Cloud edge collaborative data processing method and system based on tensor transformation

Abstract

The invention provides a cloud edge collaborative data processing method and system based on tensor transformation, which are applied to edge computing equipment, and the cloud edge collaborative data processing method comprises the following steps: carrying out tensor decomposition on original data to obtain original decomposition results, and respectively transmitting the original decomposition results to cloud computing center equipment, the cloud computing center equipment returns calculation model parameters according to the received original decomposition result; the original decomposition result comprises an original basic tensor; collecting local real-time data, reconstructing a basic tensor of the real-time calculation task according to the calculation model parameters and the local real-time data to obtain a first basic tensor adapted to the real-time calculation task, and executing the real-time calculation task according to the first basic tensor to obtain a local result; the cloud computing center device receives the data transmitted by the at least one edge computing device for global data processing; by adopting the method and the device, the real-time data interaction efficiency in the cloud edge collaborative scene can be improved.

Description

Technical Field

[0001] The present invention relates to the field of cloud-edge collaboration and tensor computing technology, and in particular to a cloud-edge collaborative data processing method and system based on tensor transformation. Background Art

[0002] With the advent of the big data era, the task of processing large-scale data and complex calculations has become increasingly important. Traditional computing methods often fail to meet the efficiency requirements of processing large-scale data. In order to solve this problem, cloud computing technology has developed rapidly. Cloud computing is mainly responsible for processing and managing remote large-scale data. However, in many application scenarios, due to the local requirements of data generation and computing, the processing efficiency of cloud computing is limited by network transmission. As a result, edge computing technology has emerged. Edge computing focuses on processing real-time data generated locally, and transfers computing tasks from the cloud to edge devices closer to users, thereby reducing network transmission delays and risks.

[0003] Although cloud computing can concentrate resources to process data transmitted remotely to the cloud via the Internet, providing powerful computing power and data storage capabilities to handle large-scale data and complex computing tasks, due to network transmission limitations, cloud computing may have certain delays in processing scenarios with high real-time requirements. Although edge computing can reduce network transmission delays and risks, due to the limited computing power and storage capacity of edge devices, edge computing may face certain challenges when processing large-scale data and complex computing tasks.

[0004] Therefore, in the cloud-edge collaborative computing environment, how to effectively combine the advantages of cloud computing and edge computing to achieve efficient and real-time data processing is an important research issue. Summary of the invention

[0005] The purpose of the present invention is to address the deficiencies of the above-mentioned prior art and to propose a cloud-edge collaborative data processing and system based on tensor transformation, which can improve the efficiency of real-time data interaction in cloud-edge collaborative scenarios.

[0006] In a first aspect, the present invention provides a cloud-edge collaborative data processing method based on tensor transformation, which is applied to an edge computing device, including:

[0007] Performing tensor decomposition on the original data to obtain original decomposition results, and transmitting the original decomposition results to the cloud computing center device respectively, so that the cloud computing center device returns the calculation model parameters according to the received original decomposition results; the original decomposition results include: original basic tensors;

[0008] Collect local real-time data, reconstruct the basic tensor of the real-time computing task according to the computing model parameters and the local real-time data, obtain a first basic tensor adapted to the real-time computing task, and execute the real-time computing task according to the first basic tensor to obtain a local result;

[0009] At least one of the local result or the first basic tensor is transmitted to the cloud computing center device, so that the cloud computing center device receives data transmitted by at least one edge computing device for global data processing.

[0010] The present invention adopts tensor decomposition at the edge computing device, and after secondary tensor decomposition at the cloud computing center device, a global joint decision is generated, thereby ensuring the transmission and low latency of the cloud computing center device through tensor decomposition; and tensor reconstruction is performed at the edge computing device, real-time tasks are calculated based on the reconstructed tensor, and the local results of the calculation are transmitted to the cloud computing center device. The cloud computing center device and the edge computing device jointly perform global data processing, thereby reducing the computing resource usage of the edge node, thereby satisfying the computing power of the edge node device to meet the calculation of local real-time tasks and ensuring the interaction efficiency of real-time data.

[0011] Further, the cloud computing center device receives data transmitted by at least one edge computing device to perform global data processing, including:

[0012] The cloud computing center device globally fine-tunes the computing model parameters based on the local results transmitted by at least one edge computing device and in combination with its own cloud results, and transmits the parameter fine-tuning results to the edge computing device, so that the edge computing device can calculate the real-time computing task for the local real-time data at the next moment based on the parameter fine-tuning results.

[0013] Further, the cloud computing center device receives data transmitted by at least one edge computing device for global data processing, and further includes: making a global first joint decision according to a first basic tensor transmitted by at least one edge computing device, including:

[0014] The cloud computing center device receives a first basic tensor sent by at least one edge computing node, and combines it with its own cloud data to generate a first joint decision based on the real-time data of the edge computing device and based on its own cloud data globally.

[0015] Furthermore, the original decomposition result also includes: an original transformation matrix; the second decomposition of the first basic tensor to obtain a final matrix includes:

[0016] After receiving the original basic tensor and the original transformation matrix, secondary decomposition is performed respectively to obtain multiple sub-tensor results and multiple sub-transformation matrices respectively, and the multiple sub-tensor results are sequentially multiplied with the multiple sub-transformation matrices at the matrix level to obtain multiple intermediate matrices, and the multiple intermediate matrices are merged to obtain the final matrix.

[0017] Further, the reconstructing the basic tensor of the real-time computing task according to the computing model parameters and the local real-time data to obtain a first basic tensor adapted to the real-time computing task includes:

[0018] The basic tensor is initially adjusted according to the computing model parameters to obtain a real-time basic tensor that adapts to the data distribution and task characteristics of the real-time computing task, and the real-time basic tensor is reconstructed according to local real-time data to obtain a reconstruction result, and the reconstruction result is post-processed according to the task requirements of the real-time computing task to obtain a first basic tensor; the post-processing includes: normalization or standardization.

[0019] The present invention can realize the merging and optimization of global results by reconstructing tensors on edge computing devices and feeding back the reconstruction results to cloud computing center devices, meeting the requirements of real-time performance and low latency. At the same time, tensor reconstruction is used to adapt to the data characteristics of real-time computing tasks, which can improve the accuracy and efficiency of real-time computing task execution, thereby improving the efficiency of cloud-edge collaborative data interaction.

[0020] Further, the real-time basic tensor is reconstructed according to the local real-time data to obtain a reconstruction result, including:

[0021] According to the local real-time data, at least one of linear transformation, nonlinear activation or pooling operation is performed on the real-time basic tensor to obtain a reconstruction result.

[0022] Furthermore, the cloud computing center device returns the calculation model parameters according to the received original decomposition result, including:

[0023] After receiving the original decomposition result, the cloud computing center device uses the original decomposition result as the input of the initial computing model, classifies or regresses the original decomposition result, outputs a predicted value, and updates the parameters of the computing model based on the predicted value and the true label of the predicted value to obtain trained computing model parameters, and transmits the computing model parameters to the edge computing device.

[0024] Further, transmitting the computing model parameters to the edge computing device includes:

[0025] The computing model parameters are compressed to obtain compressed computing model parameters, and the compressed computing model parameters are transmitted to the edge computing device through a pre-agreed transmission protocol, so that the edge computing device obtains the original computing model parameters after decompression.

[0026] The present invention can reduce the amount of data transmission in the cloud computing center by compressing the calculation model data, and can reduce data transmission interference by transmitting through a special transmission protocol, thereby improving the transmission efficiency of model parameters, reducing bandwidth consumption and transmission time. At the same time, this implementation also ensures the integrity and security of the model parameters, meeting the requirements of real-time performance and low latency.

[0027] Furthermore, the original decomposition result also includes: an original transformation matrix; in addition to returning the calculation model parameters according to the received original decomposition result, the cloud computing center device also includes:

[0028] After receiving the original basic tensor and the original transformation matrix, the cloud computing center device performs secondary decomposition respectively to obtain multiple sub-tensor results and multiple sub-transformation matrices respectively, and combines its own cloud data to generate a second joint decision based on the original data of the edge computing device and based on its own cloud data globally.

[0029] Furthermore, the combining of the own cloud data to generate a second joint decision based on the original data of the edge computing device and based on the overall cloud data of the own cloud includes:

[0030] The multiple sub-tensor results are sequentially multiplied with multiple sub-transformation matrices at the matrix level to obtain multiple intermediate matrices, and the multiple intermediate matrices are merged to obtain a final matrix. According to the final matrix and its own cloud data, a second joint decision based on the original data of the edge computing device and based on the global second joint decision of its own cloud data is generated.

[0031] In a second aspect, the present invention provides a cloud-edge collaborative data processing system based on tensor transformation, which is applied to edge computing devices and includes: a tensor decomposition unit, a tensor reconstruction unit and a global processing unit; wherein,

[0032] The tensor decomposition unit is used to perform tensor decomposition on the original data to obtain original decomposition results, and transmit the original decomposition results to the cloud computing center device respectively, so that the cloud computing center device returns the calculation model parameters according to the received original decomposition results; the original decomposition results include: original basic tensors;

[0033] Collect local real-time data, reconstruct the basic tensor of the real-time computing task according to the computing model parameters and the local real-time data, obtain a first basic tensor adapted to the real-time computing task, and execute the real-time computing task according to the first basic tensor to obtain a local result;

[0034] At least one of the local result or the first basic tensor is transmitted to the cloud computing center device, so that the cloud computing center device receives data transmitted by at least one edge computing device for global data processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a flowchart of the cloud-edge collaborative data processing method based on tensor transformation provided in this embodiment;

[0036] Figure 2 It is a flowchart of another cloud-edge collaborative data processing I based on tensor transformation provided by this embodiment;

[0037] Figure 3 It is a flowchart of another cloud-edge collaborative data processing II based on tensor transformation provided in this embodiment;

[0038] Figure 4 It is a flowchart of another cloud-edge collaborative data processing III based on tensor transformation provided in this embodiment;

[0039] Figure 5 It is a flowchart of another cloud-edge collaborative data processing IV based on tensor transformation provided in this embodiment;

[0040] Figure 6 This is a structural diagram of a cloud-edge collaborative data processing system based on tensor transformation provided in this embodiment. DETAILED DESCRIPTION

[0041] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0042] It is worth noting that the existing related technologies usually adopt the method of directly calculating the data itself in cloud-edge collaborative data processing, which makes the data processing volume of the cloud computing center device large, resulting in high transmission and computing delays, and the data processing volume at the edge computing node is small, but the computing power is limited. Based on this, the present application provides a cloud-edge collaborative data processing method and system based on tensor transformation, performs tensor decomposition at the edge computing device, and performs secondary tensor decomposition at the cloud computing center device for global decision-making, thereby ensuring the transmission and low latency of the cloud computing center device through tensor decomposition; and reconstructs tensors at the edge computing device, calculates real-time tasks based on the reconstructed tensors, and transmits the local results of the calculation to the cloud computing center device, the cloud computing center device jointly with the edge computing device updates the global parameters, and transmits the global parameters to the edge node device, and allows the edge computing node to calculate the local data through the updated parameters, so as to meet the computing power of the edge node device to meet the calculation of local real-time tasks.

[0043] In order to better demonstrate the technical solution of the present application, the following embodiments will be described in detail.

[0044] Example 1

[0045] See also Figure 1 , is a flow chart of the cloud-edge collaborative data processing method based on tensor transformation provided in this embodiment, including: steps S11 to S13, applied to edge computing devices, specifically:

[0046] Step S11, perform tensor decomposition on the original data to obtain original decomposition results, and transmit the original decomposition results to the cloud computing center equipment respectively, so that the cloud computing center equipment returns the calculation model parameters according to the received original decomposition results; the original decomposition results include: original basic tensors.

[0047] In some embodiments, the original data is subjected to tensor decomposition, and the obtained original decomposition result includes: an original basic tensor and an original decomposition matrix.

[0048] In some embodiments, tensor decomposition is performed after the original data is preprocessed; wherein the preprocessing of the original data includes: cleaning, denoising and standardizing the original data.

[0049] Next, the preprocessing of the raw data will be described in detail.

[0050] In some embodiments, the raw data is cleaned using the following formula:

[0051] X clean =X raw -mean(X raw )+1;

[0052] Among them, X clean is the cleaned data, X raw is the original data, mean(·) is the mean function. This step is used to eliminate the missing and errors of the original data.

[0053] In some embodiments, the cleaned data is denoised by the following formula:

[0054] When |X clean -mean(X clean )|<=std dev (X clean )*k, X denoised =mean(X clean );

[0055] When |X_clean-mean(X clean )|>std dev (X_clean)*k, X denoised =0;

[0056] Among them, X denoised is the denoised data, mean(X clean ) and std dev (X clean ) are the mean and standard deviation of the cleaned data, and k is the preset denoising threshold. This step is used to eliminate the noise of the cleaned data.

[0057] In some embodiments, the denoised data is normalized by the following formula:

[0058] X normalized =(X denoised -mean(X denoised )) / std dev (X denoised );

[0059] Among them, X normalized is the standardized data, mean(X denoised ) and std dev (X denoised ) are the mean and standard deviation of the denoised data.

[0060] It is worth noting that the preprocessing of the raw data is used to scale the data to a uniform standard for subsequent tensor transformation. By using the above formula for data preprocessing, the efficiency and quality of data processing can be effectively improved. This will provide reliable support for subsequent tensor transformation, data analysis and decision-making, and improve the accuracy and effectiveness of decision-making.

[0061] Next, the tensor decomposition of the original data will be explained in detail.

[0062] In some embodiments, tensor decomposition is performed on the original data using an improved tensor decomposition method to perform tensor decomposition on the preprocessed original data to obtain an original basic tensor and an original transformation matrix.

[0063] In some embodiments, the preprocessed raw data is subjected to tensor decomposition to obtain the original basic tensor and the original transformation matrix, and the tensor decomposition process can be expressed as:

[0064] X≈A×1U1×2U2× n U n ;

[0065] Among them, X is the original data after preprocessing, A is the original basic tensor, U1, U2, ..., U n is the original transformation matrix.

[0066] It is worth noting that the preprocessed raw data has multiple dimensions, and its original basis tensor and original transformation matrix also have multiple dimensions.

[0067] In some embodiments, the original basis tensor and the original transformation matrix are iterated multiple times to minimize the error between the original data and the fitting result processed by the original basis tensor and the original transformation matrix. When the error reaches a preset convergence condition, the iteration is stopped and the optimal original basis tensor and original transformation matrix are output.

[0068] In some embodiments, by iteratively updating A and U1, U2, ..., U n , so that X is equal to A×1U1×2U2× n U n The error between them is the smallest; when the preset convergence condition is reached, the iteration stops and A and U1, U2, ..., U are output. n .

[0069] In some embodiments, the iterative formulas of the original basis tensor and the original transformation matrix can be expressed as:

[0070]

[0071]

[0072] Among them, A (t) and Indicates that A and U at the tth iteration i The value of A (t+1) and Indicates that at the t+1th iteration, A and U i The new value of .

[0073] It is worth noting that by adopting the above-mentioned data preprocessing method of improved tensor decomposition, the intrinsic structure of the preprocessed data can be more effectively extracted. Compared with the traditional tensor decomposition method, this method can better process data with complex data structures and improve the accuracy of data analysis and decision-making.

[0074] Next, the acquisition of calculation model parameters will be described in detail.

[0075] In some embodiments, the cloud computing center device returns the computing model parameters based on the received original decomposition results, including: after receiving the original decomposition results, the cloud computing center device uses the original decomposition results as the input of the initial computing model, classifies or regresses the original decomposition results, outputs the predicted value, and updates the parameters of the computing model based on the predicted value and the true label of the predicted value, obtains the trained computing model parameters, and transmits the computing model parameters to the edge computing device.

[0076] In some embodiments, in a cloud computing center device, a deep learning-based computing model is trained using the transmitted original basic tensors and a large amount of training data. This computing model is used for decision-making, knowledge extraction, data cleaning or numerical prediction.

[0077] In some embodiments, in a cloud computing center device, a computing model based on deep learning is trained using the transmitted original basic tensors and a large amount of training data, including: Steps S101 to S107, specifically:

[0078] Step S101. Input the original basic tensor into the input layer of the deep learning-based computing model;

[0079] Step S101. Perform nonlinear transformation on the input layer through an activation function to obtain a feature vector;

[0080] Step S103. Input the feature vector into the hidden layer of the calculation model for classification or regression;

[0081] Step S104. Calculate the loss function based on the output of the hidden layer and the training data;

[0082] Step S105. Based on the loss function, the weights of the calculation model are updated using the back propagation algorithm;

[0083] Step S106. Repeat steps S102 to S105 until a preset number of training times is reached or the loss function reaches a preset threshold;

[0084] Step S107: Output the trained computing model.

[0085] In some embodiments, the activation function can be expressed as:

[0086] X′=f(W1A);

[0087] Where A is the base tensor, W1 is the weight matrix, and f(·) is the activation function (such as ReLU, Sigmoid, or Tanh). This formula represents the input tensor A, which is linearly transformed by W1A and then passed through the nonlinear activation function f to obtain the feature vector X′.

[0088] In some embodiments, if classification is performed, the classification function corresponding to the hidden layer can be expressed as:

[0089] Y = softmax(W2A);

[0090] Where W2 is another weight matrix used to map the obtained feature vector X′ to the classification space after the activation function. The softmax(·) function converts the output of W2A into a probability distribution Y, where each element Y i It represents the probability that the sample belongs to the i-th category.

[0091] In some embodiments, the specific form of the softmax(·) function is:

[0092]

[0093] Among them, X i ′ and X j ′ are the i-th and j-th elements of the eigenvector X′ respectively.

[0094] In some embodiments, the loss function adopts a cross entropy loss function. If the true label of the predicted value is one-hot encoded, the loss function can be expressed as:

[0095] L=-∑ i D i log(Y i );

[0096] Among them, here D i is the i-th element of the true label D (in one-hot encoding, only one element is 1 and the rest are 0), Y i is the i-th element of the probability distribution Y predicted by the computational model.

[0097] In some embodiments, the weights of the computational model are updated using a back propagation algorithm based on the loss function, including: calculating an error gradient based on the loss function, and updating the weights (ie, parameters) of the computational model through a back propagation algorithm.

[0098] In some embodiments, the error gradient δ is the partial derivative of the loss function L with respect to the weights W1 and W2. In the back propagation algorithm, this gradient needs to be calculated to update the weights. The specific form depends on the specific structure of the loss function L, W1 and W2, and usually needs to be calculated using the chain rule.

[0099] In some embodiments, since the loss function L is indirectly related to W1 and W2 through the softmax(·) function and the f(·) activation function, calculating δ requires derivation of these two functions.

[0100] In some embodiments, the weight update formula is the core of the back propagation algorithm. In standard mathematical expression, the update of weights W1 and W2 can be uniformly expressed as:

[0101]

[0102] Where W is W1 or W2; η is the learning rate, which controls the step size of weight update; is the gradient of the loss function L with respect to the weight W1 or W2.

[0103] It is worth noting that in fact W may represent multiple weight matrices (for example, the weight from the input layer to the hidden layer and the weight from the hidden layer to the output layer). The above formula is a unified abbreviation for each weight matrix, so the update process needs to be performed separately for each weight matrix.

[0104] By using basic tensors and a large amount of training data in cloud computing center devices to train a deep learning-based computing model, the performance and accuracy of the computing model can be effectively improved. At the same time, this method can also train large-scale data to meet the needs of practical applications.

[0105] In some embodiments, in a cloud computing center device, a computing model is trained using the transmitted basic tensors and a large amount of training data. The following is a detailed description of the process:

[0106] 1) Preliminary preparation

[0107] First, it is necessary to ensure that the cloud computing center equipment has built the corresponding hardware facilities, including high-performance computing nodes, large-scale storage systems, and data transmission networks. These facilities provide powerful computing power and data storage capabilities for the training of computing models.

[0108] At the same time, you also need to prepare the basic tensor A and a large amount of training data D. The basic tensor A is usually a set of pre-defined parameters or features, which will serve as the initial input of the initial calculation model. The training data D contains a large amount of sample data and its corresponding labels, which are used to train the initial calculation model.

[0109] 2) Model training process, including steps a to e, specifically:

[0110] Step a: Data input: The basic tensor A is transmitted to the cloud computing center device through the data transmission network and input into the input layer of the initial computing model M. At the same time, the training data D is also loaded into the storage system of the cloud computing center device for subsequent use.

[0111] Step b, nonlinear transformation. In the input layer, the activation function is used to perform nonlinear transformation on the basic tensor A. The selection of the activation function is crucial to the performance of the computational model. Commonly used activation functions include Sigmoid, ReLU, etc. Through nonlinear transformation, the feature vector X can be obtained, which contains the complex features and representation of the basic tensor A.

[0112] Step c, classification or regression. The feature vector X is input into the hidden layer of the computational model M. The hidden layer is usually composed of multiple neurons, which further process the feature vector X through connection weights and bias terms. According to the requirements of the specific task, an output layer can be added after the hidden layer for classification or regression operations. For classification tasks, the softmax function can be used to convert the output into a probability distribution; for regression tasks, continuous values ​​can be directly output.

[0113] Step d, calculate the loss function. During the training process, the loss function needs to be calculated based on the output of the hidden layer and the training data D. The loss function is used to measure the gap between the model prediction results and the actual labels. Commonly used loss functions include cross entropy loss, mean square error, etc. By calculating the loss function, the performance of the current model can be evaluated and a basis for subsequent weight updates can be provided.

[0114] Step d, weight update. Here, according to the loss function, the weight of the calculation model M is updated using the back propagation algorithm. The back propagation algorithm guides the adjustment direction of the weight by calculating the error gradient, so that the model gradually converges to the optimal solution during the training process. When updating the weight, it is also necessary to set the learning rate η, which determines the step size of the weight update. By continuously adjusting the weight and learning rate, the performance of the model can be gradually optimized.

[0115] Step f, iterative training. The above steps be are repeated until the preset number of training times is reached or the loss function reaches the preset threshold, and the optimal original basis tensor and original transformation matrix are output. During the iterative training process, the performance of the model will gradually improve until it converges to a stable state.

[0116] 3) Output the trained model

[0117] When the training is completed, the trained computing model M can be obtained. The model has been optimized and adjusted through a large amount of training data and has high performance and accuracy. The trained model can be exported and deployed to actual application scenarios to achieve various complex task requirements.

[0118] As for the specific formulas used, the above description has already involved the activation function, softmax function, loss function, and the related formulas of the back propagation algorithm. These formulas are the core components of the computational model training process, and together they constitute the model training and optimization process. In practical applications, appropriate formulas and parameter settings can be selected according to specific tasks and data characteristics to obtain the best model performance.

[0119] Next, the transmission process of the calculation model parameters will be described in detail.

[0120] In some embodiments, transmitting the computing model parameters to the edge computing device includes: compressing the computing model parameters to obtain compressed computing model parameters, and transmitting the compressed computing model parameters to the edge computing device through a pre-agreed transmission protocol, so that the edge computing device obtains the original computing model parameters after decompression.

[0121] In some embodiments, during implementation, the following formulas are used for compression and decompression respectively:

[0122] compress(parameters)=t1(parameters);

[0123] decompress(compressedparameters)=f2(compressedparameters);

[0124] Among them, parameters are the original model parameters, compressed parameters are the compressed model parameters; t1(·) and f2(·) are the compression function and decompression function respectively.

[0125] In some embodiments, the trained model parameters are compressed to obtain compressed model parameters; the compressed model parameters are transmitted to the edge computing device through an optimized transmission protocol; the edge computing device receives the compressed model parameters, decompresses them, and obtains the original model parameters; the original model parameters are deployed on the edge computing device for real-time prediction and decision-making.

[0126] In some embodiments, after the cloud computing center device fine-tunes the parameters of the computing model, the fine-tuned parameters are also compressed and transmitted to the edge computing device.

[0127] In some embodiments, the transmitted tensor data can also be compressed during the cloud-edge collaborative interaction process, including: the edge computing device compresses the first basic tensor and sends it to the cloud computing center device; the edge computing device compresses the original basic tensor and the original transformation matrix and sends them to the cloud computing center device.

[0128] It is worth noting that, by compressing and optimizing the transmission protocol, the transmission system and method of this embodiment can significantly improve the transmission efficiency of model parameters, reduce bandwidth consumption and transmission time. At the same time, this implementation also ensures the integrity and security of model parameters, meeting the requirements of real-time performance and low latency.

[0129] In some embodiments, in a specific scenario, after tensor decomposition, the edge computing device uploads both the original basic tensor and the original transformation matrix to the cloud computing center device. The cloud computing center device trains the initial computing model based on the original basic tensor to obtain a trained computing model, and transmits the computing model parameters to the edge computing device, so that the edge computing device performs real-time task calculations based on the computing model parameters, thereby improving the computing power of the edge computing device. When the edge computing device completes the calculation, it obtains the local result and uploads the local result to the cloud computing center device, which fine-tunes the parameters of the computing model based on the local result.

[0130] Next, the process of generating a global second joint decision based on the original data will be introduced in detail.

[0131] In some embodiments, in addition to returning the computing model parameters based on the received original decomposition results, the cloud computing center also includes: after the cloud computing center device receives the original decomposition results and the original transformation matrix, it performs secondary decomposition to obtain multiple sub-tensor results and multiple sub-transformation matrices respectively, and combines its own cloud data to generate a second joint decision based on the real-time data of the edge computing device and based on its own cloud data globally.

[0132] It is worth noting that the first joint decision is a joint decision generated by the cloud computing center device based on the basic tensor generated by real-time data and combined with its own cloud data; the second joint decision is a joint decision generated by the cloud computing center device based on the original basic tensor and original transformation matrix decomposed from the original data and combined with the original cloud data.

[0133] In some embodiments, the real-time data of the edge computing device and the second joint decision based on the global cloud data of the edge computing device are generated in combination with the cloud data of the edge computing device, including: performing matrix-level multiplication operations on multiple sub-tensor results and multiple sub-transformation matrices in sequence to obtain multiple intermediate matrices, and merging the multiple intermediate matrices to obtain a final matrix, and generating a second joint decision based on the real-time data of the edge computing device and the global cloud data of the edge computing device according to the final matrix and the cloud data of the edge computing device.

[0134] In some embodiments, after the cloud computing center device receives the original basic tensor and the original transformation matrix, performing secondary decomposition respectively includes: steps S301 to S306, specifically:

[0135] Step S301: Decompose the original basic tensor A to obtain m sub-basic tensors, which can be expressed as {A1, A2, ..., A j , ..., A m}; where m is the decomposition dimension of the original basis tensor and is a positive integer.

[0136] Step S302: for each transformation matrix U i Perform matrix decomposition to obtain n sub-transformation matrices, which can be expressed as {U i1 , U i2 , ..., U ij ..., U in}; where n is the decomposition dimension of the original transformation matrix and n is a positive integer.

[0137] Step S303: Each sub-basis tensor A j And the corresponding sub-transformation matrix U ij Multiply them together to get P intermediate matrices P ij ;

[0138] Step S304: Each intermediate matrix P ij Merge to get a final matrix F;

[0139] Step S305: Perform data analysis and decision making based on the final matrix F.

[0140] Step S306: During the implementation process, the following formula is used for matrix decomposition:

[0141] U i =U i1 +U i2 +…+U in ;

[0142] Among them, U i is the transformation matrix, U ij is the decomposed sub-transformation matrix.

[0143] By transmitting the results of tensor decomposition to cloud computing center equipment for further calculation, cloud computing resources can be effectively utilized to improve the efficiency and accuracy of data calculation. At the same time, large-scale data can also be processed to meet the needs of practical applications.

[0144] Next, the process of basic tensor reconstruction based on local real-time data will be described in detail.

[0145] It is worth noting that, in order to address the problem of low efficiency in large-scale data processing, especially under the limitation of network transmission delay, this embodiment provides a method for reconstructing basic tensors using edge computing devices to improve data processing efficiency. The following step S12 is a detailed description of the specific implementation method.

[0146] Step S12, collect local real-time data, reconstruct the basic tensor of the real-time computing task according to the computing model parameters and the local real-time data, obtain a first basic tensor adapted to the real-time computing task, and execute the real-time computing task according to the first basic tensor to obtain a local result.

[0147] It is worth noting that, first, during the data transmission and preparation process, the calculation model parameters are transmitted to the edge computing device through the network. These calculation model parameters may be part or all of the weights of the calculation model, which are used to perform computing tasks on the edge device. At the same time, the edge device collects local real-time data, which will be used together with the model parameters in the subsequent calculation process. In the process of reconstructing the basic tensor, the basic tensor is reconstructed using the received calculation model parameters and local real-time data. The basic tensor is usually a set of predefined data with a specific structure, which can be part of the input or intermediate result in the calculation model. The reconstruction process includes reorganizing, transforming or fusing the data to adapt to the current task requirements.

[0148] In some embodiments, the basic tensor of the real-time computing task is reconstructed according to the computing model parameters and local real-time data to obtain a first basic tensor adapted to the real-time computing task, including: initially adjusting the basic tensor according to the computing model parameters to obtain a real-time basic tensor adapted to the data distribution and task characteristics of the real-time computing task, and reconstructing the real-time basic tensor according to the local real-time data to obtain a reconstruction result, and post-processing the reconstruction result according to the task requirements of the real-time computing task to obtain a first basic tensor; the post-processing includes: normalization or standardization.

[0149] It is worth noting that the first basic tensor is the reconstructed basic tensor of the edge computing device in real-time computing tasks, and the original basic tensor is the basic tensor of the original data. The first basic tensor and the original basic tensor are both basic tensors, but the data sources and acquisition methods of their tensor decomposition are different.

[0150] In some embodiments, the real-time basic tensor is reconstructed according to the local real-time data to obtain a reconstruction result, including: performing at least one of linear transformation, nonlinear activation or pooling operation on the real-time basic tensor according to the local real-time data to obtain a reconstruction result.

[0151] In some embodiments, the real-time data includes: local real-time data of the edge computing device and cloud data of the cloud computing center device.

[0152] In some embodiments, the process of basic tensor reconstruction includes: steps S111 to S113, specifically:

[0153] Step S111: perform initial adjustment on the basic tensor according to the calculation model parameters to adapt it to the current data distribution and task characteristics, and obtain a real-time basic tensor.

[0154] Step S112: Combine the local real-time data and transform the adjusted real-time basic tensor to obtain a reconstruction result. This can be achieved through various mathematical operations and algorithms, such as linear transformation, nonlinear activation or pooling.

[0155] In some embodiments, the linear change of the real-time basis tensor can be expressed as:

[0156] X reconstructed =W·X base +b;

[0157] Among them, X reconstructed is the reconstructed tensor, W is the uniformly abbreviated weight matrix, which is obtained from the calculation model parameters and can be W1 or W2, X base is the base tensor and b is the bias.

[0158] In some embodiments, in order to introduce nonlinear characteristics, an activation function can be used in the reconstruction process.

[0159] For example, the ReLU (Rectified Linear Unit) activation function, the ReLU function can be expressed as:

[0160] X reconstructed =max(0, X linear );

[0161] Among them, X linear is the tensor after linear transformation.

[0162] In some embodiments, a pooling operation is performed on the reconstructed real-time basis tensor to reduce its size and extract important features. For example, the maximum pooling can be defined as:

[0163] X pooled =maxpool(X recinstructed );

[0164] Among them, maxpool(·) is the maximum pooling operation.

[0165] Step S113: Perform necessary post-processing on the reconstruction result according to task requirements, such as normalization and standardization, to obtain the first basic tensor to ensure its quality and accuracy.

[0166] In some embodiments, after reconstruction, a first basic tensor adapted to the current task is obtained. Then, using these reconstructed first basic tensors and the remaining computing model parameters, the computing task is performed on the edge computing device to generate local results. These results can be classification labels, regression values, or other forms of output, depending on the nature and requirements of the task.

[0167] In some embodiments, if there are no remaining computing model parameters, the computing task is performed directly on the edge computing device based on the first basic tensor.

[0168] By reconstructing tensors on edge computing devices and feeding back the reconstruction results to cloud computing center devices, global results can be merged and optimized to meet real-time and low-latency requirements. At the same time, this method uses formulas such as tensor decomposition, data fitting, and distributed computing to improve the accuracy and efficiency of reconstruction. In addition, this method can also be expanded and optimized according to different application scenarios.

[0169] Next, the global data processing will be described in detail.

[0170] Step S13: Transmit at least one of the local result or the first basic tensor to the cloud computing center device, so that the cloud computing center device receives data transmitted by at least one edge computing device for global data processing.

[0171] In some embodiments, the cloud computing center device receives data transmitted by at least one edge computing device for global data processing, including: the cloud computing center device globally fine-tunes the computing model parameters based on the local results transmitted by at least one edge computing device and in combination with its own cloud results, and transmits the parameter fine-tuning results to the edge computing device, so that the edge computing device calculates the real-time computing task for the local real-time data at the next moment based on the parameter fine-tuning results.

[0172] In some embodiments, the cloud computing center device receives data transmitted by at least one edge computing device for global data processing, and also includes: making a global first joint decision based on a first basic tensor transmitted by at least one edge computing device; wherein the cloud computing center device receives the first basic tensor sent by at least one edge computing node, and combines its own cloud data to generate a global first joint decision based on the real-time data of the edge computing device and based on its own cloud data.

[0173] In some embodiments, a cloud computing center device receives a first basic tensor sent by multiple edge computing nodes, decomposes the first basic tensor and the transformation matrix of the previous moment respectively, generates multiple sub-basic tensors and multiple transformation matrices, and performs matrix-level multiplication operations on the multiple sub-tensor results with multiple sub-transformation matrices in sequence to generate multiple intermediate matrices, merge the intermediate matrices to obtain the intermediate matrix, and combine the intermediate matrix with the matrix corresponding to its own cloud data to generate a first joint decision based on the real-time data of the edge computing device and based on the global cloud data of its own.

[0174] Next, a variety of different cloud-edge collaborative data processing scenarios based on tensor transformation will be given.

[0175] In some embodiments, see Figure 2 , is a flowchart of another cloud-edge collaborative data processing I based on tensor transformation provided by this embodiment. Figure 2 In the process, the edge computing device performs tensor decomposition on the original data to obtain the original basic tensor and the original transformation matrix, and transmits the original basic tensor and the original transformation matrix to the cloud computing center device; the cloud computing center device performs secondary decomposition on the original basic tensor and the original transformation matrix to obtain multiple sub-basic tensors and multiple sub-transformation matrices, and generates multiple intermediate matrices based on the multiple sub-basic tensors and multiple sub-transformation matrices, merges the multiple intermediate matrices to obtain the final matrix, and generates a global second joint decision based on the final matrix and the matrix corresponding to the original cloud data thereof, and transmits the global second joint decision to the edge computing device; the edge computing device executes the global second joint decision after receiving it.

[0176] exist Figure 2In the example, the data exchanged between the edge computing device and the cloud computing center device is in the form of tensor data, and after the edge computing device performs a tensor decomposition of the original data, the cloud computing center device performs a second decomposition to obtain the global second joint decision. The edge computing device performs tasks according to the second joint decision, which can reduce the computing power of the edge computing device for policy calculation and reduce the amount of data sent by the cloud computing center device to the edge computing device, thereby reducing the latency of cloud-edge collaboration and improving the efficiency of cloud-edge collaborative interaction. In addition, as a data center, the cloud computing center device can integrate the tensor data of multiple edge computing devices and send down the integrated joint strategy, which can improve the accuracy of controlling the edge computing device to perform tasks.

[0177] In some embodiments, see Figure 3 , is a flowchart of another cloud-edge collaborative data processing II based on tensor transformation provided in this embodiment, and Figure 3 The difference is that the edge computing device transmits the decomposed original basic tensor to the cloud computing center device. After receiving the original basic tensor, the cloud computing device trains the computing model, and compresses the computing model parameters of the trained computing model and transmits it to the edge computing device; the edge computing device reconstructs the basic tensor adapted to the real-time computing task according to the computing model parameters and local real-time data to obtain the first basic tensor, and executes the real-time computing task according to the first basic tensor to obtain the local result; the edge computing device compresses the local result and the first basic tensor, and transmits them to the cloud computing center device through a pre-determined transmission protocol; the cloud computing center device fine-tunes the parameters of the trained computing model according to the local data, combines it with its own cloud results, and compresses the fine-tuning parameters and feeds them back to the edge computing device, so that the edge computing device updates the parameters; and the cloud computing center device also performs in-depth data analysis based on the first basic tensor and the basic tensor corresponding to its own cloud data, including: generating a global joint decision based on multiple first basic tensors and its own cloud data.

[0178] exist Figure 3The edge computing device in the process reconstructs the basic tensor through the computing model parameters trained by the cloud computing center device, avoiding the computing power pressure brought by a large amount of computing model training locally, thereby releasing part of the computing resources of the edge computing device. It only needs to reconstruct the basic tensor according to the computing model parameters to perform real-time computing tasks, thereby improving the computing efficiency of the edge computing device and the cloud-edge collaborative interaction efficiency; and the edge computing device uploads the first basic tensor to the cloud computing center device. Compared with the existing related technology of directly transmitting the data itself, this embodiment can reduce the data transmission volume, thereby improving the cloud-edge collaborative interaction efficiency, and there is no need for the edge computing device to provide computing power for parameter fine-tuning, and by compressing and transmitting tensor data and parameters, the amount of cloud-edge collaborative transmission data can be reduced, thereby further reducing data transmission delay and improving data interaction efficiency.

[0179] In some embodiments, see Figure 4 , is a flow chart of another cloud-edge collaborative data processing III based on tensor transformation provided in this embodiment. In this embodiment, Figure 4 Combined with Figure 2 and Figure 3 The steps can not only perform global decision control based on the decomposed tensor, but also reduce the latency of cloud-edge collaborative data interaction and the computing pressure of edge computing devices.

[0180] In some embodiments, see Figure 5 , is a flow chart of another cloud-edge collaborative data processing IV based on tensor transformation provided in this embodiment, and Figure 4 The difference is that in Figure 5 In the process, after the edge computing device transmits the first basic tensor to the cloud computing center device, the cloud computing center device performs secondary decomposition based on the first basic tensor to obtain multiple sub-basic tensors, and combines multiple sub-transformation matrices of the previous moment to generate a final matrix, and generates a global first joint decision based on the final matrix combined with its own real-time cloud data; the edge computing node executes the task according to the global first joint decision; the cloud computing center device also fine-tunes the global parameters of the computing model based on the local results transmitted by the edge computing device and the real-time cloud results of its own cloud.

[0181] It is worth mentioning that Figures 2 to 5 Only one edge computing device is given. In fact, there can be multiple edge computing devices to interact with the cloud computing center device for data.

[0182] This embodiment adopts tensor decomposition at the edge computing device, and after secondary tensor decomposition at the cloud computing center device, a global joint decision is generated, so as to ensure the transmission and low latency of the cloud computing center device through tensor decomposition; and tensor reconstruction is performed at the edge computing device, and the real-time task is calculated based on the reconstructed tensor, and the local result of the calculation is transmitted to the cloud computing center device. The cloud computing center device and the edge computing device jointly perform global data processing, thereby reducing the computing resource usage of the edge node, thereby satisfying the computing power of the edge node device to meet the calculation of local real-time tasks and ensure the interaction efficiency of real-time data.

[0183] Example 2

[0184] See also Figure 6 , is a structural diagram of a cloud-edge collaborative data processing system based on tensor transformation provided in this embodiment, which is applied to an edge computing device 20. The cloud-edge collaborative data processing system includes: a tensor decomposition unit 21, a tensor reconstruction unit 22 and a result feedback unit 23.

[0185] It is worth noting that after decomposing the original data, the tensor decomposition unit 21 transmits the output original decomposition result to the cloud computing center device 24. The cloud computing center device 24 feeds back the calculation model parameters based on the original decomposition result. After receiving the calculation model parameters, the tensor decomposition unit 21 transmits the calculation model parameters to the tensor reconstruction unit 22; after receiving the calculation model parameters, the tensor reconstruction unit 22 outputs the local result of the real-time computing task, and transmits the local result and the first basic tensor to the result feedback unit 23; after receiving the local result and the first basic tensor, the result feedback unit 23 transmits the local result and the first basic tensor to the cloud computing center device 24; after receiving the local result and the first basic tensor, the cloud computing center device 24 outputs the fine-tuning parameters to the tensor reconstruction unit 22 according to the local result, so that the tensor reconstruction unit 22 adjusts the parameters, and outputs the global first joint decision to the tensor reconstruction unit 22 according to the first basic tensor, so that the tensor reconstruction unit 22 performs the calculation of the real-time computing task.

[0186] The tensor decomposition unit 21 is used for performing tensor decomposition on the original data to obtain original decomposition results, and transmit the original decomposition results to the cloud computing center device respectively, so that the cloud computing center device returns the calculation model parameters according to the received original decomposition results; the original decomposition results include: the original basic tensor.

[0187] In some embodiments, the cloud computing center device returns the computing model parameters based on the received original decomposition results, including: after receiving the original decomposition results, the cloud computing center device uses the original decomposition results as the input of the initial computing model, classifies or regresses the original decomposition results, outputs the predicted value, and updates the parameters of the computing model based on the predicted value and the true label of the predicted value, obtains the trained computing model parameters, and transmits the computing model parameters to the edge computing device.

[0188] In some embodiments, transmitting the computing model parameters to the edge computing device includes: compressing the computing model parameters to obtain compressed computing model parameters, and transmitting the compressed computing model parameters to the edge computing device through a pre-agreed transmission protocol, so that the edge computing device obtains the original computing model parameters after decompression.

[0189] In some embodiments, the original decomposition result also includes: the original transformation matrix; in addition to returning the calculation model parameters based on the received original decomposition result, the cloud computing center device also includes: after the cloud computing center device receives the original decomposition result and the original transformation matrix, it performs secondary decomposition on each of them, and obtains multiple sub-tensor results and multiple sub-transformation matrices respectively, and combines its own cloud data to generate a second joint decision based on the real-time data of the edge computing device and based on its own cloud data globally.

[0190] In some embodiments, the real-time data of the edge computing device and the second joint decision based on the global cloud data of the edge computing device are generated in combination with the cloud data of the edge computing device, including: performing matrix-level multiplication operations on multiple sub-tensor results and multiple sub-transformation matrices in sequence to obtain multiple intermediate matrices, and merging the multiple intermediate matrices to obtain a final matrix, and generating a second joint decision based on the real-time data of the edge computing device and the global cloud data of the edge computing device according to the final matrix and the cloud data of the edge computing device.

[0191] The tensor reconstruction unit 22 is used to collect local real-time data, reconstruct the basic tensor of the real-time computing task according to the computing model parameters and the local real-time data, obtain a first basic tensor adapted to the real-time computing task, and execute the real-time computing task according to the first basic tensor to obtain a local result.

[0192] In some embodiments, the basic tensor of the real-time computing task is reconstructed according to the computing model parameters and local real-time data to obtain a first basic tensor adapted to the real-time computing task, including: initially adjusting the basic tensor according to the computing model parameters to obtain a real-time basic tensor adapted to the data distribution and task characteristics of the real-time computing task, and reconstructing the real-time basic tensor according to the local real-time data to obtain a reconstruction result, and post-processing the reconstruction result according to the task requirements of the real-time computing task to obtain a first basic tensor; the post-processing includes: normalization or standardization.

[0193] In some embodiments, the real-time basic tensor is reconstructed according to the local real-time data to obtain a reconstruction result, including: performing at least one of linear transformation, nonlinear activation or pooling operation on the real-time basic tensor according to the local real-time data to obtain a reconstruction result.

[0194] The result feedback unit 23 is used to transmit at least one of the local result or the first basic tensor to the cloud computing center device, so that the cloud computing center device receives data transmitted by at least one edge computing device for global data processing.

[0195] In some embodiments, the cloud computing center device receives data transmitted by at least one edge computing device for global data processing, including: the cloud computing center device globally fine-tunes the computing model parameters based on the local results transmitted by at least one edge computing device and in combination with its own cloud results, and transmits the parameter fine-tuning results to the edge computing device, so that the edge computing device calculates the real-time computing task for the local real-time data at the next moment based on the parameter fine-tuning results.

[0196] In some embodiments, the cloud computing center device receives data transmitted by at least one edge computing device for global data processing, and also includes: making a global first joint decision based on a first basic tensor transmitted by at least one edge computing device, including: the cloud computing center device receives a first basic tensor sent by at least one edge computing node, and combines its own cloud data to generate a global first joint decision based on the real-time data of the edge computing device and based on its own cloud data.

[0197] This embodiment adopts tensor decomposition at the edge computing device, and after secondary tensor decomposition at the cloud computing center device, a global joint decision is generated, so as to ensure the transmission and low latency of the cloud computing center device through tensor decomposition; and tensor reconstruction is performed at the edge computing device, and the real-time task is calculated based on the reconstructed tensor, and the local result of the calculation is transmitted to the cloud computing center device. The cloud computing center device and the edge computing device jointly perform global data processing, thereby reducing the computing resource usage of the edge node, thereby satisfying the computing power of the edge node device to meet the calculation of local real-time tasks and ensure the interaction efficiency of real-time data.

[0198] Those skilled in the art will appreciate that the embodiments of the present application may also provide computer program products. Therefore, the present application may adopt the form of complete hardware embodiments, complete software embodiments, or embodiments in combination with software and hardware. Moreover, the present application may adopt the form of computer program products implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program codes.

[0199] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0200] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0201] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0202] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A cloud-edge collaborative data processing method based on tensor transformation, applied to edge computing devices, characterized in that: The cloud-edge collaborative data processing method includes: Performing tensor decomposition on the original data to obtain original decomposition results, and transmitting the original decomposition results to the cloud computing center device respectively, so that the cloud computing center device returns the calculation model parameters according to the received original decomposition results; the original decomposition results include: original basic tensors; Collect local real-time data, reconstruct the basic tensor of the real-time computing task according to the computing model parameters and the local real-time data, obtain a first basic tensor adapted to the real-time computing task, and execute the real-time computing task according to the first basic tensor to obtain a local result; At least one of the local result or the first basic tensor is transmitted to the cloud computing center device, so that the cloud computing center device receives data transmitted by at least one edge computing device for global data processing.

2. The cloud-edge collaborative data processing method according to claim 1, characterized in that: The cloud computing center device receives data transmitted by at least one edge computing device to perform global data processing, including: The cloud computing center device globally fine-tunes the computing model parameters based on the local results transmitted by at least one edge computing device and in combination with its own cloud results, and transmits the parameter fine-tuning results to the edge computing device, so that the edge computing device can calculate the real-time computing task for the local real-time data at the next moment based on the parameter fine-tuning results.

3. The cloud-edge collaborative data processing method according to claim 2, characterized in that: The cloud computing center device receives data transmitted by at least one edge computing device for global data processing, and further includes: making a global first joint decision according to a first basic tensor transmitted by at least one edge computing device, including: The cloud computing center device receives a first basic tensor sent by at least one edge computing node, and combines it with its own cloud data to generate a first joint decision based on the real-time data of the edge computing device and based on its own cloud data globally.

4. The cloud-edge collaborative data processing method according to claim 1, characterized in that: The step of reconstructing the basic tensor of the real-time computing task according to the computing model parameters and the local real-time data to obtain a first basic tensor adapted to the real-time computing task includes: The basic tensor is initially adjusted according to the computing model parameters to obtain a real-time basic tensor that adapts to the data distribution and task characteristics of the real-time computing task, and the real-time basic tensor is reconstructed according to local real-time data to obtain a reconstruction result, and the reconstruction result is post-processed according to the task requirements of the real-time computing task to obtain a first basic tensor; the post-processing includes: normalization or standardization.

5. The cloud-edge collaborative data processing method according to claim 4, characterized in that: The reconstructing the real-time basic tensor according to the local real-time data to obtain a reconstruction result includes: According to the local real-time data, at least one of linear transformation, nonlinear activation or pooling operation is performed on the real-time basic tensor to obtain a reconstruction result.

6. The cloud-edge collaborative data processing method according to claim 1, characterized in that: The cloud computing center device returns the calculation model parameters according to the received original decomposition result, including: After receiving the original decomposition result, the cloud computing center device uses the original decomposition result as the input of the initial computing model, classifies or regresses the original decomposition result, outputs a predicted value, and updates the parameters of the computing model based on the predicted value and the true label of the predicted value to obtain trained computing model parameters, and transmits the computing model parameters to the edge computing device.

7. The cloud-edge collaborative data processing method according to claim 6, characterized in that: The transmitting the computing model parameters to the edge computing device includes: The computing model parameters are compressed to obtain compressed computing model parameters, and the compressed computing model parameters are transmitted to the edge computing device through a pre-agreed transmission protocol, so that the edge computing device obtains the original computing model parameters after decompression.

8. The cloud-edge collaborative data processing method according to claim 1, characterized in that: The original decomposition result also includes: an original transformation matrix; in addition to returning the calculation model parameters according to the received original decomposition result, the cloud computing center device also includes: After receiving the original basic tensor and the original transformation matrix, the cloud computing center device performs secondary decomposition respectively to obtain multiple sub-tensor results and multiple sub-transformation matrices respectively, and combines its own cloud data to generate a second joint decision based on the original data of the edge computing device and based on its own cloud data globally.

9. The cloud-edge collaborative data processing method according to claim 8, characterized in that: The combining of the own cloud data to generate a second joint decision based on the original data of the edge computing device and based on the overall cloud data of the own cloud includes: The multiple sub-tensor results are sequentially multiplied with multiple sub-transformation matrices at the matrix level to obtain multiple intermediate matrices, and the multiple intermediate matrices are merged to obtain a final matrix. According to the final matrix and its own cloud data, a second joint decision based on the original data of the edge computing device and based on its own cloud data is generated globally.

10. A cloud-edge collaborative data processing system based on tensor transformation, applied to edge computing devices, characterized in that: include: Tensor decomposition unit, tensor reconstruction unit and global processing unit; among them, The tensor decomposition unit is used to perform tensor decomposition on the original data to obtain original decomposition results, and transmit the original decomposition results to the cloud computing center device respectively, so that the cloud computing center device returns the calculation model parameters according to the received original decomposition results; the original decomposition results include: original basic tensors; Collect local real-time data, reconstruct the basic tensor of the real-time computing task according to the computing model parameters and the local real-time data, obtain a first basic tensor adapted to the real-time computing task, and execute the real-time computing task according to the first basic tensor to obtain a local result; At least one of the local result or the first basic tensor is transmitted to the cloud computing center device, so that the cloud computing center device receives data transmitted by at least one edge computing device for global data processing.

Images