Cloud edge collaborative data processing system based on edge computing

Abstract

The invention discloses a cloud edge collaborative data processing system based on edge computing, which is composed of an edge device layer, an edge computing platform layer and a cloud platform layer, fully utilizes computing and storage capabilities of an edge side, relieves cloud pressure, reduces network bandwidth occupation, shortens data transmission experiments, and improves data transmission efficiency. The real-time performance and the localization degree of data acquisition and processing are improved, real-time acquisition and real-time processing of data are realized, the data security is enhanced, the risk that the data are attacked or stolen is reduced, and cloud unified management and efficient utilization of multi-source heterogeneous data are realized; the scheduling optimization capability is improved at the cloud by adopting a scheduling method of Q learning and a genetic algorithm; a feature extraction method of an improved GoogleNet model is designed on an edge computing platform layer, redundant features in data are filtered out, the data dimension is reduced, data analysis and processing are more efficient, and the problem that data internal features are understood insufficiently due to the fact that the data space dimension is high is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of edge computing and multi-source heterogeneous data acquisition and perception, and specifically relates to a cloud-edge collaborative data processing system based on edge computing. Background Art

[0002] With the development of emerging technologies such as the Internet of Things and artificial intelligence, various sensors and intelligent terminals are widely used in fields such as industrial production, urban management, and personal life. These devices generate a large amount of heterogeneous data, including various formats such as images, videos, sounds, texts, CSV, and sensor measurement values. How to effectively collect, integrate, and utilize this data has become a current technical problem.

[0003] Traditional data acquisition methods and processing are centralized, that is, data is transmitted from edge nodes to the cloud for storage and calculation in the cloud. This data acquisition and processing method has the following problems:

[0004] 1. High network bandwidth pressure, and a large amount of data transmission occupies network bandwidth resources.

[0005] 2. High latency, and it takes a certain amount of time for data to be transmitted from edge nodes to the cloud, and real-time processing cannot be performed.

[0006] 3. High data privacy risk, a large amount of sensitive data is transmitted through the network and is vulnerable to attacks and theft;

[0007] 4. High cloud scheduling pressure. Summary of the Invention

[0008] The purpose of the present invention is to provide a cloud-edge collaborative data processing system based on edge computing, which makes full use of the computing and storage capabilities on the edge side, reduces the cloud pressure, reduces network bandwidth occupancy, shortens the data transmission experiment, improves the real-time performance and localization degree of data acquisition and processing, realizes real-time data acquisition and real-time processing, enhances data security, reduces the risk of data being attacked or stolen, realizes unified cloud management and efficient utilization of multi-source heterogeneous data; adopts a scheduling method of Q learning + genetic algorithm in the cloud to improve the scheduling optimization ability, and designs an improved feature extraction method for the GoogleNet model to achieve weakly supervised object detection.

[0009] The present invention is implemented by the following technical solutions:

[0010] A cloud-edge collaborative data processing system based on edge computing is proposed, including:

[0011] An edge device layer for collecting physical information and generating raw data;

[0012] The edge computing platform layer is composed of multiple edge nodes, deployed at the network edge. Each edge node collects, stores, and preprocesses data from nearby edge devices, and executes local tasks under the edge node task scheduling framework.

[0013] The cloud platform layer adopts a scheduling strategy based on Q-learning and genetic algorithms for cloud tasks to obtain the optimal scheduling plan and execute remote tasks according to the optimal scheduling plan.

[0014] In some embodiments of the present invention, the system monitors the resource usage of edge nodes and the cloud platform based on platform cloud-edge collaborative resource monitoring, and divides tasks into local execution and cloud execution.

[0015] In some embodiments of the present invention, in the edge computing platform layer, the preprocessing method includes a multi-object detection method, which includes:

[0016] Pre-training the improved GoogleNet model using an open-source dataset;

[0017] Initializing the parameters of the pre-trained improved GoogleNet model through model migration;

[0018] Training the initialized model using the target dataset and optimizing the model parameters to generate the final generation model;

[0019] Performing weighted summation on the generated feature maps to generate the CAM of the image;

[0020] Selecting the corresponding loss function according to different detection targets, calculating the feature regions in the CAM using the connected component algorithm, and mapping the calculated bounding boxes to the original image to achieve the positioning of the target;

[0021] Generating the final target detection result.

[0022] In some embodiments of the present invention, in the edge computing platform layer, the improved GoogleNet model includes:

[0023] Using GoogLeNet Inception V3 as the basic network, adding a global maximum pooling layer GMP after its last Inception module, and then replacing the original sparse fully connected layer of the network with a sigmoid fully connected layer.

[0024] In some embodiments of the present invention, the objective function is selected according to the target task in the improved GoogleNet model:

[0025] For single-class object recognition, the softmax cross-entropy function is used for classification;

[0026] Multi-class object recognition uses the sigmoid cross-entropy function to approximate the probability distribution of multi-class labels, and generates class activation heatmaps accordingly.

[0027] In some embodiments of the present invention, the scheduling strategy based on Q-learning and genetic algorithm includes:

[0028] Randomly initialize the population: Define the individual encoding using real number encoding, and obtain the chromosomes for genetic algorithm operation through ranking selection.

[0029] When the number of generations is less than the generation threshold, execute the conventional genetic algorithm; the conventional genetic algorithm includes performing crossover, mutation, and reproduction operations.

[0030] When the number of generations is greater than or equal to the generation threshold and at intervals of a set number of generations, execute the immigration operation, and dynamically adjust the parameters of the immigration operation using Q-learning.

[0031] Output the scheduling result when the termination condition is met.

[0032] In some embodiments of the present invention, in the scheduling strategy based on Q-learning and genetic algorithm, the gene space of the genetic algorithm is converted into the state space that can be directly processed in Q-learning. According to the order of scheduling, the state model of the chromosome is divided by uniform clustering and used as the state pattern of the Q-learning algorithm.

[0033] In some embodiments of the present invention, in the scheduling strategy based on Q-learning and genetic algorithm, Select the performance prediction variable of the algorithm based on the task completion time, and the completion time of all tasks is the final optimization goal; evaluate the quality of the action decision of the learning system in the current state according to the performance prediction variable, and construct an immediate reward and punishment function in the phased scheduling state pattern to make the optimization directions of maximizing the Q function and minimizing the objective function consistent:

[0034] ; where R is a positive constant.

[0035] In some embodiments of the present invention, in the scheduling strategy based on Q-learning and genetic algorithm, a piecewise linear exploration rate is adopted during the Q-learning iteration process:

[0036] ;

[0037] where is the initial exploration rate, is the number of iterations, is the length of the learning period.

[0038] In some embodiments of the present invention, in the scheduling strategy based on Q-learning and genetic algorithm, the Q-learning algorithm includes:

[0039] Step 1, initialize parameters: Set control parameters, the length of the learning period step, arbitrarily initialize the value of Q(s,a), select the initial exploration rate ε0 and learning rate α, and design a counter time = 0;

[0040] Step 2, randomly select a set-string individual based on the scheduling rule;

[0041] Step 3, when time < step,

[0042] (1) Calculate the exploration rate ε, and execute the action a according to the greedy search strategy;

[0043] (2) Judge the state mode of the next decision period, calculate the makespan , and the immediate heuristic reward value r k , and update the Q value;

[0044] (3) Update the system state, judge whether the termination state of the scheduling is reached. If so, perform step 3(4), otherwise return to step 3;

[0045] (4) Reassign the counter time = time + 1;

[0046] Step 4, judge whether time ≥ step holds. If so, the learning process terminates, otherwise go to step 3.

[0047] Compared with the prior art, the advantages and positive effects of the present invention are:

[0048] 1. Make full use of the computing and storage capabilities of the edge side with a hierarchical architecture. According to different requirements for computing resources, tasks are scheduled to appropriate edge nodes for computing. The data collected by edge intelligent terminals is processed by the scheduling decision module of the data preprocessing center at the edge side. The task execution plan is divided into two schemes: local execution and cloud execution. In the local execution module, task resources process scheduling information according to priorities, which reduces the pressure on the cloud, reduces the network bandwidth occupancy, and shortens the data transmission delay.

[0049] 2. Design a feature extraction method based on the improved GoogleNet model for edge nodes. Add a GMP layer after the last Inception module, and then use a sigmoid fully connected layer to replace the sparse fully connected layer of the original network. Select the optimal feature output layer, adopt the connected component calculation method to obtain the feature position, and optimize the feature extraction result, so as to filter out redundant features in the data, reduce the data dimension, make data analysis and processing more efficient and accurate, and solve the problem of insufficient understanding of the internal features of data due to the high data space dimension.

[0050] 3. Establish a multi-objective model in the cloud execution module, design a task scheduling method that combines Q-learning and genetic algorithms, reduce the cloud task scheduling pressure, and improve the cloud task scheduling ability.

[0051] After reading the detailed description of the embodiments of the present invention in conjunction with the accompanying drawings, other features and advantages of the present invention will become clearer. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the accompanying drawings required for the description of the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

[0053] Figure 1 Schematic diagram of the design process of the cloud-edge collaborative data processing system based on edge computing proposed by the present invention;

[0054] Figure 2 Schematic diagram of the structure of the cloud-edge collaborative data processing system based on edge computing proposed by the present invention;

[0055] Figure 3 Edge node task scheduling framework in the present invention;

[0056] Figure 4 Task scheduling method flow in the present invention;

[0057] Figure 5 Task scheduling strategy flow based on Q-learning and genetic algorithms in the present invention;

[0058] Figure 6 Schematic diagram of the object detection method based on transfer learning under weak supervision conditions given by the present invention;

[0059] Figure 7 Schematic diagram of the improved GoogleNet network structure proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0060] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0061] As Figure 1As shown in the figure, the design concept of the present invention includes: First, design a cloud-edge collaborative data processing framework based on edge computing, and allocate tasks into two execution scenarios: local execution and cloud execution; on this basis, design an edge node architecture based on intelligent terminals to make full use of the computing and storage capabilities on the edge side in a hybrid scheduling manner; in the cloud, design a scheduling strategy combining Q-learning and genetic algorithms to improve the cloud scheduling optimization ability; in view of the situation that the data space dimension of edge computing is relatively high and the internal characteristics of the data are not well understood, propose a feature extraction method based on an improved GoogleNet model to filter out redundant features in the data, reduce the data dimension, and make data analysis and processing more efficient and accurate.

[0062] As Figure 2 shown, the cloud-edge collaborative data processing system based on edge computing proposed by the present invention is divided into an edge device layer, an edge computing platform layer, and a cloud platform layer:

[0063] (1) The edge device layer includes various physical devices, sensors, and human-computer interaction interfaces, which collect physical information in the real world to generate raw data. For example, image sensors (cameras, scanners, etc.) collect image and video data, sound sensors (microphones, etc.) collect sound and voice data, text sensors (character recognition, keyboard input, etc.) collect text data, and physical quantity sensors (temperature, pressure, vibration, speed, position, etc.) collect digital data. Various types of data and sensors are connected to the edge node gateway device in the edge computing platform layer through a field network.

[0064] (2) The edge computing platform layer consists of multiple edge nodes (ECN), which are respectively deployed at the network edge, close to the source of data generation. Each edge node collects, stores, and performs basic processing on the nearby sensing devices, playing the role of "edge intelligence". It has the following functions:

[0065] Execute data collection tasks to collect various heterogeneous data from nearby sensing devices.

[0066] Preprocess the collected data, including but not limited to: duplicate removal, compression, feature extraction, etc.

[0067] Complete some simple artificial intelligence calculations according to task requirements and configurations.

[0068] Use local storage space for data caching.

[0069] Conduct data interaction with other edge computing nodes or the cloud platform through the network.

[0070] The edge computing platform layer is a distributed computing architecture. Different edge nodes can cooperate to run, share computing and storage resources. By deploying multiple nodes, the fault tolerance, reliability, and scalability of the system can be improved.

[0071] (3) The cloud platform layer consists of one or more cloud servers, which are responsible for centralized data storage, high-order computing, and business applications. It has the following main functions:

[0072] Establish a connection with the edge nodes and receive the data processed by the edge nodes.

[0073] Aggregate, clean, and persistently store the collected data on disk.

[0074] Perform complex artificial intelligence calculations on the data, such as deep learning, knowledge graph construction, etc.

[0075] Provide services such as data query, visualization, and data analysis for users.

[0076] Receive human-machine interface requests, process user instructions, and distribute tasks to the edge nodes.

[0077] The cloud platform layer has powerful computing, storage, and network resources, which can support the operation of the entire system and play a role in command and coordination. In this invention application, not all raw data is uploaded to the cloud, and the cloud only receives the preprocessed and streamlined data.

[0078] This invention is directed to the edge computing platform layer, and designs the edge nodes as a multi-source heterogeneous data acquisition node (an edge computing node) that includes computing, storage, and network communication functions and is cross-platform. This node is deployed on the edge side of various intelligent devices or sensors for local data acquisition and processing. Its hardware architecture includes: microcontroller (MCU), RAM, ROM, memory, communication module (WIFI, Bluetooth, 5G, etc.). Its software architecture includes: real-time operating system (RTOS), data acquisition module, preprocessing module, model deployment module, and network communication module, etc.

[0079] The edge computing node is designed with the following key features:

[0080] (1) Highly integrated, small in size, low in power consumption, and convenient for embedding and deployment.

[0081] (2) Support the acquisition of multiple heterogeneous data types, such as images, videos, texts, sounds, sensor data, etc.

[0082] (3) Support local intelligent computing, can deploy simple artificial intelligence models, and perform tasks such as heterogeneous data preprocessing, compression, and classification.

[0083] (4) Support multiple network communication protocols and can flexibly communicate with the cloud platform or other edge nodes.

[0084] (5) It supports multiple working modes and can dynamically switch between local computing and cloud platform computing according to network conditions and task requirements.

[0085] Taking the production of discrete manufacturing industry as an example, the edge node obtains static information and dynamic data in the industrial network, and uses services such as algorithm models, event management, and message routing deployed on the edge node to perform real-time processing and analysis, on-site reasoning and decision-making, budget management, and business analysis on the collected data. In the edge control module of discrete manufacturing industry, ECNController is used to connect multiple ECNs, process task requests, dynamically allocate virtual machine resources for ECNs, and transmit the preprocessed data that cannot be processed on local devices to the remote cloud platform layer through the task scheduling model for further analysis and processing; the cloud platform layer receives the information uploaded by the edge node, provides Internet of Things applications such as device management, intelligent production applications such as virtual factories, and intelligent service applications based on big data analysis, etc., and transmits the processed information back to the edge computing platform layer, and finally feedbacks to the corresponding devices in the edge device layer to realize the supervision and scheduling of on-site resources.

[0086] In the system proposed by the present invention, the edge node can be flexibly deployed and updated, providing an integrated service from data collection to cloud applications. Through the edge computing platform layer, the production plan management system and the edge control system of the manufacturing process in the discrete manufacturing industry realize the unified access, storage, and edge analysis and processing of various types of data on the edge side, and collect the operation parameters of the equipment through the Internet of Things integration of on-site devices (such as production equipment, logistics equipment, and detection equipment), transmit the data to the cloud platform layer, and at the same time receive the control information feedback from the cloud platform layer in real time, and finally feedback to the corresponding device, so as to realize the digital management of the device, reduce the requirements for network bandwidth, transmission delay, and cloud platform resource cost, which is of great significance to the production process control and process optimization of discrete manufacturing industry.

[0087] In the present invention application, in order to improve the utilization rate of network bandwidth by the edge node under high traffic conditions and enhance data real-time performance, such as Figure 3 shown, an edge node task scheduling framework is designed. The intelligent terminal collects multi-source data, monitors the resource usage of the edge terminal and the cloud platform according to the cloud-edge collaborative resource monitoring of the platform, and divides the execution plan of the task into two schemes: local execution and cloud execution, so as to effectively reduce the communication overhead and network delay of task processing and relieve the processing pressure on the core network and data center. In the local execution module, a hybrid scheduling method is adopted, combining static and dynamic scheduling algorithms, and different scheduling strategies are selected according to the characteristics of the task and environmental conditions; in the cloud execution module, a multi-objective model is established, and a scheduling strategy generated by combining Q learning and genetic algorithm is used to schedule the cloud tasks. The whole process of the task scheduling algorithm is as Figure 4 shown.

[0088] As Figure 5 shown, the scheduling algorithm of the present invention combining Q - learning and genetic algorithm includes:

[0089] S1: Determine the control function.

[0090] This control function is the core control logic for the collaborative optimization of the present invention combining Q - learning and genetic algorithm, and is used for dynamically deciding when to trigger the immigration operation of the genetic algorithm, when to trigger Q - learning optimization, and when to perform conventional genetic operations, so as to balance the exploration and exploitation capabilities of the algorithm.

[0091] The immigration operation is used to promote population diversity, and Q - learning optimization is used to adjust genetic operator parameters (such as crossover rate, mutation rate, etc.) to improve the convergence efficiency.

[0092] S2: Randomly initialize the population, and record the genetic generation as g = 0.

[0093] Taking the port tugboat scheduling as an example, in this embodiment, the real - number coding method is adopted to define the individual coding, as shown in Table 1 below:

[0094] Table 1

[0095]

[0096] In Table 1, the first column is the number of the scheduling task, the second column represents the abbreviation of the scheduling task name, and the third and fourth columns represent the resource situation and the corresponding quantity allocated to this task. For example, for task T3, the two resource allocation schemes are 2 G2 terminals and 3 G3 terminals respectively; among them, G1 - G4 represent the current classification of device resource occupancy: G1 has more idle resources, and G4 has the least idle resources. The matching rules for other tasks also adopt a similar method.

[0097] According to the above real - number coding, the chromosome for the genetic algorithm operation can be obtained through the sorting - selection method, and is used to execute the operation steps of crossover, mutation, reproduction, and reinforcement learning.

[0098] S3: Judge whether the genetic generation satisfies g > G and the set generation interval.

[0099] If the above conditions are met, execute S4, otherwise execute S5. The present invention implements stage - by - stage control through the generation threshold G. In the early stage (when g < G), conventional genetic operations are mainly used to quickly converge to the local optimal solution. In the later stage (when g > G and the set generation interval), immigration and Q - learning are periodically introduced, and Q - learning is used to dynamically adjust parameters such as the crossover rate and mutation rate of the genetic algorithm to jump out of the local optimum.

[0100] Here, G is usually set to 20% - 40% of the total number of generations MG, and the set generation is 10% - 20% of the population.

[0101] S4: Perform the immigration operation and dynamically adjust the parameters of the immigration operation using Q-learning so that it selects the optimal migration strategy according to the current search state.

[0102] The immigration operation is an information exchange mechanism between populations in the genetic algorithm. By regularly migrating some individuals from one sub-population to another, it introduces new genetic diversity, avoids premature convergence, and enhances the global search ability.

[0103] The key parameters of the immigration operation include:

[0104] Migration frequency: How many generations to perform a migration at intervals.

[0105] Migration ratio: The proportion of the number of individuals migrated each time to the population.

[0106] Migration direction: Which sub-populations to migrate between (for example, elite individuals migrate to the disadvantaged population).

[0107] Migration selection strategy: How to select the individuals to be migrated (for example, the optimal individual, random individual, etc.).

[0108] Traditional migration operations are usually fixed-frequency or random migrations, while the present invention realizes dynamic optimization of the migration strategy through the Q-learning algorithm. Q-learning dynamically adjusts the parameters of the immigration operation so that the algorithm can adaptively select the optimal migration strategy according to the current search state. The goal of Q-learning is to maximize the long-term reward (i.e., the algorithm convergence speed and the quality of the global optimal solution).

[0109] Specifically, the present invention converts the gene space of the genetic algorithm into a state space that can be directly processed in Q-learning. According to the scheduling sequence, it uses the uniform clustering method to divide the state model of the chromosome, which is used as the state pattern of the Q-learning algorithm.

[0110] Generally speaking, the reward and punishment function established by the Q-learning algorithm is a rank-based method:

[0111] ;

[0112] If the performance of the scheduling is measured by the evaluation criteria of the above formula, then in the middle stage of the scheduling process, the reward value returned by the learning system will only be set to empty, resulting in long-delay rewards, causing random search in the scheduling steps of non-terminal states.

[0113] The present invention selects the performance prediction variable of the algorithm based on the task completion time, and the completion time of all tasks is the ultimate optimization goal of the system.

[0114] During the scheduling process, the k-th task sequence number position (1, 2, 3, …, k, … R) is the k-th scheduling step of the learning system (the process of an agent interacting with the environment in Q-learning). The present invention evaluates the quality of the action decision of the learning system in the current state according to the performance prediction variable (task completion time ), constructs an immediate reward function in the phased scheduling state mode in Q-learning, as shown in the following formula, to make the optimization directions of maximizing the Q function and minimizing the objective function consistent:

[0115] .

[0116] where R is a relatively large positive constant. The above formula transforms the minimization problem into the maximization problem of the reward function. In the non-terminal state of the scheduling system, this immediate heuristic reward function can relatively accurately evaluate the quality of actions, provide reward information for the learning system in a timely manner, and thus guide the reinforcement learning algorithm to learn the optimal strategy faster.

[0117] Based on the establishment of the phased state mode, the present invention divides the state space. Different scheduling stages belong to different state sub-spaces, while the sub-spaces in the same phased state mode are the same. Evaluating the scheduling performance of the current stage according to the prediction variable of the scheduling performance is comparable and can achieve the purpose of optimizing the intermediate-stage scheduling.

[0118] To enable the search of Q-learning to find the local optimum according to the existing strategy and at the same time have good exploration ability to achieve the balance between "exploration" and "exploitation" of Q-learning, the present invention adopts a greedy search strategy based on a dynamic exploration rate. When making an action selection, the learning system will select the optimal strategy tending to the target with a probability of 1−ε, as shown in the following formula, and randomly select any strategy in the strategy space with a probability of ε:

[0119] ;

[0120] where the policy function represents the probability distribution of selecting action a at state s and time step k; is the Q-value function, representing the long-term expected return of executing action a’ in state s.

[0121] ε is a piecewise linear exploration rate adopted in the order of the Q-learning iteration process:

[0122] ;

[0123] where is the initial exploration rate, is the number of iterations, is the length of the learning period.

[0124] In Q-learning, Piecewise Design is a method of dividing different training stages and adjusting algorithm parameters or strategies accordingly, aiming to more efficiently balance exploration and exploitation, accelerate convergence, or adapt to complex environments. It has the following advantages: high exploration in the early stage to avoid local optima, pure exploitation in the later stage to accelerate convergence; flexible adaptation to different tasks by adjusting the piecewise points (such as 0.6step, 0.8step) and the initial value ε0; avoiding the problem of premature termination of exploration that may be caused by the traditional exponentially decaying exploration rate.

[0125] The steps of the Q-learning algorithm are as follows:

[0126] Step 1, Initialize parameters - Set control parameters, the length of the learning period step, arbitrarily initialize the Q(s,a) value, select the initial exploration rate ε0 and the learning rate α, etc. Design a counter time = 0.

[0127] Step 2, Randomly select a set of string individuals based on the scheduling rules.

[0128] Step 3, When time < step,

[0129] (1) Calculate the exploration rate ε and execute the action a according to the greedy search strategy.

[0130] (2) Judge the state mode of the next decision period and calculate the makespan , the immediate heuristic reward value r k , and update the Q value.

[0131] (3) Update the system state, judge whether the termination state of the scheduling is reached. If so, proceed to step 3(4); otherwise, return to step 3.

[0132] (4) Reassign the counter time = time + 1.

[0133] Step 4, Judge whether time ≥ step holds. If so, the learning process terminates; otherwise, go to step 3.

[0134] Introducing an immigration operation into the improved genetic algorithm of Q-learning can enhance the global search ability of the algorithm, avoid premature convergence, and promote population diversity.

[0135] S5: Perform crossover, mutation, and reproduction operations.

[0136] "Exploration" and "exploitation" are two important aspects of the search strategy of intelligent optimization algorithms. In the genetic algorithm designed by the present invention, reproduction, crossover, mutation, reinforcement learning and their related parameters embody the balance between exploration and exploitation. The role of reproduction is to exploit the search space in order to make full use of the existing information of the current population. Crossover and mutation are to explore the search space in order to find those potentially optimal regions. Crossover is considered to be the most important factor contributing to the global search performance of the genetic algorithm. Q-learning is to exploit the search area by learning statistical knowledge according to the optimization objective to find the local optimal solution. The Q-learning operation searches in the pattern state space and can also be regarded as a crossover operator with learning and adaptive capabilities of the genetic algorithm, which will clearly guide the search direction of the genetic algorithm.

[0137] In the genetic algorithm, fitness is used to measure the goodness of each individual in the population to reach the optimal solution in the optimization operation. The higher the fitness of an individual, the greater the probability of being inherited to the next generation; while the lower the fitness of an individual, the relatively smaller the probability of being inherited to the next generation. For the fitness of the individual, it can be calculated according to the following formula:

[0138]

[0139] where and are the highest and lowest objective function values in the population respectively.

[0140] When the termination condition g>MG is not satisfied, return to S3, and output the scheduling result when satisfied.

[0141] In the above scheduling strategy, based on the condition trigger of generations, by coordinating the immigration operation and Q-learning to dynamically switch the exploration and exploitation strategies at different stages of the genetic algorithm, it converges quickly with conventional genetic operations in the early stage, and then combines the immigration operation and Q-learning to achieve global optimization in the later stage. This design performs excellently in complex problems.

[0142] Aiming at the characteristics of discrete, multi-source, heterogeneous and massive data resources in the manufacturing process, in order to improve the effectiveness and response time of edge node data analysis and reduce the network load, the present invention also proposes a method for processing edge computing data sets.

[0143] The curse of dimensionality of data will reduce the performance of data processing and increase the task processing time. The main reason for the curse of dimensionality is the existence of redundant features in high-dimensional data. Feature selection is an important method for data set processing, which can filter out redundant features in the data, reduce the data dimension, and make data analysis and processing more efficient and accurate.

[0144] In the case where the data space dimension is relatively high and the internal characteristics of the data are not well understood, the present invention proposes a feature extraction method based on an improved GoogLeNet model. The network is improved using Class Activation Map (CAM). After processing the generated model through softmax and sigmoid, the CAM technology is used to locate the regions in the image that play a key role in class judgment. With the help of the connected region localization algorithm, single-category list detection and multi-category object detection are carried out, realizing weakly supervised object detection. As Figure 6 shown, the model is applicable to image data. For text data analysis, it needs to be converted into high-dimensional dense vectors.

[0145] Principle of the CAM object detection method: A CNN-GAP network composed of several convolutional layers and a Global Average Pooling (GAP) layer is constructed to generate CAM by weighting. Since no fully connected layer is used for classification, the position information of the feature map is retained. CAM is generated by weighting, taking into account the information of each feature map, and at the same time, class and position information can be obtained.

[0146] The target data set is image data with class labels; single-category and multi-category object detection are achieved through different loss functions; the connected region algorithm locates the detected objects by generating the feature connected regions in the class activation map CAM. The specific process is as follows:

[0147] 1. Improve the network structure of the GooLeNet model according to the CAM method.

[0148] 2. Use an open-source data set to train the improved model, using the ImageNet open-source data set.

[0149] 3. Initialize the parameters of the improved model through model migration.

[0150] 4. Use the target data set to train the initialized model and optimize the model parameters. The target data set is the historical data of the operation site collected by the system.

[0151] 5. Perform weighted summation on the generated feature maps to generate the CAM of the image.

[0152] 6. According to different detection targets, select the corresponding loss function. For example, Softmax cross-entropy is used for single-object detection, and Simgoid cross-entropy is used for multi-object detection.

[0153] 7. Use the connected region algorithm to calculate the feature regions in the CAM and map the calculated generated bounding boxes to the original image.

[0154] 8. Generate object detection results.

[0155] The GoogLeNet-GMP network structure is as Figure 7 shown. It uses GoogLeNet Inception V3 as the base network, adds a global max pooling layer (Global Max Pooling, GMP) after the last Inception module, and then replaces the sparse fully connected layer of the original network with a sigmoid fully connected layer. Let F k (x, y) represent the k-th feature map of the last convolutional layer. After passing through GMP, m k is as follows:

[0156] .

[0157] According to different target tasks, different objective functions are used. For single-class object recognition, the softmax cross-entropy function is used as the objective function for classification. The score Si for class i is as follows:

[0158] .

[0159] Training with the Sigmoid cross-entropy function as the objective function to approximate the probability distribution of multi-class labels. The expression for generating the class activation heat map Mi is as follows:

[0160] .

[0161] The GoogLeNet-GMP network structure parameters are shown in Table II. Among them, the Inception module is divided into 5 categories according to structural differences. After the feature map passes through GMP, the output size is 1x1x2048.

[0162] Table II

[0163]

[0164] The GoogLeNet network finally uses the softmax function for classification. If multiple objects need to be detected during the detection process, the sigmoid cross-entropy function can be used for replacement, and the average sigmoid cross-entropy function value of all samples is used as the objective function to train the network to achieve multi-class prediction.

[0165] After obtaining the CAM of the image to be tested, extract the feature localization label data, use the connected component algorithm to obtain the set of boundary coordinates of the feature, and then apply a bounding box on the image to be tested to mark the target.

[0166] The connected component algorithm is as follows:

[0167] Suppose the input is the activation feature map Mi of a certain type in the GMP layer, the learned feature map offset bi, the feature map threshold is θ, the binary threshold is δ, the size of the original image is Size, and the output is the set Li of the bounding box coordinates of all activation regions of the feature map Mi.

[0168] It should be noted that in the specific implementation process, the above-mentioned method part can be implemented by a processor in the form of hardware executing computer-executable instructions in software form stored in a memory, which will not be elaborated here. And the programs corresponding to the executed actions can all be stored in the computer-readable storage medium of the system in software form, so as to facilitate the processor to call and execute the operations corresponding to each above module.

[0169] The computer-readable storage medium mentioned above can include volatile memory, such as random access memory; it can also include non-volatile memory, such as read-only memory, flash memory, hard disk or solid-state drive; it can also include a combination of the above types of memory.

[0170] The processor mentioned above can also be a general term for multiple processing elements. For example, the processor can be a central processing unit, or it can be other general-purpose processors, digital signal processors, application-specific integrated circuits, field-programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc., and can also be a dedicated processor.

[0171] It should be pointed out that the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by those of ordinary skill in the art within the essence of the present invention should also fall within the protection scope of the present invention.

Claims

1. A cloud-edge collaborative data processing system based on edge computing, characterized in that: include: The edge device layer is used to collect physical information and generate raw data; The edge computing platform layer consists of multiple edge nodes deployed at the edge of the network. Each edge node collects, stores, and preprocesses data from nearby edge devices; it executes local tasks under the edge node task scheduling framework. At the cloud platform layer, a scheduling strategy based on Q learning and genetic algorithm is used to obtain the optimal scheduling plan for cloud tasks, and remote tasks are executed with the optimal scheduling plan.

2. The cloud-edge collaborative data processing system based on edge computing according to claim 1 is characterized in that: The system monitors the resource usage of edge nodes and cloud platforms based on platform cloud-edge collaborative resource monitoring, and divides tasks into local execution and cloud execution.

3. The cloud-edge collaborative data processing system based on edge computing according to claim 1 is characterized in that: In the edge computing platform layer, the preprocessing method includes a multi-target detection method, which includes: Use open source datasets to pre-train the improved GoogleNet model; Initialize the parameters of the pre-trained improved GoogleNet model through model migration; Use the target data set to train the initialization model and optimize the model parameters to generate the final generative model; Perform weighted summation on the generated feature maps to generate the CAM of the image; Select the corresponding loss function according to the different detection targets, use the connected region algorithm to calculate the feature area in CAM, and map the generated bounding box to the original image to achieve target positioning; Generate the final target detection result.

4. The cloud-edge collaborative data processing system based on edge computing according to claim 1 or 3, characterized in that: In the edge computing platform layer, the improvements to the GoogleNet model include: GoogLeNet Inception V3 is used as the basic network. A global maximum pooling layer GMP is added after the last Inception module, and then the sigmoid fully connected layer is used to replace the sparse fully connected layer of the original network.

5. The cloud-edge collaborative data processing system based on edge computing according to claim 4 is characterized in that: Improve the selection of objective function according to the target task in the GoogleNet model: Single-category target recognition uses the softmax cross entropy function for classification; Multi-category target recognition uses the sigmoid cross entropy function to approximate the probability distribution of multi-class labels and activates heat maps based on generated categories.

6. The cloud-edge collaborative data processing system based on edge computing according to claim 1 is characterized in that: Scheduling strategies based on Q-learning and genetic algorithms include: Randomly initialize the population: use real number coding to define individual codes, and obtain chromosomes for genetic algorithm operations through sorting and selection; When the number of genetic generations is less than the generation threshold, a conventional genetic algorithm is executed; the conventional genetic algorithm includes executing crossover, mutation and reproduction operations; When the genetic generation is greater than or equal to the generation threshold and the interval generation is set, the immigration operation is performed, and the parameters of the immigration operation are dynamically adjusted using Q learning; Output the scheduling result when the termination condition is met.

7. The cloud-edge collaborative data processing system based on edge computing according to claim 6 is characterized in that: In the scheduling strategy based on Q learning and genetic algorithm, the gene space of the genetic algorithm is converted into a state space that can be directly processed in Q learning. According to the order of scheduling, the state model of the chromosome is divided by uniform clustering, which is used as the state mode of the Q learning algorithm.

8. The cloud-edge collaborative data processing system based on edge computing according to claim 6 is characterized in that: In the scheduling strategy based on Q learning and genetic algorithm, the task completion time is Based on the performance prediction variables of the selection algorithm, the completion time of all tasks is the final optimization goal; the quality of the action decision of the learning system in the current state is evaluated according to the performance prediction variables, and an immediate reward and punishment function under the staged scheduling state mode is constructed to make the optimization direction of maximizing the Q function and minimizing the objective function consistent: ; Among them, R is always on when it is a positive number.

9. The cloud-edge collaborative data processing system based on edge computing according to claim 6, characterized in that: In the scheduling strategy based on Q learning and genetic algorithm, a piecewise linear exploration rate is used in the Q learning iteration process: ； in, is the initial exploration rate, is the number of iterations, is the length of the learning cycle.

10. The cloud-edge collaborative data processing system based on edge computing according to claim 8, characterized in that: In the scheduling strategy based on Q learning and genetic algorithm, the Q learning algorithm includes: Step 1, Initialize parameters: Set control parameters, the length of the learning period step, arbitrarily initialize the Q(s,a) value, select the initial exploration rate ε0 and the learning rate α, and design a counter time = 0; Step 2, Randomly select a set-string individual based on the scheduling rule; Step 3, When time < step, (1) Calculate the exploration rate ε and execute the action a according to the greedy search strategy; (2) Determine the state mode of the next decision period and calculate the maximum completion time , immediate heuristic reward value r k , update the Q value; (3) Update the system state, judge whether the termination state of the scheduling is reached. If so, perform step 3(4), otherwise return to step 3; (4) Reset the counter to time = time + 1; Step 4, Judge whether time ≥ step holds. If so, the learning process terminates, otherwise go to step 3.

Images