CN107871164B - A deep learning method for personalized fog computing environment - Google Patents

Abstract

The invention discloses a personalized deep learning method for a fog computing environment. Cloud nodes train a general model through massive training data, distribute the general model obtained by training to each fog computing node, and then utilize the computing and storage capabilities of the fog computing nodes. Train a deep learning model that meets the needs of the edge-side industry; collect data from smart sensing devices, conduct real-time inference at fog computing nodes, and output results in real time; identify errors in inference, continuously train and optimize the model, and integrate industry personality The customized model is selectively transmitted to the cloud, and the general model of the cloud is received and continuously improved and optimized. Compared with the prior art, the present invention improves the efficiency of real-time business execution, and at the same time, the personalized deep learning model is stored in the fog computing node, which ensures the security of the model. On the other hand, the model can be shared to the cloud according to the user's permission. , is that the model is open.

Description

A deep learning method for personalized fog computing environment

technical field

The invention relates to cloud computing, Internet of Things, artificial intelligence and deep learning technologies, in particular to a personalized deep learning method for fog computing environments.

Background technique

In recent years, artificial intelligence technology has developed rapidly, and its commercialization speed has exceeded expectations. Artificial intelligence will bring disruptive changes to the entire society and has become an important development strategy for countries in the future. In particular, the evolution of algorithms with deep learning as the core, its super evolution ability, with the support of big data, through training to build a large-scale convolutional neural network similar to the structure of the human brain, has been able to solve various problems.

Deep learning requires a large amount of data and computing resources for training, and cloud services can meet the requirements to a certain extent. The cloud center aggregates a large number of physical hardware resources, and adopts virtualization technology to realize the unified allocation and scheduling of computing resources in heterogeneous networks And management, centralized construction of data centers greatly reduces the cost of computing and storage. However, with the increasing amount of data, the transmission rate is declining, and sometimes there is a large network delay, and computing and storage cannot all be placed in the remote cloud. At this time, the emergence of fog computing has greatly improved this situation, especially for the edge side requirements such as real-time services, data optimization, bandwidth limitation, application intelligence, security and privacy, etc., which has accelerated the development of "fog computing". "Cloud computing + fog computing" brings new possibilities.

With the development of deep learning technology, cloud training and learning will generate general data models such as object detection. However, in specific application scenarios, general models cannot meet the individual needs of their industries. Training can be completed in the cloud, while fog computing nodes run through Between the cloud and the device, it becomes a bridge between the cloud and the device, and can provide near-end training and reasoning computing services. In this case, how to effectively use the "cloud computing + fog computing" capabilities to provide personalized deep learning computing capabilities in a fog computing environment has become an urgent problem to be solved.

SUMMARY OF THE INVENTION

The technical task of the present invention is to provide a personalized deep learning method for fog computing environment.

The technical task of the present invention is achieved in the following manner:

A personalized deep learning method for fog computing environments. Cloud nodes train general models through massive training data, distribute the trained general models to each fog computing node, and then use the computing and storage capabilities of the fog computing nodes to train to meet the edge side. The deep learning model required by the industry; by collecting data from intelligent sensing devices, real-time inference at fog computing nodes, and real-time output results; identify errors in inference, continuously train and optimize models, and can select industry-specific models The incoming cloud, while receiving the cloud's general model and continuous improvement and optimization.

The cloud node is responsible for continuously training and optimizing the general deep learning model, distributing the general model obtained by training to each fog computing node, and collecting and storing industry-specific deep learning models from the fog computing nodes.

The fog computing node is a bridge between the cloud and the device, and is responsible for receiving the general deep learning model from the cloud node, and adding industry-specific data on the edge side for training to generate an industry-specific deep learning model.

The fog computing node provides the inference function, and processes and feeds back the input data from the edge-side intelligent sensing device in real time. The fog computing node conducts training according to the errors in the inference, and compares with the general model in the cloud to continuously optimize the industry personality. Deep learning models can be customized, and can interact with the cloud to upload their industry-specific models according to customer needs.

The intelligent sensing device collects environmental data in real time, uses fog computing nodes to perform real-time deep learning calculations, and obtains results that are promptly fed back to users or take actions.

The operation steps of this method are as follows:

Step 1) The cloud node performs cloud training through a large number of general deep learning training sets, and the training obtains a basic model with general cognitive recognition ability;

Step 2) The fog computing node requests the deep learning model from the cloud node according to the specific industry requirements close to the edge side and the reliability of the local personalized deep learning model;

Step 3) The cloud node distributes the general deep learning model to the fog computing node;

Step 4) The cloud node inquires according to the specific industry requirements of the fog computing node; if there is a personalized deep learning model shared by the user, the model is returned to the fog computing node;

Step 5) If the fog computing node receives the personalized deep learning model shared by the cloud user, compare whether the model already exists locally, and if so, go to step 7); otherwise, go to step 6);

Step 6) The fog computing node is trained according to the specific industry requirements close to the edge side, based on the model shared by the user, using industry application data and locally stored training data to generate an industry-specific deep learning model and save it locally;

The fog computing node in step 7) receives the general deep learning model from the cloud, and compares whether the model already exists locally, if so, go to step 9); otherwise, go to step 8);

Step 8) The fog computing node uses industry application data and locally stored training data for training based on the general model according to the specific industry requirements close to the edge side, to generate an industry-specific deep learning model, and save it locally;

Step 9) The fog computing node compares the reliability of each local model, and selects the optimal model for reasoning;

Step 10) The intelligent sensing device collects data according to industry requirements and sends it to the fog computing node for reasoning;

Step 11) The fog computing node uses the industry-specific deep learning model to reason about the collected data, completes cognitive services, realizes industry requirements, and sends the real-time output results to the intelligent sensing device;

Step 12) The intelligent sensing device feeds back the result to the end user, and executes the corresponding work instruction;

Step 13) The fog computing node forms a training sample from the collected error data and saves it in the local storage;

Step 14) The fog computing node selects idle time, regularly trains the training samples stored in the local storage based on the existing personalized deep learning model, forms a new deep learning training model, and saves it in the local cache;

Step 15) The fog computing node uploads and shares the personalized deep learning model that meets the needs of the specific industry to the cloud according to the user's permission;

Step 16) Execute step 1) to step 15) in a circular manner, continue to optimize the model, improve the computational reasoning ability of deep learning, and meet the individual needs of the industry.

The cognitive ability of the basic model in the described step 1) is suitable for general scenarios.

In the step 11), the real-time output results are sent to the intelligent sensing device, including forming training samples from the collected data and inference results at the same time, and storing them in the local cache storage of the fog computing node.

In the step 12), the corresponding work instructions are executed, including, if an error occurs, uploading the error and the correction result to the fog computing node.

In the step 13), saving in the local storage, including and calculating the credibility of the multiple deep learning models in the local storage.

Compared with the prior art, the personalized deep learning method of the fog computing environment of the present invention makes full use of the characteristics of fog computing nodes as a bridge between the cloud and the device side, and distributes the deep learning computing in the cloud and the fog computing nodes, and the cloud is responsible for the general purpose Model training, and then distribute the general model to each fog computing node, using the computing and storage capacity of the fog computing node and combining with the specific industry application requirements on the edge side of fog computing. This method fully considers the individual needs of specific industries, and provides efficient cognitive computing capabilities that meet specific industry applications. Since deep learning calculations are completed at the fog computing nodes, close to the equipment demand side, the efficiency of real-time business execution is improved, and the depth of personalization is at the same time. The learning model is stored in the fog computing node, which ensures the security of the model. On the other hand, the model can be shared to the cloud according to the user's permission, which means that the model is open. In addition, the fog computing node selects idle time to continuously optimize the model through faulty reasoning feedback from actual applications, effectively utilizing computing resources, and the cloud general model is also continuously updated, choosing to use models with higher reliability, providing continuous Optimized industry-specific deep learning computing power.

Description of drawings

Figure 1 is a schematic diagram of the composition of a deep learning computing node of a fog computing environment personalized deep learning method;

FIG. 2 is a schematic flowchart of a deep learning method for personalizing a fog computing environment.

Detailed ways

Example 1:

The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but it is not intended to limit the present invention.

As shown in Figure 1, the cloud gathers a large number of computing resources, trains the general model through massive training data, distributes the general model obtained by training to each fog computing node, and then uses the computing and storage capabilities of the fog computing nodes to train to meet the edge side. The deep learning model required by the industry; by collecting data from intelligent sensing devices, real-time inference at fog computing nodes, and real-time output results; identify errors in inference, continuously train and optimize models, and can select industry-specific models The incoming cloud, while receiving the cloud's general model and continuous improvement and optimization. in,

The cloud node is responsible for continuously training and optimizing the general deep learning model, distributing the general model obtained by training to each fog computing node, and is responsible for collecting and storing industry-specific deep learning models from the fog computing nodes; the fog computing The node is the bridge between the cloud and the device. It is responsible for receiving general deep learning models from the cloud, adding industry-specific data on the edge side for training, generating industry-specific deep learning models, and providing reasoning functions. Real-time processing feedback comes from the edge-side intelligent transmission. Sensing the input data of the device, the fog computing nodes are trained according to the errors in the inference, and compared with the general model in the cloud, continuously optimize the industry-specific deep learning model, and can interact with the cloud to upload their industry according to customer needs. Personalized model; the intelligent sensing device collects environmental data in real time, uses fog computing nodes to perform real-time deep learning calculations, and obtains results that are promptly fed back to users or take actions.

Referring to Figure 2, personalized deep learning computation includes the following steps:

Step 1) The cloud node performs cloud training through a large number of general deep learning training sets, and the training obtains a basic model with general cognitive recognition ability, and the cognitive ability of the model is suitable for general scenarios;

Step 2) The fog computing node requests the deep learning model from the cloud node according to the specific industry requirements close to the edge side and the reliability of the local personalized deep learning model;

Step 3) The cloud node distributes the general deep learning model to the fog computing node;

Step 4) The cloud node inquires according to the specific industry requirements of the fog computing node, and if there is a personalized deep learning model shared by the user, the model is returned to the fog computing node;

Step 5) If the fog computing node receives the personalized deep learning model shared by the cloud user, compare whether the model already exists locally, and if so, go to step 7); otherwise, go to step 6);

Step 6) The fog computing node is trained according to the specific industry requirements close to the edge side, based on the model shared by the user, using industry application data and locally stored training data to generate an industry-specific deep learning model and save it locally;

The fog computing node in step 7) receives the general deep learning model from the cloud, and compares whether the model already exists locally, if so, go to step 9); otherwise, go to step 8);

Step 8) The fog computing node uses industry application data and locally stored training data for training based on the general model according to the specific industry requirements close to the edge side, to generate an industry-specific deep learning model, and save it locally;

Step 9) The fog computing node compares the reliability of each local model, and selects the optimal model for reasoning;

Step 10) The intelligent sensing device collects data according to industry requirements and sends it to the fog computing node for reasoning;

Step 11) The fog computing node uses the industry-specific deep learning model to infer the collected data, completes cognitive services, realizes industry requirements, sends the real-time output results to the intelligent sensing device, and simultaneously transmits the collected data and reasoning The result forms a training sample, which is stored in the local cache storage of the fog computing node;

Step 12) The intelligent sensing device feeds back the result to the end user, executes the corresponding work instruction, and if an error occurs, uploads the error and the correction result to the fog computing node;

Step 13) The fog computing node forms a training sample from the collected error data, saves it in the local storage, and calculates the credibility of the multiple deep learning models in the local storage;

Step 14) The fog computing node selects idle time, regularly trains the training samples stored in the local storage based on the existing personalized deep learning model, forms a new deep learning training model, and saves it in the local cache;

Step 15) The fog computing node uploads and shares the personalized deep learning model that meets the needs of the specific industry to the cloud according to the user's permission;

Step 16) Execute step 1) to step 15) in a circular manner, continue to optimize the model, improve the computational reasoning ability of deep learning, and meet the individual needs of the industry.

Through the above specific embodiments, those skilled in the art can easily implement the present invention. However, it should be understood that the present invention is not limited to the above-mentioned specific embodiments. On the basis of the disclosed embodiments, those skilled in the technical field can arbitrarily combine different technical features to realize different technical solutions.

Claims (9)

1. A personalized deep learning method for a fog computing environment, characterized in that a cloud node trains a general model through massive training data, distributes the general model obtained by training to each fog computing node, and then utilizes the computing and computing power of the fog computing node. Storage capacity trains deep learning models that meet the needs of edge-side industries; collects data from smart sensing devices, conducts real-time inference at fog computing nodes, and outputs results in real time; identifies errors in inference, continuously trains and optimizes models, and can The industry-specific models are selectively transmitted to the cloud, while the general models in the cloud are received and continuously improved and optimized; The operation steps of this method are as follows: Step 1) The cloud node performs cloud training through a large number of general deep learning training sets, and the training obtains a basic model with general cognitive recognition ability; Step 2) The fog computing node requests the deep learning model from the cloud node according to the specific industry requirements close to the edge side and the reliability of the local personalized deep learning model; Step 3) The cloud node distributes the general deep learning model to the fog computing node; Step 4) The cloud node inquires according to the specific industry requirements of the fog computing node; if there is a personalized deep learning model shared by the user, the model is returned to the fog computing node; Step 5) If the fog computing node receives the personalized deep learning model shared by the cloud user, compare whether the model already exists locally, and if so, go to step 7); otherwise, go to step 6); Step 6) The fog computing node is trained according to the specific industry requirements close to the edge side, based on the model shared by the user, using industry application data and locally stored training data to generate an industry-specific deep learning model and save it locally; The fog computing node in step 7) receives the general deep learning model from the cloud, and compares whether the model already exists locally, if so, go to step 9); otherwise, go to step 8); Step 8) The fog computing node uses industry application data and locally stored training data for training based on the general model according to the specific industry requirements close to the edge side, to generate an industry-specific deep learning model, and save it locally; Step 9) The fog computing node compares the reliability of each local model, and selects the optimal model for reasoning; Step 10) The intelligent sensing device collects data according to industry requirements and sends it to the fog computing node for reasoning; Step 11) The fog computing node uses the industry-specific deep learning model to reason about the collected data, completes cognitive services, realizes industry requirements, and sends the real-time output results to the intelligent sensing device; Step 12) The intelligent sensing device feeds back the result to the end user, and executes the corresponding work instruction; Step 13) The fog computing node forms a training sample from the collected error data and saves it in the local storage; Step 14) The fog computing node selects idle time, regularly trains the training samples stored in the local storage based on the existing personalized deep learning model, forms a new deep learning training model, and saves it in the local cache; Step 15) The fog computing node uploads and shares the personalized deep learning model that meets the needs of the specific industry to the cloud according to the user's permission; Step 16) Execute step 1) to step 15) in a circular manner, continue to optimize the model, improve the computational reasoning ability of deep learning, and meet the individual needs of the industry. 2. The method according to claim 1, wherein the cloud node is responsible for continuously training and optimizing the general deep learning model, distributing the general model obtained by training to each fog computing node, and is responsible for collecting and storing data from fog computing. The node's industry-specific deep learning model. 3. The method according to claim 1 or 2, wherein the fog computing node is a bridge between the cloud and the device end, and is responsible for receiving a general deep learning model from the cloud node, and adding industry-specific data on the edge side for processing. Training to generate industry-specific deep learning models. 4. The method according to claim 1 or 2, wherein the fog computing node provides an inference function, processes and feeds back the input data from the edge-side intelligent sensing device in real time, and the fog computing node according to the error occurred in the inference , conduct training, and compare with the general models in the cloud, continuously optimize the industry-specific deep learning models, and interact with the cloud to upload their industry-specific models according to customer needs. 5 . The method according to claim 1 , wherein the intelligent sensing device collects environmental data in real time, uses fog computing nodes to perform real-time deep learning calculations, and obtains results that are promptly fed back to users or take actions. 6 . 6 . The method according to claim 1 , wherein the cognitive ability of the basic model in the step 1) is suitable for general scenarios. 7 . 7. The method according to claim 1, characterized in that in step 11), the real-time output results are sent to the intelligent sensing device, comprising, simultaneously forming training samples from the collected data and inference results, and saving them in a Fog computing node local cache storage. 8 . The method according to claim 1 , wherein executing the corresponding work instruction in the step 12) includes, if an error occurs, uploading the error and the correction result to the fog computing node. 9 . 9 . The method according to claim 1 , wherein, in the step 13), saving in the local storage includes and calculating the credibility of multiple deep learning models in the local storage. 10 .

Images