Satellite-borne processing neural network lightweight system

Abstract

The invention discloses a satellite-borne processing neural network lightweight system, the system is provided with a computer program, and the program is used for executing to realize a method comprising the following steps: S1, obtaining an index for measuring the redundancy degree of a neural network model by analyzing the network structure of the satellite-borne neural network model; s2, comprehensively considering inclusion and complementarity of a redundant part based on a mutual relationship among the indexes, and constructing a comprehensive redundancy index of the overall neural network; s3, performing dynamic weight reduction on the neural network according to the comprehensive redundancy fusion model; and uploading the lightened neural network model to a satellite so as to execute an in-orbit AI reasoning task in a satellite-borne scene. According to the method, a multi-level redundancy index construction and fusion mechanism can be realized, so that the redundancy degree in the neural network is quantitatively analyzed, the lightweight potential of the model is evaluated, the lightweight of the satellite-borne neural network is realized, and reasoning application of the neural network in a satellite-borne environment with limited resources is facilitated.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of satellite data intelligent processing, in particular to a lightweight system for on-board processing neural network. BACKGROUND

[0002] The rapid development of deep learning technology has promoted the innovation of artificial intelligence applications, but the scale expansion of neural networks and the resource limitations of edge devices have formed a sharp contradiction. With the development of satellite intelligence, deep learning technology is closely integrated with satellites. Satellites can perform intelligent data processing on-orbit with the help of deep learning models, greatly improving the timeliness of data. However, the scale of modern neural network parameters grows exponentially, evolving from millions to hundreds of millions, which poses a serious challenge to model deployment. On-board devices generally have limited computing power (usually only a few TOPS), limited memory (1-4GB), and sensitive power consumption (milliwatt level), making it difficult to meet the real-time inference requirements of complex models. In this scenario, model lightweight technology is particularly important.

[0003] Existing research mainly focuses on four major technical routes: pruning, quantization, knowledge distillation, and lightweight architecture design. However, current research shows a clear "technology practice ahead, method support lag" feature. Most work focuses on performance improvement of lightweight specific methods, but ignores the exploration of the nature of lightweight.

[0004] From a technical research perspective, current model lightweight research focuses on lightweight technical methods, lacking systematic guidance for cause and effect exploration. Existing methods (such as channel pruning and parameter quantization) can reduce model complexity, but they cannot answer fundamental questions such as "why can be compressed", "where can be compressed", and "how much can be compressed". If the redundancy of the current model is not measured, developers will have difficulty determining whether the model can be further optimized. The current lack of explanation is due to the "black box" nature of neural networks: the model redundancy representation lacks interpretability, leading to a disconnect between pruning strategies and task requirements, and the reliability of the compressed model in extreme scenarios is questionable. SUMMARY

[0005] The purpose of the present application is to solve the technical problem that the prior art cannot efficiently implement neural network model lightweight, and a lightweight system for on-board processing neural network is proposed.

[0006] The technical problem of the present application is solved by the following technical solution: A lightweight system for on-board processing neural network, the system is installed with a computer program, the program is used to execute to realize a method comprising the following steps: S1, by analyzing the network structure of the on-board neural network model, obtaining a redundancy quantization index for measuring the redundancy degree of the neural network model; S2, based on the correlation between the redundancy quantification indexes, comprehensively considering the inclusiveness and complementarity of the redundant part, a comprehensive redundancy fusion model of the overall neural network is constructed; S3, dynamically lightening the neural network model according to the comprehensive redundancy fusion model; S4, the lightened neural network model is injected into the satellite to execute the on-orbit AI inference task in the on-board scene.

[0007] In some embodiments, the following technical features are further included: The redundancy degree indicators include hierarchical redundancy, channel redundancy, and convolution kernel redundancy, and step S1 specifically includes 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.

[0008] In some embodiments, the following technical features are further included: Step S11 specifically includes 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 put into the hyperbolic tangent function for scaling to obtain the hierarchical redundancy.

[0009] In some embodiments, the following technical features are further included: Step S12 specifically includes the following steps: S121, the Euclidean L1 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.

[0010] S123, according to the weight calculated according to the parameter amount of the fully connected layer, the layer channel redundancies of all fully connected layers are weighted to obtain the channel redundancy.

[0011] In some embodiments, the following technical features are further included: S131, the average absolute gradient of the convolution kernel is counted; S132, input the average absolute gradient into a scaled hyperbolic tangent function for normalization to obtain the redundancy of the convolution kernel.

[0012] In some embodiments, the technical features further include: The centralized kernel alignment index in step S114 The expression for scaling into a hyperbolic tangent function tanh is: Wherein, is the scaled centralized kernel alignment index, and β is a scaling factor, is a translation parameter, and 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; the hierarchical redundancy The expression is: Wherein, is the scaled centralized kernel alignment index, and i and n are constants.

[0013] In some embodiments, the technical features further include: The channel redundancy in step S123 The expression is: Wherein, is a weight, is a layer channel redundancy, and L is the total number of fully connected layers; the convolution kernel redundancy in step S132 The expression is: Wherein, N is the total number of convolution layers, is a layer convolution kernel redundancy.

[0014] In some embodiments, the technical features further include: The expression of the comprehensive redundancy fusion model in step S2 is: ; Step S3 specifically includes: performing multi-dimensional diagnosis on the input neural network according to the comprehensive redundancy fusion model, analyzing the redundancy of the neural network model, releasing redundant channel resources precisely and enhancing key layers to perform neural network lightweight.

[0015] The beneficial effects of the present application compared with the prior art include: The satellite-borne processing neural network lightweight system provided by the application can realize multi-level redundancy index construction and fusion mechanism by setting the technical features of neural network lightweight according to the comprehensive redundancy fusion model, thereby quantitatively analyzing the degree of redundancy existing in the neural network and evaluating the lightweight potential of the model, so as to realize dynamic lightweight of the satellite-borne neural network, and facilitate the inference application of the neural network in the resource-limited satellite-borne environment.

[0016] Other beneficial effects in the embodiments of the application will be further described below. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is a method flowchart of program execution of a satellite-borne processing neural network lightweight system in an embodiment of the application; Figure 2 is a method flowchart of program execution of a satellite-borne processing neural network lightweight system in another embodiment of the application. DETAILED DESCRIPTION

[0018] The application will be further described below by comparing with the drawings and combining with the preferred embodiments. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0019] It should be noted that the left, right, up, down, top, bottom and other orientation terms in the embodiments are only relative concepts or references to the normal use state of the product, and should not be considered as limiting.

[0020] The embodiments of the application focus on the problems of difficult quantification of satellite-borne neural network redundancy and difficult representation of lightweight degree, and according to the idea of "simplifying, from bottom to top, classification and layering, and evaluation and quantification", the network level structure, connection channel, convolution unit and other key components are started to develop the neural network redundancy representation and quantification model to measure the redundancy of the network. The embodiments of the application aim to build a network redundancy quantification model to more comprehensively evaluate the lightweight potential of the model, and then form a network redundancy quantification index to provide a technical framework for neural network lightweight, a methodological basis for adaptive pruning strategy, dynamic parameter quantization and other lightweight key technologies, promote the paradigm shift of lightweight design from experience-driven to method-driven, and significantly enhance the explainability and controllability of the model optimization process.

[0021] Around the hierarchy of the network, the connection channel, the convolution unit and other core components, the redundancy of the network is analyzed layer by layer. The hierarchical redundancy, channel redundancy and convolution kernel redundancy are designed to more comprehensively evaluate the lightweight potential of the model, and then the comprehensive redundancy fusion model is obtained to provide a technical basis for neural network lightweight. The neural network redundancy representation and quantization technology route are as shown in Figure 2 The first step is to analyze the network structure of the neural network, and calculate the hierarchical redundancy, channel redundancy and convolution kernel redundancy to measure the redundancy degree of the network from different aspects. The second step is to construct the comprehensive redundancy fusion model of the overall network based on the relationship among the three, and comprehensively consider the inclusiveness and complementarity of the redundant parts.

[0022] The embodiment of the application provides a satellite processing neural network lightweight system, which is installed with a computer program, as shown in Figure 1 The computer program is used for executing to realize a method comprising the following steps: Step S1: by analyzing the network structure of the satellite neural network model, redundancy quantization indexes for measuring the redundancy degree of the neural network model are obtained, and the redundancy quantization indexes comprise hierarchical redundancy, channel redundancy and convolution kernel redundancy. Wherein, step S1 specifically comprises the following steps: S11, based on the similarity of the output feature maps of each level, the hierarchical redundancy is obtained; S12, based on the distribution sparsity of the full connection layer channel weight, the channel redundancy is obtained; S13, based on the activation response ability of the convolution kernel, the convolution kernel redundancy is obtained.

[0023] The following specifically explains how to calculate the hierarchical redundancy, channel redundancy and convolution kernel redundancy respectively: I. The hierarchical redundancy calculation specifically comprises the following steps: S111, the feature map is mapped to a kernel matrix to eliminate the size difference; the input data is extracted by different levels of the model, and gradually forms a high-level feature representation. Each level outputs an intermediate feature map, and the intermediate feature maps learned by the 6th to 8th layers of ResNet50 are very similar (whether visually or quantitatively), which indicates that these layers may be redundant when extracting useful information.

[0024] Based on this, we use the similarity of the intermediate feature maps to reflect the hierarchical redundancy of the deep learning model. We reflect the contribution degree of different levels of the model to information extraction by quantitatively calculating the similarity between two feature maps. Since the sizes of different feature maps are inconsistent, it is unrealistic to use traditional indicators such as cosine similarity and Euclidean distance to measure the similarity between two feature maps.

[0025] S112, measure the similarity of two feature maps by the centered kernel alignment method, calculate the similarity value of the kernel matrix, first, there are feature maps and feature maps , , batch_size represents the input data, dx or dy represents the number of neurons, and the multiplication sign represents the matrix dimension, corresponding to the length of the one-dimensional vector formed by the feature map expansion. Calculate the kernel matrix of two feature maps: (1) This step makes the resulting , The size of the matrix is fixed as n x n, regardless of dx and dy, which solves the problem of comparing feature maps of different sizes. Then we use the Hilbert-Schmidt Independence Criterion (HSIC) to calculate the similarity value of , The matrix: (2) Where Tr(·) represents the trace of the matrix, and H is the centering matrix. HSIC can reflect the similarity of two features. However, the above formula has limitations, which has bias, not a strict unbiased estimate, by using the following formula to calculate it away from the batch size (batch size) n: (3) Where are the biased , Matrices.

[0026] S113, normalize the similarity value to get the centered kernel alignment index: (4) Its value range is 0 to 1, the larger the value, the more relevant the two intermediate features.

[0027] S114, then put the centered kernel alignment index into the hyperbolic tangent function (tanh) for scaling: (5) Where is the scaled centered kernel alignment index, β is the scaling factor, is the translation parameter, the scaling factor controls the distribution score of low similarity values. With the increase of the scaling factor, the influence of low similarity values on redundancy gradually decreases. As for the translation parameter of a specific model We use different [methods] for different model structures. .

[0028] 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: (6) 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.

[0029] II. The channel redundancy calculation includes the following steps: 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 yields 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".

[0030] (1) Calculate the channel importance score for each channel in each fully connected layer. 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: where, is the Euclidean norm of the cth output dimension vector of the lth layer weight matrix, and the formula means taking all elements in this dimension, is the absolute value of the ith element in the cth output dimension vector of the lth layer weight matrix, in order to prevent all zero vectors or division by zero, a very small value is added in this embodiment for numerical stability: where, is the channel importance score after numerical stability.

[0031] (2) Calculate the layer channel redundancy of each fully connected layer The channel importance score of each channel of this layer is normalized to a probability distribution:

[0032] Calculate the information entropy of its natural base: Use information entropy to measure the similarity of each channel, and calculate the effective channel number of this fully connected layer according to the following formula :

[0033] where, the exp function is the exponential function with natural constant e as the base, if the channel weight is completely uniform, then , If the weight is highly concentrated in a few channels, then will be much smaller than . The layer channel redundancy is as follows: If all output units are equally important (no channel redundancy); if only one unit is dominant (maximum redundancy).

[0034] (3) Weight the layer channel redundancy according to the parameter quantity to form the overall channel redundancy Give each layer a weight ( ) The overall channel redundancy is calculated as follows:

[0035] where, is the weight, is the layer channel redundancy, and L is the total number of fully connected layers. The weight is determined by the parameter quantity of each fully connected layer: where, is the parameter quantity of the fully connected layer.

[0036] Fully connected layer parameter quantity calculation method:

[0037] III. Redundancy calculation of convolution kernel includes the following steps: The importance of kernel is measured by the average gradient amplitude of the kernel. The gradient reflects the sensitivity of the current loss to the parameter. The kernel with small average absolute gradient has less impact on the loss under the current data / task, and can be regarded as a candidate for "deletion / redundancy".

[0038] (1) Average absolute gradient statistics The absolute gradient sum of each output kernel of each convolution layer is calculated and then averaged at batch level: wherein, is the average absolute gradient scalar of the convolution kernel of the convolution layer l, is the number of samples in a batch, t is the current sample, is the gradient calculated for the th sample, is the th parameter of the th convolution kernel of the th layer, is the number of kernel elements, Dividing by can eliminate the influence of kernel size on the total amount of absolute gradient. Then, the of all layers is flattened into a gradient vector of the whole network, wherein is the total number of kernels in the whole network.

[0039] (2) Mapping using scaled hyperbolic tangent function The scaled hyperbolic tangent function is used as the mapping function to map the average absolute gradient of different scales to the interval [0, 1]. The scaled hyperbolic tangent function is normalized by inputting it. The formula of the scaled hyperbolic tangent function is as follows:

[0040] wherein, represents the translation parameter, represents the slope parameter, and the method of Bayesian optimization is used to automatically select the appropriate value. The translation parameter controls the "center position" of the function, which is the threshold for determining the importance of the kernel. The slope parameter determines the transition speed from redundancy to importance near . Then the redundancy of the convolution kernel i is defined as: Through the mapping function, the gradient of different scales can be mapped to the interval [0, 1]. The method of Bayesian optimization is used to automatically select the appropriate value of the translation parameter and the slope parameter. The optimization objective is set as: wherein variance of the redundancy distribution of the convolution kernel, the greater the variance, the stronger the distinguishability of the redundancy distribution, so add a negative sign. is the mean of the redundancy distribution, and the number of high and low redundancy kernels is balanced to reduce the sensitivity of the extreme value of the redundancy, so it is expected to be close to 0.5. The search space of Bayesian optimization is (Q1, Q3 are the first quartile and the third quartile of all convolution kernels g, respectively), ). Use Gaussian Process and Expected Improvement (EI) for Bayesian optimization, and iterate a number of times to find the optimal point.

[0041] (3) Calculate the redundancy of the convolution kernel All convolution kernels are considered equally important, and the average value of the redundancy of the convolution kernel of the layer is calculated: wherein, is the total number of convolution layers, is the redundancy of the convolution kernel of the layer.

[0042] Step S2: Based on the mutual relationship between the redundancy quantification indexes, the comprehensiveness and complementarity of the redundant part are comprehensively considered, and a comprehensive redundancy fusion model of the overall neural network is constructed, so as to comprehensively evaluate the lightweight of the neural network. The following specifically explains how to calculate the comprehensive redundancy fusion model: In the target detection and recognition task, most of the models are composed of convolution layers and fully connected layers. The hierarchical redundancy , the channel redundancy and the convolution kernel redundancy respectively embody the module redundancy, the fully connected layer redundancy and the convolution layer redundancy of the model, and they have certain inclusion and complementary relationship. Channel redundancy and convolution kernel redundancy are complementary, because the former measures the redundancy of the fully connected layer, and the latter measures the redundancy of the convolution layer. The hierarchical redundancy implies the convolution kernel redundancy and the channel redundancy, because it represents the redundancy of all levels in the network, which includes both the fully connected layer and the convolution layer. The comprehensive redundancy fusion model needs to express both the inclusion and the complementary relationship, avoid redundant calculation of the redundant part, and also be between [0, 1]. The formula of the comprehensive redundancy fusion model we determine is as follows: (12) Next, the above formula is analyzed. The above formula is divided into two parts by the plus sign. The first part represents the hierarchical redundancy, which is the basis of the overall redundancy. The second part , wherein indicates that if the hierarchical redundancy is already high, the additional and contribution is small.​ The idea of set is introduced to represent and the joint redundancy between them, remove their intersection part, and ensure no repeated calculation.

[0043] Rationality verification Consider the limit case to further understand the rationality of the formula: 1, if , that is, the entire network level redundancy has reached 100%, then , which means that the overall redundancy has been completely determined, and there is no need to consider and .

[0044] 2, if , that is, the level redundancy is completely 0, but the convolution kernel or channel may still have redundancy, then it degenerates to the union calculation, that is , which means that when the level redundancy is insufficient, the overall redundancy is mainly contributed by and .

[0045] 3, if , , that is, there is no redundancy in the convolution kernel and the channel alone, then , which means that at this time the overall redundancy is only determined by the level redundancy.

[0046] Finally, we verify whether the formula always remains in [0, 1]. All parameters , and are non-negative, then is non-negative, and is non-negative, so , the lower bound is established. We expand the formula as follows: wherein and the maximum value is 1.

[0047] Therefore, , the upper bound is established.

[0048] In the research process, the embodiment of the invention found that: the level redundancy is positively correlated with the network depth, because deep features are easy to transmit redundancy; and is negatively correlated with the residual connection density, because the jump connection can let the deep layer focus on learning incremental features, avoid the deep layer being forced to repeat learning shallow features due to the over-deep neural network, and reduce the inter-layer dependence. For example, the resnet network with residual connection learns more stably, the gradient update is smooth, and the difference between inter-layer features is significant, while the VGG network without residual connection has serious gradient disappearance in the deep layer, and the output similarity of the last few layers is high (high level redundancy). The channel redundancy Positive correlation with channel expansion rate, because the wide structure is easy to exist redundant channels; negative correlation with group convolution proportion, because group design naturally suppresses channel redundancy. Convolution kernel redundancy Positive correlation with standard convolution proportion, because dense convolution kernel is easy to be redundant; negative correlation with depth separable convolution proportion, because the operator has decomposed redundancy.

[0049] In this embodiment, taking ResNet50 as the object, if the calculated hierarchical redundancy , channel redundancy , convolution kernel redundancy , the comprehensive redundancy calculated by formula (12) is 0.79, which indicates that Resnet50 has higher hierarchical redundancy and convolution kernel redundancy, and in the current scene, the depth can be reduced and the depth separable convolution is used instead of the standard convolution. Taking MobileNet as the object, if the calculated hierarchical redundancy , channel redundancy , convolution kernel redundancy , the comprehensive redundancy calculated by formula (12) is 0.608, which indicates that MobileNet reduces the redundancy by using depth separable convolution, and compared with ResNet, the lightweight potential of ResNet is greater.

[0050] Step S3, dynamically lightening the neural network according to the comprehensive redundancy fusion model. Specifically, the input neural network is diagnosed in multiple dimensions according to the comprehensive redundancy fusion model, the redundancy of the neural network model is analyzed, the redundant channel resources are released accurately, and the key layer is enhanced, so as to lighten the neural network.

[0051] Specifically, the input neural network is diagnosed in multiple dimensions according to the comprehensive redundancy fusion model, and the lightening is realized by the following operations.

[0052] S31, redundancy diagnosis: synchronously calculating the hierarchical redundancy , channel redundancy , and convolution kernel redundancy of the neural network.

[0053] S32, dynamic lightening operation: S321, hierarchical pruning: when the of a certain layer exceeds the preset threshold , the redundant layer is removed.

[0054] S322, channel selection: retaining the channels with the top a% importance scores in each layer.

[0055] S323, convolution kernel replacement: when the of a certain convolution kernel exceeds the preset threshold At this time, it is replaced by a depth separable convolution.

[0056] S33, precision recovery: after completing the lightweight operation, the model performance is recovered by using a fine-tuning mechanism.

[0057] S331, retraining the compressed model using the original training data; S332, setting the learning rate to 10% of the initial training value (typical value 0.001); S333, when the compression rate exceeds 50%, the fine-tuning rounds are not less than 20 rounds; S334, in a high compression scenario, it can be extended to knowledge distillation, and the original model is used as a teacher model to guide the lightweight model training.

[0058] The three redundancies correspond to three different components of the neural network, covering the whole stack redundancy from the macro structure (level) to the medium data flow (channel) to the micro computing unit (convolution kernel). The lightweight model (such as MobileNet (Mobile Neural Network)) often causes precision bottleneck due to excessive compression, but it is difficult to determine whether it is caused by insufficient depth, too few channels or low efficiency of convolution kernel. The embodiment of the present application can perform multi-dimensional diagnosis on the neural network according to the integrated redundancy fusion model: Level diagnosis: find that the shallow redundancy is low (need to retain feature extraction ability) → prohibit layer pruning, increase Skip (jump) connection; Channel diagnosis: detect that the SE (Squeeze-and-Excitation, a channel attention module) module output channel redundancy is high → reduce the number of SE channels to avoid excessive calculation; Convolution kernel diagnosis: find that the Depthwise (depth separable) convolution kernel redundancy is low (already efficient) → retain the structure and focus on optimizing the Pointwise (pointwise) convolution kernel quantization.

[0059] Step S4, the lightweight neural network model is annotated with a satellite to perform on-orbit AI inference tasks in a satellite-borne scenario. The embodiment of the present application can realize accurate release of redundant channel resources and enhancement of key layers, accurately and effectively guide the compression position, have little effect on precision, so as to realize lightweight of the neural network model.

[0060] Compared with the prior art, the embodiment of the present application has the following advantages: Most of the existing researches focus on the innovation of lightweight means and ignore the mechanism behind lightweight. The existing lightweight research is an empirical driven paradigm, which often relies on the intuition of engineers, a large number of experiments (such as trying different pruning rates, quantization bit widths, searching different lightweight architectures), experience of benchmark models (such as MobileNet series) and trial-and-error cost of neural architecture search. The process is usually: try compression → evaluate effect → adjust → try again, which is low in efficiency and difficult to generalize. In the embodiment of the present application, the above process is optimized as: first, quantitatively evaluate redundancy → according to the redundancy, predict compression potential and guide compression strategy → efficiently perform compression → verify the result. The redundancy index provides prior knowledge and guidelines, significantly reducing blindness.

[0061] The embodiment of the present application proposes a method for quantitatively analyzing network redundancy, which builds multi-layer redundancy quantization indicators (hierarchical feature similarity, channel importance distribution, convolution kernel response intensity), establishes a comprehensive redundancy fusion model, realizes multi-dimensional analysis and global evaluation of neural network redundancy, provides quantitative evaluation benchmark and compression boundary prediction for lightweight, promotes the paradigm upgrade of pruning strategy from empirical optimization to quantitative driving; can explain why the model can be lightweight, how much part can be lightweight, can provide method support for lightweight key technologies, and promote the paradigm change of lightweight design from empirical driving to method driving. In the implementation process, the embodiment of the present application calculates the redundancy degree of the network from three parts, considers the network in different levels, and can more comprehensively mine the redundancy of the network. Considering that there is a containing and complementary relationship between different levels, the embodiment of the present application proposes a comprehensive redundancy calculation formula, which better solves this problem. The embodiment of the present application proposes a method for quantitatively analyzing network redundancy, which answers the question of "how much redundancy does the network have", and can be used as a reliable reference in the lightweight process, significantly enhancing the explainability and controllability of the model optimization process; the embodiment of the present application calculates the redundancy degree of the network from three different levels, and scientifically fuses them to form a comprehensive redundancy, which is the overall representation of network redundancy.

[0062] The application scenarios of the embodiment of the present application include: ①Automated model selection and lightweight deployment pipeline: In the deployment scenario of edge devices with limited computing power (such as IoT terminals, embedded systems), the system receives multiple candidate neural network models. The system automatically performs multi-dimensional redundancy profiling on each candidate model according to the embodiments of the present application, and calculates the comprehensive redundancy score. The system automatically identifies the model with the highest comprehensive redundancy score, and then triggers the automatic lightweight process (such as applying a preset pruning, quantization algorithm) for the model. The lightweight model after the lightweight process is automatically deployed on the target edge device for running. This process is completely automated, avoiding the resource consumption of deploying multiple candidate models for real measurement on real devices, directly outputting a refined and executable model for deployment, significantly shortening the selection and deployment cycle, and reducing the storage requirements and computing load of edge devices.

[0063] ②Automated 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 the embodiments of the present application, generating specific redundancy quantization results (such as identifying specific layers, channels or convolution kernels with redundancy exceeding a preset threshold). Based on this quantization result, the engine automatically matches and performs the most suitable compression operation: automatically performing layer pruning or module replacement on high-redundancy layers; automatically performing channel pruning algorithm on high-redundancy channels; automatically performing kernel pruning, low-rank decomposition or replacing with depth separable convolution operator on high-redundancy convolution kernels. The selection and execution of the entire compression strategy are automatically completed by the engine according to the profiling results, without relying on expert experience configuration or time-consuming NAS search. Finally, a significantly reduced volume and significantly reduced computational load of the refined model is output, directly improving the inference speed and energy efficiency ratio of the model on the target hardware.

[0064] ③Spaceborne model lightweight system based on dynamic redundancy optimization according to the embodiments of the present application: To meet the demand of satellites for real-time execution of AI tasks such as earth observation target identification and space event detection on orbit, model lightweight compression needs to be performed on the ground, and then the AI task is executed on the satellite. At the ground station, the system receives the target spaceborne AI model. First, the system automatically performs multi-dimensional redundancy profiling (layer, channel, convolution kernel) on the model, and accurately quantizes the redundancy degree of each dimension. Based on the profiling results, the system automatically generates and executes the optimal compression strategy. The compression process strictly constrains the model accuracy loss within a preset threshold (such as ≤5%), ensuring task reliability. The refined model after lightweight compression has a significantly reduced volume and a significantly reduced computational load. The refined model is automatically packaged and uploaded to the satellite through the narrow space-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 scheme directly overcomes the space-ground transmission bottleneck, shortens the model update window, reduces the computing power and memory occupancy on the satellite, and improves the long-term autonomous operation capability and real-time response efficiency of the satellite in complex space environments, which is a key technology to support future intelligent satellite constellations.

[0065] Those skilled in the art will appreciate that embodiments of the present application can be readily used as a method, a system or a computer program product. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) embodying computer readable program code.

[0066] The present application is described in reference to the flowchart illustrations and / or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the application. 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 processing device 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, create means for implementing the functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 means for carrying out each of the

[0067] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 means for carrying out each of the

[0068] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 means for carrying out each of the

[0069] The above further describes the present application in conjunction with specific / preferred embodiments, and cannot be deemed to limit the specific implementation of the present application to these descriptions. For those skilled in the art to which the present application belongs, without departing from the concept of the present application, they can make several substitutions or variations to the described embodiments, and these substitutions or variations shall be deemed to belong to the protection scope of the present application. In the description of the present specification, the description of the terms "an embodiment", "some embodiments", "a preferred embodiment", "an example", "a specific example" or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are contained in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In the case of no mutual contradiction, those skilled in the art can combine and combine the different embodiments or examples described in the present specification and the features of the different embodiments or examples. Although the embodiments of the present application 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 satellite-borne processing neural network lightweight system, characterized in that, The system is installed with a computer program for performing a method comprising the following steps: S1, by analyzing the network structure of the satellite-borne neural network model, obtaining a redundancy quantification index for measuring the redundancy degree of the neural network model; S2, based on the mutual relationship between the redundancy quantification indexes, comprehensively considering the inclusiveness and complementarity of the redundant parts, constructing a comprehensive redundancy fusion model of the overall neural network; S3, according to the comprehensive redundancy fusion model, dynamically lightening the neural network model; S4, uploading the lightened neural network model to the satellite to perform on-orbit AI inference tasks in the satellite-borne scene.

2. The system of claim 1, wherein, The redundancy degree index includes hierarchical redundancy, channel redundancy, and convolution kernel redundancy, and step S1 specifically includes the following steps: S11, based on the similarity of the output feature maps of each level, obtaining the hierarchical redundancy; S12, based on the distribution sparsity of the channel weights of the fully connected layer, obtaining the channel redundancy; S13, based on the average gradient amplitude of the convolution kernel, obtaining the convolution kernel redundancy.

3. The system of claim 2, wherein, Step S11 specifically includes the following steps: S111, kernel matrix mapping is performed on the feature maps to eliminate size differences; 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 put into the hyperbolic tangent function for scaling to obtain the hierarchical redundancy.

4. The system of claim 2, wherein, Step S12 specifically includes 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, according to the weight calculated based on the parameter amount of the fully connected layer, the layer channel redundancy of all fully connected layers is weighted to obtain the channel redundancy.

5. The system of claim 2, wherein, Step S13 specifically includes 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.

6. The system of claim 3, wherein, The centered kernel alignment index in step S114 The expression for scaling by the hyperbolic tangent function tanh is: wherein, is the scaled centered kernel alignment index, and β is a scaling factor, is a translation parameter, and 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.

7. The system of claim 6, wherein, The hierarchical redundancy The expression is: wherein, is the scaled centralized kernel alignment indicator, i, n are constants.

8. The system of claim 4, wherein, Channel redundancy of step S123 The expression of the channel redundancy is: Wherein, is the weight, is the layer channel redundancy, and L is the total number of fully connected layers; Convolution kernel redundancy of step S132 The expression of the convolution kernel redundancy is: Wherein, N is the total number of convolution layers, is the layer convolution kernel redundancy.

9. The system of claim 1, wherein, The integrated redundancy fusion model in step S2 The expression is: .

10. The system of claim 1, wherein, Step S3 specifically includes: according to the comprehensive redundancy fusion model, multi-dimensional diagnosis is performed on the input neural network, the redundancy degree of the neural network model is analyzed, the redundant channel resources are released accurately, and the key layers are enhanced to lighten the neural network.

Images