A lightweight system for onboard neural network processing

Abstract

This invention discloses a lightweight system for spaceborne neural network processing. The system is equipped with a computer program that executes a method comprising the following steps: S1, analyzing the network structure of the spaceborne neural network model to obtain an index for measuring the redundancy of the neural network model; S2, based on the interrelationships between various indices, comprehensively considering the inclusion and complementarity of redundant parts, constructing a comprehensive redundancy index for the overall neural network; S3, dynamically lightweighting the neural network according to the comprehensive redundancy fusion model; and uploading the lightweight neural network model to a satellite to perform on-orbit AI inference tasks in a spaceborne environment. This invention enables the construction and fusion mechanism of multi-level redundancy indices, thereby quantitatively analyzing the degree of redundancy in the neural network, evaluating the model's lightweighting potential, and thus achieving lightweighting of the spaceborne neural network, which is beneficial for neural networks to perform inference applications in resource-constrained spaceborne environments.

Description

Technical Field

[0001] This invention relates to the field of intelligent satellite data processing, and in particular to a lightweight system for onboard processing neural networks. Background Technology

[0002] The rapid development of deep learning technology has driven innovation in artificial intelligence applications, but the exponential growth of neural network scale and the resource constraints of edge devices have created a sharp contradiction. With the intelligent development of satellites, deep learning technology is closely integrated with satellites, enabling them to perform intelligent data processing in orbit using deep learning models, greatly improving data timeliness. However, the exponential growth in the parameter scale of modern neural networks, evolving from millions to hundreds of billions, poses a severe challenge to model deployment. Onboard equipment generally suffers from limited computing power (typically only a few TOPS), limited memory (1-4GB), and power sensitivity (milliwatts), making it difficult to support the real-time inference needs of complex models. In this scenario, lightweight model technology becomes particularly important.

[0003] Existing research mainly revolves around four major technical approaches: pruning, quantization, knowledge distillation, and lightweight architecture design. However, current research exhibits a clear characteristic of "technical practice leading the way, while methodological support lags behind." Most works focus on improving the performance of specific lightweighting methods, while neglecting the exploration of the essential principles of lightweighting.

[0004] From a technical research perspective, current research on lightweight models focuses on lightweight technical methods, lacking systematic guidance and relying on causal exploration. While existing methods (such as channel pruning and parameter quantization) can reduce model complexity, they struggle to answer fundamental questions like "why is it compressible?", "where is it compressible?", and "how much can it be compressed?". Without measuring the redundancy of current models, developers find it difficult to determine whether further optimization is possible. This current lack of interpretability stems from the "black box" nature of neural networks: redundant model representations lack interpretability, leading to a disconnect between pruning strategies and task requirements, and raising doubts about the reliability of compressed models in extreme scenarios. Summary of the Invention

[0005] The purpose of this invention is to solve the technical problem that existing technologies cannot efficiently achieve lightweight neural network models, and to propose a spaceborne lightweight neural network processing system.

[0006] The technical problem of this invention is solved by the following technical solution:

[0007] A lightweight spaceborne neural network processing system, the system being equipped with a computer program that executes a method comprising the following steps:

[0008] S1. By analyzing the network structure of the onboard neural network model, a redundancy quantification index is obtained to measure the redundancy of the neural network model.

[0009] S2. Based on the interrelationships between redundancy measurement indicators, and taking into account the inclusion and complementarity of the redundant parts, a comprehensive redundancy fusion model for the overall neural network is constructed.

[0010] S3. Dynamically lightweight the neural network model based on the comprehensive redundancy fusion model;

[0011] S4. Upload the lightweight neural network model to the satellite to perform on-orbit AI inference tasks in a spaceborne scenario.

[0012] In some embodiments, the following technical features are also included:

[0013] The redundancy metrics include hierarchical redundancy, channel redundancy, and convolutional kernel redundancy. Step S1 specifically includes the following steps:

[0014] S11. Based on the similarity of the output feature maps of each level, the redundancy of the level is obtained;

[0015] S12. Based on the sparse distribution of channel weights in the fully connected layer, the channel redundancy is obtained;

[0016] S13. Based on the average gradient magnitude of the convolution kernel, the redundancy of the convolution kernel is obtained.

[0017] In some embodiments, the following technical features are also included:

[0018] Step S11 specifically includes the following steps:

[0019] S111. Perform kernel matrix mapping on the feature map to eliminate size differences;

[0020] S112. Calculate the similarity value of the kernel matrix using the centered kernel alignment method;

[0021] S113. Normalize the similarity values ​​to obtain a centered kernel alignment index;

[0022] S114. The centralized kernel alignment index is scaled using a hyperbolic tangent function to obtain the hierarchical redundancy.

[0023] In some embodiments, the following technical features are also included:

[0024] Step S12 specifically includes the following steps:

[0025] S121. Calculate the Euclidean L1 norm of the channel weight vector of each fully connected layer to obtain the importance score of each channel;

[0026] S122. Use information entropy to measure the similarity of each channel, calculate the effective number of channels in the fully connected layer, and then obtain the layer channel redundancy.

[0027] S123. Calculate the weights based on the parameters of the fully connected layer, and weight the redundancy of the channel channels of all fully connected layers to obtain the channel redundancy.

[0028] In some embodiments, the following technical features are also included:

[0029] S131, the average absolute gradient of the statistical convolution kernel;

[0030] S132. Normalize the mean absolute gradient input scaling hyperbolic tangent function to obtain the convolution kernel redundancy.

[0031] In some embodiments, the following technical features are also included:

[0032] In step S114, the centered kernel alignment index is... The expression for scaling using the hyperbolic tangent function tanh is: in, Here, β is the scaled, centered kernel alignment metric, and β is the scaling factor. The scaling factor, acting as a translation parameter, controls the allocation of low similarity values; as the scaling factor increases, the impact of low similarity values ​​on redundancy gradually decreases; hierarchical redundancy. The expression is: in, is the scaled, centralized kernel alignment index, where i and n are constants.

[0033] In some embodiments, the following technical features are also included:

[0034] Channel redundancy in step S123 The expression is: in, As weight, The redundancy is defined as the channel redundancy, where L is the total number of fully connected layers; the redundancy is defined as the kernel redundancy in step S132. The expression is: Where N is the total number of convolutional layers. This represents the redundancy of the convolutional kernel.

[0035] In some embodiments, the following technical features are also included:

[0036] The comprehensive redundancy fusion model in step S2 The expression is: ;

[0037] Step S3 specifically includes: performing multi-dimensional diagnosis on the input neural network based on the comprehensive redundancy fusion model, analyzing the redundancy of the neural network model, and lightweighting the neural network by accurately releasing redundant channel resources and enhancing key layers.

[0038] The beneficial effects of this invention compared to the prior art include:

[0039] This invention proposes a lightweight system for spaceborne neural networks. By analyzing the network structure of a spaceborne neural network model, a redundancy quantification index is obtained to measure the redundancy degree of the neural network model. Based on the interrelationships between the redundancy quantification indices, and considering the inclusion and complementarity of redundant parts, a comprehensive redundancy fusion model of the overall neural network is constructed. According to the comprehensive redundancy fusion model, lightweight neural network technical features are set, enabling the construction and fusion mechanism of multi-level redundancy indices. This allows for the quantitative analysis of the degree of redundancy in the neural network, the evaluation of the model's lightweight potential, and the realization of dynamic lightweighting of the spaceborne neural network, which is beneficial for the inference application of the neural network in the resource-constrained spaceborne environment.

[0040] Other beneficial effects of the embodiments of the present invention will be further described below. Attached Figure Description

[0041] Figure 1 This is a flowchart of a program execution method for a lightweight onboard neural network processing system according to an embodiment of the present invention;

[0042] Figure 2 This is a flowchart of a method for program execution in a lightweight system for processing neural networks on a spaceborne platform, according to another embodiment of the present invention. Detailed Implementation

[0043] The present invention will be further described below with reference to the accompanying drawings and preferred embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0044] It should be noted that the directional terms such as left, right, up, down, top, and bottom used in this embodiment are only relative concepts or are based on the normal use of the product, and should not be considered as restrictive.

[0045] This invention focuses on the challenges of quantifying redundancy and characterizing the degree of lightweighting in spaceborne neural networks. Following the principles of "simplification, bottom-up approach, classification and hierarchical structure, and quantitative evaluation," it aims to develop a model for characterizing and quantifying neural network redundancy, starting with key components such as the hierarchical structure, connection channels, and convolutional units, to measure network redundancy. This invention aims to construct a network redundancy quantification model to more comprehensively evaluate the model's lightweighting potential, thereby forming a network redundancy quantification index. This provides a technical framework for neural network lightweighting, a methodological basis for key lightweighting technologies such as adaptive pruning strategies and dynamic parameter quantification, and promotes a paradigm shift in lightweight design from experience-driven to method-driven approaches, significantly enhancing the interpretability and controllability of the model optimization process.

[0046] This paper analyzes the redundancy of a network layer by layer, focusing on its core components such as hierarchical structure, connection channels, and convolutional units. Layer redundancy, channel redundancy, and convolutional kernel redundancy are designed to more comprehensively evaluate the model's lightweight potential, leading to a comprehensive redundancy fusion model that provides a technical basis for lightweighting neural networks. The technical route for characterizing and quantifying neural network redundancy is as follows: Figure 2 As shown: First, the network structure of the neural network is analyzed, and the redundancy of layers, channels, and convolutional kernels is calculated to measure the redundancy of the network from different levels. Second, based on the interrelationship of these three factors, and considering the inclusion and complementarity of the redundant parts, a comprehensive redundancy fusion model of the overall network is constructed.

[0047] This invention provides a lightweight system for onboard neural network processing, which includes computer programs such as... Figure 1 As shown, a computer program is used to execute a method that includes the following steps:

[0048] Step S1: By analyzing the network structure of the onboard neural network model, redundancy metrics are obtained to measure the redundancy level of the neural network model. These redundancy metrics include layer redundancy, channel redundancy, and convolutional kernel redundancy. Step S1 specifically includes the following steps:

[0049] S11. Based on the similarity of the output feature maps of each level, the level redundancy is obtained;

[0050] S12. Based on the sparse distribution of channel weights in the fully connected layer, the channel redundancy is obtained.

[0051] S13. Based on the activation response capability of the convolution kernel, the redundancy of the convolution kernel is obtained.

[0052] The following explains in detail how to calculate layer redundancy, channel redundancy, and convolution kernel redundancy respectively:

[0053] I. The calculation of hierarchical redundancy includes the following steps:

[0054] S111. Kernel matrix mapping is performed on the feature maps to eliminate size differences; the input data is gradually formed into high-level feature representations through feature extraction at different levels of the model. Each level outputs intermediate feature maps. The intermediate feature maps learned by layers 6 to 8 of ResNet50 are very similar (both visually and quantitatively), which indicates that these layers may be redundant in extracting useful information.

[0055] Based on this, we utilize the similarity of intermediate feature maps to reflect the hierarchical redundancy of deep learning models. We quantitatively calculate the similarity between two feature maps to reflect the contribution of different levels of the model to information extraction. Since different feature maps have varying sizes, using traditional metrics such as cosine similarity and Euclidean distance to measure the similarity between two feature maps is impractical.

[0056] S112. The similarity between two feature maps is measured using a centered kernel alignment method, and the similarity value of the kernel matrix is ​​calculated. First, there are feature maps... and feature map , , The batch size represents the input data, and dx or dy represents the number of neurons. and The multiplication sign in This represents the matrix dimension, corresponding to the length of the one-dimensional vector formed by flattening the feature maps. The kernel matrix is ​​calculated for the two feature maps: (1)

[0057] This step allows for the acquisition , The matrix size is fixed at n×n and is independent of dx and dy, thus resolving the problem of incomparability between feature maps of different sizes. Then, we use the Hilbert-Schmidt Independence Criterion (HSIC) to calculate... , Matrix similarity values:

[0058] (2)

[0059] Where Tr(·) represents the trace of the matrix, and H is the centering matrix. HSIC can reflect the similarity between two features. However, the above formula has limitations; it has... This deviation, while not a strictly unbiased estimate, can be removed from the influence of batch size n by using the following formula:

[0060] (3)

[0061] in, They are respectively the ones with deviations , matrix.

[0062] S113. The similarity values ​​will be normalized to obtain the centered kernel alignment index: (4)

[0063] Its value ranges from 0 to 1, and the larger the value, the more correlated the two intermediate features are.

[0064] S114. Next, the centered kernel alignment index is scaled using the hyperbolic tangent function (tanh):

[0065] (5)

[0066] in, Here, β is the scaled, centered kernel alignment metric, and β is the scaling factor. The scaling factor, acting as a translation parameter, controls the allocation of low similarity values. As the scaling factor increases, the impact of low similarity values ​​on redundancy gradually decreases. Regarding the translation parameters for a specific model... We use different [methods] for different model structures. .

[0067] Regardless of the complexity of the model's branching structure, we can divide the model into different levels for consideration, that is, divide the model into blocks. In this way, the input features before each level and the output features after each level can be calculated. This reflects the redundancy at the current level. A higher number indicates higher feature similarity, meaning that the level contributes less to feature improvement and has higher redundancy. Assuming there are n levels to be evaluated, the model's level redundancy is:

[0068] (6)

[0069] in, For the scaled centralized kernel alignment metric, This reflects the hierarchical redundancy of the model, with values ​​between 0 and 1. A higher value indicates a higher level of hierarchical redundancy.

[0070] II. The channel redundancy calculation includes the following steps:

[0071] Treating each output neuron of each fully connected layer as a "channel", for the first... Importance score of each output neuron Take the Euclidean L1 norm of the weight vector corresponding to the output neuron (i.e., the sum of the absolute values ​​of the weights of the output neuron). Next, assign importance scores to all channels of each fully connected layer. Normalization to probability distribution Using information entropy index Characterizing the equivalent non-redundant channel count (effective channel count), layer channel redundancy Depend on With the number of channels The comparison yielded the channel redundancy. Layer weight aggregation yields results in the range [0,1]. This embodiment directly quantifies the intuition that "the more concentrated the weight sparsity (L1) → the fewer effective channels → the higher the redundancy".

[0072] (1) Calculate the channel importance score for each channel in each fully connected layer.

[0073] For the A fully connected layer, the weight matrix typically has the following shape: , It is the output dimension (number of output neurons). The input dimension (number of input features) is used to design channel importance scores. The importance of each output channel can be measured by calculating the Euclidean norm of each output dimension vector, which is defined as follows:

[0074] in, Let be the Euclidean norm of the c-th output dimension vector of the l-th layer weight matrix. In the formula, : denotes taking all elements in this dimension. This is the absolute value of the i-th element in the c-th output dimension vector of the l-th layer weight matrix. To prevent all-zero vectors or division by zero, this embodiment adds a very small value. For numerical stabilization:

[0075] in, This represents the channel importance score after the values ​​have stabilized.

[0076] (2) Calculate the layer channel redundancy of each fully connected layer.

[0077] The importance score of each channel in this layer is normalized to a probability distribution:

[0078] Calculate the information entropy of its natural base: The similarity of each channel is measured using information entropy, and the effective number of channels in the fully connected layer is calculated using the following formula. :

[0079] Here, the exp function is an exponential function with the natural constant e as its base. If the channel weights are perfectly uniform, ,but , If the weights are highly concentrated in a few channels, then It will be much smaller The layer channel redundancy is shown in the following formula: If all output units are equally important (No channel redundancy); if only one unit dominates (Maximum redundancy).

[0080] (3) Weight the redundancy of the layer channels according to the parameter quantity to form the overall channel redundancy.

[0081] Give each layer a weight ( The overall channel redundancy is calculated as follows:

[0082] in, As weight, Where L is the layer channel redundancy, and L is the total number of fully connected layers. Weights Determined by the number of parameters in each fully connected layer: in, These are the parameters of the fully connected layer.

[0083] Method for calculating parameters of fully connected layers:

[0084] III. The calculation of convolution kernel redundancy includes the following steps:

[0085] The importance of a kernel is measured by the average gradient magnitude of the convolution kernel. The gradient reflects the sensitivity of the current loss to this parameter. Kernels with small average absolute gradients have less impact on the loss under the current data / task and can often be regarded as "deletable / redundant" candidates.

[0086] (1) Statistics of mean absolute gradient

[0087] For each output kernel of each convolutional layer, compute the batch-level absolute gradient and then average it:

[0088] in, For convolutional layers, the convolutional kernel is l. The mean absolute gradient scalar, The number of samples in a batch. t For the current sample, To indicate that the first The gradient calculated from each sample For the first Layer The first convolutional kernel One parameter, The number of nuclear elements, By dividing by This can eliminate the influence of kernel size on the total absolute gradient. Then, all layers... Flatten into a full network gradient vector ,in It represents the total number of cores across the entire network.

[0089] (2) Mapping using the scaled hyperbolic tangent function

[0090] Using the scaled hyperbolic tangent function as the mapping function, the average absolute gradients at different scales are mapped onto the [0,1] interval, and then normalized by inputting the scaled hyperbolic tangent function. The formula for the scaled hyperbolic tangent function is as follows:

[0091] in, Represents the translation parameter. The slope parameter is represented by a Bayesian optimization method that automatically selects an appropriate value. Translation parameter. The "center position" of the function is controlled, serving as a threshold for determining kernel importance. Slope parameter. Decided to be in The transition speed from redundancy to importance in the vicinity. Then the redundancy of convolution kernel i is defined as: By using a mapping function, gradients of different scales can be mapped to the [0,1] interval. We employ Bayesian optimization to automatically select appropriate values ​​for the translation and slope parameters. Let the optimization objective be:

[0092] in This represents the variance of the redundancy distribution of the convolution kernel. We want the larger the variance to be, the stronger the distinguishability of the redundancy distribution, so we add a negative sign. Let be the mean of the redundancy distribution. We aim to balance the number of high and low redundancy kernels to reduce the sensitivity to extreme values ​​of redundancy, so we want it to be close to 0.5. The search space for Bayesian optimization is... (Q1 and Q3 are the first and third quartiles of all convolution kernels g, respectively). ( Bayesian optimization is performed using a Gaussian process and Expected Improvement (EI) to iterate through several rounds to find the optimal point.

[0093] (3) Calculate the redundancy of the convolution kernel.

[0094] All convolutional kernels are considered equally important; the average redundancy of the convolutional kernels in each layer is calculated.

[0095] in, This represents the total number of convolutional layers. This represents the redundancy of the convolutional kernel.

[0096] Step S2: Based on the interrelationships between redundancy metrics, and considering the inclusion and complementarity of redundant parts, a comprehensive redundancy fusion model of the overall neural network is constructed to comprehensively evaluate the lightweight nature of the neural network. The following details how to calculate the comprehensive redundancy fusion model: In object detection and recognition tasks, most models consist of convolutional layers and fully connected layers. Layer redundancy... Channel redundancy and convolution kernel redundancy These respectively reflect the module redundancy, fully connected layer redundancy, and convolutional layer redundancy of the model, which have certain inclusion and complementary relationships. Channel redundancy and convolutional kernel redundancy are complementary because the former measures the redundancy of fully connected layers, while the latter measures the redundancy of convolutional layers. Layer redundancy implicitly includes convolutional kernel redundancy and channel redundancy, as it represents the redundancy of all layers in the network, including both fully connected layers and convolutional layers. The comprehensive redundancy fusion model needs to express the inclusion and complementary relationships, avoid the repeated calculation of redundant parts, and be within the range [0, 1]. The formula for our proposed comprehensive redundancy fusion model is as follows: (12)

[0097] Next, we will analyze the above equation. The above equation is divided into two parts by the plus sign. The first part... This represents hierarchical redundancy, serving as the basis for overall redundancy. Part Two middle, This indicates that if the hierarchical redundancy is already high, then additional... and The contribution will be small. The concept of sets is introduced to represent and The joint redundancy between them was eliminated, and their overlapping parts were removed to ensure that there was no duplicate calculation.

[0098] Reasonableness verification

[0099] Consider the limiting case to further understand the rationality of this formula: 1. If That is, the redundancy of the entire network hierarchy has reached 100%. This indicates that the overall redundancy has been completely determined and does not need to be considered. and .

[0100] 2. If If the hierarchical redundancy is completely zero, but the convolution kernel or channels may still have redundancy, then it degenerates into union computation, i.e. This indicates that when hierarchical redundancy is insufficient, overall redundancy is mainly due to... and contribute.

[0101] 3. If , That is, individual convolutional kernels and channels have no redundancy. This indicates that the overall redundancy at this point is determined only by the hierarchical redundancy.

[0102] Finally, let's verify whether the formula always stays between [0, 1]. All parameters , and If all are non-negative, then Non-negative, plus Non-negative, therefore we have The lower bound holds. We expand the formula as follows:

[0103] in, And the maximum value is 1.

[0104] so, The Upper Realm was established.

[0105] During the research process, the embodiments of this invention discovered: hierarchical redundancy. Positively correlated with network depth, as deeper features easily propagate redundancy; negatively correlated with residual connection density, as skip connections allow deeper layers to focus on learning incremental features, preventing excessively deep neural networks from being forced to repeatedly learn shallow features, thus reducing inter-layer dependencies. For example, ResNet networks with residual connections exhibit more stable learning, smoother gradient updates, and significant differences in features between layers, while VGG networks without residual connections suffer from severe gradient vanishing in deeper layers, resulting in high similarity in the outputs of later layers (high layer redundancy). Channel redundancy Positively correlated with channel expansion rate, due to the tendency for redundant channels in wide structures; negatively correlated with the proportion of grouped convolutions, as grouped design naturally suppresses channel redundancy. Convolution kernel redundancy. It is positively correlated with the proportion of standard convolutions because dense convolution kernels are prone to redundancy; it is negatively correlated with the proportion of depthwise separable convolutions because the operator has decomposed redundancy.

[0106] In this embodiment, ResNet50 is used as the object, and the calculated hierarchical redundancy is... Channel redundancy convolution kernel redundancy The overall redundancy calculated using formula (12) is 0.79, indicating that ResNet50 has high layer redundancy and kernel redundancy. In the current scenario, the depth can be reduced, and depthwise separable convolutions can be used instead of standard convolutions. Taking MobileNet as an example, if the calculated layer redundancy... Channel redundancy convolution kernel redundancy The overall redundancy calculated using formula (12) is 0.608, indicating that MobileNet effectively reduces redundancy through depthwise separable convolutions. Compared with ResNet, ResNet has greater potential for lightweighting.

[0107] Step S3: Dynamically lightweight the neural network based on the comprehensive redundancy fusion model. Specifically, the input neural network is subjected to multi-dimensional diagnosis based on the comprehensive redundancy fusion model to analyze the redundancy of the neural network model. By accurately releasing redundant channel resources and enhancing key layers, the neural network is lightweighted.

[0108] Specifically, the input neural network is subjected to multidimensional diagnosis based on the comprehensive redundancy fusion model, and lightweighting is achieved through the following operations.

[0109] S31. Redundancy Diagnosis: Synchronously calculate the hierarchical redundancy of the neural network. Channel redundancy Redundancy of convolution kernels .

[0110] S32, Perform dynamic lightweighting operation:

[0111] S321, Hierarchical pruning: When a certain level of pruning... Exceeding the preset threshold When this happens, remove the redundant layer.

[0112] S322, Channel Selection: For each layer, retain the top α% of channels by importance score.

[0113] S323, Kernel Replacement: When a certain convolution kernel is replaced... Exceeding the preset threshold When that happens, replace it with a depthwise separable convolution.

[0114] S33. Accuracy Restoration: After completing the lightweighting operation, a fine-tuning mechanism is used to restore the model performance.

[0115] S331. Retrain the compressed model using the original training data;

[0116] S332, Set the learning rate to 10% of the initial training value (typical value 0.001);

[0117] S333. When the compression ratio exceeds 50%, the number of fine-tuning rounds shall not be less than 20.

[0118] S334, in high compression scenarios, can be extended to knowledge distillation, using the original model as a teacher model to guide the training of lightweight models.

[0119] The three redundancies correspond to three different components of the neural network, covering the full stack redundancy from macroscopic structure (layers) → mesoscopic data flow (channels) → microscopic computational units (convolutional kernels). Lightweight models (such as MobileNet) often suffer from accuracy bottlenecks due to excessive compression, but it is difficult to pinpoint whether this is caused by insufficient depth, too few channels, or low convolutional kernel efficiency. This invention can perform multidimensional diagnosis of the neural network based on a comprehensive redundancy fusion model:

[0120] Hierarchical diagnosis: If shallow layers are found to have low redundancy (feature extraction capability needs to be preserved), layer pruning should be prohibited, and skip connections should be added;

[0121] Channel diagnostics: High redundancy in the output channels of the SE (Squeeze-and-Excitation, a channel attention module) module was detected → reduce the number of SE channels to avoid overcomputation;

[0122] Convolutional kernel diagnosis: Depthwise convolutional kernels were found to have low redundancy (already highly efficient) → Preserve the structure and focus on optimizing pointwise convolutional kernel quantization.

[0123] Step S4: Upload the lightweight neural network model to the satellite to perform on-orbit AI inference tasks in a spaceborne scenario. This embodiment of the invention achieves precise and effective guidance for compression by accurately releasing redundant channel resources and enhancing key layers, with minimal impact on accuracy, thereby realizing the lightweighting of the neural network model.

[0124] The embodiments of the present invention have the following advantages compared with the prior art:

[0125] Most existing research focuses on innovations in lightweighting techniques while neglecting the underlying mechanisms. Current lightweighting research follows an experience-driven paradigm, often relying on engineers' intuition, extensive experimentation (such as trying different pruning rates, quantization bit widths, and searching for different lightweight architectures), experience with benchmark models (such as the MobileNet series), and the costly trial-and-error process of neural architecture search. The process typically involves: attempting compression → evaluating the effect → adjusting → re-attempting, which is inefficient and the results are difficult to generalize. In this embodiment of the invention, the above process is optimized as follows: first, quantitatively evaluate redundancy → predict compression potential and guide compression strategies based on redundancy → efficiently execute compression → verify the results. The redundancy metric provides prior knowledge and guiding principles, significantly reducing blind spots.

[0126] This invention proposes a method for quantitatively analyzing network redundancy. By constructing multi-layered redundancy quantification indices (layer feature similarity, channel importance distribution, and convolutional kernel response intensity), a comprehensive redundancy fusion model is established to achieve multi-dimensional analysis and global evaluation of neural network redundancy. This provides a quantitative evaluation benchmark and compression boundary prediction for lightweighting, promoting the paradigm shift of pruning strategies from experience-driven optimization to vectorization-driven approaches. It can explain why models can be lightweighted and how much can be lightweighted, providing methodological support for key lightweighting technologies and driving the paradigm shift of lightweight design from experience-driven to method-driven. In terms of implementation, this invention calculates the network redundancy from three parts, considering the network at different levels to more comprehensively uncover network redundancy. Considering the inclusion and complementarity relationships between different levels, this invention proposes a comprehensive redundancy calculation formula to effectively address this issue. This invention proposes a method for quantitatively analyzing network redundancy, answering the question "how much redundancy exists in the network". This method can serve as a reliable reference in the lightweighting process, significantly enhancing the interpretability and controllability of the model optimization process. This invention calculates the redundancy of the network from three different levels and scientifically integrates them to form a comprehensive redundancy, which serves as an overall characterization of network redundancy.

[0127] Application scenarios of this invention include:

[0128] ① Automated Model Selection and Lightweight Deployment Pipeline: In deployment scenarios on edge devices with limited computing power (such as IoT terminals and embedded systems), the system receives multiple candidate neural network models. It automatically performs multi-dimensional redundancy analysis according to this invention on each candidate model, calculating its comprehensive redundancy score. The system automatically identifies the model with the highest comprehensive redundancy score and then triggers an automated lightweighting process for that model (such as applying preset pruning and quantization algorithms). The streamlined model after lightweighting is automatically deployed to the target edge device for operation. This process is fully automated, avoiding the resource consumption of deploying multiple candidate models on real devices for testing, and directly outputting a streamlined and runnable model for deployment, significantly shortening the selection and deployment cycle and reducing the storage requirements and computing load of edge devices.

[0129] ② Automated Model Compression Engine Based on Redundancy Profiling: In the model compression process, the system first automatically performs redundancy profiling on the input target neural network model according to this embodiment of the invention, generating specific redundancy quantification results (e.g., identifying specific layers, channels, or convolutional kernels with redundancy exceeding a preset threshold). Based on this quantification result, the engine automatically matches and executes the most suitable compression operation: automatically performing layer pruning or module replacement on highly redundant layers; automatically performing channel pruning algorithms on highly redundant channels; and automatically performing kernel pruning, low-rank decomposition, or replacement with depthwise separable convolutional operators on highly redundant convolutional kernels. The selection and execution of the entire compression strategy are automatically completed by the engine based on the profiling results, without relying on expert experience for configuration or time-consuming NAS searches. The final output is a simplified model with significantly reduced size and computational cost, directly improving the model's inference speed and energy efficiency on the target hardware.

[0130] ③ A lightweight spaceborne model system for dynamic redundancy optimization based on embodiments of the present invention: To meet the requirements of satellites performing AI tasks such as Earth observation target identification and space event detection in real time on orbit, the model needs to be lightweighted and compressed on the ground before being uploaded to the satellite to execute AI tasks. At the ground station, the system receives the target spaceborne AI model. First, it automatically performs a multi-dimensional redundancy analysis of the model (hierarchy, channels, convolution kernels) to accurately quantify the redundancy level of each dimension. Based on the analysis results, the system automatically generates and executes the optimal compression strategy. The compression process strictly constrains the model accuracy loss to within a preset threshold (e.g., ≤5%) to ensure mission reliability. After lightweighting, the simplified model has a significantly reduced size and a greatly reduced computational load. This simplified model is automatically packaged and uploaded to the satellite via a narrow space-to-ground link. After receiving the model, the satellite automatically loads and replaces the old model on the spaceborne computing platform. The new model is immediately put into operation to perform on-orbit AI inference tasks. This solution directly overcomes the bottleneck of satellite-to-ground transmission, shortens the model update window, reduces on-board computing power consumption and memory usage, and improves the satellite's long-term autonomous operation capability and real-time response efficiency in complex space environments. It is a key technology supporting future intelligent satellite constellations.

[0131] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0132] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0133] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0134] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0135] The above description provides a further detailed explanation of the present invention in conjunction with specific / preferred embodiments, and it should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various substitutions or modifications can be made to these described embodiments without departing from the concept of the present invention, and all such substitutions or modifications should be considered within the scope of protection of the present invention. In the description of this specification, the reference to terms such as "an embodiment," "some embodiments," "preferred embodiment," "example," "specific example," or "some examples," etc., indicates that the specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in any suitable manner in one or more embodiments or examples. Without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification and the features of different embodiments or examples. Although the embodiments of the present invention and their advantages have been described in detail, it should be understood that various changes, substitutions, and modifications can be made herein without departing from the scope of protection of the patent application.

Claims

1. A neural network lightening system for on-board platform redundancy dynamic optimization, characterized in that, The system is installed with a computer program for performing a method comprising the following steps: S1. In the deployment scene of edge devices with limited computing power, the system receives multiple candidate on-board neural network models at a ground station, and automatically performs multi-dimensional redundancy profiling on each candidate on-board neural network model, obtains a redundancy quantification index for measuring the redundancy degree of the neural network model by analyzing the network structure of the on-board neural network model; the redundancy quantification index includes hierarchical redundancy, channel redundancy, and convolution kernel redundancy; S2. Based on the interrelationship between the redundancy quantification indexes, the comprehensiveness and complementarity of the redundant parts are comprehensively considered to construct a comprehensive redundancy fusion model of the overall neural network; S3. The automatic model compression engine automatically matches and performs compression operations, dynamically lightens the neural network model according to the comprehensive redundancy fusion model; wherein the compression process constrains the model accuracy loss within a preset threshold; S4. The lightened neural network model is uploaded to the satellite through the satellite-ground link, and the satellite automatically loads and replaces the old neural network model on the on-board platform to shorten the selection and deployment cycle, reduce the storage requirement and computing load of the edge device, and perform on-orbit AI inference tasks in the on-board scene; Wherein, step S1 specifically comprises the following steps: S11. Based on the similarity of the output feature maps of each layer, the hierarchical redundancy is obtained; S12. Based on the distribution sparsity of the channel weights of the fully connected layer, the channel redundancy is obtained; S13. Based on the average gradient amplitude of the convolution kernel, the convolution kernel redundancy is obtained; Step S11 specifically comprises the following steps: S111. Kernel matrix mapping is performed on the feature map to eliminate size difference; S112. The similarity value of the kernel matrix is calculated by the centralized kernel alignment method; S113. The similarity value is normalized to obtain the centralized kernel alignment index; S114. The centralized kernel alignment index is scaled by the hyperbolic tangent function to obtain the hierarchical redundancy.

2. The system of claim 1, wherein, Step S12 specifically comprises the following steps: S121. The Euclidean norm of each fully connected layer channel weight vector is calculated to obtain the importance score of each channel; S122. The similarity of each channel is measured using information entropy, the effective channel number of the fully connected layer is calculated, and then the layer channel redundancy is obtained; S123. The weight of the parameter amount of the fully connected layer is calculated, and the layer channel redundancy of all fully connected layers is weighted to obtain the channel redundancy.

3. The system of claim 1, wherein, Step S13 specifically comprises the following steps: S131. The average absolute gradient of the convolution kernel is counted; S132. The average absolute gradient is input into the scaling hyperbolic tangent function for normalization to obtain the convolution kernel redundancy.

4. The system of claim 1, wherein, In step S114, the centralised core alignment indicator is calculated The expression that is put into the hyperbolic tangent function tanh for scaling is: wherein, is the scaled centered kernel alignment indicator, is a scaling factor, is a translation parameter, the scaling factor controls the distribution score of low similarity values; as the scaling factor increases, the influence of low similarity values on redundancy gradually decreases.

5. The system of claim 4, wherein, The hierarchical redundancy The expression is: wherein, is the scaled centered kernel alignment metric, and n is a constant.

6. The system of claim 5, wherein, Channel redundancy of step S123 The expression is: wherein, is a weight, is a layer channel redundancy, L is a total number of full connection layers; the convolution kernel redundancy of step S132 The expression of is: Wherein, N is the total number of convolution layers, is the redundancy of layer convolution kernel.

7. The system of claim 6, wherein, The integrated redundancy fusion model in step S2 The expression is: 。 8. The system of claim 1, wherein, Step S3 specifically comprises: multi-dimensional diagnosis of the input neural network according to the comprehensive redundancy fusion model, profiling the redundancy of the neural network model, releasing redundant channel resources accurately and enhancing key layers to lighten the neural network.

Images