End-side cloud collaborative intelligent optimization and resource scheduling method and system for industrial Internet of Things

Abstract

The invention relates to an end-side cloud collaborative intelligent optimization and resource scheduling method and system for an industrial Internet of Things. The method comprises the steps that sparse processing and quantization are conducted on model parameters on end equipment through accelerator hardware, the storage and transmission requirements of a model are reduced, and meanwhile the accuracy of the model is kept; through the knowledge extraction technology, the end equipment can extract key information in the data, and high-quality input is provided for subsequent data processing and model training; by deploying the lightweight reasoning model and the routing model in the edge device, the uploaded data can be preliminarily classified, and only the data which is difficult to process is sent to the cloud, so that the overall delay is reduced to the greatest extent while the accuracy is kept; performance monitoring, bandwidth allocation and task segmentation are carried out on end equipment and edge equipment in real time through a cloud end, and global model optimization and resource scheduling are responsible for realizing optimal performance.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial Internet of Things, and in particular to a method and system for end-edge-cloud collaborative intelligent optimization and resource scheduling for industrial Internet of Things. Background Art

[0002] In the context of the rapid development of the current Industrial Internet of Things (IIoT), the end-edge-cloud collaborative architecture has become one of the key technologies for realizing intelligent and automated production. In this architecture, the end device is responsible for collecting real-time data from the production site, the edge device provides local data processing and rapid response, and the cloud is responsible for large-scale data processing and complex decision support. However, the existing end-edge-cloud collaborative system still faces many challenges when dealing with high-real-time tasks, unbalanced resource allocation, and insufficient model generalization capabilities.

[0003] Specifically, end devices play an important role in data collection and preliminary processing, but their computing and storage capabilities are limited, making it difficult to handle large-scale or highly complex data analysis tasks. Although edge devices can alleviate some of the computing pressure, their resource allocation and job scheduling capabilities are limited and cannot adapt to the rapidly changing needs of the production site, resulting in inefficient resource allocation and suboptimal job scheduling. In addition, when the cloud center performs data analysis and model training, existing technologies fail to fully utilize the collaborative computing resources of the end-edge cloud, which not only affects the training efficiency, but also limits the generalization potential of the model.

[0004] In terms of resource allocation, traditional resource allocation strategies are usually based on static resource allocation plans, which lack adaptability to real-time data characteristics and changes in network status. This makes it difficult for the system to respond quickly and effectively when faced with emergencies or resource competition in the production process, thus affecting production efficiency and product quality. In terms of model generalization, since data in industrial environments are often non-independent and identically distributed, existing technologies make it difficult to achieve effective migration and application of models between different devices and scenarios, limiting the generalization ability of the model and the accuracy of intelligent decision-making. In addition, with the circulation and sharing of data, how to achieve efficient use of data and intelligent decision-making while ensuring data privacy is an urgent problem to be solved in technological development. Existing data protection and privacy protection mechanisms often sacrifice data processing efficiency, and it is difficult to meet the dual requirements of the Industrial Internet of Things for real-time and privacy. Summary of the invention

[0005] Based on this, in order to solve the above technical problems, a method and system for end-edge-cloud collaborative intelligent optimization and resource scheduling for the Industrial Internet of Things is provided, which can solve the problem of efficient collaboration between end devices, edge devices and the cloud in the Industrial Internet of Things.

[0006] A method for intelligent optimization and resource scheduling of end-edge-cloud collaboration for industrial Internet of Things, the method comprising:

[0007] The end device uses the neural network accelerator hardware to perform sparse and quantization processing on the deep learning model to obtain a compressed model; the end device collects raw data and inputs the raw data into the compressed model for preprocessing to obtain preprocessed data;

[0008] The terminal device uses a feature extraction network and a knowledge extraction technology to extract feature data from the preprocessed data, and uploads the feature data;

[0009] A lightweight inference model and a routing model are deployed in the edge device; the edge device receives the feature data transmitted from the end device, and preliminarily classifies the feature data through the lightweight inference model to obtain classified data;

[0010] The edge device inputs the classified data into the routing model; the difficult data is identified through the routing model and sent to the cloud, the cloud processes the difficult data, and updates the global model stored in the cloud according to the processed difficult data;

[0011] The cloud is connected to the terminal device and the edge device respectively, monitors the performance of the terminal device and the edge device in real time, dynamically adjusts the operation of the terminal device and the edge device by executing deep learning algorithms to allocate bandwidth and split tasks, and optimizes the global model.

[0012] In one embodiment, the terminal device uses neural network accelerator hardware to perform sparse and quantized processing on the deep learning model to obtain a compressed model, including:

[0013] The terminal device uses a neural network accelerator hardware to introduce a number of zero values ​​into the deep learning model, and adjusts the original gradient of the deep learning model through each of the zero values ​​to complete the sparse processing;

[0014] The terminal device uses neural network accelerator hardware to reduce the precision of the floating-point numbers of the deep learning model and complete the quantization process.

[0015] In one embodiment, the terminal device uses a feature extraction network and a knowledge extraction technique to extract feature data from the preprocessed data, including:

[0016] The terminal device inputs the preprocessed data into a feature extraction network to determine the number of data points and feature dimensions in the preprocessed data;

[0017] Determining the selected original feature data based on the number of data points and feature dimensions through the feature extraction network to generate an original feature data matrix;

[0018] The knowledge extraction technology is used to extract feature data based on the preprocessed data and the original feature data matrix.

[0019] In one embodiment, the edge device receives the feature data transmitted from the end device, and performs preliminary classification on the feature data using the lightweight reasoning model, including:

[0020] The edge device receives the feature data transmitted from the end device, and receives the original completed model of the global model transmitted from the cloud;

[0021] The edge device determines a pruning threshold according to the feature data, and performs a pruning operation on the original completed model according to the pruning threshold to obtain a lightweight inference model;

[0022] The feature data is preliminarily classified using the lightweight inference model.

[0023] In one embodiment, the method further comprises:

[0024] The edge device determines the difficulty of the feature data, and determines the confidence of the model for the feature data based on the routing model;

[0025] The edge device obtains a confidence threshold and completes the decision of the routing model according to the difficulty and confidence.

[0026] In one embodiment, the method comprises:

[0027] The edge device identifies difficult data in the feature data based on the confidence threshold and Boolean variables, and sends the difficult data to the cloud.

[0028] In one embodiment, the method further comprises:

[0029] The cloud determines each set of edge devices and obtains the computing capacity and current task requirements of each edge device;

[0030] The cloud dynamically determines the tasks to be executed on the edge device and the tasks to be executed on the cloud according to the association function and the computational cost function based on the cloud computing resources, the computing capability of the edge device and the current task requirements;

[0031] The cloud determines a bandwidth threshold according to the bandwidth of the edge device, and according to the bandwidth threshold, offloads the tasks executed on the edge device to the edge device based on the offloading function, the bandwidth function, and the task set.

[0032] In one embodiment, the cloud dynamically adjusts the operation of the terminal device and the edge device, including:

[0033] The cloud collects the performance data of the edge device in real time and determines the performance evaluation index;

[0034] The cloud obtains a performance evaluation weight coefficient, and updates the performance evaluation indicator based on the performance evaluation weight coefficient to obtain a change trend corresponding to the performance evaluation indicator;

[0035] The cloud uses a deep learning algorithm to dynamically adjust the operation of the terminal device and the edge device according to the change trend.

[0036] In one embodiment, the cloud optimizes the global model, including:

[0037] A local model is deployed in the edge device, and a global model is deployed in the cloud;

[0038] The cloud determines a distance metric between the local model and the global model, and performs global model optimization based on the distance metric.

[0039] A device-edge-cloud collaborative intelligent optimization and resource scheduling system for industrial Internet of Things, the system comprising:

[0040] The end device is used to perform sparse and quantization processing on the deep learning model using the neural network accelerator hardware to obtain a compressed model; the end device collects raw data and inputs the raw data into the compressed model for preprocessing to obtain preprocessed data;

[0041] The terminal device is also used to utilize a feature extraction network and a knowledge extraction technology to extract feature data from the preprocessed data, and upload the feature data;

[0042] An edge device deployed with a lightweight inference model and a routing model is used to receive the feature data transmitted from the end device, and preliminarily classify the feature data through the lightweight inference model to obtain classified data;

[0043] The edge device is also used to input the classified data into the routing model; identify the difficult data through the routing model and send it to the cloud, the cloud processes the difficult data, and updates the global model stored in the cloud according to the processed difficult data;

[0044] The cloud is connected to the terminal device and the edge device respectively, monitors the performance of the terminal device and the edge device in real time, dynamically adjusts the operation of the terminal device and the edge device by executing deep learning algorithms to allocate bandwidth and split tasks, and optimizes the global model.

[0045] The above-mentioned end-edge-cloud collaborative intelligent optimization and resource scheduling method and system for industrial Internet of Things, by using accelerator hardware on the end device to sparsely process and quantify the model parameters, reduce the storage and transmission requirements of the model while maintaining the accuracy of the model; through knowledge extraction technology, the end device can extract key information in the data and provide high-quality input for subsequent data processing and model training; by deploying lightweight inference models and routing models in edge devices, it can preliminarily classify the uploaded data and only send difficult-to-process data to the cloud, thereby minimizing overall latency while maintaining accuracy; through the cloud, real-time performance monitoring of end devices and edge devices, as well as bandwidth allocation and task segmentation, is performed, and is responsible for global model optimization and resource scheduling to achieve optimal performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is an application environment diagram of a device-edge-cloud collaborative intelligent optimization and resource scheduling method for the industrial Internet of Things in one embodiment;

[0047] Figure 2 It is a flowchart of a method for end-edge-cloud collaborative intelligent optimization and resource scheduling for industrial Internet of Things in one embodiment;

[0048] Figure 3 It is a structural block diagram of a device-edge-cloud collaborative intelligent optimization and resource scheduling system for the industrial Internet of Things in one embodiment;

[0049] Figure 4 It is a structural block diagram of a device-edge-cloud collaborative intelligent optimization and resource scheduling system for industrial Internet of Things in another embodiment;

[0050] Figure 5 A schematic diagram of a terminal device model compression framework combining model sparsification, quantization and knowledge extraction in one embodiment;

[0051] Figure 6 A schematic diagram of an end-to-end hybrid learning framework in one embodiment;

[0052] Figure 7 A schematic diagram of a structural framework of cloud collaborative computing and global optimization in one embodiment;

[0053] Figure 8 A schematic diagram of a process structure in an equipment maintenance prediction system applied in the industrial Internet of Things in one embodiment;

[0054] Fig. 9 The figure is a schematic diagram of a process structure applied to an energy management system in an embodiment. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0056] The end-edge-cloud collaborative intelligent optimization and resource scheduling method for industrial Internet of Things provided in the embodiments of the present application can be applied to Figure 1 In the application environment shown. Figure 1 As shown, the application environment includes a terminal device 110, an edge device 120, and a cloud 130, wherein the cloud 130 is connected to and communicates with the terminal device 110 and the edge device 120 respectively; the terminal device 110 uses the neural network accelerator hardware to perform sparse and quantized processing on the deep learning model to obtain a compressed model; the terminal device 110 collects raw data and inputs the raw data into the compressed model for preprocessing to obtain preprocessed data; the terminal device 110 uses the feature extraction network and the knowledge extraction technology to extract feature data from the preprocessed data and upload the feature data; a lightweight inference model and a routing model are deployed in the edge device 120; the edge device 120 receives the data from the terminal device The feature data transmitted by the edge device 110 is preliminarily classified by a lightweight inference model to obtain classified data; the edge device 120 inputs the classified data into the routing model; the difficult data is identified by the routing model and sent to the cloud 130, the cloud 130 processes the difficult data, and updates the global model stored in the cloud 130 according to the processed difficult data; the cloud 130 is connected to the end device 110 and the edge device 120 respectively, and performs performance monitoring on the end device 110 and the edge device 120 in real time, and dynamically adjusts the operation of the end device 110 and the edge device 120 by executing a deep learning algorithm to allocate bandwidth and split tasks, and optimizes the global model.

[0057] In one embodiment, Figure 2 As shown, a method for end-edge-cloud collaborative intelligent optimization and resource scheduling for industrial Internet of Things is provided, including the following steps:

[0058] In step 202, the terminal device uses the neural network accelerator hardware to perform sparsification and quantization processing on the deep learning model to obtain a compressed model; the terminal device collects raw data and inputs the raw data into the compressed model for preprocessing to obtain preprocessed data.

[0059] Among them, performing model compression on the end device is a key step to achieve efficient collaborative computing. In this embodiment, the end device can use the characteristics of the neural network accelerator hardware to perform sparse and quantized processing on the deep learning model, thereby reducing the storage and transmission requirements of the model while maintaining the prediction accuracy of the model.

[0060] Specifically, in this embodiment, in order to efficiently perform model compression and knowledge extraction on the end device, specific hardware acceleration technology can be used. The hardware acceleration technology can include a dedicated digital signal processor, tensor core or other customized hardware unit to accelerate matrix operations and data processing. Hardware acceleration can be achieved in the following ways: HW =H(O(X),S(W),Q(W)); where, O HW is the output of the hardware accelerated operation, H(·,·,·) represents the hardware accelerated function, O(X) is the predicted output of the model, S(W) is the sparse weight set, and Q(W) is the quantized parameter. The hardware characteristics are used to efficiently perform model compression and knowledge extraction.

[0061] In one embodiment, a provided end-edge-cloud collaborative intelligent optimization and resource scheduling method for the industrial Internet of Things may also include a process of model sparsification and parameter quantization. The specific process includes: the end device uses a neural network accelerator hardware to introduce a number of zero values ​​into the deep learning model, and adjusts the original gradient of the deep learning model through each zero value to complete the sparsification processing; the end device uses a neural network accelerator hardware to reduce the precision of the floating-point numbers of the deep learning model to complete the quantization processing.

[0062] Among them, model sparsification refers to the introduction of a large number of zero values ​​in the model parameters, so that the model can be stored and transmitted more efficiently; parameter quantization converts the model's floating-point parameters into low-precision representations, further reducing the model's storage size and computing requirements. In this embodiment, the sparsification technology used is based on the importance of gradients, which can be expressed by the formula: Among them, g is the original gradient of the model, S T (g) is the gradient after sparseness, It is a set of indices selected according to the absolute value of the gradient. Parameter quantization is the process of converting model parameters from high-precision representation to low-precision representation to reduce the storage and transmission requirements of the model; specifically, the quantization process can be expressed as: Among them, W oriare the original model parameters, is the quantized model parameter, f quantize (·,·) is a quantization operation that reduces the precision of the model parameters to Where b is the number of quantization bits.

[0063] Step 204: The terminal device uses the feature extraction network and knowledge extraction technology to extract feature data from the preprocessed data, and uploads the feature data.

[0064] Performing knowledge extraction on the end device is a key step in achieving efficient collaborative computing. Through knowledge extraction technology, the end device can extract key information from the raw data and provide high-quality input for subsequent data processing and model training. The purpose of knowledge extraction is to extract information from the raw data that is useful for model training and prediction.

[0065] Specifically, in one embodiment, a method for end-edge-cloud collaborative intelligent optimization and resource scheduling for industrial Internet of Things can also include a knowledge extraction process, and the specific process includes: the end device inputs the preprocessed data into the feature extraction network to determine the number of data points and the feature dimension in the preprocessed data; the feature extraction network determines the selected original feature data based on the number of data points and the feature dimension, generates an original feature data matrix, and uses knowledge extraction technology to extract feature data based on the preprocessed data and the original feature data matrix.

[0066] The feature extraction network consists of multiple layers, including multiple convolutional layers, pooling layers, and fully connected layers, which are responsible for extracting key patterns and trends from the raw data, that is, the preprocessed data. Where n is the number of data points, d is the feature dimension of each data point, and the output of the knowledge extraction network is Where d' is the feature dimension after extraction. The knowledge extraction process can be expressed as: in, It is a knowledge extraction operation that extracts key information from the data through a series of transformations; I is a d×d' matrix, which is used to indicate which original features are selected and participate in the knowledge extraction process. Each row represents an output feature, and each column represents an input feature. The elements in the matrix represent the contribution of the input feature to the output feature.

[0067] At the end device, model compression and knowledge extraction are jointly optimized to simultaneously reduce the size of the model and improve the quality of data representation. Specifically, this can be achieved by jointly training the sparsification, quantization, and feature extraction networks. The joint optimization problem can be expressed as: in, is the prediction loss of the model, is the sparsification loss, is the quantization loss, λ and μ are regularization weights, Y is the true label of the data, O(X) is the predicted output of the model, S(W) is the sparse weight set, and Q(W) is the quantized parameter.

[0068] The end device uses the neural processing unit hardware to achieve model compression, and through specific hardware acceleration technology, the model parameters are sparsely processed and quantified to reduce the storage and transmission requirements of the model while maintaining the accuracy of the model. In addition, through knowledge extraction technology, the end device can extract key information from the data and provide high-quality input for subsequent data processing and model training.

[0069] In step 206, a lightweight inference model and a routing model are deployed in the edge device; the edge device receives the feature data transmitted from the end device, and preliminarily classifies the feature data through the lightweight inference model to obtain classified data.

[0070] The fast and responsive data processing of edge devices is a key link in achieving intelligent optimization and resource scheduling of edge-cloud collaboration. Through the end-to-end hybrid learning framework, edge devices can identify and selectively upload difficult data examples that can be correctly classified by the cloud. This process involves the deployment of lightweight inference models and routing models, as well as the preliminary classification of uploaded data, thereby minimizing overall latency while maintaining accuracy.

[0071] Specifically, in one embodiment, a method for end-edge-cloud collaborative intelligent optimization and resource scheduling for industrial Internet of Things can also include a process of deploying a lightweight inference model for preliminary classification. The specific process includes: the edge device receives feature data transmitted from the end device, and receives the original completed model of the global model transmitted from the cloud; the edge device determines a pruning threshold based on the feature data, and performs pruning operations on the original completed model based on the pruning threshold to obtain a lightweight inference model; and uses the lightweight inference model to perform preliminary classification of the feature data.

[0072] In order to achieve fast data processing on edge devices, a lightweight inference model is used in this embodiment to reduce the computational complexity and model size of the cloud global model, so that edge devices can process data quickly. The general extraction process of the lightweight model can be expressed as: light =f prune (M g ,θ pru ), where M g It is the original complete model of the global model sent by the cloud server. light is a pruned lightweight model, f prune is the pruning operation, θpru is the pruning threshold.

[0073] The edge device performs preliminary classification of the uploaded data, which can reduce the amount of data that needs to be uploaded to the cloud. The accuracy of the preliminary classification is crucial to the performance of the entire system. The preliminary classification process can be expressed as: Among them, C edge (X) is the classification result of the data X input from the edge device peer side, c is a possible category, P(c|X,M light ) is given a lightweight model M light The probability that data X belongs to category c is:

[0074] In one embodiment, a method for end-edge-cloud collaborative intelligent optimization and resource scheduling for industrial Internet of Things can also include a process for selecting a routing model. The specific process includes: the edge device determines the difficulty of feature data, and determines the confidence of the model in the feature data based on the routing model; the edge device obtains the confidence threshold, and completes the decision of the routing model based on the difficulty and confidence.

[0075] In this embodiment, the routing model can be responsible for deciding which data should be uploaded to the cloud and which can be processed by the edge device itself. The routing model makes decisions based on the difficulty of the data and the confidence of the model. The decision-making process of the routing model is expressed by the formula: Among them, R(X) is the decision of the routing model on the data X, Con(X) is the confidence of the model on X, and τ is the confidence threshold.

[0076] Since the routing model makes decisions based on the difficulty of the data, in one embodiment, a method for end-edge-cloud collaborative intelligent optimization and resource scheduling for the industrial Internet of Things may also include a process of identifying difficult data. The specific process includes: the edge device identifies the difficult data in the feature data based on Boolean variables according to a confidence threshold, and sends the difficult data to the cloud.

[0077] That is, to further improve the efficiency of the system, the edge device needs to identify data that is difficult to classify and upload these examples to the cloud. Difficult data can be defined in many ways, such as by confidence threshold or prediction uncertainty. The identification process is expressed as: where U(X) is a Boolean variable indicating whether the data x is considered difficult, and τ is the confidence threshold.

[0078] In step 208, the edge device inputs the classified data into the routing model; the difficult data is identified through the routing model and sent to the cloud, the cloud processes the difficult data, and updates the global model stored in the cloud based on the processed difficult data.

[0079] The edge device and the cloud work together, and the collaboration between the cloud and the edge device is achieved through the routing model and the lightweight inference model. The cloud focuses on processing difficult data that the edge device cannot accurately classify and updates the global model. The formula is expressed as: Among them, M g is the global model in the cloud, f update is the model update operation, Indicates that the cloud aggregates a difficult set of data from all n edge devices Update the global model.

[0080] Collaborative computing and global optimization in the cloud are the core links to achieve intelligent optimization and resource scheduling of edge-cloud collaboration, which minimizes system costs by optimizing device association, offloading mode, bandwidth allocation and task split ratio. The cloud not only executes complex deep learning algorithms, but is also responsible for global model optimization and resource scheduling to achieve optimal performance.

[0081] Specifically, in one embodiment, a method for end-edge-cloud collaborative intelligent optimization and resource scheduling for the industrial Internet of Things can also include a process of device association and offloading decision-making, and the specific process includes: the cloud determines a set of edge devices to obtain the computing power of each edge device and the current task requirements; the cloud dynamically determines the tasks to be executed on the edge device and the tasks to be executed on the cloud based on the cloud computing resources, the computing power of the edge device and the current task requirements according to the association function and the computing cost function; the cloud determines the bandwidth threshold according to the bandwidth of the edge device, and according to the bandwidth threshold, offloads the tasks executed on the edge device to the edge device based on the offloading function, the bandwidth function and the task set.

[0082] Among them, device association and offloading decision is the first step of cloud collaborative computing, which determines whether the task should be processed on the end device, edge device or cloud. Device association optimization means that the cloud dynamically determines which tasks should be executed on the edge device and which tasks should be uploaded to the cloud for processing based on the computing power of the edge device and the current task requirements. This process can be expressed by the formula: Among them, E is the edge device set, C is the cloud computing resource, and F Assoc is the correlation function, is the computational cost function.

[0083] The offloading mode decision involves deciding whether the task is fully executed on the edge device, partially executed, or fully uploaded to the cloud for execution. The offloading mode can be determined through the decision process: Among them, Offload is the unloading function, T is the task set, is the bandwidth function, B th is the bandwidth threshold.

[0084] In this embodiment, bandwidth allocation and task segmentation are two other key aspects of cloud optimization. The bandwidth allocation strategy aims to optimize limited bandwidth resources to meet the data transmission requirements of different tasks. Bandwidth allocation can be achieved by the following optimization problem: max∑ t∈T BW(t)st∑ t∈T BW(t)≤B total ; Where BW(t) is the bandwidth assigned to task t, B total is the total bandwidth. The task split ratio refers to the workload of allocating tasks between edge devices and the cloud. The optimization of the split ratio can be expressed by the formula: Among them, Split is the splitting function, is the performance function, is the cost function and s is the split ratio.

[0085] In step 210, the cloud connects to the end device and the edge device respectively, monitors the performance of the end device and the edge device in real time, dynamically adjusts the operation of the end device and the edge device by executing deep learning algorithms to allocate bandwidth and split tasks, and optimizes the global model.

[0086] The cloud plays a core role in real-time performance monitoring and dynamic adjustment. It is responsible for comprehensively evaluating the operating status of the entire system, including the data processing efficiency and accuracy of edge devices. By collecting performance data of edge devices in real time, the cloud can gain insight into the performance of the system under different operating conditions and implement dynamic adjustment strategies accordingly.

[0087] In one embodiment, a method for end-edge-cloud collaborative intelligent optimization and resource scheduling for the industrial Internet of Things can also include a process of real-time performance feedback and dynamic adjustment in the cloud. The specific process includes: the cloud collects performance data of edge devices in real time and determines performance evaluation indicators; the cloud obtains performance evaluation weight coefficients, and updates the performance evaluation indicators based on the performance evaluation weight coefficients to obtain change trends corresponding to the performance evaluation indicators; the cloud uses deep learning algorithms to dynamically adjust the operation of end devices and edge devices based on the change trends.

[0088] Specifically, the cloud uses a comprehensive performance evaluation index Ψ, which combines multiple dimensions such as latency, accuracy, and resource utilization to quantify the overall performance of the system. The performance evaluation index Ψ is defined and updated by the formula: Ψ = λ1·Z latency +λ2·Z accuracy +λ3·Z resource; Among them, λ1,λ2,λ3 are weight coefficients, which are adjusted according to different optimization goals to ensure that Ψ accurately reflects the system performance focus at the current stage. The cloud uses machine learning algorithms to optimize model parameters and resource scheduling strategies based on the changing trend of Ψ, thereby achieving dynamic adjustment of edge devices and the entire system. This adaptive adjustment mechanism ensures that the system always maintains optimal performance under changing workloads and network conditions. Through centralized intelligent optimization in the cloud, the entire end-edge-cloud collaborative system can flexibly respond to various challenges, achieve efficient resource utilization and rapid response to tasks.

[0089] In one embodiment, a method for end-edge-cloud collaborative intelligent optimization and resource scheduling for industrial Internet of Things is provided, which may also include an end-edge-cloud collaborative optimization process, and the specific process includes: a local model is deployed in the edge device, and a global model is deployed in the cloud; the cloud determines the distance measurement between the local model and the global model, and optimizes the global model based on the distance measurement.

[0090] The ultimate goal of the cloud is to achieve close coordination among the three layers of the device, edge, and cloud, and to improve the performance and adaptability of the intelligent learning system through global optimization, while reducing energy consumption and latency. The global model optimization of the cloud is achieved by aggregating the local models uploaded by edge devices. e is the local model on the end device e, M g is the global model on the cloud server, then the global model optimization problem can be expressed as: Among them, l(·,·) is the distance measure between models. In addition, it is necessary to comprehensively consider the backup association, offloading mode, bandwidth allocation and task split ratio to achieve the optimization of the overall system performance. The optimization can be expressed as: Among them, E i represents the i-th end-side device, P th It is a performance threshold manually set based on historical performance data. The collaborative computing and global optimization modules in the cloud can effectively schedule the resources of the entire system, optimize the efficiency of task execution, and ensure the stability and response speed of the system.

[0091] The application claims a method for intelligent optimization and resource scheduling of end-edge-cloud collaboration for industrial IoT, which can improve data processing efficiency, save network resources, and enhance system intelligence. Specifically, it can be mentioned in:

[0092] Improved communication efficiency: Model compression techniques on the end device, such as sparsification and quantization, significantly reduce the amount of model parameter transmission; experiments show that while maintaining model performance, the amount of communication is reduced by about 10% compared to existing methods;

[0093] Improved data processing efficiency: Model compression and knowledge extraction technology implemented on the end device effectively reduces the storage and transmission requirements of the model. Under the same network environment, this method reduces the model transmission time by about 30%, greatly improving the efficiency of data processing;

[0094] Data privacy protection: By performing model compression and knowledge extraction on the end device, the amount of raw data transmission is reduced, thereby reducing the risk of data leakage. Combined with the preliminary data processing of the edge device, data privacy is better protected;

[0095] Effective cost reduction: The global optimization strategy in the cloud significantly reduces the operating cost of the system by precisely controlling task offloading, bandwidth allocation, and resource scheduling;

[0096] Enhanced model generalization capability: The close collaboration of the three layers of device, edge, and cloud enables continuous optimization and intelligent decision-making of the model, and improves the generalization capability of the model in diverse scenarios. Cross-device and cross-scenario tests show that this method performs better in terms of model accuracy and robustness.

[0097] Heterogeneous data processing capability: Ability to process heterogeneous data from different sources. The unified optimization strategy of the cloud enhances the processing capability of such data and improves the overall robustness of the system.

[0098] Rapid adaptation to changing business needs: This technical solution is suitable for scenarios with high real-time requirements such as autonomous driving, smart cities, and smart manufacturing. It can quickly respond to changes in business needs by dynamically adjusting collaboration strategies to ensure efficient operation of the system;

[0099] Improved system adaptability: The close collaboration of the three layers of end, edge and cloud enables the system to quickly adapt to changing business needs and scenarios with high real-time requirements. Experiments have shown that this method can maintain stable performance in different industrial IoT application scenarios, showing excellent adaptability.

[0100] Reduction of energy consumption and delay: By reducing data transmission and optimizing the allocation of computing tasks, this technical solution effectively reduces the system's energy consumption and processing delay.

[0101] It should be understood that, although the various steps in the above-mentioned flow chart are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least a part of the steps in the above-mentioned flow chart may include a plurality of sub-steps or a plurality of stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the sub-steps or stages of other steps.

[0102] In one embodiment, Figure 1 , Figure 3 As shown, a device-edge-cloud collaborative intelligent optimization and resource scheduling system for industrial Internet of Things is provided, including: a terminal device 110, an edge device 120 and a cloud 130, wherein:

[0103] The end device 110 is used to perform sparse and quantization processing on the deep learning model using the neural network accelerator hardware to obtain a compressed model; the end device collects raw data and inputs the raw data into the compressed model for preprocessing to obtain preprocessed data;

[0104] The terminal device 110 is also used to use the feature extraction network and the knowledge extraction technology to extract feature data from the preprocessed data, and upload the feature data;

[0105] The edge device 120 deployed with a lightweight inference model and a routing model is used to receive feature data transmitted from the end device 110, and preliminarily classify the feature data through the lightweight inference model to obtain classified data;

[0106] The edge device 120 is also used to input the classified data into the routing model; the difficult data is identified through the routing model and sent to the cloud 130, and the cloud 130 processes the difficult data and updates the global model stored in the cloud according to the processed difficult data;

[0107] The cloud 130 is connected to the terminal device 110 and the edge device 120 respectively, monitors the performance of the terminal device 110 and the edge device 130 in real time, dynamically adjusts the operation of the terminal device 110 and the edge device 120 by executing deep learning algorithms to allocate bandwidth and split tasks, and optimizes the global model.

[0108] In one embodiment, the terminal device 110 is also used to use the neural network accelerator hardware to introduce a number of zero values ​​into the deep learning model, adjust the original gradient of the deep learning model through each zero value, and complete the sparsification processing; use the neural network accelerator hardware to reduce the precision of the floating-point numbers of the deep learning model to complete the quantization processing.

[0109] In one embodiment, the terminal device 110 is also used to input the preprocessed data into the feature extraction network to determine the number of data points and feature dimensions in the preprocessed data; determine the selected original feature data based on the number of data points and feature dimensions through the feature extraction network to generate an original feature data matrix; and use knowledge extraction technology to extract feature data based on the preprocessed data and the original feature data matrix.

[0110] In one embodiment, the edge device 120 is also used to receive feature data transmitted from the end device, and receive the original completed model of the global model transmitted from the cloud; determine a pruning threshold based on the feature data, and perform pruning operations on the original completed model based on the pruning threshold to obtain a lightweight inference model; and use the lightweight inference model to perform preliminarily classification on the feature data.

[0111] In one embodiment, the edge device 120 is also used to determine the difficulty of the feature data, and determine the confidence of the model in the feature data based on the routing model; obtain the confidence threshold, and complete the decision of the routing model based on the difficulty and confidence.

[0112] In one embodiment, the edge device 120 is further configured to identify difficult data in the feature data based on a Boolean variable according to a confidence threshold, and send the difficult data to the cloud 130 .

[0113] In one embodiment, the cloud 130 is also used to determine each edge device set, obtain the computing power of each edge device and the current task requirements; based on the cloud computing resources, the computing power of the edge device and the current task requirements, dynamically determine the tasks to be executed on the edge device and the tasks to be executed on the cloud according to the association function and the computing cost function; determine the bandwidth threshold according to the bandwidth of the edge device, and according to the bandwidth threshold, offload the tasks executed on the edge device to the edge device based on the offload function, the bandwidth function, and the task set.

[0114] In one embodiment, the cloud 130 is also used to collect performance data of edge devices in real time and determine performance evaluation indicators; obtain performance evaluation weight coefficients, and update the performance evaluation indicators based on the performance evaluation weight coefficients to obtain change trends corresponding to the performance evaluation indicators; and use deep learning algorithms to dynamically adjust the operation of terminal devices and edge devices based on the change trends.

[0115] In one embodiment, a local model is deployed in the edge device 120, and a global model is deployed in the cloud 130; the cloud 130 is also used to determine a distance metric between the local model and the global model, and optimize the global model based on the distance metric.

[0116] In one embodiment, a collaborative intelligent optimization and resource scheduling framework combining end devices, edge devices and the cloud is provided, such as Figure 4 As shown, it can specifically include a model compression module and a knowledge extraction module set in the terminal device; a data processing module set in the edge device; a collaborative computing module and a system performance evaluation module set in the cloud.

[0117] Among them, the model compression module is responsible for realizing the sparse processing and parameter quantization of the model, reducing the storage and transmission requirements of the model; the knowledge extraction module is used to extract key information from the original data and provide high-quality input data; the data processing module deploys lightweight inference models and routing models, performs preliminary classification of data and decides whether to upload it to the cloud; the collaborative computing module executes complex deep learning algorithms and is responsible for global model optimization and resource scheduling; the system performance evaluation module is used to evaluate the performance of the entire system, including accuracy, latency, and energy consumption.

[0118] like Figure 4 As shown in the figure, arrows are used to represent data flow and information transmission in each part. The output of the model compression module points to the knowledge extraction module through arrows, indicating the feature transmission after model compression; the output of the knowledge extraction module is transmitted to the data processing module through arrows, indicating the transmission of knowledge extraction results; the data processing module uploads data to the collaborative computing module based on preliminary classification and routing decisions, indicating the transmission of data processing and routing decisions; after the collaborative computing module performs cloud collaborative computing, it feeds back the optimization results to the data processing module and the model compression module, indicating the transmission of cloud computing results; the system performance evaluation module evaluates the system performance and feeds back the evaluation results to each module to guide dynamic adjustment, indicating the transmission of performance feedback.

[0119] In one embodiment, a device model compression framework combining model sparsification, quantization, and knowledge extraction is provided, such as Figure 5As shown, it can specifically include a model sparsification unit: responsible for sparse processing of model parameters, by selecting important parameters to retain and unimportant parameters to zero, so as to reduce the storage and transmission requirements of the model; a model quantization unit: quantizing the sparsified model parameters to reduce the accuracy of the parameters, and further reducing the storage size and computing requirements of the model; a knowledge extraction unit: extracting key information from the original data to provide high-quality input for model training and data processing; a joint optimization controller: responsible for coordinating the processes of model sparsification, quantization and knowledge extraction to ensure that the model is compressed while maintaining accuracy; a hardware acceleration unit: using dedicated hardware to accelerate the inference process of the model to improve the efficiency of model compression and knowledge extraction.

[0120] like Figure 5 As shown in the figure, each part represents data flow and information transmission through arrows. The output of the model sparsification unit points to the model quantization unit through an arrow, indicating the transmission of sparsification parameters, and points to the joint optimization controller, indicating the transmission of sparsification processing; the output of the model quantization unit also points to the joint optimization controller, indicating the transmission of quantization processing; the output of the knowledge extraction unit points to the joint optimization controller, indicating the transmission of knowledge extraction; the output of the joint optimization controller is transmitted to the hardware acceleration unit, indicating the transmission of the joint optimization strategy; finally, the output of the hardware acceleration unit indicates the transmission of hardware acceleration operations.

[0121] In one embodiment, an end-to-end hybrid learning framework is provided to enable edge devices to identify and selectively upload difficult data examples that the cloud can correctly classify, such as Figure 6 As shown, it can specifically include a lightweight inference model deployment unit: responsible for deploying lightweight inference models on edge devices to quickly process data; a routing model decision unit: based on the inference results of the lightweight model, it decides whether the data is processed on the edge device or uploaded to the cloud; a difficult data identification unit: used to identify data that is difficult for the lightweight model to process, and mark this data for upload to the cloud; a cloud and edge device collaboration unit: responsible for data transmission and model updates between the cloud and edge devices, ensuring that the cloud model can process difficult data uploaded by the edge device.

[0122] like Figure 6As shown in the figure, each part represents data flow and information transmission through arrows; the output of the lightweight inference model deployment unit points to the routing model decision unit through an arrow, indicating the transmission of the inference result; the output of the routing model decision unit is divided into two parts, one part points to the local processing flow, and the other part points to the difficult data identification unit, indicating the identification of difficult data; the output of the difficult data identification unit points to the cloud and edge device collaboration unit, indicating that the difficult data is uploaded to the cloud; the cloud and edge device collaboration unit is also responsible for feeding back the cloud model update to the lightweight inference model deployment unit and the routing model decision unit, forming a closed-loop data processing and learning process.

[0123] In one embodiment, a cloud collaborative computing and global optimization framework is provided, such as Figure 7 As shown, it specifically includes a device association optimization unit: responsible for dynamically determining the execution location of the task, that is, allocating tasks between edge devices or the cloud; an offloading mode decision unit: determining the offloading mode of the task according to the computing power of the edge device and the resource status of the cloud, including full offloading, partial offloading or no offloading; a bandwidth allocation strategy unit: responsible for optimizing the allocation of bandwidth resources to ensure that the data transmission requirements of different tasks are met; a task split ratio optimization unit: determining the split ratio of the task between the edge device and the cloud to achieve optimal execution efficiency; a real-time performance feedback and dynamic adjustment unit: monitoring the performance of the cloud and edge devices, and dynamically adjusting resource allocation and task execution strategies based on real-time feedback; an end-edge-cloud collaborative optimization unit: comprehensively considering the decision results of each unit, and performing end-edge-cloud collaborative optimization to achieve the optimal performance of the entire system.

[0124] like Figure 7 As shown, each part uses arrows to represent data flow and information transmission; the output of the device association optimization unit points to the offloading mode decision unit and the bandwidth allocation strategy unit through an arrow, indicating the transmission of the device association optimization result; the output of the offloading mode decision unit points to the bandwidth allocation strategy unit through an arrow, indicating the transmission of the offloading decision information; the output of the bandwidth allocation strategy unit points to the task splitting ratio optimization unit through an arrow, indicating the transmission of the bandwidth allocation information; the output of the task splitting ratio optimization unit points to the end-edge-cloud collaborative optimization unit through an arrow, indicating the transmission of the task splitting ratio; the real-time performance feedback and dynamic adjustment unit points to all units through arrows, indicating the transmission of real-time performance feedback and dynamic adjustment.

[0125] In one embodiment, the technical solution is applied to the equipment maintenance prediction system in the industrial Internet of Things to improve the accuracy and real-time performance of equipment status monitoring and reduce maintenance costs. Figure 8 The specific implementation process is as follows:

[0126] 1. Equipment status data collection: In industrial environments, sensors installed on equipment collect real-time data on its operating status, such as temperature, vibration frequency, and pressure;

[0127] 2. Data preprocessing and compression: The end device preprocesses the collected raw data, including filtering and denoising, and reduces the data transmission volume through model compression technology;

[0128] 3. Feature extraction and knowledge distillation: The end device uses a self-distillation strategy to extract key features from the preprocessed data and enhance the model's understanding of the device status;

[0129] 4. Preliminary analysis and anomaly detection: The edge device receives data from the end device and uses a lightweight inference model to perform preliminary analysis to quickly identify potential abnormal conditions;

[0130] 5. Cloud model training and optimization: For data that requires more in-depth analysis, edge devices upload it to the cloud, where cloud servers use advanced machine learning models for training and optimization;

[0131] 6. Fault prediction and maintenance decision-making: The cloud server comprehensively analyzes the results, uses the global optimization algorithm to predict possible equipment failures, and makes corresponding maintenance decisions;

[0132] 7. Maintenance strategy delivery: The cloud delivers maintenance decisions to edge devices in the form of instructions, and the edge devices execute or guide on-site maintenance personnel to perform operations based on actual conditions;

[0133] 8. Maintenance execution and feedback: After the maintenance work is completed, the on-site personnel will feedback the maintenance results to the edge device, and the edge device will upload it to the cloud;

[0134] 9. Performance evaluation and model iteration: The system performance evaluation module collects maintenance results and performance data, evaluates the effectiveness of maintenance strategies, and uses feedback for model iteration and optimization.

[0135] In one embodiment, the technical solution is applied to an energy management system to improve energy efficiency and reduce energy consumption costs, such as Fig. 9 The specific implementation process is as follows:

[0136] 1. Energy consumption data collection: Smart meters and sensors deployed in buildings or industrial facilities collect energy consumption data in real time;

[0137] 2. Data preprocessing and compression: The end device preprocesses the collected energy consumption data, such as data cleaning and normalization, and reduces the amount of data transmission through model compression technology;

[0138] 3. Energy consumption pattern analysis: The edge device receives the data uploaded by the end device and uses a lightweight inference model to analyze the energy consumption pattern and identify energy saving potential;

[0139] 4. Energy saving opportunity identification: The edge device identifies energy saving opportunities and optimization points based on the results of energy consumption pattern analysis;

[0140] 5. Cloud energy consumption prediction and optimization: The edge device uploads the analysis results to the cloud, and the cloud server uses machine learning models to predict energy consumption trends and develop optimization strategies;

[0141] 6. Energy-saving strategy generation and distribution: The cloud server generates specific energy-saving strategies and distributes them to edge devices;

[0142] 7. Energy consumption monitoring and real-time adjustment: Edge devices monitor actual energy consumption and make real-time adjustments based on the energy-saving strategies issued to ensure optimal energy use;

[0143] 8. System performance evaluation: The system performance evaluation module collects energy consumption data and the execution results of energy-saving strategies to evaluate the effectiveness of energy-saving measures.

[0144] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0145] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A method for intelligent optimization and resource scheduling of end-edge-cloud collaboration for industrial Internet of Things, characterized in that: The method comprises: The end device uses the neural network accelerator hardware to perform sparse and quantization processing on the deep learning model to obtain a compressed model; the end device collects raw data and inputs the raw data into the compressed model for preprocessing to obtain preprocessed data; The terminal device uses a feature extraction network and a knowledge extraction technology to extract feature data from the preprocessed data, and uploads the feature data; A lightweight inference model and a routing model are deployed in the edge device; the edge device receives the feature data transmitted from the end device, and preliminarily classifies the feature data through the lightweight inference model to obtain classified data; The edge device inputs the classified data into the routing model; the difficult data is identified through the routing model and sent to the cloud, the cloud processes the difficult data, and updates the global model stored in the cloud according to the processed difficult data; The cloud is connected to the terminal device and the edge device respectively, monitors the performance of the terminal device and the edge device in real time, dynamically adjusts the operation of the terminal device and the edge device by executing deep learning algorithms to allocate bandwidth and split tasks, and optimizes the global model.

2. The method for end-edge-cloud collaborative intelligent optimization and resource scheduling for industrial Internet of Things according to claim 1 is characterized in that: The end device uses the neural network accelerator hardware to perform sparse and quantized processing on the deep learning model to obtain a compressed model, including: The terminal device uses a neural network accelerator hardware to introduce a number of zero values ​​into the deep learning model, and adjusts the original gradient of the deep learning model through each of the zero values ​​to complete the sparse processing; The terminal device uses neural network accelerator hardware to reduce the precision of the floating-point numbers of the deep learning model and complete the quantization process.

3. The method for end-edge-cloud collaborative intelligent optimization and resource scheduling for industrial Internet of Things according to claim 1 is characterized in that: The terminal device uses a feature extraction network and a knowledge extraction technology to extract feature data from the preprocessed data, including: The terminal device inputs the preprocessed data into a feature extraction network to determine the number of data points and feature dimensions in the preprocessed data; Determining the selected original feature data based on the number of data points and feature dimensions through the feature extraction network to generate an original feature data matrix; The knowledge extraction technology is used to extract feature data based on the preprocessed data and the original feature data matrix.

4. The method for end-edge-cloud collaborative intelligent optimization and resource scheduling for industrial Internet of Things according to claim 1 is characterized in that: The edge device receives the feature data transmitted from the end device, and preliminarily classifies the feature data using the lightweight inference model, including: The edge device receives the feature data transmitted from the end device, and receives the original completed model of the global model transmitted from the cloud; The edge device determines a pruning threshold according to the feature data, and performs a pruning operation on the original completed model according to the pruning threshold to obtain a lightweight inference model; The feature data is preliminarily classified using the lightweight inference model.

5. The method for end-edge-cloud collaborative intelligent optimization and resource scheduling for industrial Internet of Things according to claim 1 is characterized in that: The method further comprises: The edge device determines the difficulty of the feature data, and determines the confidence of the model for the feature data based on the routing model; The edge device obtains a confidence threshold and completes the decision of the routing model according to the difficulty and confidence.

6. The method for end-edge-cloud collaborative intelligent optimization and resource scheduling for industrial Internet of Things according to claim 5 is characterized in that: The method comprises: The edge device identifies difficult data in the feature data based on the confidence threshold and Boolean variables, and sends the difficult data to the cloud.

7. The method for end-edge-cloud collaborative intelligent optimization and resource scheduling for industrial Internet of Things according to claim 1 is characterized in that: The method further comprises: The cloud determines each set of edge devices and obtains the computing capacity and current task requirements of each edge device; The cloud dynamically determines the tasks to be executed on the edge device and the tasks to be executed on the cloud according to the association function and the computational cost function based on the cloud computing resources, the computing capability of the edge device and the current task requirements; The cloud determines a bandwidth threshold according to the bandwidth of the edge device, and according to the bandwidth threshold, offloads the tasks executed on the edge device to the edge device based on the offloading function, the bandwidth function, and the task set.

8. The method for end-edge-cloud collaborative intelligent optimization and resource scheduling for industrial Internet of Things according to claim 7 is characterized in that: The cloud dynamically adjusts the operation of the terminal device and the edge device, including: The cloud collects the performance data of the edge device in real time and determines the performance evaluation index; The cloud obtains a performance evaluation weight coefficient, and updates the performance evaluation indicator based on the performance evaluation weight coefficient to obtain a change trend corresponding to the performance evaluation indicator; The cloud uses a deep learning algorithm to dynamically adjust the operation of the terminal device and the edge device according to the change trend.

9. The method for end-edge-cloud collaborative intelligent optimization and resource scheduling for industrial Internet of Things according to claim 1 is characterized in that: The cloud optimizes the global model, including: A local model is deployed in the edge device, and a global model is deployed in the cloud; The cloud determines a distance metric between the local model and the global model, and performs global model optimization based on the distance metric.

10. A device-edge-cloud collaborative intelligent optimization and resource scheduling system for industrial Internet of Things, characterized in that: The system comprises: The end device is used to perform sparse and quantization processing on the deep learning model using the neural network accelerator hardware to obtain a compressed model; the end device collects raw data and inputs the raw data into the compressed model for preprocessing to obtain preprocessed data; The terminal device is also used to utilize a feature extraction network and a knowledge extraction technology to extract feature data from the preprocessed data, and upload the feature data; An edge device deployed with a lightweight inference model and a routing model is used to receive the feature data transmitted from the end device, and preliminarily classify the feature data through the lightweight inference model to obtain classified data; The edge device is also used to input the classified data into the routing model; identify the difficult data through the routing model and send it to the cloud, the cloud processes the difficult data, and updates the global model stored in the cloud according to the processed difficult data; The cloud is connected to the terminal device and the edge device respectively, monitors the performance of the terminal device and the edge device in real time, dynamically adjusts the operation of the terminal device and the edge device by executing deep learning algorithms to allocate bandwidth and split tasks, and optimizes the global model.

Images