced pointing deviation prediction.

[0102] (II) Construction of Training Dataset

[0103] In this embodiment of the invention, the training dataset comes from real, stable tracking operation records of a large reflector antenna.

[0104] First, the stable tracking operation record of the large reflector antenna is obtained. This record includes the following seven types of variables: antenna attitude variables (azimuth, elevation, command angle), angle error variables (actual azimuth angle difference, actual elevation angle difference), motion state variables (azimuth angular velocity, elevation angular velocity), servo state variables (motor current, motor speed), wind field variables (wind speed, wind direction), environmental variables (temperature, humidity, air pressure), and pointing deviation label value (…). , Of the seven types of variables mentioned above, the first six are input features, and the last type (pointing to the bias label value) is the supervision label.

[0105] The data from the aforementioned multi-source heterogeneous sensors are sequentially sampled using a fixed time window length T. Each window contains the first six variable values ​​of T consecutive time points as input samples, and the pointing deviation label at the last time point of the same window is used as a supervision label. The window slides along the time axis with a sliding step size S to obtain multiple window samples. In this embodiment, the sampling interval is 50ms, resulting in 359,956 windows that meet the requirement of continuous input sequence.

[0106] All window samples were divided into training, validation, and test sets according to a preset ratio: 251,969 windows (approximately 70%) for the training set, 35,997 windows (approximately 10%) for the validation set, and 71,990 windows (approximately 20%) for the test set. The training set was used for training the student network at each stage, the validation set was used for early stopping judgment and hyperparameter selection between stages, and the test set was used for final performance evaluation.

[0107] The wind field variables (wind speed and wind direction) in the aforementioned multi-source heterogeneous sensor data reflect the impact of external wind disturbances on antenna pointing; the antenna attitude variables and angle error variables reflect the current spatial state and control deviation of the antenna; the servo state variables reflect the real-time response of the drive system; and the environmental variables provide background operating conditions. These seven types of variables together constitute a complete physical scenario for predicting wind-induced pointing deviation.

[0108] (III) Specific Implementation of the TriKD-Seg Progressive Three-Stage Distillation Strategy

[0109] like Figure 2 As shown, the distillation training process of the present invention is divided into three stages.

[0110] Phase 1: Local Semantic Learning

[0111] This phase aims to establish a robust spatial prior and geometric detail foundation for the student network. DDRNet is used as the local semantic teacher network to guide the student network model. Figure 2 As shown, at this stage, both the backbone and the head of the student network are in full-parameter training.

[0112] Window samples from the training dataset are sequentially input into the student network and the local semantic teacher network to extract specified layer features from the student network and the local semantic teacher network. This invention designs two alignment constraints: Multi-Prototype Dynamic Center Alignment (DCA) and Structured Geometric Alignment (SGA). DCA operates on hierarchical features to capture the compact multimodal distribution of feature responses within the same physical condition (e.g., the same wind speed range). SGA operates on mid-level features to maintain the topological integrity of scene elements such as wind pressure propagation paths and gust spatiotemporal evolution sequences.

[0113] Define the total loss in the first stage as:

[0114]

[0115] in The regression loss is based on the true pointing deviation value. The vector of prediction bias values ​​output by the student network. This is the vector of true pointing deviation values. For multi-prototype dynamic center alignment loss, For structured geometric alignment loss, , For the weighting coefficient, in this embodiment, we take... , This stage of training, until convergence, establishes a robust spatial prior and geometric detail foundation for the student network, enabling it to effectively capture the local transient spatial response patterns generated by gusts of wind impacting the antenna panel.

[0116] Phase Two: Learning Global Association Patterns

[0117] After the student network has a solid foundation in local structure, a global teacher network (SegFormer) is introduced using window samples from the training dataset to inject long-range temporal dependencies. Due to the significant feature heterogeneity gap between CNNs and Transformers (SegFormer), direct full-parameter fine-tuning would result in significant gradient differences that would disrupt the stable geometric feature manifold established in the first stage. Therefore, the second stage is defined as a temporal correlation remapping process. Freezing part of the backbone aims to preserve prior geometric structures. In this stage, the first three layers of the backbone network near the input are frozen, while only the last layer near the output (the last layer in this embodiment) and the decoder head are unfrozen for fine-tuning to adapt to the global semantic signal.

[0118] This stage achieves heterogeneous feature alignment through two methods: Backbone Feature Alignment (BFA) and Shared Decoder Head Alignment (SDHA). BFA constrains the consistency of the distribution of intermediate layer feature maps between the student and teacher networks, compensating for the inherent limitations of CNNs in modeling global temporal dependencies. SDHA constrains the consistency of the final output predictions between the student and teacher networks, ensuring that the student network's predictions closely approximate the teacher network's prediction distribution. These two alignment methods form a complete supervision chain from intermediate representations to the final output.

[0119]

[0120] The global distillation loss is defined as follows:

[0121]

[0122] The backbone feature alignment loss is calculated as follows:

[0123]

[0124] In the formula and The student network and the teacher network are respectively in the first... Feature maps output by the layer; This indicates the number of elements in the feature map.

[0125] The shared decoder alignment loss is calculated as follows:

[0126]

[0127] In the formula and These are the pointing deviation prediction vectors output by the student network and the teacher network, respectively. MSE represents the mean squared error, which is used to measure the difference between the two prediction vectors.

[0128] In this stage, through the joint constraint of the two alignment losses mentioned above, the student network gradually absorbs global temporal correlation knowledge while maintaining the prior knowledge of local structure learned in the first stage. This makes up for the inherent shortcomings of CNN in large-scale temporal data modeling, enabling the network to effectively learn the cumulative impact of the continuous evolution trend of meteorological conditions on antenna pointing deviation on a time scale of several seconds to tens of seconds.

[0129] Phase 3: Collaborative Integration and Fine-tuning

[0130] Using windowed samples from the training dataset, the entire student network is unfrozen for full parameter fine-tuning, while simultaneously enabling both the local and global semantic teacher networks. To address potential knowledge conflicts during joint optimization, a Dynamic Weight Balancing Module (DWBM) is introduced in this stage to adaptively adjust the gradient contributions from both teachers. This module uses two learnable parameters... and The weights are dynamically calculated, corresponding to the confidence levels of the local and global semantic teacher networks, respectively. The Softmax function is used to map these two parameters, and the formula for calculating the dynamic weights is defined as follows:

[0131]

[0132] In this embodiment, initialization Initial weights This means that the weights of the two teacher networks are equal at the initial moment. During backpropagation, these two parameters, as part of the network, are jointly optimized with the backbone parameters of the student network, and are automatically updated by minimizing the total loss. For example... The gradient is calculated using the chain rule:

[0133]

[0134] in:

[0135]

[0136] The total loss in the third phase is:

[0137]

[0138] in , Includes local structural constraints such as DCA and SGA. This represents the global semantic distillation loss. The total loss function combines the regression loss with the weighted dual-teacher distillation loss, updating all parameters of the student network simultaneously through backpropagation. and .

[0139] This update mechanism resulted in losses for local teachers. Significantly lower than the overall teacher loss hour, Automatic scaling enables adaptive teacher selection based on gradient direction. This mechanism allows student networks to adaptively adjust their focus on "structure" or "semantics" according to the difficulty and gradient direction of different training stages. When the gradient direction of a teacher network deviates significantly from the optimization direction of the main task, DWBM will automatically reduce the weight of that teacher network through gradient descent, thereby smoothing the optimization trajectory and achieving synergistic complementarity between local and global knowledge.

[0140] Training stops when the validation set loss no longer decreases for E consecutive epochs (E=10 in this example), resulting in the final student network. This student network retains only a single-branch CNN structure during the inference phase; all teacher networks and distillation modules are removed, introducing no additional computational overhead.

[0141] (iv) Adaptive Feature Refinement Module

[0142] In the task of predicting wind-induced pointing deviation of large reflector antennas, the local abrupt changes caused by gusts (such as instantaneous jumps in antenna azimuth error and pitch response caused by sudden increases in wind speed) are crucial for accurately predicting the evolution trend of pointing deviation. However, as the depth of lightweight student networks increases, continuous downsampling operations lead to severe attenuation of contextual correlations and spatial details in deep features, causing the aforementioned local abrupt changes to be submerged in deep features, thus affecting prediction accuracy. To address this issue, this invention designs an Adaptive Feature Refinement Block (AFRBlock). Current lightweight methods mostly extract contextual information based on Global Average Pooling (GAP), but GAP leads to the complete loss of spatial details. The core design of AFRBlock lies in the Restore operation: after reconstructing a low-resolution spatialized grid through interpolation, a weak spatial prior structure is generated. This weak spatial prior structure allows the network to apply differentiated adjustment weights according to different domains, thereby effectively restoring the ability to perceive abrupt features and fine-grained change patterns in antenna wind-induced deviation signals while maintaining the global receptive field.

[0143] AFRBlock consists of two key phases:

[0144] (1) Global Context Modeling

[0145] Given As the input feature map of a certain deep layer of the student network, in this embodiment, AFRBlock is deployed after the last three convolutional blocks in the deep layer of the student network backbone. First, channel-level semantic description vectors are extracted through global average pooling:

[0146]

[0147] in Each dimension represents the average activation intensity of the corresponding channel in the spatial dimension. A lightweight nonlinear mapping is then constructed using two 1×1 convolutional layers:

[0148]

[0149] in The learnable weight matrix for a 1×1 convolutional layer. This indicates a batch normalization operation. Represents the ReLU activation function. Represents the Sigmoid mapping. This is the channel attention weight vector.

[0150] To enhance the perception of spatial information and avoid the problem of traditional channel attention completely losing spatial information, this invention employs a core Restore operation: rebroadcasting channel descriptors in the spatial dimension using an interpolation algorithm and generating a low-resolution spatial grid. ,in In this embodiment, S=7 is chosen as the preset low-resolution grid size. That is:

[0151]

[0152] This allows it to contain both global statistical information and a weakly spatialized structure. Among them... This represents the spatialized channel attention weight tensor generated after the Restore operation. The Restore operation reconstructs the data through channel broadcasting and bilinear interpolation, resulting in spatial attention weights that contain both global statistics and weak spatial structure characteristics. To apply the generated spatial attention back to the original features, a Resize operation is performed:

[0153]

[0154] Then, spatial attention is used to modulate the original features element-wise, resulting in the modulated feature map. :

[0155]

[0156] in This indicates element-wise multiplication.

[0157] (2) Local Feature Refinement

[0158] To further supplement local structural information, a 3×3 convolution is used to improve spatial continuity, resulting in the final response:

[0159]

[0160] in It is a 3×3 convolution. To normalize the batch size.

[0161] The AFRBlock module simultaneously models global context and encodes global dependencies and local details through spatial attention. The reconstructed... The weak spatial prior structure of the mesh makes AFRBlock more sensitive to boundary changes, abrupt changes, and fine-grained variation patterns in antenna wind-induced bias signals. This module is characterized by its lightweight nature yet strong semantic expressiveness, using only 1×1 and 3×3 convolution operations, eliminating the need to compute a self-attention matrix, and exhibiting extremely low computational overhead. AFRBlock effectively compensates for performance bottlenecks and significantly alleviates deep information degradation, enabling the lightweight student network to maintain sensitivity to wind-induced pointing bias features even at deeper layers. As an inherent structure of the student network, AFRBlock is retained and participates in forward propagation during the inference phase, and its lightweight design ensures that no additional computational overhead is introduced.

[0162] (v) Global-Local Equilibrium Module

[0163] In the task of predicting wind-induced pointing deviation for large reflector antennas, the local transient response caused by gust impacts and the global trend information generated by the evolution of large-scale meteorological conditions coexist and are coupled with each other. In order to effectively integrate heterogeneous knowledge from the global semantic teacher network and the local semantic teacher network in a lightweight single-branch architecture, and to achieve a dynamic balance between local transient features and global temporal dependencies, this invention proposes a Global-Local Balanced Block (GLBBlock) as a core bridge for cross-architecture correlation alignment.

[0164] At the macro level, GLBBlock adopts a Transformer-like meta-structure. It abandons traditional BatchNormalization and instead employs Group Normalization (GN) and a Feed-Forward Network (FFN) to mimic the feature statistical distribution of the Transformer to align with the feature statistical characteristics of the SegFormer, thus harmonizing these heterogeneous feature statistics. Let... Given the input feature map, its overall mapping process can be represented as:

[0165]

[0166]

[0167] in This is an intermediate feature map. This is the final output feature map of the GLBBlock module. This represents the core, lightweight local-global bottleneck structure. This is for Group Normalization.

[0168] At the micro-modeling level, to address the contradiction that a single-branch network cannot handle high-frequency details and long-range correlations, LGLBlock internally constructs three parallel lightweight branches:

[0169] The Local Branch acts as a structured carrier of local features. This branch includes Convolutional Attention (Conv Attention) and Spatial Attention (Spa Attention) modules to extract and preserve the fine spatial topology of features, such as the local jump patterns of azimuth and elevation errors within a short time window during a gust of wind impacting the antenna panel, and the spatial correlation between sudden wind speed changes and the angle error response. Its output is represented as:

[0170]

[0171] in The local features output by the local perception branch.

[0172] The Global Branch acts as a low-rank approximation of the self-attention mechanism. Through channel compression and depthwise separable convolution, it achieves efficient long-range dependency modeling without constructing an explicit attention matrix. This simulates the global receptive field of the SegFormer with extremely low computational cost, capturing the long-term impact of large-scale meteorological conditions (such as continuous wind direction changes and the overall trend of gust sequences) on antenna pointing deviation.

[0173]

[0174] in This indicates a channel compression operation. Indicates global average pooling. This represents depthwise separable convolution. The global features output by the global modeling branch.

[0175] The adaptive gate branch is used to dynamically balance the contributions of local and global features. This branch generates spatial weights through global average pooling and 1×1 convolutions, which are used to weight and arbitrate global features. Specifically, this gate mechanism dynamically determines whether to rely more on local transient details or global trend information based on the current input feature distribution: when gust patterns are obvious in the features, the contribution of the local perception branch is increased; when the features reflect slow weather evolution, the contribution of the global modeling branch is increased.

[0176]

[0177]

[0178] in Indicates global average pooling. Represents a 1×1 convolution. For normalized exponential functions, The spatial weights of the adaptive gated branch output. This represents element-wise multiplication. This is the final output feature map of the GLBBlock module.

[0179] GLBBlock achieves a dynamic balance between local transient response information and global temporal trend information without introducing an explicit self-attention matrix and high computational complexity. This provides lightweight networks with stable and efficient context modeling capabilities, enabling them to handle both sudden gusts of wind and continuous weather changes. As an inherent structure of the student network, GLBBlock is retained and participates in forward propagation during the inference phase, and its lightweight design ensures that no additional computational overhead is introduced.

[0180] (vi) Multi-prototype dynamic center alignment module

[0181] As mentioned earlier, in the first stage of training, this invention aligns the features of the local semantic teacher network and the student network using a Multi-prototype Dynamic Center Alignment (DCA) module. This section elaborates on the specific implementation of this module.

[0182] Existing feature distillation methods generally use the class mean as the alignment target, which implicitly assumes that "features within the domain follow a unimodal distribution." However, this invention, through analysis of large-scale operational data of wind disturbances on large reflector antennas, reveals that even at the same wind speed, the response mode of antenna pointing deviation still exhibits significant multimodal distribution characteristics due to the coupling effects of factors such as wind direction changes, antenna elevation angle differences, and local turbulence intensity fluctuations. For example, at wind speeds of 5-10 m / s, the wind pressure distribution pattern of the antenna panel differs between downwind and crosswind conditions, resulting in multiple distinct clusters in the feature response. Based on this observation, this invention proposes a DCA module to construct a unified multi-prototype semantic structure across models within the domain, thereby more precisely characterizing the multimodal distribution within various physical conditions under wind disturbance and avoiding the loss of differentiated wind disturbance response information caused by single-class center alignment.

[0183] like Figure 5 As shown, this module contains the following processes:

[0184] (1) Input and multi-layer feature extraction: The teacher network (fixed parameters) and the student network (trainable) extract multi-layer features respectively;

[0185] (2) Layer-by-layer clustering of working condition category elements: Based on the physical working condition category, the elements are clustered into layers. Layer features by category Perform element-level aggregation to obtain the teacher element set and the student element set respectively;

[0186] (3) K-Means clustering: for each category K-Means clustering is performed independently in the feature sets of the teacher network and the feature sets of the student network to generate K teacher prototypes and K student prototypes.

[0187] (4) Alignment loss calculation: Establish the best correspondence between teacher and student prototypes through optimal matching mapping, and accumulate all deployment layers, all physical condition categories, and all prototype centers. Distance, forming a multi-prototype dynamic center alignment loss.

[0188] In the Layers, for categories Teacher characteristics are defined as ,in, For the first The characteristics of each element For the physical condition category label of the element, Indicates the first The layer belongs to the category The set of all teacher characteristic elements. The physical condition category is a discrete label obtained by discretizing a continuous physical quantity into intervals. For example, wind speed is divided into P wind speed level intervals according to a preset threshold, wind direction is divided into Q wind direction sector intervals according to angle, and antenna elevation angle is divided into U elevation angle intervals according to angle.

[0189] To eliminate the magnitude bias between heterogeneous models, all features are subjected to [further processing / processing] before clustering. Normalization. To characterize the multimodal distribution within the domain, normalization is performed by normalizing the distribution in the feature space. conduct Clustering, to obtain A representative teacher prototype The calculation formula is as follows:

[0190]

[0191] in Assign variables to clustering, when elements Assigned to the If the center is a single point, take 1; otherwise, take 0. For the first The teacher element features. Similarly, the student prototype can be obtained. This embodiment sets... That is, three representative response sub-patterns are learned under each physical condition category, and the K-means algorithm is used to cluster the category features in each mini-batch.

[0192] To establish a correspondence between teacher and student prototypes, and considering the randomness of cluster center indices, this invention uses cosine similarity as a metric to find the optimal matching mapping between teacher and student prototypes. Its formula is expressed as follows:

[0193]

[0194] in This is a permutation mapping from the teacher prototype to the student prototype. The set of all possible K permutations of the mapping. The optimal permutation mapping is to maximize the sum of the cosine similarities between K pairs of matching prototypes.

[0195] DCA, by constructing a fine-grained semantic structure covering different wind disturbance response modes, better reflects the internal organization of local teachers. The final multi-layer alignment loss... Defined as:

[0196]

[0197] in, For deploying the network layer set of DCA modules, This represents the total number of physical condition categories. For optimal permutation mapping Next The index of student prototypes corresponding to each teacher prototype.

[0198] By aligning multimodal structures across layers under various physical conditions, the DCA module effectively avoids the information loss problem caused by lightweight models forcibly compressing different wind disturbance response modes into a single representation. The student network can learn the basic statistical features of wind field data at a shallow layer, capture the differentiated response structures under different conditions at a mid-layer, and maintain semantic consistency across conditions at a deep layer. As a training-dedicated distillation module, DCA is only enabled during the training phase and completely removed during the inference phase, without introducing any additional computational overhead.

[0199] (vii) Structured Geometric Alignment Module

[0200] As mentioned earlier, in the first stage of training, this invention also performs geometric topological alignment of the mid-layer features of the local semantic teacher network and student network through the Structured Geometric Alignment (SGA) module. This section elaborates on the specific implementation of this module.

[0201] While DCA can characterize the multimodal semantic distribution within a domain under various physical conditions, it essentially only constrains first-order statistical properties and cannot explicitly describe the higher-order geometric topological structures inherent in the relationships between elements. Examples include differences in wind pressure propagation paths across different regions of the antenna aperture, the temporal correlation patterns between azimuth and elevation errors during the same gust event, weak correlations of wind disturbances between distant time steps, and the connectivity of wind-induced responses within the same antenna pointing region. This higher-order structural information is crucial for depicting a complete physical picture of wind-induced pointing deviations. To address this issue, this invention proposes the SGA module, designed to model and align the element relationship structure of the teacher network in the feature space. Unlike directly aligning element-level features, SGA guides the student network to capture the implicit gust propagation paths and wind-structure coupling topological organization within the teacher network by learning the relationship graph between elements.

[0202] Due to the high computational complexity of SGA, it is only introduced into the mid-level association space of the network. Mid-level features achieve a good balance between preserving spatial structural details and extracting high-level information, thus being more conducive to the modeling and alignment of structured geometric information.

[0203] Extract the intermediate layer feature maps of the student network and the local semantic teacher network respectively, and expand each feature map into N element features. and perform analysis on the characteristics of each element. Normalization;

[0204]

[0205] Unlike existing methods, this invention abandons the row-normalized Softmax definition of the relation matrix because Softmax forces each row to sum to 1, causing the degree matrix to degenerate into an identity matrix and the Laplace matrix to lose second-order geometric information. This invention uses a non-negative Gaussian kernel function to define the semantic association strength between elements:

[0206]

[0207] in The i-th and j-th are respectively Normalized elemental characteristics The Gaussian kernel bandwidth parameter is set to 0.5 in this embodiment; this definition has the following properties: (1) Non-negativity: This ensures that the diagonal elements of the degree matrix D are strictly positive; (2) Symmetry: This ensures the real symmetry of the Laplace matrix; (3) Self-similarity: , indicating that the correlation strength between an element and itself is at its maximum; (4) more importantly, this definition does not impose row normalization constraints, therefore the diagonal elements of the degree matrix D The normalized graph Laplace matrix is ​​a variable rather than a constant. It can preserve complete second-order geometric topological information, namely the smoothness and connectivity structure between elements in the feature space, specifically including structural features such as boundary directionality consistency, preservation of long-distance weak associations, and connectivity within regions.

[0208] Further construct the degree matrix Let be a diagonal matrix, and its first... The diagonal elements are ,in Represents element nodes The total association strength between the graph and all other element nodes. Based on the relation matrix and degree matrix, the normalized graph Laplacian matrix is ​​defined as follows: ,in It is the identity matrix. The inverse square root matrix of the degree matrix D is used to normalize the degree of different nodes.

[0209] SGA ultimately calculates the alignment loss of the student-teacher relationship matrix and the alignment loss of the Laplacian matrix, respectively, and then weights them to obtain the structured geometric alignment loss, as follows:

[0210]

[0211] in These are the element relationship matrices for the student network and the teacher network, respectively. These are the normalized graph Laplacian matrices for the student network and the teacher network, respectively. Denotes the Frobenius norm. These are weighting coefficients used to balance the contributions of the relation matrix alignment term and the Laplacian matrix alignment term to the total loss. In this embodiment, we take... , .

[0212] Through the dual alignment of the Gaussian kernel relation matrix and the normalized graph Laplacian matrix, the SGA module can effectively guide the student network to learn the higher-order geometric topological relationships between elements in the teacher network feature space. Specifically, in the task of predicting wind-induced pointing deviation for large reflector antennas, this includes:

[0213] (a) Consistency of the directionality of wind pressure propagation path on the antenna aperture surface: This is achieved by aligning the eigenvectors of the Laplace matrix to ensure that the wind pressure propagation mode is consistent in different spatial regions;

[0214] (b) The weak correlation between azimuth error and pitch error in the same gust of wind event is maintained: this is achieved through global structural alignment of the relation matrix, ensuring that the temporal coupling relationship between multiple variables is preserved;

[0215] (c) Continuity of gust evolution sequence: achieved by smoothness constraint of graph Laplace, ensuring that the temporal evolution trajectory of pointing deviation is continuous and smooth.

[0216] These high-order structural information are crucial for multivariate temporal context modeling in wind-induced pointing deviation prediction. SGA, as a training-specific distillation module, operates only in the mid-level semantic feature space, is enabled only during the training phase, and is completely removed during the inference phase, introducing no additional computational overhead.

[0217] (viii) Dynamic weight balancing module

[0218] As mentioned earlier, in the collaborative fusion fine-tuning stage, this invention introduces a Dynamic Weight Balancing Module (DWBM) to adaptively adjust the contributions of the local semantic teacher network and the global semantic teacher network to the total loss. This section provides further explanation of the design motivation and working mechanism of this module.

[0219] DWBM uses two learnable parameters and The weights are dynamically calculated, corresponding to the confidence levels of the local and global semantic teacher networks, respectively. During training, and Automatic updates are achieved through gradient descent. The physical meaning is that when a teacher's gradient direction aligns with the main task's optimization direction, its corresponding learnable parameters are updated in the direction of increase, and the teacher's weight increases; conversely, when the gradient direction deviates, the corresponding learnable parameters decrease, and the teacher's weight decreases. Therefore, the magnitudes of the two learnable parameters reflect the confidence level of the teacher's current contribution to the main task's optimization direction.

[0220] During training, DWBM automatically determines which type of knowledge should be emphasized based on the statistical characteristics of the current batch in the training data. In actual operation, large reflector antennas experience significant differences in wind disturbance characteristics at different times, sometimes dominated by sudden gusts. In such cases, the transient changes in azimuth and elevation errors caused by gusts are more pronounced, while local distillation losses are relatively small. Automatic scaling enhances the learning of spatial details; sometimes the focus is on the global trend of large-scale meteorological conditions, in which case the long-term evolution trend of the pointing bias is more critical, and the global distillation loss is smaller. Automatically increase to enhance the learning of trend patterns.

[0221] Through this adaptive balancing mechanism, DWBM effectively alleviates gradient conflicts during heterogeneous teacher joint optimization, enabling a single student network to simultaneously handle different types of wind disturbance conditions during the inference phase, achieving synergistic complementarity between local transient knowledge and global temporal knowledge. As a training-dedicated distillation module, DWBM is only enabled during the training phase and completely removed during the inference phase, introducing no additional computational overhead.

[0222] (ix) Regression Loss Function

[0223] In this invention, regression loss The Smooth L1 loss function is used and is defined as follows:

[0224]

[0225] in For the regression loss value, The vector of prediction bias values ​​output by the student network. This is the vector of true pointing deviation values. The number of samples in a training batch. For the first The prediction bias value for each sample. For the first The true pointing deviation label value for each sample.

[0226] The reason for choosing the Smooth L1 loss function is that it is robust to outliers and is suitable for scenarios where there are occasional large fluctuations in wind-induced pointing deviation prediction (such as instantaneous deviation spikes caused by sudden gusts). When the prediction error is small, the smooth L2 loss is used to ensure convergence stability, and when the prediction error is large, the L1 loss is used to reduce the excessive influence of outliers on the gradient.

[0227] (x) Experimental verification

[0228] This embodiment has been extensively tested on a large-scale dataset of wind disturbances.

[0229] 1. Dataset Setup

[0230] The data source is the operational records of stable tracking of a large radio antenna. The raw data has been organized into a standardized stable tracking database. The main sensors and variable types used in this stage are as described in Part (II), “Construction of Training Dataset”, and will not be repeated here.

[0231] The standard library was cleaned up at 50 ms intervals. There were 359,956 windows that met the requirements for continuous input sequences in the experiments. All available windows were stratified and randomly divided according to the X / Y direction trend: 251,969 windows (approximately 70%) for the training set, 35,997 windows (approximately 10%) for the validation set, and 71,990 windows (approximately 20%) for the test set.

[0232] 2. Evaluation Indicators

[0233] This phase employs both accuracy and inference speed metrics for comprehensive evaluation. The accuracy metric uses the root mean square error (RMSE), calculated as the square root of the mean square of the differences between predicted and actual values. This metric is more sensitive to large errors and effectively reflects the stability of the prediction results. The inference speed metric includes inference latency, which is the average inference time per window, measured in milliseconds per window (ms / window), and is used to measure the real-time response capability of a single prediction.

[0234] 3. Training parameter settings

[0235] In this embodiment, the training parameters are set as follows: the optimizer uses AdamW, and the initial learning rate is set to... It employs a cosine annealing learning rate scheduling strategy, with a minimum learning rate of [missing information]. Batch size set to 64, training epochs set to 200; weight decay coefficient is... The maximum norm of gradient clipping is set to 1.0. The hyperparameters for each module are as follows: Gaussian kernel bandwidth is 0.5 (SGA module), number of clusters K is 3 (DCA module), low-resolution grid size S is 7 (AFRBlock module); the loss weight coefficients are respectively... , , , , To ensure the reproducibility of the experiment, the random seed was set to 42.

[0236] 4. Hardware Environment

[0237] The experimental hardware environment is as follows: the GPU is an NVIDIA GeForce RTX 5060 Ti with 16311 MiB of video memory; the CPU is an AMD Ryzen 7 H 255 with Radeon 780M Graphics, featuring 8 cores and 16 threads; the system memory is 30.8 GB; and the operating system is Windows 11 10.0.26200 SP0. For the software environment, Python version 3.12.7 is used, the deep learning framework is PyTorch 2.8.0+cu129, CUDA version is 12.9, and dependencies include scikit-learn 1.8.0, NumPy 1.26.4, and Pandas 2.3.3.

[0238] 5. Quantitative Experimental Results

[0239] Table 1: Quantitative Experiment Results

[0240]

[0241] This method combines excellent accuracy and real-time performance, demonstrating satisfactory performance in both RMSE and inference delay, representing an optimal balance between prediction accuracy and efficiency. The inference delay of 0.2101 ms / window means that this method can achieve real-time inference on the front-end controller of large reflector antennas, meeting the stringent millisecond-level prediction delay requirements of antenna control systems.

[0242] 6. Ablation test

[0243] To comprehensively verify the effectiveness of each module, this experiment included multiple ablation control groups. The "end-to-end joint distillation" control group employed dual instructors simultaneously without using the three-stage progressive strategy. Local and global distillation losses were calculated and jointly optimized throughout the entire training process to test the necessity of the TriKD-Seg progressive strategy. The experimental results are as follows:

[0244] Table 2: Ablation Experiment Results

[0245]

[0246] Note: The baseline refers to a student network trained using only ground truth regression loss, without any teacher supervision. DCA, SGA, TriKD-Seg, and end-to-end joint distillation are all training-only strategies that do not introduce any computational overhead during the inference phase, thus maintaining the same latency.

[0247] Experimental results show that: (1) the introduction of DCA and SGA resulted in a decrease in RMSE of 0.0010 and 0.0022 respectively, proving the effectiveness of intra-class multimodal distribution alignment and geometric topology alignment in the antenna wind-induced pointing deviation task; (2) Alignment Both (DCA+SGA) reduced the RMSE from 0.0533 to 0.0488, a reduction of 0.0045, verifying the synergistic effect of local structural constraints and geometric topology constraints; (3) the RMSE of end-to-end joint distillation (dual teachers activated simultaneously but without a three-stage strategy) was 0.0469, significantly better than Alignment. Both, indicating that the introduction of global teachers can supplement global temporal correlation information for student networks, which is necessary for capturing the long-term impact of large-scale weather evolution on antenna pointing; (4) TriKD-Seg (Full) further reduced RMSE from 0.0469 to 0.0431 compared to end-to-end joint distillation, with a reduction of 0.0038, proving that the three-stage progressive distillation strategy can effectively alleviate the gradient conflict between heterogeneous teachers of CNN and Transformer, and has obvious performance advantages compared to direct joint training, enabling student networks to maintain high prediction accuracy under different wind disturbance conditions of large reflector antennas.

[0248] The above ablation experiments fully validated the necessity of each design module in this method and the superiority of the TriKD-Seg progressive strategy over the traditional end-to-end combined distillation.

[0249] Experimental results demonstrate that this method achieves a good balance between pointing prediction accuracy and inference efficiency, enabling the output of short-term pointing deviation trends with millisecond-level inference overhead. Within typical physical operating conditions, the model exhibits good recognition ability for the evolution direction of wind-induced pointing deviation, providing accurate short-term predictions for both rapid pointing drift under sudden gusts and slow deviation accumulation under continuous weather changes. This provides a reliable methodological foundation for subsequent online confidence determination, conservative compensation output, and engineering verification of large reflector antennas.

[0250] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention without departing from the principles and spirit of the present invention.

Claims

1. A method for predicting pointing deviation of a large reflector antenna, characterized in that, Includes the following steps: Step S1: Construct a heterogeneous dual-teacher network, student network, and training dataset; The heterogeneous dual-teacher network includes a local semantic teacher network and a global semantic teacher network. The local semantic teacher network is used to extract the fine spatial structure and geometric topological features of the antenna pointing deviation under wind disturbance, and the global semantic teacher network is used to model the long-range dependencies between multi-sensor time series data. The student network is a single-branch lightweight CNN network. The student network embeds global-local balance modules at multiple levels. During the inference stage, only the student network is retained for forward prediction, and outputs short-term pointing bias prediction values. The global-local balancing module adopts a Transformer-like meta-structure, containing three parallel branches that process the input feature map X and output the feature map Y. Local Perception Branch: Processes the input feature map X using convolutional attention and spatial attention to output local features. A fine spatial topology used to preserve features; Global modeling branch: Processes the input feature map X using channel compression and depthwise separable convolution to output global features. It is used to implement long-range dependency modeling; Adaptive gated branch: The input feature map X is processed by global average pooling and 1×1 convolution, and the output spatial weights are obtained. This is used for weighted arbitration of global features; The output feature map of the global-local balancing module ,in This represents element-wise multiplication; The training dataset is constructed by sampling multi-source heterogeneous sensor data sequences with a fixed time window length, and the pointing deviation label value is used as the supervision label. Step S2: First stage training: Local semantic learning Enable the local semantic teacher network and freeze the global semantic teacher network; set the backbone and decoder head of the student network to a trainable state. The window samples in the training dataset are sequentially input into the student network and the local semantic teacher network. The specified layer features of the student network and the local semantic teacher network are extracted. The local distillation loss is calculated by the multi-prototype dynamic center alignment module and the structured geometric alignment module, respectively. The local distillation loss is calculated using a multi-prototype dynamic center alignment module, including the following sub-steps: In the specified layer, the feature maps of the student network and the local semantic teacher network are clustered element-wise according to category c, based on physical condition categories; the physical condition categories are discrete labels obtained by discretizing continuous physical quantities into intervals. Perform K-Means clustering on the feature set of each category c to generate K representative prototype centers; Using cosine similarity as a metric, an optimal matching mapping is established between student prototypes and teacher prototypes. ; Calculate the match between the student prototype and the teacher prototype Distance, accumulated to form a multi-prototype dynamic center alignment loss; The calculation of local distillation loss using the structured geometry alignment module includes the following sub-steps: Extract the intermediate layer feature maps of the student network and the local semantic teacher network, expand each feature map into N element features, and then perform... Normalization; The relationship matrix R between any two elements in the N feature elements is defined using a non-negative Gaussian kernel function, where This represents the semantic association strength between the i-th element and the j-th element; Construct a degree matrix D based on the relation matrix R, whose diagonal elements are: ; Calculate the normalized graph Laplacian matrix ,in It is the identity matrix; Calculate the alignment loss of the student-teacher relationship matrix and the alignment loss of the Laplacian matrix, respectively. The regression loss is combined to form the total loss for the first stage, and all parameters of the student network are updated accordingly. Step S3: Second Stage Training: Global Association Pattern Learning Using window samples from the training dataset, the global semantic teacher network is introduced, the first part of the student network backbone near the input end is frozen to protect the established geometric feature manifold, and only the last part of the layer near the output end and the decoding head are unfrozen to adapt to the global semantic signal. The backbone feature alignment loss and the shared decoder head alignment loss are calculated to form the global distillation loss; The regression loss is combined to form the total loss for the second stage, and the unfrozen parameters in the student network are updated accordingly. Step S4: Third Phase Training: Co-fusion Fine-tuning Using window samples from the training dataset, the entire student network is unfrozen, while the local semantic teacher network and the global semantic teacher network are enabled. A dynamic weight balancing module is introduced, which obtains local teacher weights and global teacher weights through two learnable parameters via Softmax mapping. The total loss is the sum of regression loss and the weighted local distillation loss and global distillation loss. The student network is obtained by simultaneously updating all parameters of the student network and the two learnable parameters through backpropagation, and the training converges. Step S5: Real-time Inference: The real-time acquired multi-source heterogeneous sensor data is input into the trained student network, and the short-time pointing deviation prediction result is output.

2. The method according to claim 1, characterized in that, The student network deploys an adaptive feature refinement module at a deep layer to refine the deep feature map of the student network; the adaptive feature refinement module performs the following operations on the deep feature map: Global context modeling: Channel-level description vectors are extracted through global average pooling, and channel attention weights are generated through two layers of 1×1 convolution and Sigmoid mapping; Restore operation: Reconstructs a low-resolution spatial grid by broadcasting the channel attention weights in the spatial dimension through bilinear interpolation; Resize operation: restores the spatial grid to the resolution of the input feature map and performs element-wise modulation; Local feature refinement: Output after 3×3 convolution, synchronous batch normalization, and ReLU activation.

3. The method according to claim 1, characterized in that, The physical condition categories include at least one of the following: dividing wind speed into P wind speed level intervals according to a preset threshold, dividing wind direction into Q wind direction sector intervals according to angle, and dividing antenna elevation angle into U elevation angle intervals according to angle, where P, Q, and U are preset positive integers.

4. The method according to claim 1, characterized in that, The nonnegative Gaussian kernel function is defined as follows: in The i-th and j-th are respectively Normalized elemental characteristics This is the Gaussian kernel bandwidth parameter.

5. The method according to claim 1, characterized in that, In the dynamic weight balancing module, two learnable parameters are: and Local teacher weight and global teacher weight The calculation formula is: initialization Initial weights ; During training, the two learnable parameters are automatically updated via gradient descent.

6. The method according to claim 1, characterized in that, The training dataset described in step S1 is constructed as follows: The stable tracking operation record of a large reflector antenna is obtained. The operation record includes the following seven types of variables: antenna attitude variables, angle error variables, motion state variables, servo state variables, wind field variables, environmental variables, and pointing deviation label values. The data from multi-source heterogeneous sensors are serialized and sampled with a fixed time window length T. Each window contains the first six types of variable values ​​at T consecutive time points as input samples, and the pointing deviation label value at the last time point of the same window is used as a supervision label. The window uses a sliding step size S to slide and sample on the time axis to obtain multiple window samples.

7. The method according to claim 1, characterized in that, The regression loss uses the Smooth L1 loss function, defined as: in For the regression loss value, The vector of prediction bias values ​​output by the student network. This is the vector of true pointing deviation values. The number of samples in a training batch. For the first The prediction bias value for each sample. For the first The true pointing deviation label value for each sample.

Images