45 results about "Tucker decomposition" patented technology

The invention provides a multi-mode ultrasonic fusion pressure vessel weld defect nondestructive testing method and system, and relates to the technical field of nondestructive testing. According to the method, geometric parameters of a welding seam are obtained through three-dimensional laser scanning, and an optimal scanning parameter set is generated; driving ultrasonic phased array equipment to scan for one time and synchronously acquire shear wave full-matrix capture and longitudinal wave linear scanning data; performing energy flow angular spectrum analysis and envelope analysis on the bimodal data, extracting defect feature parameters and constructing a three-dimensional feature tensor; carrying out multi-dimensional feature fusion by adopting Tucker decomposition, and enhancing a core tensor through physical modeling; generating three types of defect indication diagrams including a defect existence possibility diagram, a defect relative scale diagram and a defect space orientation diagram from the enhanced feature tensor; and the three types of indication diagrams are visually presented for comprehensive interpretation of detection personnel. Through multi-modal data fusion and physical modeling enhancement, the defect identification accuracy and detection efficiency are remarkably improved, the false alarm rate is reduced, and reliable technical support is provided for pressure vessel welding seam safety detection.

The invention discloses an energy consumption prediction monitoring system based on BIM, and relates to the technical field of intelligent building energy management, which comprises the steps of defining an optimization objective function based on modal weight, performing iterative updating by using an OIALM algorithm, obtaining a low-rank matrix, outputting initial prediction energy consumption through a prediction model, and optimizing model parameters by using a golden vehicle optimization algorithm. Generating final predicted energy consumption; through a three-mode alignment and sparse suppression tensor modeling method based on BIM and Internet of Things data, building physical semantics and dynamic environment features are effectively fused, the accuracy and robustness of energy consumption feature expression are improved, a prediction framework fusing key time node rearrangement, attention enhancement Tucker decomposition and OIALM low-rank optimization is constructed, and the prediction efficiency is improved. And the modeling capability and the energy consumption prediction precision of the unsteady-state thermal process are improved.

The invention provides a multi-parameter water quality data fusion analysis method and system. The method comprises the steps that water quality parameters are collected to form a three-dimensional data cube; constructing a space-time tensor model by using Tucker decomposition and a graph convolution network, and generating a core tensor matrix; constructing a dynamic constraint library and embedding the generative adversarial network; training a generative adversarial network by using Transform and physical constraint loss, generating synthetic data and verifying the synthetic data; performing space-time fusion by using meta learning weight distribution and Bayesian deep learning to generate a weight matrix and a confidence interval; missing data are restored through physical constraint interpolation and Gaussian process regression, and SHAP and LIME interpretation and path diagrams are generated; and performing real-time analysis by using an edge-cloud collaborative architecture to generate an intelligent report. Through physical constraint modeling, dynamic weight distribution and edge-cloud collaborative architecture, the problems that synthetic data violates physical laws, weight staticization, response lag and insufficient interpretability are solved.

The invention discloses a road network traffic state prediction method based on large language model spatial-temporal feature embedding, and belongs to the technical field of calculation, reckoning or counting. According to the method, semantic embedding of space and time is extracted by pre-training a large language model, and the space-time heterogeneity of a traffic system is quantified; a dynamic adjacent matrix is constructed by using Tucker decomposition, and combined modeling is performed on historical traffic data in combination with a dynamic space-time diagram convolutional network, so that high-precision prediction of a future traffic state is realized. The method does not need manual feature design, has good generalization ability and multi-mode adaptability, can be widely applied to scenes of road traffic, rail traffic demand prediction and the like, and significantly improves prediction accuracy and control response ability in a complex traffic environment.

The invention discloses a multi-modal denoising method for a hyperspectral image, which is characterized by comprising the following steps of: S1, respectively performing low-rank tensor representation on the hyperspectral image and a registered multispectral image by utilizing Tucker decomposition, and extracting core tensors of the hyperspectral image and the registered multispectral image; s2, establishing a correlation model between the hyperspectral core tensor and the multispectral core tensor through model-driven linear mapping or data-driven multi-layer perceptron network; s3, iteratively solving a core tensor, a factor matrix and correlation model parameters by adopting an alternating direction multiplier method; and S4, reconstructing a denoised hyperspectral image by using the optimized hyperspectral core tensor and factor matrix. Compared with the prior art, the method has the advantages that the spectral details of the hyperspectral image and the high-signal-to-noise-ratio spatial information of the multispectral image are fully mined and utilized, the restoration precision in the mixed noise scene is effectively improved through the double-Tucker decomposition framework and the core tensor association strategy, and the high-quality hyperspectral image is obtained.

The application relates to an abnormality detection method based on a hyperspectral image. The main body is based on a space-spectrum combined feature extraction method of filtering and low-rank decomposition to perform abnormality detection on the hyperspectral image. The specific method comprises the following steps: firstly, in the spatial dimension, a reduced dimension image is obtained through a data dimension reduction and eigenvalue weighted fusion method, and then an improved spatial filtering method is used to extract the spatial features of the image to obtain an initial spatial feature image. In the spectral dimension, a background reconstruction image of the approximate background is obtained by using a Tucker decomposition method on the original hyperspectral image, and a background dictionary of the image is obtained by using an improved k-means clustering method, then the background dictionary is input into a low-rank decomposition model to obtain a sparse matrix, and an initial spectral feature image is obtained, finally, the initial spectral feature image is fused with the spatial feature image to realize abnormality detection.

The invention discloses a production line equipment comprehensive efficiency monitoring method and system based on machine learning, and relates to the technical field of intelligent manufacturing, and the method comprises the steps: carrying out the space-time alignment and heterogeneous feature fusion of a production equipment coupling data set, and obtaining a synchronous data cube; extracting a low-rank nuclear tensor of the synchronous data cube by using a Tucker decomposition algorithm, and performing multi-modal projection to form an equipment fusion feature matrix; and inputting the equipment fusion feature matrix into a causal inference model, performing potential semantic projection and subspace division by the feature decoupling layer, performing contribution degree quantification and root cause influence analysis by the causal inference layer, and outputting equipment root cause association features. Through the causal inference model and the Pareto optimization algorithm, the accuracy and interpretability of the production line efficiency anomaly diagnosis are improved, and the relevance between the scheduling decision and the underlying causal mechanism is enhanced.

The invention discloses a hyperspectral image unsupervised change detection method based on a tensor convolution structure, and belongs to the technical field of remote sensing image intelligent analysis, and the method comprises the following steps: S1, constructing a training tensor for a dual-temporal hyperspectral image; s2, performing Tucker decomposition on the training tensor, and determining a spectrum rank of a spectrum mode factor matrix; s3, determining a layer-by-layer stacking structure according to the spectrum rank of the spectrum mode factor matrix, and outputting a dual-time-phase depth feature map; s4, obtaining an independent spectral difference chart and a local space average chart according to the double-time-phase depth feature map; and S5, obtaining a final change chart according to the obtained independent spectrum difference chart and the local space average chart. The double-view change detector constructed by the invention can effectively suppress noise interference, and has good resistance to illumination change, seasonal fluctuation, system noise and the like.

The invention discloses an analysis scheduling method and system based on multi-label routing. The method comprises the following steps: receiving multi-label hypergraph data, performing tuck decomposition on a hypergraph adjacency tensor, applying sparse constraint, extracting a core multi-label dependency structure, and generating a low-rank factor matrix; extracting hyperedge embedding vectors according to the low-rank factor matrix, compressing the hyperedge embedding vectors into binary hash codes through a locality sensitive hash module, and splicing the binary hash codes with the residual floating point features to form mixed hyperedge embedding vectors; for nodes in the graph neural network, attention weights are calculated through a multi-label compatibility function based on mixed embedding vectors, neighborhood aggregation is carried out, node states are updated through a gating circulation unit, and node embedding representation is obtained; and performing routing scheme generation on the node embedded representation through a path search algorithm, evaluating different schemes through a path scoring function, and outputting an optimal scheduling route according to a scoring result. According to the method, the calculation and message transmission efficiency of the large-scale multi-label graph is improved.

The invention discloses a charging station-oriented intelligent drainage cabinet advertisement effect analysis method and system, and relates to the technical field of digital outdoor advertisement intelligent analysis, and the method comprises the steps: carrying out the subfield division of an advertisement people stream response field, calculating the space center and scale difference of a subfield, constructing an affine mapping matrix, and generating a standard response field through iterative optimization. Constructing a causal graph, calculating an average processing effect, generating a comprehensive response index, and setting an advertisement putting strategy; through fusion of normalized multi-modal data, the accuracy of advertisement crowd response analysis is improved, through Tucker decomposition and residual enhancement, stable extraction of spatio-temporal characteristics of an advertisement effect is realized, and in combination with causal graph construction and return on investment calculation, the decision-making precision of an advertisement putting strategy is improved.

The invention discloses a heat map sequence downsampling reconstruction method based on tensor space constraint, and relates to the field of photo-thermal science and detection and signal processing. The invention aims to solve the problem that an active infrared thermal wave imaging detection technology cannot meet the requirements of high temporal-spatial resolution, low data volume transmission and real-time monitoring in engineering application. According to the method, the core tensor of the infrared image sequence is obtained through Tucker decomposition and is updated until the convergence condition is met, and the infrared image sequence is reconstructed through the updated core tensor; the updating process comprises the following steps of: performing time mode expansion on a core tensor of an infrared image sequence, and obtaining a spatial reconstruction field under each time mode; applying thermal diffusion physical constraint on the spatial reconstruction field by adopting a lattice Boltzmann method to obtain a physical evolution field; and projecting the difference between the physical evolution field and the corresponding space reconstruction field to the corresponding position of the core tensor, and updating the core tensor slice through Adam optimization so as to update the core tensor.

The application discloses a drug synergistic effect prediction method, system, device and medium based on heterogeneous graph tensor decomposition, and the prediction method comprises the following steps: obtaining a SMILES sequence of a drug, extracting a molecular structure feature of the drug, and obtaining a molecular structure feature representation of the drug; constructing a drug pair heterogeneous graph in each cell line, and obtaining a local interaction feature of the drug through a heterogeneous graph conversion network; performing Tucker decomposition on a three-channel heterogeneous graph relationship tensor, splicing the decomposition result with the local interaction feature of the drug, and extracting a global interaction feature vector of the drug; and predicting a synergistic score of a current drug-drug combination in a cell line according to the molecular structure feature representation of the drug and the global interaction feature vector of the drug. The application integrates the SMILES sequence of the drug and gene expression information of the drug pair in the cell line, and also converts the SMILES sequence into a drug molecular structure graph by using a graph convolution network, so as to extract the molecular structure feature of the drug. Such a design enables TensoGraph to more accurately reflect the complexity and diversity of the drug interaction network in the real world, thereby improving the prediction performance.

Embodiments of the present application provide a parallel computing method based on tensor Tucker decomposition result and related equipment, without reconstructing the original tensor from the Tucker decomposition result for operation, and the parallel computing mode used in operation greatly improves the computing efficiency of the Tucker decomposition result. The method comprises: performing tensor Tucker decomposition on the target tensor data which has been constructed to obtain the core tensor and the factor matrix corresponding to the target tensor data; determining the operation rule corresponding to the target tensor data; determining the parallel computing mode of the target tensor data according to the operation rule, wherein the parallel computing mode comprises inter-core parallel computing mode and intra-core parallel computing mode; and performing parallel computing on the core tensor and the factor matrix based on the parallel computing mode and the operation rule to obtain the computing result.

The application discloses a direct-current system grounding fault safety early warning method based on current characteristics and belongs to the technical field of power system fault diagnosis. The method comprises the following steps: firstly, injecting a safe micro-current disturbance signal into a direct-current system and synchronously collecting bus voltage and branch current response; then, solving a branch coupling admittance matrix at multiple frequency points, performing tensor Tucker decomposition and principal component analysis on the branch coupling admittance matrix, and extracting a first principal component direction vector; next, calculating the variance contribution rate of the vector, combining a dynamic reference and a fault probability accumulator model, and generating a system insulation state index; finally, using a dynamic decision threshold and a confidence positioning mechanism to realize accurate determination and positioning of a grounding fault; the method solves the problems of early warning, anti-interference and complex fault positioning of traditional methods, and improves the safety and reliability of direct-current system operation.

The invention provides an irradiation hazardous chemical trace multi-mode joint inspection method and system, and relates to the technical field of intelligent monitoring of hazardous chemical trace detection.The method comprises the steps that three paths of original response signals of electrochemical impedance, surface enhanced Raman and plasma-assisted ionization mass spectrometry arranged in an irradiation environment are obtained, and a multi-mode characteristic matrix is obtained; performing Tucker decomposition to obtain a core tensor and a factor matrix, and performing interpretability screening and vectorization to form a low-dimensional fusion feature vector; constructing an attribute graph by taking the low-dimensional fusion feature vector as mutual information, and outputting a trace hazardous chemical category label and a concentration estimated value through three-layer graph attention network embedding; substituting the category label and the concentration estimated value into the ERPG parameter to generate a digital twinning risk map; and driving ventilation, isolation and spraying actuators to complete closed-loop response through a federal learning consensus risk map. The invention provides an efficient and reliable technical scheme for hazardous chemical trace detection and safety management and control in the irradiation environment.

The invention discloses a method for analyzing an influence value of operating parameters of a coal-fired unit on power supply coal consumption. The method comprises the following steps: acquiring operation parameter data and a power supply coal consumption value from a coal-fired unit DCS (Distributed Control System), and preprocessing and discretizing; a mutual information matrix M belongs to the set is constructed, and conditional mutual information I (X; x [delta] C) quantizes the coupling relationship between the parameters; establishing a three-order tensor T which belongs to a characteristic'parameter-parameter-coal consumption 'nonlinear coupling relation; carrying out Tucker decomposition on the tensor; and calculating an independent influence item and a coupling influence item of the parameter based on a decomposition result to realize quantitative separation of the coupling effect. The method further comprises a dynamic updating mechanism, incremental updating is carried out every 8 hours, and working condition changes are rapidly adapted through combination of CP decomposition and self-adaptive weight adjustment. According to the method, the defect that the parameter coupling effect is ignored in a traditional method is overcome, the coupling effect recognition accuracy is improved by 35-40%, the power supply coal consumption prediction error is reduced to + / -0.8 g / (kW.h), and a reliable decision basis is provided for optimized operation of a unit.

The application discloses a high-voltage cable grounding loop resistance prediction method based on data driving, collects historical grounding loop resistance detection records of high-voltage cable sections with different static attribute labels under the same voltage level; adopts iterative threshold variational mode decomposition (ITVMD) to decouple irregular interval multi-loop resistance time series into a plurality of intrinsic mode components with different frequency characteristics; splices frequency domain characteristics of loop resistance mode components with static attributes to construct a comprehensive feature vector, and clusters the comprehensive feature vector through a K-means algorithm; combines a two-way delay embedding transformation (TDT) algorithm, and based on Tucker decomposition, adopts a block Hankel tensor autoregressive integrated moving average (BHT-ARIMA) multi-sequence prediction model to predict the high-voltage cable grounding loop resistance. Compared with the prior art, the application effectively solves the technical problems of low prediction accuracy caused by sparse grounding loop resistance data, complex coupling and missing values, and realizes accurate prediction of the loop resistance.

This invention relates to the field of big data analysis and public safety monitoring technology, and discloses a method for identifying abnormal behavior in student communities based on big data. The method includes constructing a dynamic spatiotemporal hypergraph sequence from multi-source data slices, mapping it to a higher-order tensor and performing non-negative Tucker decomposition, determining semantic impedance values ​​based on core tensor matching degrees, calculating instantaneous energy dissipation values ​​by combining behavior intensity and semantic impedance, and generating a relative dissipation index; calculating global structural entropy based on node degree distribution, and generating a structural sensitivity index through counterfactual inference by virtually removing nodes to be detected; and weighted fusion of the above indicators to determine explicit violations or implicit structural anomalies. This invention, by introducing semantic impedance and structural entropy counterfactual inference, effectively distinguishes between high-intensity normal activities and violations, and can identify implicit topological risks, achieving refined hierarchical identification of abnormal behavior.

The present application relates to the technical field of audio signal processing, and particularly relates to an audio scene classification method and device, the present application uses an audio classification model of Tucker decomposition and tensor regression, the model first enhances and expands a data set for log-mel data; redundancy of convolution weight is reduced by using Tucker decomposition, so that data can be more effectively and intuitively trained and feature extracted; the amount of calculation and the calculation complexity are reduced by using spatial separable convolution, and the running speed of the network is accelerated; then, a tensor regression layer is used to replace a traditional fully connected layer, the multi-mode structure of the data itself is retained, and the parameter quantity is reduced. The problems of high redundancy and parameter quantity of the traditional convolutional neural network, not enough intuitive and effective data features and loss of multi-mode structure information are solved, so that the accuracy of audio scene classification is improved.

The embodiment of the application discloses a structural color design method and device based on a tensor completion algorithm, which comprises the following steps: obtaining spectral data and geometric data of a dielectric array constituting a structural color, and combining the spectral data and the geometric data into multi-dimensional tensor data, wherein the tensor data comprises a to-be-completed tensor and a complete tensor containing all known entries; applying an alternating least squares method to Tucker decomposition of the to-be-completed tensor according to the tensor data, so as to obtain a minimum rank tensor after tensor completion, which is the same as known entries in the to-be-completed tensor; and converting the minimum rank tensor into spectral data and geometric data to obtain corresponding structural color. Through the above method, the embodiment of the application can quickly and accurately design structural color, avoids complex electromagnetic theory, and improves the device design efficiency of a small number of features.

The application relates to the technical field of electric data processing, in particular to a data processing method for a specimen digital management platform; the method comprises the following steps: performing multi-scale Tucker decomposition on a three-order data tensor to obtain a low-rank component representing a tissue structure, and taking the low-rank component as a correction image; extracting a semantic vector of unstructured clinical text by using a preset bidirectional long short-term memory network; inputting a graph embedding vector, a topological feature vector and the semantic vector into a multi-modal fusion network, and calculating a unified multi-modal index code by using a hash function; and encapsulating an identifier of the correction image, the graph embedding vector, the semantic vector, the multi-modal index code and specimen metadata into a structured data entity. The application has the effect of facilitating accurate analysis of data.

The application discloses a kind of energy consumption prediction monitoring systems based on BIM, it is related to intelligent building energy management technical field, including based on modal weight, definition optimization objective function, using OIALM algorithm is iterated and updated, obtain low rank matrix, through prediction model output initial predicted energy consumption, using golden jackal optimization algorithm is optimized to model parameter, generate final predicted energy consumption;Through three modal alignment and sparse inhibition tensor modeling method based on BIM and internet of things data, effectively fuse building physical semantics and dynamic environment characteristics, improve the accuracy and robustness of energy consumption feature expression, construct the prediction framework of fusion key time node rearrangement, attention-enhanced Tucker decomposition and OIALM low rank optimization, improve the modeling ability and energy consumption prediction accuracy to non-steady thermal process.

The invention discloses an open vocabulary object detection method based on tensor decomposition, belongs to the technical field of computer vision, and aims to solve the problems that an existing open vocabulary object detection method is not accurate enough in semantic and visual alignment and has redundancy in knowledge distillation. The method comprises the following steps: acquiring an input image and generating an initial candidate box; performing singular value decomposition on the candidate frame feature map through a low-rank candidate frame filtering module, and screening out a high-quality candidate frame by taking the sum of singular values as an integrity index; performing regional semantic alignment on the visual features of the high-quality candidate box and category text embedding, and calculating semantic alignment loss; through a core tensor knowledge distillation module, Tucker decomposition is carried out on feature tensors of the teacher and student models, and L1 distillation loss is calculated in a core tensor space; and combining the alignment loss and the distillation loss to train a model.

The invention provides an image feature detection method based on Tucker decomposition and a self-attention mechanism, and the method comprises the steps: collecting an aerial image, and making an aerial image data set; with the help of Tucker decomposition, initial parameters of a model convolution layer are determined; introducing a sparse self-attention module and initializing parameters of the self-attention module; constructing an image feature detection network model; training the image feature detection network model based on the aerial image data set; deploying the image feature detection network model to an airborne computer of the unmanned aerial vehicle; and the aerial image is processed by the image feature detection network model to obtain feature points and descriptors of the real-time aerial image in the flight process of the unmanned aerial vehicle. The channel dimension of the network model weight is reduced based on Tucker decomposition, a self-attention mechanism is introduced to enhance key feature point location information, and the problem that an image feature detection network model is poor in instantaneity and robustness when applied to small terminal equipment is solved.

The invention relates to a deep sea direct sound area target depth estimation method based on Tucker decomposition noise reduction and a storage medium, and the method comprises the following steps: 1, calculating the phase change in a specified frequency band from array data collected by a vertical array; and step 2, performing conventional beam forming on the array data, constructing noisy beam tensors in three dimensions of frequency-glancing angle-distance, and performing Tucker decomposition and reconstruction on the noisy beam tensors to obtain de-noised beam tensors. Most noise signals in the target signals can be effectively removed, and the method has great advantages when the problem of large extraction error of the sound field broadband interference structure under the condition of low signal-to-noise ratio is solved.

This invention provides a method and system for identifying electricity data. It acquires time-series electricity consumption data from a target enterprise over multiple consecutive analysis periods, transforms the data into a basic electricity consumption feature vector, constructs a three-dimensional tensor based on the historical period's basic electricity consumption feature vector, performs Tucker decomposition, forms an electricity consumption correlation network, and calculates the structural potential time series of each enterprise node. A nonlinear autoregressive model is used to fit the structural potential time series, and the absolute value of the predicted residual is defined as the node's evolutionary activity. During the identification phase, the feature vector of the current period is weighted according to the evolutionary activity, the electricity consumption correlation network is reconstructed, and the node structural potential is calculated. When the average evolutionary activity of the internal nodes of a certain enterprise cluster exceeds a first threshold, and the average structural potential of the cluster shows a monotonically increasing trend over the most recent N periods with a total increase exceeding a second threshold, the enterprise set corresponding to the cluster is identified as an advanced enterprise unit.

The invention provides a power data identification method and system, and the method comprises the steps: obtaining the time sequence power utilization data of a target enterprise in a plurality of continuous analysis periods, converting the obtained data into a basic power utilization feature vector, constructing a three-dimensional tensor based on the basic power utilization feature vector of a historical period, and carrying out the Tuck decomposition, the method comprises the following steps of: forming a power utilization association network, calculating a structure potential time sequence of each enterprise node, fitting the structure potential time sequence by adopting a nonlinear autoregression model, defining an absolute value of a prediction residual error as evolution activity of the node, weighting a current period feature vector according to the evolution activity in an identification stage, and identifying the current period feature vector. And reconstructing a power utilization association network and calculating node structure potential, and when the average evolution activity of internal nodes of a certain enterprise cluster exceeds a first threshold value, the average structure potential of the cluster is monotonically increased in the latest N periods, and the total amplification exceeds a second threshold value, identifying the enterprise set corresponding to the cluster as an advanced enterprise unit.

The invention discloses a traffic data interpolation method based on third-order Tucker decomposition, and the method comprises the steps: firstly constructing a three-dimensional traffic flow tensor according to the features of original traffic spatio-temporal data, so as to comprehensively represent the distribution conditions of traffic flows in space and time dimensions; and then, by constructing a third-order Tucker decomposition model, the third-order kernel tensor is decomposed into three third-order cubic tensors, so that the calculation complexity is effectively reduced, and the spatial-temporal characteristics in the traffic data are accurately extracted. On this basis, an optimization objective function based on third-order Tucker decomposition is established, an efficient solving algorithm is designed based on a least square method, and an approximate tensor is obtained to realize interpolation of a traffic flow missing value. According to the method, the accuracy of traffic flow data interpolation is improved on the whole, the interpolation performance is still excellent under the condition that the data missing rate is high, and the method has good application and popularization values.

The application discloses a hyperspectral image unsupervised change detection method based on a tensor convolution structure and belongs to the technical field of remote sensing image intelligent analysis, and comprises the following steps: S1, constructing a training tensor for double-time-phase hyperspectral images; S2, performing Tucker decomposition on the training tensor to determine the spectral rank of a spectral mode factor matrix; S3, determining a layer-by-layer stacking structure according to the spectral rank of the spectral mode factor matrix, and outputting double-time-phase deep feature maps; S4, obtaining independent spectral difference maps and local spatial average maps according to the double-time-phase deep feature maps; and S5, obtaining a final change map according to the independent spectral difference maps and the local spatial average maps. The double-view change detector constructed by the application can effectively inhibit noise interference and has good anti-interference ability to illumination changes, seasonal fluctuations and system noise.

The invention discloses a convolutional neural network collaborative optimization method and system for resource-constrained edge devices. The method comprises the steps that a Tucker decomposition method based on singular value decomposition and a hyper-parameter reconstruction rate is adopted to compress a CNN model needing to be optimized, and the compression rate and precision loss are flexibly controlled by adjusting the reconstruction rate; performing data quality labeling by using the compressed model, and training a lightweight data filtering module based on depth separable convolution; on the edge device, input data are screened through a filtering module, and only high-quality data enter the compressed trunk model for reasoning tasks. Through the collaborative optimization, the reasoning time delay, the calculation overhead and the storage occupation of the convolutional neural network on the edge equipment are reduced, and relatively high precision is kept.